xref: /titanic_51/usr/src/uts/common/dtrace/dtrace.c (revision 15a5f64a0e54ccfff87908c70d561a14d8b63c16)
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_TOUPPER:
3908 	case DIF_SUBR_TOLOWER: {
3909 		uintptr_t s = tupregs[0].dttk_value;
3910 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3911 		char *dest = (char *)mstate->dtms_scratch_ptr, c;
3912 		size_t len = dtrace_strlen((char *)s, size);
3913 		char lower, upper, convert;
3914 		int64_t i;
3915 
3916 		if (subr == DIF_SUBR_TOUPPER) {
3917 			lower = 'a';
3918 			upper = 'z';
3919 			convert = 'A';
3920 		} else {
3921 			lower = 'A';
3922 			upper = 'Z';
3923 			convert = 'a';
3924 		}
3925 
3926 		if (!dtrace_canload(s, len + 1, mstate, vstate)) {
3927 			regs[rd] = NULL;
3928 			break;
3929 		}
3930 
3931 		if (!DTRACE_INSCRATCH(mstate, size)) {
3932 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3933 			regs[rd] = NULL;
3934 			break;
3935 		}
3936 
3937 		for (i = 0; i < size - 1; i++) {
3938 			if ((c = dtrace_load8(s + i)) == '\0')
3939 				break;
3940 
3941 			if (c >= lower && c <= upper)
3942 				c = convert + (c - lower);
3943 
3944 			dest[i] = c;
3945 		}
3946 
3947 		ASSERT(i < size);
3948 		dest[i] = '\0';
3949 		regs[rd] = (uintptr_t)dest;
3950 		mstate->dtms_scratch_ptr += size;
3951 		break;
3952 	}
3953 
3954 case DIF_SUBR_GETMAJOR:
3955 #ifdef _LP64
3956 		regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
3957 #else
3958 		regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
3959 #endif
3960 		break;
3961 
3962 	case DIF_SUBR_GETMINOR:
3963 #ifdef _LP64
3964 		regs[rd] = tupregs[0].dttk_value & MAXMIN64;
3965 #else
3966 		regs[rd] = tupregs[0].dttk_value & MAXMIN;
3967 #endif
3968 		break;
3969 
3970 	case DIF_SUBR_DDI_PATHNAME: {
3971 		/*
3972 		 * This one is a galactic mess.  We are going to roughly
3973 		 * emulate ddi_pathname(), but it's made more complicated
3974 		 * by the fact that we (a) want to include the minor name and
3975 		 * (b) must proceed iteratively instead of recursively.
3976 		 */
3977 		uintptr_t dest = mstate->dtms_scratch_ptr;
3978 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3979 		char *start = (char *)dest, *end = start + size - 1;
3980 		uintptr_t daddr = tupregs[0].dttk_value;
3981 		int64_t minor = (int64_t)tupregs[1].dttk_value;
3982 		char *s;
3983 		int i, len, depth = 0;
3984 
3985 		/*
3986 		 * Due to all the pointer jumping we do and context we must
3987 		 * rely upon, we just mandate that the user must have kernel
3988 		 * read privileges to use this routine.
3989 		 */
3990 		if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
3991 			*flags |= CPU_DTRACE_KPRIV;
3992 			*illval = daddr;
3993 			regs[rd] = NULL;
3994 		}
3995 
3996 		if (!DTRACE_INSCRATCH(mstate, size)) {
3997 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3998 			regs[rd] = NULL;
3999 			break;
4000 		}
4001 
4002 		*end = '\0';
4003 
4004 		/*
4005 		 * We want to have a name for the minor.  In order to do this,
4006 		 * we need to walk the minor list from the devinfo.  We want
4007 		 * to be sure that we don't infinitely walk a circular list,
4008 		 * so we check for circularity by sending a scout pointer
4009 		 * ahead two elements for every element that we iterate over;
4010 		 * if the list is circular, these will ultimately point to the
4011 		 * same element.  You may recognize this little trick as the
4012 		 * answer to a stupid interview question -- one that always
4013 		 * seems to be asked by those who had to have it laboriously
4014 		 * explained to them, and who can't even concisely describe
4015 		 * the conditions under which one would be forced to resort to
4016 		 * this technique.  Needless to say, those conditions are
4017 		 * found here -- and probably only here.  Is this the only use
4018 		 * of this infamous trick in shipping, production code?  If it
4019 		 * isn't, it probably should be...
4020 		 */
4021 		if (minor != -1) {
4022 			uintptr_t maddr = dtrace_loadptr(daddr +
4023 			    offsetof(struct dev_info, devi_minor));
4024 
4025 			uintptr_t next = offsetof(struct ddi_minor_data, next);
4026 			uintptr_t name = offsetof(struct ddi_minor_data,
4027 			    d_minor) + offsetof(struct ddi_minor, name);
4028 			uintptr_t dev = offsetof(struct ddi_minor_data,
4029 			    d_minor) + offsetof(struct ddi_minor, dev);
4030 			uintptr_t scout;
4031 
4032 			if (maddr != NULL)
4033 				scout = dtrace_loadptr(maddr + next);
4034 
4035 			while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4036 				uint64_t m;
4037 #ifdef _LP64
4038 				m = dtrace_load64(maddr + dev) & MAXMIN64;
4039 #else
4040 				m = dtrace_load32(maddr + dev) & MAXMIN;
4041 #endif
4042 				if (m != minor) {
4043 					maddr = dtrace_loadptr(maddr + next);
4044 
4045 					if (scout == NULL)
4046 						continue;
4047 
4048 					scout = dtrace_loadptr(scout + next);
4049 
4050 					if (scout == NULL)
4051 						continue;
4052 
4053 					scout = dtrace_loadptr(scout + next);
4054 
4055 					if (scout == NULL)
4056 						continue;
4057 
4058 					if (scout == maddr) {
4059 						*flags |= CPU_DTRACE_ILLOP;
4060 						break;
4061 					}
4062 
4063 					continue;
4064 				}
4065 
4066 				/*
4067 				 * We have the minor data.  Now we need to
4068 				 * copy the minor's name into the end of the
4069 				 * pathname.
4070 				 */
4071 				s = (char *)dtrace_loadptr(maddr + name);
4072 				len = dtrace_strlen(s, size);
4073 
4074 				if (*flags & CPU_DTRACE_FAULT)
4075 					break;
4076 
4077 				if (len != 0) {
4078 					if ((end -= (len + 1)) < start)
4079 						break;
4080 
4081 					*end = ':';
4082 				}
4083 
4084 				for (i = 1; i <= len; i++)
4085 					end[i] = dtrace_load8((uintptr_t)s++);
4086 				break;
4087 			}
4088 		}
4089 
4090 		while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4091 			ddi_node_state_t devi_state;
4092 
4093 			devi_state = dtrace_load32(daddr +
4094 			    offsetof(struct dev_info, devi_node_state));
4095 
4096 			if (*flags & CPU_DTRACE_FAULT)
4097 				break;
4098 
4099 			if (devi_state >= DS_INITIALIZED) {
4100 				s = (char *)dtrace_loadptr(daddr +
4101 				    offsetof(struct dev_info, devi_addr));
4102 				len = dtrace_strlen(s, size);
4103 
4104 				if (*flags & CPU_DTRACE_FAULT)
4105 					break;
4106 
4107 				if (len != 0) {
4108 					if ((end -= (len + 1)) < start)
4109 						break;
4110 
4111 					*end = '@';
4112 				}
4113 
4114 				for (i = 1; i <= len; i++)
4115 					end[i] = dtrace_load8((uintptr_t)s++);
4116 			}
4117 
4118 			/*
4119 			 * Now for the node name...
4120 			 */
4121 			s = (char *)dtrace_loadptr(daddr +
4122 			    offsetof(struct dev_info, devi_node_name));
4123 
4124 			daddr = dtrace_loadptr(daddr +
4125 			    offsetof(struct dev_info, devi_parent));
4126 
4127 			/*
4128 			 * If our parent is NULL (that is, if we're the root
4129 			 * node), we're going to use the special path
4130 			 * "devices".
4131 			 */
4132 			if (daddr == NULL)
4133 				s = "devices";
4134 
4135 			len = dtrace_strlen(s, size);
4136 			if (*flags & CPU_DTRACE_FAULT)
4137 				break;
4138 
4139 			if ((end -= (len + 1)) < start)
4140 				break;
4141 
4142 			for (i = 1; i <= len; i++)
4143 				end[i] = dtrace_load8((uintptr_t)s++);
4144 			*end = '/';
4145 
4146 			if (depth++ > dtrace_devdepth_max) {
4147 				*flags |= CPU_DTRACE_ILLOP;
4148 				break;
4149 			}
4150 		}
4151 
4152 		if (end < start)
4153 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4154 
4155 		if (daddr == NULL) {
4156 			regs[rd] = (uintptr_t)end;
4157 			mstate->dtms_scratch_ptr += size;
4158 		}
4159 
4160 		break;
4161 	}
4162 
4163 	case DIF_SUBR_STRJOIN: {
4164 		char *d = (char *)mstate->dtms_scratch_ptr;
4165 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4166 		uintptr_t s1 = tupregs[0].dttk_value;
4167 		uintptr_t s2 = tupregs[1].dttk_value;
4168 		int i = 0;
4169 
4170 		if (!dtrace_strcanload(s1, size, mstate, vstate) ||
4171 		    !dtrace_strcanload(s2, size, mstate, vstate)) {
4172 			regs[rd] = NULL;
4173 			break;
4174 		}
4175 
4176 		if (!DTRACE_INSCRATCH(mstate, size)) {
4177 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4178 			regs[rd] = NULL;
4179 			break;
4180 		}
4181 
4182 		for (;;) {
4183 			if (i >= size) {
4184 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4185 				regs[rd] = NULL;
4186 				break;
4187 			}
4188 
4189 			if ((d[i++] = dtrace_load8(s1++)) == '\0') {
4190 				i--;
4191 				break;
4192 			}
4193 		}
4194 
4195 		for (;;) {
4196 			if (i >= size) {
4197 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4198 				regs[rd] = NULL;
4199 				break;
4200 			}
4201 
4202 			if ((d[i++] = dtrace_load8(s2++)) == '\0')
4203 				break;
4204 		}
4205 
4206 		if (i < size) {
4207 			mstate->dtms_scratch_ptr += i;
4208 			regs[rd] = (uintptr_t)d;
4209 		}
4210 
4211 		break;
4212 	}
4213 
4214 	case DIF_SUBR_LLTOSTR: {
4215 		int64_t i = (int64_t)tupregs[0].dttk_value;
4216 		uint64_t val, digit;
4217 		uint64_t size = 65;	/* enough room for 2^64 in binary */
4218 		char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
4219 		int base = 10;
4220 
4221 		if (nargs > 1) {
4222 			if ((base = tupregs[1].dttk_value) <= 1 ||
4223 			    base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
4224 				*flags |= CPU_DTRACE_ILLOP;
4225 				break;
4226 			}
4227 		}
4228 
4229 		val = (base == 10 && i < 0) ? i * -1 : i;
4230 
4231 		if (!DTRACE_INSCRATCH(mstate, size)) {
4232 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4233 			regs[rd] = NULL;
4234 			break;
4235 		}
4236 
4237 		for (*end-- = '\0'; val; val /= base) {
4238 			if ((digit = val % base) <= '9' - '0') {
4239 				*end-- = '0' + digit;
4240 			} else {
4241 				*end-- = 'a' + (digit - ('9' - '0') - 1);
4242 			}
4243 		}
4244 
4245 		if (i == 0 && base == 16)
4246 			*end-- = '0';
4247 
4248 		if (base == 16)
4249 			*end-- = 'x';
4250 
4251 		if (i == 0 || base == 8 || base == 16)
4252 			*end-- = '0';
4253 
4254 		if (i < 0 && base == 10)
4255 			*end-- = '-';
4256 
4257 		regs[rd] = (uintptr_t)end + 1;
4258 		mstate->dtms_scratch_ptr += size;
4259 		break;
4260 	}
4261 
4262 	case DIF_SUBR_HTONS:
4263 	case DIF_SUBR_NTOHS:
4264 #ifdef _BIG_ENDIAN
4265 		regs[rd] = (uint16_t)tupregs[0].dttk_value;
4266 #else
4267 		regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
4268 #endif
4269 		break;
4270 
4271 
4272 	case DIF_SUBR_HTONL:
4273 	case DIF_SUBR_NTOHL:
4274 #ifdef _BIG_ENDIAN
4275 		regs[rd] = (uint32_t)tupregs[0].dttk_value;
4276 #else
4277 		regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
4278 #endif
4279 		break;
4280 
4281 
4282 	case DIF_SUBR_HTONLL:
4283 	case DIF_SUBR_NTOHLL:
4284 #ifdef _BIG_ENDIAN
4285 		regs[rd] = (uint64_t)tupregs[0].dttk_value;
4286 #else
4287 		regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
4288 #endif
4289 		break;
4290 
4291 
4292 	case DIF_SUBR_DIRNAME:
4293 	case DIF_SUBR_BASENAME: {
4294 		char *dest = (char *)mstate->dtms_scratch_ptr;
4295 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4296 		uintptr_t src = tupregs[0].dttk_value;
4297 		int i, j, len = dtrace_strlen((char *)src, size);
4298 		int lastbase = -1, firstbase = -1, lastdir = -1;
4299 		int start, end;
4300 
4301 		if (!dtrace_canload(src, len + 1, mstate, vstate)) {
4302 			regs[rd] = NULL;
4303 			break;
4304 		}
4305 
4306 		if (!DTRACE_INSCRATCH(mstate, size)) {
4307 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4308 			regs[rd] = NULL;
4309 			break;
4310 		}
4311 
4312 		/*
4313 		 * The basename and dirname for a zero-length string is
4314 		 * defined to be "."
4315 		 */
4316 		if (len == 0) {
4317 			len = 1;
4318 			src = (uintptr_t)".";
4319 		}
4320 
4321 		/*
4322 		 * Start from the back of the string, moving back toward the
4323 		 * front until we see a character that isn't a slash.  That
4324 		 * character is the last character in the basename.
4325 		 */
4326 		for (i = len - 1; i >= 0; i--) {
4327 			if (dtrace_load8(src + i) != '/')
4328 				break;
4329 		}
4330 
4331 		if (i >= 0)
4332 			lastbase = i;
4333 
4334 		/*
4335 		 * Starting from the last character in the basename, move
4336 		 * towards the front until we find a slash.  The character
4337 		 * that we processed immediately before that is the first
4338 		 * character in the basename.
4339 		 */
4340 		for (; i >= 0; i--) {
4341 			if (dtrace_load8(src + i) == '/')
4342 				break;
4343 		}
4344 
4345 		if (i >= 0)
4346 			firstbase = i + 1;
4347 
4348 		/*
4349 		 * Now keep going until we find a non-slash character.  That
4350 		 * character is the last character in the dirname.
4351 		 */
4352 		for (; i >= 0; i--) {
4353 			if (dtrace_load8(src + i) != '/')
4354 				break;
4355 		}
4356 
4357 		if (i >= 0)
4358 			lastdir = i;
4359 
4360 		ASSERT(!(lastbase == -1 && firstbase != -1));
4361 		ASSERT(!(firstbase == -1 && lastdir != -1));
4362 
4363 		if (lastbase == -1) {
4364 			/*
4365 			 * We didn't find a non-slash character.  We know that
4366 			 * the length is non-zero, so the whole string must be
4367 			 * slashes.  In either the dirname or the basename
4368 			 * case, we return '/'.
4369 			 */
4370 			ASSERT(firstbase == -1);
4371 			firstbase = lastbase = lastdir = 0;
4372 		}
4373 
4374 		if (firstbase == -1) {
4375 			/*
4376 			 * The entire string consists only of a basename
4377 			 * component.  If we're looking for dirname, we need
4378 			 * to change our string to be just "."; if we're
4379 			 * looking for a basename, we'll just set the first
4380 			 * character of the basename to be 0.
4381 			 */
4382 			if (subr == DIF_SUBR_DIRNAME) {
4383 				ASSERT(lastdir == -1);
4384 				src = (uintptr_t)".";
4385 				lastdir = 0;
4386 			} else {
4387 				firstbase = 0;
4388 			}
4389 		}
4390 
4391 		if (subr == DIF_SUBR_DIRNAME) {
4392 			if (lastdir == -1) {
4393 				/*
4394 				 * We know that we have a slash in the name --
4395 				 * or lastdir would be set to 0, above.  And
4396 				 * because lastdir is -1, we know that this
4397 				 * slash must be the first character.  (That
4398 				 * is, the full string must be of the form
4399 				 * "/basename".)  In this case, the last
4400 				 * character of the directory name is 0.
4401 				 */
4402 				lastdir = 0;
4403 			}
4404 
4405 			start = 0;
4406 			end = lastdir;
4407 		} else {
4408 			ASSERT(subr == DIF_SUBR_BASENAME);
4409 			ASSERT(firstbase != -1 && lastbase != -1);
4410 			start = firstbase;
4411 			end = lastbase;
4412 		}
4413 
4414 		for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
4415 			dest[j] = dtrace_load8(src + i);
4416 
4417 		dest[j] = '\0';
4418 		regs[rd] = (uintptr_t)dest;
4419 		mstate->dtms_scratch_ptr += size;
4420 		break;
4421 	}
4422 
4423 	case DIF_SUBR_CLEANPATH: {
4424 		char *dest = (char *)mstate->dtms_scratch_ptr, c;
4425 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4426 		uintptr_t src = tupregs[0].dttk_value;
4427 		int i = 0, j = 0;
4428 
4429 		if (!dtrace_strcanload(src, size, mstate, vstate)) {
4430 			regs[rd] = NULL;
4431 			break;
4432 		}
4433 
4434 		if (!DTRACE_INSCRATCH(mstate, size)) {
4435 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4436 			regs[rd] = NULL;
4437 			break;
4438 		}
4439 
4440 		/*
4441 		 * Move forward, loading each character.
4442 		 */
4443 		do {
4444 			c = dtrace_load8(src + i++);
4445 next:
4446 			if (j + 5 >= size)	/* 5 = strlen("/..c\0") */
4447 				break;
4448 
4449 			if (c != '/') {
4450 				dest[j++] = c;
4451 				continue;
4452 			}
4453 
4454 			c = dtrace_load8(src + i++);
4455 
4456 			if (c == '/') {
4457 				/*
4458 				 * We have two slashes -- we can just advance
4459 				 * to the next character.
4460 				 */
4461 				goto next;
4462 			}
4463 
4464 			if (c != '.') {
4465 				/*
4466 				 * This is not "." and it's not ".." -- we can
4467 				 * just store the "/" and this character and
4468 				 * drive on.
4469 				 */
4470 				dest[j++] = '/';
4471 				dest[j++] = c;
4472 				continue;
4473 			}
4474 
4475 			c = dtrace_load8(src + i++);
4476 
4477 			if (c == '/') {
4478 				/*
4479 				 * This is a "/./" component.  We're not going
4480 				 * to store anything in the destination buffer;
4481 				 * we're just going to go to the next component.
4482 				 */
4483 				goto next;
4484 			}
4485 
4486 			if (c != '.') {
4487 				/*
4488 				 * This is not ".." -- we can just store the
4489 				 * "/." and this character and continue
4490 				 * processing.
4491 				 */
4492 				dest[j++] = '/';
4493 				dest[j++] = '.';
4494 				dest[j++] = c;
4495 				continue;
4496 			}
4497 
4498 			c = dtrace_load8(src + i++);
4499 
4500 			if (c != '/' && c != '\0') {
4501 				/*
4502 				 * This is not ".." -- it's "..[mumble]".
4503 				 * We'll store the "/.." and this character
4504 				 * and continue processing.
4505 				 */
4506 				dest[j++] = '/';
4507 				dest[j++] = '.';
4508 				dest[j++] = '.';
4509 				dest[j++] = c;
4510 				continue;
4511 			}
4512 
4513 			/*
4514 			 * This is "/../" or "/..\0".  We need to back up
4515 			 * our destination pointer until we find a "/".
4516 			 */
4517 			i--;
4518 			while (j != 0 && dest[--j] != '/')
4519 				continue;
4520 
4521 			if (c == '\0')
4522 				dest[++j] = '/';
4523 		} while (c != '\0');
4524 
4525 		dest[j] = '\0';
4526 		regs[rd] = (uintptr_t)dest;
4527 		mstate->dtms_scratch_ptr += size;
4528 		break;
4529 	}
4530 
4531 	case DIF_SUBR_INET_NTOA:
4532 	case DIF_SUBR_INET_NTOA6:
4533 	case DIF_SUBR_INET_NTOP: {
4534 		size_t size;
4535 		int af, argi, i;
4536 		char *base, *end;
4537 
4538 		if (subr == DIF_SUBR_INET_NTOP) {
4539 			af = (int)tupregs[0].dttk_value;
4540 			argi = 1;
4541 		} else {
4542 			af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
4543 			argi = 0;
4544 		}
4545 
4546 		if (af == AF_INET) {
4547 			ipaddr_t ip4;
4548 			uint8_t *ptr8, val;
4549 
4550 			/*
4551 			 * Safely load the IPv4 address.
4552 			 */
4553 			ip4 = dtrace_load32(tupregs[argi].dttk_value);
4554 
4555 			/*
4556 			 * Check an IPv4 string will fit in scratch.
4557 			 */
4558 			size = INET_ADDRSTRLEN;
4559 			if (!DTRACE_INSCRATCH(mstate, size)) {
4560 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4561 				regs[rd] = NULL;
4562 				break;
4563 			}
4564 			base = (char *)mstate->dtms_scratch_ptr;
4565 			end = (char *)mstate->dtms_scratch_ptr + size - 1;
4566 
4567 			/*
4568 			 * Stringify as a dotted decimal quad.
4569 			 */
4570 			*end-- = '\0';
4571 			ptr8 = (uint8_t *)&ip4;
4572 			for (i = 3; i >= 0; i--) {
4573 				val = ptr8[i];
4574 
4575 				if (val == 0) {
4576 					*end-- = '0';
4577 				} else {
4578 					for (; val; val /= 10) {
4579 						*end-- = '0' + (val % 10);
4580 					}
4581 				}
4582 
4583 				if (i > 0)
4584 					*end-- = '.';
4585 			}
4586 			ASSERT(end + 1 >= base);
4587 
4588 		} else if (af == AF_INET6) {
4589 			struct in6_addr ip6;
4590 			int firstzero, tryzero, numzero, v6end;
4591 			uint16_t val;
4592 			const char digits[] = "0123456789abcdef";
4593 
4594 			/*
4595 			 * Stringify using RFC 1884 convention 2 - 16 bit
4596 			 * hexadecimal values with a zero-run compression.
4597 			 * Lower case hexadecimal digits are used.
4598 			 * 	eg, fe80::214:4fff:fe0b:76c8.
4599 			 * The IPv4 embedded form is returned for inet_ntop,
4600 			 * just the IPv4 string is returned for inet_ntoa6.
4601 			 */
4602 
4603 			/*
4604 			 * Safely load the IPv6 address.
4605 			 */
4606 			dtrace_bcopy(
4607 			    (void *)(uintptr_t)tupregs[argi].dttk_value,
4608 			    (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));
4609 
4610 			/*
4611 			 * Check an IPv6 string will fit in scratch.
4612 			 */
4613 			size = INET6_ADDRSTRLEN;
4614 			if (!DTRACE_INSCRATCH(mstate, size)) {
4615 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4616 				regs[rd] = NULL;
4617 				break;
4618 			}
4619 			base = (char *)mstate->dtms_scratch_ptr;
4620 			end = (char *)mstate->dtms_scratch_ptr + size - 1;
4621 			*end-- = '\0';
4622 
4623 			/*
4624 			 * Find the longest run of 16 bit zero values
4625 			 * for the single allowed zero compression - "::".
4626 			 */
4627 			firstzero = -1;
4628 			tryzero = -1;
4629 			numzero = 1;
4630 			for (i = 0; i < sizeof (struct in6_addr); i++) {
4631 				if (ip6._S6_un._S6_u8[i] == 0 &&
4632 				    tryzero == -1 && i % 2 == 0) {
4633 					tryzero = i;
4634 					continue;
4635 				}
4636 
4637 				if (tryzero != -1 &&
4638 				    (ip6._S6_un._S6_u8[i] != 0 ||
4639 				    i == sizeof (struct in6_addr) - 1)) {
4640 
4641 					if (i - tryzero <= numzero) {
4642 						tryzero = -1;
4643 						continue;
4644 					}
4645 
4646 					firstzero = tryzero;
4647 					numzero = i - i % 2 - tryzero;
4648 					tryzero = -1;
4649 
4650 					if (ip6._S6_un._S6_u8[i] == 0 &&
4651 					    i == sizeof (struct in6_addr) - 1)
4652 						numzero += 2;
4653 				}
4654 			}
4655 			ASSERT(firstzero + numzero <= sizeof (struct in6_addr));
4656 
4657 			/*
4658 			 * Check for an IPv4 embedded address.
4659 			 */
4660 			v6end = sizeof (struct in6_addr) - 2;
4661 			if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
4662 			    IN6_IS_ADDR_V4COMPAT(&ip6)) {
4663 				for (i = sizeof (struct in6_addr) - 1;
4664 				    i >= DTRACE_V4MAPPED_OFFSET; i--) {
4665 					ASSERT(end >= base);
4666 
4667 					val = ip6._S6_un._S6_u8[i];
4668 
4669 					if (val == 0) {
4670 						*end-- = '0';
4671 					} else {
4672 						for (; val; val /= 10) {
4673 							*end-- = '0' + val % 10;
4674 						}
4675 					}
4676 
4677 					if (i > DTRACE_V4MAPPED_OFFSET)
4678 						*end-- = '.';
4679 				}
4680 
4681 				if (subr == DIF_SUBR_INET_NTOA6)
4682 					goto inetout;
4683 
4684 				/*
4685 				 * Set v6end to skip the IPv4 address that
4686 				 * we have already stringified.
4687 				 */
4688 				v6end = 10;
4689 			}
4690 
4691 			/*
4692 			 * Build the IPv6 string by working through the
4693 			 * address in reverse.
4694 			 */
4695 			for (i = v6end; i >= 0; i -= 2) {
4696 				ASSERT(end >= base);
4697 
4698 				if (i == firstzero + numzero - 2) {
4699 					*end-- = ':';
4700 					*end-- = ':';
4701 					i -= numzero - 2;
4702 					continue;
4703 				}
4704 
4705 				if (i < 14 && i != firstzero - 2)
4706 					*end-- = ':';
4707 
4708 				val = (ip6._S6_un._S6_u8[i] << 8) +
4709 				    ip6._S6_un._S6_u8[i + 1];
4710 
4711 				if (val == 0) {
4712 					*end-- = '0';
4713 				} else {
4714 					for (; val; val /= 16) {
4715 						*end-- = digits[val % 16];
4716 					}
4717 				}
4718 			}
4719 			ASSERT(end + 1 >= base);
4720 
4721 		} else {
4722 			/*
4723 			 * The user didn't use AH_INET or AH_INET6.
4724 			 */
4725 			DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
4726 			regs[rd] = NULL;
4727 			break;
4728 		}
4729 
4730 inetout:	regs[rd] = (uintptr_t)end + 1;
4731 		mstate->dtms_scratch_ptr += size;
4732 		break;
4733 	}
4734 
4735 	}
4736 }
4737 
4738 /*
4739  * Emulate the execution of DTrace IR instructions specified by the given
4740  * DIF object.  This function is deliberately void of assertions as all of
4741  * the necessary checks are handled by a call to dtrace_difo_validate().
4742  */
4743 static uint64_t
4744 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
4745     dtrace_vstate_t *vstate, dtrace_state_t *state)
4746 {
4747 	const dif_instr_t *text = difo->dtdo_buf;
4748 	const uint_t textlen = difo->dtdo_len;
4749 	const char *strtab = difo->dtdo_strtab;
4750 	const uint64_t *inttab = difo->dtdo_inttab;
4751 
4752 	uint64_t rval = 0;
4753 	dtrace_statvar_t *svar;
4754 	dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
4755 	dtrace_difv_t *v;
4756 	volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
4757 	volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
4758 
4759 	dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
4760 	uint64_t regs[DIF_DIR_NREGS];
4761 	uint64_t *tmp;
4762 
4763 	uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
4764 	int64_t cc_r;
4765 	uint_t pc = 0, id, opc;
4766 	uint8_t ttop = 0;
4767 	dif_instr_t instr;
4768 	uint_t r1, r2, rd;
4769 
4770 	/*
4771 	 * We stash the current DIF object into the machine state: we need it
4772 	 * for subsequent access checking.
4773 	 */
4774 	mstate->dtms_difo = difo;
4775 
4776 	regs[DIF_REG_R0] = 0; 		/* %r0 is fixed at zero */
4777 
4778 	while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
4779 		opc = pc;
4780 
4781 		instr = text[pc++];
4782 		r1 = DIF_INSTR_R1(instr);
4783 		r2 = DIF_INSTR_R2(instr);
4784 		rd = DIF_INSTR_RD(instr);
4785 
4786 		switch (DIF_INSTR_OP(instr)) {
4787 		case DIF_OP_OR:
4788 			regs[rd] = regs[r1] | regs[r2];
4789 			break;
4790 		case DIF_OP_XOR:
4791 			regs[rd] = regs[r1] ^ regs[r2];
4792 			break;
4793 		case DIF_OP_AND:
4794 			regs[rd] = regs[r1] & regs[r2];
4795 			break;
4796 		case DIF_OP_SLL:
4797 			regs[rd] = regs[r1] << regs[r2];
4798 			break;
4799 		case DIF_OP_SRL:
4800 			regs[rd] = regs[r1] >> regs[r2];
4801 			break;
4802 		case DIF_OP_SUB:
4803 			regs[rd] = regs[r1] - regs[r2];
4804 			break;
4805 		case DIF_OP_ADD:
4806 			regs[rd] = regs[r1] + regs[r2];
4807 			break;
4808 		case DIF_OP_MUL:
4809 			regs[rd] = regs[r1] * regs[r2];
4810 			break;
4811 		case DIF_OP_SDIV:
4812 			if (regs[r2] == 0) {
4813 				regs[rd] = 0;
4814 				*flags |= CPU_DTRACE_DIVZERO;
4815 			} else {
4816 				regs[rd] = (int64_t)regs[r1] /
4817 				    (int64_t)regs[r2];
4818 			}
4819 			break;
4820 
4821 		case DIF_OP_UDIV:
4822 			if (regs[r2] == 0) {
4823 				regs[rd] = 0;
4824 				*flags |= CPU_DTRACE_DIVZERO;
4825 			} else {
4826 				regs[rd] = regs[r1] / regs[r2];
4827 			}
4828 			break;
4829 
4830 		case DIF_OP_SREM:
4831 			if (regs[r2] == 0) {
4832 				regs[rd] = 0;
4833 				*flags |= CPU_DTRACE_DIVZERO;
4834 			} else {
4835 				regs[rd] = (int64_t)regs[r1] %
4836 				    (int64_t)regs[r2];
4837 			}
4838 			break;
4839 
4840 		case DIF_OP_UREM:
4841 			if (regs[r2] == 0) {
4842 				regs[rd] = 0;
4843 				*flags |= CPU_DTRACE_DIVZERO;
4844 			} else {
4845 				regs[rd] = regs[r1] % regs[r2];
4846 			}
4847 			break;
4848 
4849 		case DIF_OP_NOT:
4850 			regs[rd] = ~regs[r1];
4851 			break;
4852 		case DIF_OP_MOV:
4853 			regs[rd] = regs[r1];
4854 			break;
4855 		case DIF_OP_CMP:
4856 			cc_r = regs[r1] - regs[r2];
4857 			cc_n = cc_r < 0;
4858 			cc_z = cc_r == 0;
4859 			cc_v = 0;
4860 			cc_c = regs[r1] < regs[r2];
4861 			break;
4862 		case DIF_OP_TST:
4863 			cc_n = cc_v = cc_c = 0;
4864 			cc_z = regs[r1] == 0;
4865 			break;
4866 		case DIF_OP_BA:
4867 			pc = DIF_INSTR_LABEL(instr);
4868 			break;
4869 		case DIF_OP_BE:
4870 			if (cc_z)
4871 				pc = DIF_INSTR_LABEL(instr);
4872 			break;
4873 		case DIF_OP_BNE:
4874 			if (cc_z == 0)
4875 				pc = DIF_INSTR_LABEL(instr);
4876 			break;
4877 		case DIF_OP_BG:
4878 			if ((cc_z | (cc_n ^ cc_v)) == 0)
4879 				pc = DIF_INSTR_LABEL(instr);
4880 			break;
4881 		case DIF_OP_BGU:
4882 			if ((cc_c | cc_z) == 0)
4883 				pc = DIF_INSTR_LABEL(instr);
4884 			break;
4885 		case DIF_OP_BGE:
4886 			if ((cc_n ^ cc_v) == 0)
4887 				pc = DIF_INSTR_LABEL(instr);
4888 			break;
4889 		case DIF_OP_BGEU:
4890 			if (cc_c == 0)
4891 				pc = DIF_INSTR_LABEL(instr);
4892 			break;
4893 		case DIF_OP_BL:
4894 			if (cc_n ^ cc_v)
4895 				pc = DIF_INSTR_LABEL(instr);
4896 			break;
4897 		case DIF_OP_BLU:
4898 			if (cc_c)
4899 				pc = DIF_INSTR_LABEL(instr);
4900 			break;
4901 		case DIF_OP_BLE:
4902 			if (cc_z | (cc_n ^ cc_v))
4903 				pc = DIF_INSTR_LABEL(instr);
4904 			break;
4905 		case DIF_OP_BLEU:
4906 			if (cc_c | cc_z)
4907 				pc = DIF_INSTR_LABEL(instr);
4908 			break;
4909 		case DIF_OP_RLDSB:
4910 			if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
4911 				*flags |= CPU_DTRACE_KPRIV;
4912 				*illval = regs[r1];
4913 				break;
4914 			}
4915 			/*FALLTHROUGH*/
4916 		case DIF_OP_LDSB:
4917 			regs[rd] = (int8_t)dtrace_load8(regs[r1]);
4918 			break;
4919 		case DIF_OP_RLDSH:
4920 			if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
4921 				*flags |= CPU_DTRACE_KPRIV;
4922 				*illval = regs[r1];
4923 				break;
4924 			}
4925 			/*FALLTHROUGH*/
4926 		case DIF_OP_LDSH:
4927 			regs[rd] = (int16_t)dtrace_load16(regs[r1]);
4928 			break;
4929 		case DIF_OP_RLDSW:
4930 			if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
4931 				*flags |= CPU_DTRACE_KPRIV;
4932 				*illval = regs[r1];
4933 				break;
4934 			}
4935 			/*FALLTHROUGH*/
4936 		case DIF_OP_LDSW:
4937 			regs[rd] = (int32_t)dtrace_load32(regs[r1]);
4938 			break;
4939 		case DIF_OP_RLDUB:
4940 			if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
4941 				*flags |= CPU_DTRACE_KPRIV;
4942 				*illval = regs[r1];
4943 				break;
4944 			}
4945 			/*FALLTHROUGH*/
4946 		case DIF_OP_LDUB:
4947 			regs[rd] = dtrace_load8(regs[r1]);
4948 			break;
4949 		case DIF_OP_RLDUH:
4950 			if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
4951 				*flags |= CPU_DTRACE_KPRIV;
4952 				*illval = regs[r1];
4953 				break;
4954 			}
4955 			/*FALLTHROUGH*/
4956 		case DIF_OP_LDUH:
4957 			regs[rd] = dtrace_load16(regs[r1]);
4958 			break;
4959 		case DIF_OP_RLDUW:
4960 			if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
4961 				*flags |= CPU_DTRACE_KPRIV;
4962 				*illval = regs[r1];
4963 				break;
4964 			}
4965 			/*FALLTHROUGH*/
4966 		case DIF_OP_LDUW:
4967 			regs[rd] = dtrace_load32(regs[r1]);
4968 			break;
4969 		case DIF_OP_RLDX:
4970 			if (!dtrace_canstore(regs[r1], 8, mstate, vstate)) {
4971 				*flags |= CPU_DTRACE_KPRIV;
4972 				*illval = regs[r1];
4973 				break;
4974 			}
4975 			/*FALLTHROUGH*/
4976 		case DIF_OP_LDX:
4977 			regs[rd] = dtrace_load64(regs[r1]);
4978 			break;
4979 		case DIF_OP_ULDSB:
4980 			regs[rd] = (int8_t)
4981 			    dtrace_fuword8((void *)(uintptr_t)regs[r1]);
4982 			break;
4983 		case DIF_OP_ULDSH:
4984 			regs[rd] = (int16_t)
4985 			    dtrace_fuword16((void *)(uintptr_t)regs[r1]);
4986 			break;
4987 		case DIF_OP_ULDSW:
4988 			regs[rd] = (int32_t)
4989 			    dtrace_fuword32((void *)(uintptr_t)regs[r1]);
4990 			break;
4991 		case DIF_OP_ULDUB:
4992 			regs[rd] =
4993 			    dtrace_fuword8((void *)(uintptr_t)regs[r1]);
4994 			break;
4995 		case DIF_OP_ULDUH:
4996 			regs[rd] =
4997 			    dtrace_fuword16((void *)(uintptr_t)regs[r1]);
4998 			break;
4999 		case DIF_OP_ULDUW:
5000 			regs[rd] =
5001 			    dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5002 			break;
5003 		case DIF_OP_ULDX:
5004 			regs[rd] =
5005 			    dtrace_fuword64((void *)(uintptr_t)regs[r1]);
5006 			break;
5007 		case DIF_OP_RET:
5008 			rval = regs[rd];
5009 			pc = textlen;
5010 			break;
5011 		case DIF_OP_NOP:
5012 			break;
5013 		case DIF_OP_SETX:
5014 			regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
5015 			break;
5016 		case DIF_OP_SETS:
5017 			regs[rd] = (uint64_t)(uintptr_t)
5018 			    (strtab + DIF_INSTR_STRING(instr));
5019 			break;
5020 		case DIF_OP_SCMP: {
5021 			size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
5022 			uintptr_t s1 = regs[r1];
5023 			uintptr_t s2 = regs[r2];
5024 
5025 			if (s1 != NULL &&
5026 			    !dtrace_strcanload(s1, sz, mstate, vstate))
5027 				break;
5028 			if (s2 != NULL &&
5029 			    !dtrace_strcanload(s2, sz, mstate, vstate))
5030 				break;
5031 
5032 			cc_r = dtrace_strncmp((char *)s1, (char *)s2, sz);
5033 
5034 			cc_n = cc_r < 0;
5035 			cc_z = cc_r == 0;
5036 			cc_v = cc_c = 0;
5037 			break;
5038 		}
5039 		case DIF_OP_LDGA:
5040 			regs[rd] = dtrace_dif_variable(mstate, state,
5041 			    r1, regs[r2]);
5042 			break;
5043 		case DIF_OP_LDGS:
5044 			id = DIF_INSTR_VAR(instr);
5045 
5046 			if (id >= DIF_VAR_OTHER_UBASE) {
5047 				uintptr_t a;
5048 
5049 				id -= DIF_VAR_OTHER_UBASE;
5050 				svar = vstate->dtvs_globals[id];
5051 				ASSERT(svar != NULL);
5052 				v = &svar->dtsv_var;
5053 
5054 				if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
5055 					regs[rd] = svar->dtsv_data;
5056 					break;
5057 				}
5058 
5059 				a = (uintptr_t)svar->dtsv_data;
5060 
5061 				if (*(uint8_t *)a == UINT8_MAX) {
5062 					/*
5063 					 * If the 0th byte is set to UINT8_MAX
5064 					 * then this is to be treated as a
5065 					 * reference to a NULL variable.
5066 					 */
5067 					regs[rd] = NULL;
5068 				} else {
5069 					regs[rd] = a + sizeof (uint64_t);
5070 				}
5071 
5072 				break;
5073 			}
5074 
5075 			regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
5076 			break;
5077 
5078 		case DIF_OP_STGS:
5079 			id = DIF_INSTR_VAR(instr);
5080 
5081 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5082 			id -= DIF_VAR_OTHER_UBASE;
5083 
5084 			svar = vstate->dtvs_globals[id];
5085 			ASSERT(svar != NULL);
5086 			v = &svar->dtsv_var;
5087 
5088 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5089 				uintptr_t a = (uintptr_t)svar->dtsv_data;
5090 
5091 				ASSERT(a != NULL);
5092 				ASSERT(svar->dtsv_size != 0);
5093 
5094 				if (regs[rd] == NULL) {
5095 					*(uint8_t *)a = UINT8_MAX;
5096 					break;
5097 				} else {
5098 					*(uint8_t *)a = 0;
5099 					a += sizeof (uint64_t);
5100 				}
5101 				if (!dtrace_vcanload(
5102 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5103 				    mstate, vstate))
5104 					break;
5105 
5106 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
5107 				    (void *)a, &v->dtdv_type);
5108 				break;
5109 			}
5110 
5111 			svar->dtsv_data = regs[rd];
5112 			break;
5113 
5114 		case DIF_OP_LDTA:
5115 			/*
5116 			 * There are no DTrace built-in thread-local arrays at
5117 			 * present.  This opcode is saved for future work.
5118 			 */
5119 			*flags |= CPU_DTRACE_ILLOP;
5120 			regs[rd] = 0;
5121 			break;
5122 
5123 		case DIF_OP_LDLS:
5124 			id = DIF_INSTR_VAR(instr);
5125 
5126 			if (id < DIF_VAR_OTHER_UBASE) {
5127 				/*
5128 				 * For now, this has no meaning.
5129 				 */
5130 				regs[rd] = 0;
5131 				break;
5132 			}
5133 
5134 			id -= DIF_VAR_OTHER_UBASE;
5135 
5136 			ASSERT(id < vstate->dtvs_nlocals);
5137 			ASSERT(vstate->dtvs_locals != NULL);
5138 
5139 			svar = vstate->dtvs_locals[id];
5140 			ASSERT(svar != NULL);
5141 			v = &svar->dtsv_var;
5142 
5143 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5144 				uintptr_t a = (uintptr_t)svar->dtsv_data;
5145 				size_t sz = v->dtdv_type.dtdt_size;
5146 
5147 				sz += sizeof (uint64_t);
5148 				ASSERT(svar->dtsv_size == NCPU * sz);
5149 				a += CPU->cpu_id * sz;
5150 
5151 				if (*(uint8_t *)a == UINT8_MAX) {
5152 					/*
5153 					 * If the 0th byte is set to UINT8_MAX
5154 					 * then this is to be treated as a
5155 					 * reference to a NULL variable.
5156 					 */
5157 					regs[rd] = NULL;
5158 				} else {
5159 					regs[rd] = a + sizeof (uint64_t);
5160 				}
5161 
5162 				break;
5163 			}
5164 
5165 			ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5166 			tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5167 			regs[rd] = tmp[CPU->cpu_id];
5168 			break;
5169 
5170 		case DIF_OP_STLS:
5171 			id = DIF_INSTR_VAR(instr);
5172 
5173 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5174 			id -= DIF_VAR_OTHER_UBASE;
5175 			ASSERT(id < vstate->dtvs_nlocals);
5176 
5177 			ASSERT(vstate->dtvs_locals != NULL);
5178 			svar = vstate->dtvs_locals[id];
5179 			ASSERT(svar != NULL);
5180 			v = &svar->dtsv_var;
5181 
5182 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5183 				uintptr_t a = (uintptr_t)svar->dtsv_data;
5184 				size_t sz = v->dtdv_type.dtdt_size;
5185 
5186 				sz += sizeof (uint64_t);
5187 				ASSERT(svar->dtsv_size == NCPU * sz);
5188 				a += CPU->cpu_id * sz;
5189 
5190 				if (regs[rd] == NULL) {
5191 					*(uint8_t *)a = UINT8_MAX;
5192 					break;
5193 				} else {
5194 					*(uint8_t *)a = 0;
5195 					a += sizeof (uint64_t);
5196 				}
5197 
5198 				if (!dtrace_vcanload(
5199 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5200 				    mstate, vstate))
5201 					break;
5202 
5203 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
5204 				    (void *)a, &v->dtdv_type);
5205 				break;
5206 			}
5207 
5208 			ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5209 			tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5210 			tmp[CPU->cpu_id] = regs[rd];
5211 			break;
5212 
5213 		case DIF_OP_LDTS: {
5214 			dtrace_dynvar_t *dvar;
5215 			dtrace_key_t *key;
5216 
5217 			id = DIF_INSTR_VAR(instr);
5218 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5219 			id -= DIF_VAR_OTHER_UBASE;
5220 			v = &vstate->dtvs_tlocals[id];
5221 
5222 			key = &tupregs[DIF_DTR_NREGS];
5223 			key[0].dttk_value = (uint64_t)id;
5224 			key[0].dttk_size = 0;
5225 			DTRACE_TLS_THRKEY(key[1].dttk_value);
5226 			key[1].dttk_size = 0;
5227 
5228 			dvar = dtrace_dynvar(dstate, 2, key,
5229 			    sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
5230 			    mstate, vstate);
5231 
5232 			if (dvar == NULL) {
5233 				regs[rd] = 0;
5234 				break;
5235 			}
5236 
5237 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5238 				regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5239 			} else {
5240 				regs[rd] = *((uint64_t *)dvar->dtdv_data);
5241 			}
5242 
5243 			break;
5244 		}
5245 
5246 		case DIF_OP_STTS: {
5247 			dtrace_dynvar_t *dvar;
5248 			dtrace_key_t *key;
5249 
5250 			id = DIF_INSTR_VAR(instr);
5251 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5252 			id -= DIF_VAR_OTHER_UBASE;
5253 
5254 			key = &tupregs[DIF_DTR_NREGS];
5255 			key[0].dttk_value = (uint64_t)id;
5256 			key[0].dttk_size = 0;
5257 			DTRACE_TLS_THRKEY(key[1].dttk_value);
5258 			key[1].dttk_size = 0;
5259 			v = &vstate->dtvs_tlocals[id];
5260 
5261 			dvar = dtrace_dynvar(dstate, 2, key,
5262 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5263 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
5264 			    regs[rd] ? DTRACE_DYNVAR_ALLOC :
5265 			    DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5266 
5267 			/*
5268 			 * Given that we're storing to thread-local data,
5269 			 * we need to flush our predicate cache.
5270 			 */
5271 			curthread->t_predcache = NULL;
5272 
5273 			if (dvar == NULL)
5274 				break;
5275 
5276 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5277 				if (!dtrace_vcanload(
5278 				    (void *)(uintptr_t)regs[rd],
5279 				    &v->dtdv_type, mstate, vstate))
5280 					break;
5281 
5282 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
5283 				    dvar->dtdv_data, &v->dtdv_type);
5284 			} else {
5285 				*((uint64_t *)dvar->dtdv_data) = regs[rd];
5286 			}
5287 
5288 			break;
5289 		}
5290 
5291 		case DIF_OP_SRA:
5292 			regs[rd] = (int64_t)regs[r1] >> regs[r2];
5293 			break;
5294 
5295 		case DIF_OP_CALL:
5296 			dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
5297 			    regs, tupregs, ttop, mstate, state);
5298 			break;
5299 
5300 		case DIF_OP_PUSHTR:
5301 			if (ttop == DIF_DTR_NREGS) {
5302 				*flags |= CPU_DTRACE_TUPOFLOW;
5303 				break;
5304 			}
5305 
5306 			if (r1 == DIF_TYPE_STRING) {
5307 				/*
5308 				 * If this is a string type and the size is 0,
5309 				 * we'll use the system-wide default string
5310 				 * size.  Note that we are _not_ looking at
5311 				 * the value of the DTRACEOPT_STRSIZE option;
5312 				 * had this been set, we would expect to have
5313 				 * a non-zero size value in the "pushtr".
5314 				 */
5315 				tupregs[ttop].dttk_size =
5316 				    dtrace_strlen((char *)(uintptr_t)regs[rd],
5317 				    regs[r2] ? regs[r2] :
5318 				    dtrace_strsize_default) + 1;
5319 			} else {
5320 				tupregs[ttop].dttk_size = regs[r2];
5321 			}
5322 
5323 			tupregs[ttop++].dttk_value = regs[rd];
5324 			break;
5325 
5326 		case DIF_OP_PUSHTV:
5327 			if (ttop == DIF_DTR_NREGS) {
5328 				*flags |= CPU_DTRACE_TUPOFLOW;
5329 				break;
5330 			}
5331 
5332 			tupregs[ttop].dttk_value = regs[rd];
5333 			tupregs[ttop++].dttk_size = 0;
5334 			break;
5335 
5336 		case DIF_OP_POPTS:
5337 			if (ttop != 0)
5338 				ttop--;
5339 			break;
5340 
5341 		case DIF_OP_FLUSHTS:
5342 			ttop = 0;
5343 			break;
5344 
5345 		case DIF_OP_LDGAA:
5346 		case DIF_OP_LDTAA: {
5347 			dtrace_dynvar_t *dvar;
5348 			dtrace_key_t *key = tupregs;
5349 			uint_t nkeys = ttop;
5350 
5351 			id = DIF_INSTR_VAR(instr);
5352 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5353 			id -= DIF_VAR_OTHER_UBASE;
5354 
5355 			key[nkeys].dttk_value = (uint64_t)id;
5356 			key[nkeys++].dttk_size = 0;
5357 
5358 			if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
5359 				DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5360 				key[nkeys++].dttk_size = 0;
5361 				v = &vstate->dtvs_tlocals[id];
5362 			} else {
5363 				v = &vstate->dtvs_globals[id]->dtsv_var;
5364 			}
5365 
5366 			dvar = dtrace_dynvar(dstate, nkeys, key,
5367 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5368 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
5369 			    DTRACE_DYNVAR_NOALLOC, mstate, vstate);
5370 
5371 			if (dvar == NULL) {
5372 				regs[rd] = 0;
5373 				break;
5374 			}
5375 
5376 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5377 				regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5378 			} else {
5379 				regs[rd] = *((uint64_t *)dvar->dtdv_data);
5380 			}
5381 
5382 			break;
5383 		}
5384 
5385 		case DIF_OP_STGAA:
5386 		case DIF_OP_STTAA: {
5387 			dtrace_dynvar_t *dvar;
5388 			dtrace_key_t *key = tupregs;
5389 			uint_t nkeys = ttop;
5390 
5391 			id = DIF_INSTR_VAR(instr);
5392 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5393 			id -= DIF_VAR_OTHER_UBASE;
5394 
5395 			key[nkeys].dttk_value = (uint64_t)id;
5396 			key[nkeys++].dttk_size = 0;
5397 
5398 			if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
5399 				DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5400 				key[nkeys++].dttk_size = 0;
5401 				v = &vstate->dtvs_tlocals[id];
5402 			} else {
5403 				v = &vstate->dtvs_globals[id]->dtsv_var;
5404 			}
5405 
5406 			dvar = dtrace_dynvar(dstate, nkeys, key,
5407 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5408 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
5409 			    regs[rd] ? DTRACE_DYNVAR_ALLOC :
5410 			    DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5411 
5412 			if (dvar == NULL)
5413 				break;
5414 
5415 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5416 				if (!dtrace_vcanload(
5417 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5418 				    mstate, vstate))
5419 					break;
5420 
5421 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
5422 				    dvar->dtdv_data, &v->dtdv_type);
5423 			} else {
5424 				*((uint64_t *)dvar->dtdv_data) = regs[rd];
5425 			}
5426 
5427 			break;
5428 		}
5429 
5430 		case DIF_OP_ALLOCS: {
5431 			uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5432 			size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
5433 
5434 			/*
5435 			 * Rounding up the user allocation size could have
5436 			 * overflowed large, bogus allocations (like -1ULL) to
5437 			 * 0.
5438 			 */
5439 			if (size < regs[r1] ||
5440 			    !DTRACE_INSCRATCH(mstate, size)) {
5441 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5442 				regs[rd] = NULL;
5443 				break;
5444 			}
5445 
5446 			dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
5447 			mstate->dtms_scratch_ptr += size;
5448 			regs[rd] = ptr;
5449 			break;
5450 		}
5451 
5452 		case DIF_OP_COPYS:
5453 			if (!dtrace_canstore(regs[rd], regs[r2],
5454 			    mstate, vstate)) {
5455 				*flags |= CPU_DTRACE_BADADDR;
5456 				*illval = regs[rd];
5457 				break;
5458 			}
5459 
5460 			if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
5461 				break;
5462 
5463 			dtrace_bcopy((void *)(uintptr_t)regs[r1],
5464 			    (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
5465 			break;
5466 
5467 		case DIF_OP_STB:
5468 			if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
5469 				*flags |= CPU_DTRACE_BADADDR;
5470 				*illval = regs[rd];
5471 				break;
5472 			}
5473 			*((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
5474 			break;
5475 
5476 		case DIF_OP_STH:
5477 			if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
5478 				*flags |= CPU_DTRACE_BADADDR;
5479 				*illval = regs[rd];
5480 				break;
5481 			}
5482 			if (regs[rd] & 1) {
5483 				*flags |= CPU_DTRACE_BADALIGN;
5484 				*illval = regs[rd];
5485 				break;
5486 			}
5487 			*((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
5488 			break;
5489 
5490 		case DIF_OP_STW:
5491 			if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
5492 				*flags |= CPU_DTRACE_BADADDR;
5493 				*illval = regs[rd];
5494 				break;
5495 			}
5496 			if (regs[rd] & 3) {
5497 				*flags |= CPU_DTRACE_BADALIGN;
5498 				*illval = regs[rd];
5499 				break;
5500 			}
5501 			*((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
5502 			break;
5503 
5504 		case DIF_OP_STX:
5505 			if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
5506 				*flags |= CPU_DTRACE_BADADDR;
5507 				*illval = regs[rd];
5508 				break;
5509 			}
5510 			if (regs[rd] & 7) {
5511 				*flags |= CPU_DTRACE_BADALIGN;
5512 				*illval = regs[rd];
5513 				break;
5514 			}
5515 			*((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
5516 			break;
5517 		}
5518 	}
5519 
5520 	if (!(*flags & CPU_DTRACE_FAULT))
5521 		return (rval);
5522 
5523 	mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
5524 	mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
5525 
5526 	return (0);
5527 }
5528 
5529 static void
5530 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
5531 {
5532 	dtrace_probe_t *probe = ecb->dte_probe;
5533 	dtrace_provider_t *prov = probe->dtpr_provider;
5534 	char c[DTRACE_FULLNAMELEN + 80], *str;
5535 	char *msg = "dtrace: breakpoint action at probe ";
5536 	char *ecbmsg = " (ecb ";
5537 	uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
5538 	uintptr_t val = (uintptr_t)ecb;
5539 	int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
5540 
5541 	if (dtrace_destructive_disallow)
5542 		return;
5543 
5544 	/*
5545 	 * It's impossible to be taking action on the NULL probe.
5546 	 */
5547 	ASSERT(probe != NULL);
5548 
5549 	/*
5550 	 * This is a poor man's (destitute man's?) sprintf():  we want to
5551 	 * print the provider name, module name, function name and name of
5552 	 * the probe, along with the hex address of the ECB with the breakpoint
5553 	 * action -- all of which we must place in the character buffer by
5554 	 * hand.
5555 	 */
5556 	while (*msg != '\0')
5557 		c[i++] = *msg++;
5558 
5559 	for (str = prov->dtpv_name; *str != '\0'; str++)
5560 		c[i++] = *str;
5561 	c[i++] = ':';
5562 
5563 	for (str = probe->dtpr_mod; *str != '\0'; str++)
5564 		c[i++] = *str;
5565 	c[i++] = ':';
5566 
5567 	for (str = probe->dtpr_func; *str != '\0'; str++)
5568 		c[i++] = *str;
5569 	c[i++] = ':';
5570 
5571 	for (str = probe->dtpr_name; *str != '\0'; str++)
5572 		c[i++] = *str;
5573 
5574 	while (*ecbmsg != '\0')
5575 		c[i++] = *ecbmsg++;
5576 
5577 	while (shift >= 0) {
5578 		mask = (uintptr_t)0xf << shift;
5579 
5580 		if (val >= ((uintptr_t)1 << shift))
5581 			c[i++] = "0123456789abcdef"[(val & mask) >> shift];
5582 		shift -= 4;
5583 	}
5584 
5585 	c[i++] = ')';
5586 	c[i] = '\0';
5587 
5588 	debug_enter(c);
5589 }
5590 
5591 static void
5592 dtrace_action_panic(dtrace_ecb_t *ecb)
5593 {
5594 	dtrace_probe_t *probe = ecb->dte_probe;
5595 
5596 	/*
5597 	 * It's impossible to be taking action on the NULL probe.
5598 	 */
5599 	ASSERT(probe != NULL);
5600 
5601 	if (dtrace_destructive_disallow)
5602 		return;
5603 
5604 	if (dtrace_panicked != NULL)
5605 		return;
5606 
5607 	if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
5608 		return;
5609 
5610 	/*
5611 	 * We won the right to panic.  (We want to be sure that only one
5612 	 * thread calls panic() from dtrace_probe(), and that panic() is
5613 	 * called exactly once.)
5614 	 */
5615 	dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
5616 	    probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
5617 	    probe->dtpr_func, probe->dtpr_name, (void *)ecb);
5618 }
5619 
5620 static void
5621 dtrace_action_raise(uint64_t sig)
5622 {
5623 	if (dtrace_destructive_disallow)
5624 		return;
5625 
5626 	if (sig >= NSIG) {
5627 		DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
5628 		return;
5629 	}
5630 
5631 	/*
5632 	 * raise() has a queue depth of 1 -- we ignore all subsequent
5633 	 * invocations of the raise() action.
5634 	 */
5635 	if (curthread->t_dtrace_sig == 0)
5636 		curthread->t_dtrace_sig = (uint8_t)sig;
5637 
5638 	curthread->t_sig_check = 1;
5639 	aston(curthread);
5640 }
5641 
5642 static void
5643 dtrace_action_stop(void)
5644 {
5645 	if (dtrace_destructive_disallow)
5646 		return;
5647 
5648 	if (!curthread->t_dtrace_stop) {
5649 		curthread->t_dtrace_stop = 1;
5650 		curthread->t_sig_check = 1;
5651 		aston(curthread);
5652 	}
5653 }
5654 
5655 static void
5656 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
5657 {
5658 	hrtime_t now;
5659 	volatile uint16_t *flags;
5660 	cpu_t *cpu = CPU;
5661 
5662 	if (dtrace_destructive_disallow)
5663 		return;
5664 
5665 	flags = (volatile uint16_t *)&cpu_core[cpu->cpu_id].cpuc_dtrace_flags;
5666 
5667 	now = dtrace_gethrtime();
5668 
5669 	if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
5670 		/*
5671 		 * We need to advance the mark to the current time.
5672 		 */
5673 		cpu->cpu_dtrace_chillmark = now;
5674 		cpu->cpu_dtrace_chilled = 0;
5675 	}
5676 
5677 	/*
5678 	 * Now check to see if the requested chill time would take us over
5679 	 * the maximum amount of time allowed in the chill interval.  (Or
5680 	 * worse, if the calculation itself induces overflow.)
5681 	 */
5682 	if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
5683 	    cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
5684 		*flags |= CPU_DTRACE_ILLOP;
5685 		return;
5686 	}
5687 
5688 	while (dtrace_gethrtime() - now < val)
5689 		continue;
5690 
5691 	/*
5692 	 * Normally, we assure that the value of the variable "timestamp" does
5693 	 * not change within an ECB.  The presence of chill() represents an
5694 	 * exception to this rule, however.
5695 	 */
5696 	mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
5697 	cpu->cpu_dtrace_chilled += val;
5698 }
5699 
5700 static void
5701 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
5702     uint64_t *buf, uint64_t arg)
5703 {
5704 	int nframes = DTRACE_USTACK_NFRAMES(arg);
5705 	int strsize = DTRACE_USTACK_STRSIZE(arg);
5706 	uint64_t *pcs = &buf[1], *fps;
5707 	char *str = (char *)&pcs[nframes];
5708 	int size, offs = 0, i, j;
5709 	uintptr_t old = mstate->dtms_scratch_ptr, saved;
5710 	uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
5711 	char *sym;
5712 
5713 	/*
5714 	 * Should be taking a faster path if string space has not been
5715 	 * allocated.
5716 	 */
5717 	ASSERT(strsize != 0);
5718 
5719 	/*
5720 	 * We will first allocate some temporary space for the frame pointers.
5721 	 */
5722 	fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5723 	size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
5724 	    (nframes * sizeof (uint64_t));
5725 
5726 	if (!DTRACE_INSCRATCH(mstate, size)) {
5727 		/*
5728 		 * Not enough room for our frame pointers -- need to indicate
5729 		 * that we ran out of scratch space.
5730 		 */
5731 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5732 		return;
5733 	}
5734 
5735 	mstate->dtms_scratch_ptr += size;
5736 	saved = mstate->dtms_scratch_ptr;
5737 
5738 	/*
5739 	 * Now get a stack with both program counters and frame pointers.
5740 	 */
5741 	DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5742 	dtrace_getufpstack(buf, fps, nframes + 1);
5743 	DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5744 
5745 	/*
5746 	 * If that faulted, we're cooked.
5747 	 */
5748 	if (*flags & CPU_DTRACE_FAULT)
5749 		goto out;
5750 
5751 	/*
5752 	 * Now we want to walk up the stack, calling the USTACK helper.  For
5753 	 * each iteration, we restore the scratch pointer.
5754 	 */
5755 	for (i = 0; i < nframes; i++) {
5756 		mstate->dtms_scratch_ptr = saved;
5757 
5758 		if (offs >= strsize)
5759 			break;
5760 
5761 		sym = (char *)(uintptr_t)dtrace_helper(
5762 		    DTRACE_HELPER_ACTION_USTACK,
5763 		    mstate, state, pcs[i], fps[i]);
5764 
5765 		/*
5766 		 * If we faulted while running the helper, we're going to
5767 		 * clear the fault and null out the corresponding string.
5768 		 */
5769 		if (*flags & CPU_DTRACE_FAULT) {
5770 			*flags &= ~CPU_DTRACE_FAULT;
5771 			str[offs++] = '\0';
5772 			continue;
5773 		}
5774 
5775 		if (sym == NULL) {
5776 			str[offs++] = '\0';
5777 			continue;
5778 		}
5779 
5780 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5781 
5782 		/*
5783 		 * Now copy in the string that the helper returned to us.
5784 		 */
5785 		for (j = 0; offs + j < strsize; j++) {
5786 			if ((str[offs + j] = sym[j]) == '\0')
5787 				break;
5788 		}
5789 
5790 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5791 
5792 		offs += j + 1;
5793 	}
5794 
5795 	if (offs >= strsize) {
5796 		/*
5797 		 * If we didn't have room for all of the strings, we don't
5798 		 * abort processing -- this needn't be a fatal error -- but we
5799 		 * still want to increment a counter (dts_stkstroverflows) to
5800 		 * allow this condition to be warned about.  (If this is from
5801 		 * a jstack() action, it is easily tuned via jstackstrsize.)
5802 		 */
5803 		dtrace_error(&state->dts_stkstroverflows);
5804 	}
5805 
5806 	while (offs < strsize)
5807 		str[offs++] = '\0';
5808 
5809 out:
5810 	mstate->dtms_scratch_ptr = old;
5811 }
5812 
5813 /*
5814  * If you're looking for the epicenter of DTrace, you just found it.  This
5815  * is the function called by the provider to fire a probe -- from which all
5816  * subsequent probe-context DTrace activity emanates.
5817  */
5818 void
5819 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
5820     uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
5821 {
5822 	processorid_t cpuid;
5823 	dtrace_icookie_t cookie;
5824 	dtrace_probe_t *probe;
5825 	dtrace_mstate_t mstate;
5826 	dtrace_ecb_t *ecb;
5827 	dtrace_action_t *act;
5828 	intptr_t offs;
5829 	size_t size;
5830 	int vtime, onintr;
5831 	volatile uint16_t *flags;
5832 	hrtime_t now;
5833 
5834 	/*
5835 	 * Kick out immediately if this CPU is still being born (in which case
5836 	 * curthread will be set to -1) or the current thread can't allow
5837 	 * probes in its current context.
5838 	 */
5839 	if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
5840 		return;
5841 
5842 	cookie = dtrace_interrupt_disable();
5843 	probe = dtrace_probes[id - 1];
5844 	cpuid = CPU->cpu_id;
5845 	onintr = CPU_ON_INTR(CPU);
5846 
5847 	if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
5848 	    probe->dtpr_predcache == curthread->t_predcache) {
5849 		/*
5850 		 * We have hit in the predicate cache; we know that
5851 		 * this predicate would evaluate to be false.
5852 		 */
5853 		dtrace_interrupt_enable(cookie);
5854 		return;
5855 	}
5856 
5857 	if (panic_quiesce) {
5858 		/*
5859 		 * We don't trace anything if we're panicking.
5860 		 */
5861 		dtrace_interrupt_enable(cookie);
5862 		return;
5863 	}
5864 
5865 	now = dtrace_gethrtime();
5866 	vtime = dtrace_vtime_references != 0;
5867 
5868 	if (vtime && curthread->t_dtrace_start)
5869 		curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
5870 
5871 	mstate.dtms_difo = NULL;
5872 	mstate.dtms_probe = probe;
5873 	mstate.dtms_strtok = NULL;
5874 	mstate.dtms_arg[0] = arg0;
5875 	mstate.dtms_arg[1] = arg1;
5876 	mstate.dtms_arg[2] = arg2;
5877 	mstate.dtms_arg[3] = arg3;
5878 	mstate.dtms_arg[4] = arg4;
5879 
5880 	flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
5881 
5882 	for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
5883 		dtrace_predicate_t *pred = ecb->dte_predicate;
5884 		dtrace_state_t *state = ecb->dte_state;
5885 		dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
5886 		dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
5887 		dtrace_vstate_t *vstate = &state->dts_vstate;
5888 		dtrace_provider_t *prov = probe->dtpr_provider;
5889 		uint64_t tracememsize = 0;
5890 		int committed = 0;
5891 		caddr_t tomax;
5892 
5893 		/*
5894 		 * A little subtlety with the following (seemingly innocuous)
5895 		 * declaration of the automatic 'val':  by looking at the
5896 		 * code, you might think that it could be declared in the
5897 		 * action processing loop, below.  (That is, it's only used in
5898 		 * the action processing loop.)  However, it must be declared
5899 		 * out of that scope because in the case of DIF expression
5900 		 * arguments to aggregating actions, one iteration of the
5901 		 * action loop will use the last iteration's value.
5902 		 */
5903 #ifdef lint
5904 		uint64_t val = 0;
5905 #else
5906 		uint64_t val;
5907 #endif
5908 
5909 		mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
5910 		mstate.dtms_access = DTRACE_ACCESS_ARGS | DTRACE_ACCESS_PROC;
5911 		*flags &= ~CPU_DTRACE_ERROR;
5912 
5913 		if (prov == dtrace_provider) {
5914 			/*
5915 			 * If dtrace itself is the provider of this probe,
5916 			 * we're only going to continue processing the ECB if
5917 			 * arg0 (the dtrace_state_t) is equal to the ECB's
5918 			 * creating state.  (This prevents disjoint consumers
5919 			 * from seeing one another's metaprobes.)
5920 			 */
5921 			if (arg0 != (uint64_t)(uintptr_t)state)
5922 				continue;
5923 		}
5924 
5925 		if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
5926 			/*
5927 			 * We're not currently active.  If our provider isn't
5928 			 * the dtrace pseudo provider, we're not interested.
5929 			 */
5930 			if (prov != dtrace_provider)
5931 				continue;
5932 
5933 			/*
5934 			 * Now we must further check if we are in the BEGIN
5935 			 * probe.  If we are, we will only continue processing
5936 			 * if we're still in WARMUP -- if one BEGIN enabling
5937 			 * has invoked the exit() action, we don't want to
5938 			 * evaluate subsequent BEGIN enablings.
5939 			 */
5940 			if (probe->dtpr_id == dtrace_probeid_begin &&
5941 			    state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
5942 				ASSERT(state->dts_activity ==
5943 				    DTRACE_ACTIVITY_DRAINING);
5944 				continue;
5945 			}
5946 		}
5947 
5948 		if (ecb->dte_cond && !dtrace_priv_probe(state, &mstate, ecb))
5949 			continue;
5950 
5951 		if (now - state->dts_alive > dtrace_deadman_timeout) {
5952 			/*
5953 			 * We seem to be dead.  Unless we (a) have kernel
5954 			 * destructive permissions (b) have expicitly enabled
5955 			 * destructive actions and (c) destructive actions have
5956 			 * not been disabled, we're going to transition into
5957 			 * the KILLED state, from which no further processing
5958 			 * on this state will be performed.
5959 			 */
5960 			if (!dtrace_priv_kernel_destructive(state) ||
5961 			    !state->dts_cred.dcr_destructive ||
5962 			    dtrace_destructive_disallow) {
5963 				void *activity = &state->dts_activity;
5964 				dtrace_activity_t current;
5965 
5966 				do {
5967 					current = state->dts_activity;
5968 				} while (dtrace_cas32(activity, current,
5969 				    DTRACE_ACTIVITY_KILLED) != current);
5970 
5971 				continue;
5972 			}
5973 		}
5974 
5975 		if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
5976 		    ecb->dte_alignment, state, &mstate)) < 0)
5977 			continue;
5978 
5979 		tomax = buf->dtb_tomax;
5980 		ASSERT(tomax != NULL);
5981 
5982 		if (ecb->dte_size != 0)
5983 			DTRACE_STORE(uint32_t, tomax, offs, ecb->dte_epid);
5984 
5985 		mstate.dtms_epid = ecb->dte_epid;
5986 		mstate.dtms_present |= DTRACE_MSTATE_EPID;
5987 
5988 		if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
5989 			mstate.dtms_access |= DTRACE_ACCESS_KERNEL;
5990 
5991 		if (pred != NULL) {
5992 			dtrace_difo_t *dp = pred->dtp_difo;
5993 			int rval;
5994 
5995 			rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
5996 
5997 			if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
5998 				dtrace_cacheid_t cid = probe->dtpr_predcache;
5999 
6000 				if (cid != DTRACE_CACHEIDNONE && !onintr) {
6001 					/*
6002 					 * Update the predicate cache...
6003 					 */
6004 					ASSERT(cid == pred->dtp_cacheid);
6005 					curthread->t_predcache = cid;
6006 				}
6007 
6008 				continue;
6009 			}
6010 		}
6011 
6012 		for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
6013 		    act != NULL; act = act->dta_next) {
6014 			size_t valoffs;
6015 			dtrace_difo_t *dp;
6016 			dtrace_recdesc_t *rec = &act->dta_rec;
6017 
6018 			size = rec->dtrd_size;
6019 			valoffs = offs + rec->dtrd_offset;
6020 
6021 			if (DTRACEACT_ISAGG(act->dta_kind)) {
6022 				uint64_t v = 0xbad;
6023 				dtrace_aggregation_t *agg;
6024 
6025 				agg = (dtrace_aggregation_t *)act;
6026 
6027 				if ((dp = act->dta_difo) != NULL)
6028 					v = dtrace_dif_emulate(dp,
6029 					    &mstate, vstate, state);
6030 
6031 				if (*flags & CPU_DTRACE_ERROR)
6032 					continue;
6033 
6034 				/*
6035 				 * Note that we always pass the expression
6036 				 * value from the previous iteration of the
6037 				 * action loop.  This value will only be used
6038 				 * if there is an expression argument to the
6039 				 * aggregating action, denoted by the
6040 				 * dtag_hasarg field.
6041 				 */
6042 				dtrace_aggregate(agg, buf,
6043 				    offs, aggbuf, v, val);
6044 				continue;
6045 			}
6046 
6047 			switch (act->dta_kind) {
6048 			case DTRACEACT_STOP:
6049 				if (dtrace_priv_proc_destructive(state,
6050 				    &mstate))
6051 					dtrace_action_stop();
6052 				continue;
6053 
6054 			case DTRACEACT_BREAKPOINT:
6055 				if (dtrace_priv_kernel_destructive(state))
6056 					dtrace_action_breakpoint(ecb);
6057 				continue;
6058 
6059 			case DTRACEACT_PANIC:
6060 				if (dtrace_priv_kernel_destructive(state))
6061 					dtrace_action_panic(ecb);
6062 				continue;
6063 
6064 			case DTRACEACT_STACK:
6065 				if (!dtrace_priv_kernel(state))
6066 					continue;
6067 
6068 				dtrace_getpcstack((pc_t *)(tomax + valoffs),
6069 				    size / sizeof (pc_t), probe->dtpr_aframes,
6070 				    DTRACE_ANCHORED(probe) ? NULL :
6071 				    (uint32_t *)arg0);
6072 
6073 				continue;
6074 
6075 			case DTRACEACT_JSTACK:
6076 			case DTRACEACT_USTACK:
6077 				if (!dtrace_priv_proc(state, &mstate))
6078 					continue;
6079 
6080 				/*
6081 				 * See comment in DIF_VAR_PID.
6082 				 */
6083 				if (DTRACE_ANCHORED(mstate.dtms_probe) &&
6084 				    CPU_ON_INTR(CPU)) {
6085 					int depth = DTRACE_USTACK_NFRAMES(
6086 					    rec->dtrd_arg) + 1;
6087 
6088 					dtrace_bzero((void *)(tomax + valoffs),
6089 					    DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
6090 					    + depth * sizeof (uint64_t));
6091 
6092 					continue;
6093 				}
6094 
6095 				if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
6096 				    curproc->p_dtrace_helpers != NULL) {
6097 					/*
6098 					 * This is the slow path -- we have
6099 					 * allocated string space, and we're
6100 					 * getting the stack of a process that
6101 					 * has helpers.  Call into a separate
6102 					 * routine to perform this processing.
6103 					 */
6104 					dtrace_action_ustack(&mstate, state,
6105 					    (uint64_t *)(tomax + valoffs),
6106 					    rec->dtrd_arg);
6107 					continue;
6108 				}
6109 
6110 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6111 				dtrace_getupcstack((uint64_t *)
6112 				    (tomax + valoffs),
6113 				    DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
6114 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6115 				continue;
6116 
6117 			default:
6118 				break;
6119 			}
6120 
6121 			dp = act->dta_difo;
6122 			ASSERT(dp != NULL);
6123 
6124 			val = dtrace_dif_emulate(dp, &mstate, vstate, state);
6125 
6126 			if (*flags & CPU_DTRACE_ERROR)
6127 				continue;
6128 
6129 			switch (act->dta_kind) {
6130 			case DTRACEACT_SPECULATE:
6131 				ASSERT(buf == &state->dts_buffer[cpuid]);
6132 				buf = dtrace_speculation_buffer(state,
6133 				    cpuid, val);
6134 
6135 				if (buf == NULL) {
6136 					*flags |= CPU_DTRACE_DROP;
6137 					continue;
6138 				}
6139 
6140 				offs = dtrace_buffer_reserve(buf,
6141 				    ecb->dte_needed, ecb->dte_alignment,
6142 				    state, NULL);
6143 
6144 				if (offs < 0) {
6145 					*flags |= CPU_DTRACE_DROP;
6146 					continue;
6147 				}
6148 
6149 				tomax = buf->dtb_tomax;
6150 				ASSERT(tomax != NULL);
6151 
6152 				if (ecb->dte_size != 0)
6153 					DTRACE_STORE(uint32_t, tomax, offs,
6154 					    ecb->dte_epid);
6155 				continue;
6156 
6157 			case DTRACEACT_CHILL:
6158 				if (dtrace_priv_kernel_destructive(state))
6159 					dtrace_action_chill(&mstate, val);
6160 				continue;
6161 
6162 			case DTRACEACT_RAISE:
6163 				if (dtrace_priv_proc_destructive(state,
6164 				    &mstate))
6165 					dtrace_action_raise(val);
6166 				continue;
6167 
6168 			case DTRACEACT_COMMIT:
6169 				ASSERT(!committed);
6170 
6171 				/*
6172 				 * We need to commit our buffer state.
6173 				 */
6174 				if (ecb->dte_size)
6175 					buf->dtb_offset = offs + ecb->dte_size;
6176 				buf = &state->dts_buffer[cpuid];
6177 				dtrace_speculation_commit(state, cpuid, val);
6178 				committed = 1;
6179 				continue;
6180 
6181 			case DTRACEACT_DISCARD:
6182 				dtrace_speculation_discard(state, cpuid, val);
6183 				continue;
6184 
6185 			case DTRACEACT_DIFEXPR:
6186 			case DTRACEACT_LIBACT:
6187 			case DTRACEACT_PRINTF:
6188 			case DTRACEACT_PRINTA:
6189 			case DTRACEACT_SYSTEM:
6190 			case DTRACEACT_FREOPEN:
6191 			case DTRACEACT_TRACEMEM:
6192 				break;
6193 
6194 			case DTRACEACT_TRACEMEM_DYNSIZE:
6195 				tracememsize = val;
6196 				break;
6197 
6198 			case DTRACEACT_SYM:
6199 			case DTRACEACT_MOD:
6200 				if (!dtrace_priv_kernel(state))
6201 					continue;
6202 				break;
6203 
6204 			case DTRACEACT_USYM:
6205 			case DTRACEACT_UMOD:
6206 			case DTRACEACT_UADDR: {
6207 				struct pid *pid = curthread->t_procp->p_pidp;
6208 
6209 				if (!dtrace_priv_proc(state, &mstate))
6210 					continue;
6211 
6212 				DTRACE_STORE(uint64_t, tomax,
6213 				    valoffs, (uint64_t)pid->pid_id);
6214 				DTRACE_STORE(uint64_t, tomax,
6215 				    valoffs + sizeof (uint64_t), val);
6216 
6217 				continue;
6218 			}
6219 
6220 			case DTRACEACT_EXIT: {
6221 				/*
6222 				 * For the exit action, we are going to attempt
6223 				 * to atomically set our activity to be
6224 				 * draining.  If this fails (either because
6225 				 * another CPU has beat us to the exit action,
6226 				 * or because our current activity is something
6227 				 * other than ACTIVE or WARMUP), we will
6228 				 * continue.  This assures that the exit action
6229 				 * can be successfully recorded at most once
6230 				 * when we're in the ACTIVE state.  If we're
6231 				 * encountering the exit() action while in
6232 				 * COOLDOWN, however, we want to honor the new
6233 				 * status code.  (We know that we're the only
6234 				 * thread in COOLDOWN, so there is no race.)
6235 				 */
6236 				void *activity = &state->dts_activity;
6237 				dtrace_activity_t current = state->dts_activity;
6238 
6239 				if (current == DTRACE_ACTIVITY_COOLDOWN)
6240 					break;
6241 
6242 				if (current != DTRACE_ACTIVITY_WARMUP)
6243 					current = DTRACE_ACTIVITY_ACTIVE;
6244 
6245 				if (dtrace_cas32(activity, current,
6246 				    DTRACE_ACTIVITY_DRAINING) != current) {
6247 					*flags |= CPU_DTRACE_DROP;
6248 					continue;
6249 				}
6250 
6251 				break;
6252 			}
6253 
6254 			default:
6255 				ASSERT(0);
6256 			}
6257 
6258 			if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF) {
6259 				uintptr_t end = valoffs + size;
6260 
6261 				if (tracememsize != 0 &&
6262 				    valoffs + tracememsize < end) {
6263 					end = valoffs + tracememsize;
6264 					tracememsize = 0;
6265 				}
6266 
6267 				if (!dtrace_vcanload((void *)(uintptr_t)val,
6268 				    &dp->dtdo_rtype, &mstate, vstate))
6269 					continue;
6270 
6271 				/*
6272 				 * If this is a string, we're going to only
6273 				 * load until we find the zero byte -- after
6274 				 * which we'll store zero bytes.
6275 				 */
6276 				if (dp->dtdo_rtype.dtdt_kind ==
6277 				    DIF_TYPE_STRING) {
6278 					char c = '\0' + 1;
6279 					int intuple = act->dta_intuple;
6280 					size_t s;
6281 
6282 					for (s = 0; s < size; s++) {
6283 						if (c != '\0')
6284 							c = dtrace_load8(val++);
6285 
6286 						DTRACE_STORE(uint8_t, tomax,
6287 						    valoffs++, c);
6288 
6289 						if (c == '\0' && intuple)
6290 							break;
6291 					}
6292 
6293 					continue;
6294 				}
6295 
6296 				while (valoffs < end) {
6297 					DTRACE_STORE(uint8_t, tomax, valoffs++,
6298 					    dtrace_load8(val++));
6299 				}
6300 
6301 				continue;
6302 			}
6303 
6304 			switch (size) {
6305 			case 0:
6306 				break;
6307 
6308 			case sizeof (uint8_t):
6309 				DTRACE_STORE(uint8_t, tomax, valoffs, val);
6310 				break;
6311 			case sizeof (uint16_t):
6312 				DTRACE_STORE(uint16_t, tomax, valoffs, val);
6313 				break;
6314 			case sizeof (uint32_t):
6315 				DTRACE_STORE(uint32_t, tomax, valoffs, val);
6316 				break;
6317 			case sizeof (uint64_t):
6318 				DTRACE_STORE(uint64_t, tomax, valoffs, val);
6319 				break;
6320 			default:
6321 				/*
6322 				 * Any other size should have been returned by
6323 				 * reference, not by value.
6324 				 */
6325 				ASSERT(0);
6326 				break;
6327 			}
6328 		}
6329 
6330 		if (*flags & CPU_DTRACE_DROP)
6331 			continue;
6332 
6333 		if (*flags & CPU_DTRACE_FAULT) {
6334 			int ndx;
6335 			dtrace_action_t *err;
6336 
6337 			buf->dtb_errors++;
6338 
6339 			if (probe->dtpr_id == dtrace_probeid_error) {
6340 				/*
6341 				 * There's nothing we can do -- we had an
6342 				 * error on the error probe.  We bump an
6343 				 * error counter to at least indicate that
6344 				 * this condition happened.
6345 				 */
6346 				dtrace_error(&state->dts_dblerrors);
6347 				continue;
6348 			}
6349 
6350 			if (vtime) {
6351 				/*
6352 				 * Before recursing on dtrace_probe(), we
6353 				 * need to explicitly clear out our start
6354 				 * time to prevent it from being accumulated
6355 				 * into t_dtrace_vtime.
6356 				 */
6357 				curthread->t_dtrace_start = 0;
6358 			}
6359 
6360 			/*
6361 			 * Iterate over the actions to figure out which action
6362 			 * we were processing when we experienced the error.
6363 			 * Note that act points _past_ the faulting action; if
6364 			 * act is ecb->dte_action, the fault was in the
6365 			 * predicate, if it's ecb->dte_action->dta_next it's
6366 			 * in action #1, and so on.
6367 			 */
6368 			for (err = ecb->dte_action, ndx = 0;
6369 			    err != act; err = err->dta_next, ndx++)
6370 				continue;
6371 
6372 			dtrace_probe_error(state, ecb->dte_epid, ndx,
6373 			    (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
6374 			    mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
6375 			    cpu_core[cpuid].cpuc_dtrace_illval);
6376 
6377 			continue;
6378 		}
6379 
6380 		if (!committed)
6381 			buf->dtb_offset = offs + ecb->dte_size;
6382 	}
6383 
6384 	if (vtime)
6385 		curthread->t_dtrace_start = dtrace_gethrtime();
6386 
6387 	dtrace_interrupt_enable(cookie);
6388 }
6389 
6390 /*
6391  * DTrace Probe Hashing Functions
6392  *
6393  * The functions in this section (and indeed, the functions in remaining
6394  * sections) are not _called_ from probe context.  (Any exceptions to this are
6395  * marked with a "Note:".)  Rather, they are called from elsewhere in the
6396  * DTrace framework to look-up probes in, add probes to and remove probes from
6397  * the DTrace probe hashes.  (Each probe is hashed by each element of the
6398  * probe tuple -- allowing for fast lookups, regardless of what was
6399  * specified.)
6400  */
6401 static uint_t
6402 dtrace_hash_str(char *p)
6403 {
6404 	unsigned int g;
6405 	uint_t hval = 0;
6406 
6407 	while (*p) {
6408 		hval = (hval << 4) + *p++;
6409 		if ((g = (hval & 0xf0000000)) != 0)
6410 			hval ^= g >> 24;
6411 		hval &= ~g;
6412 	}
6413 	return (hval);
6414 }
6415 
6416 static dtrace_hash_t *
6417 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
6418 {
6419 	dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
6420 
6421 	hash->dth_stroffs = stroffs;
6422 	hash->dth_nextoffs = nextoffs;
6423 	hash->dth_prevoffs = prevoffs;
6424 
6425 	hash->dth_size = 1;
6426 	hash->dth_mask = hash->dth_size - 1;
6427 
6428 	hash->dth_tab = kmem_zalloc(hash->dth_size *
6429 	    sizeof (dtrace_hashbucket_t *), KM_SLEEP);
6430 
6431 	return (hash);
6432 }
6433 
6434 static void
6435 dtrace_hash_destroy(dtrace_hash_t *hash)
6436 {
6437 #ifdef DEBUG
6438 	int i;
6439 
6440 	for (i = 0; i < hash->dth_size; i++)
6441 		ASSERT(hash->dth_tab[i] == NULL);
6442 #endif
6443 
6444 	kmem_free(hash->dth_tab,
6445 	    hash->dth_size * sizeof (dtrace_hashbucket_t *));
6446 	kmem_free(hash, sizeof (dtrace_hash_t));
6447 }
6448 
6449 static void
6450 dtrace_hash_resize(dtrace_hash_t *hash)
6451 {
6452 	int size = hash->dth_size, i, ndx;
6453 	int new_size = hash->dth_size << 1;
6454 	int new_mask = new_size - 1;
6455 	dtrace_hashbucket_t **new_tab, *bucket, *next;
6456 
6457 	ASSERT((new_size & new_mask) == 0);
6458 
6459 	new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
6460 
6461 	for (i = 0; i < size; i++) {
6462 		for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
6463 			dtrace_probe_t *probe = bucket->dthb_chain;
6464 
6465 			ASSERT(probe != NULL);
6466 			ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
6467 
6468 			next = bucket->dthb_next;
6469 			bucket->dthb_next = new_tab[ndx];
6470 			new_tab[ndx] = bucket;
6471 		}
6472 	}
6473 
6474 	kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
6475 	hash->dth_tab = new_tab;
6476 	hash->dth_size = new_size;
6477 	hash->dth_mask = new_mask;
6478 }
6479 
6480 static void
6481 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
6482 {
6483 	int hashval = DTRACE_HASHSTR(hash, new);
6484 	int ndx = hashval & hash->dth_mask;
6485 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6486 	dtrace_probe_t **nextp, **prevp;
6487 
6488 	for (; bucket != NULL; bucket = bucket->dthb_next) {
6489 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
6490 			goto add;
6491 	}
6492 
6493 	if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
6494 		dtrace_hash_resize(hash);
6495 		dtrace_hash_add(hash, new);
6496 		return;
6497 	}
6498 
6499 	bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
6500 	bucket->dthb_next = hash->dth_tab[ndx];
6501 	hash->dth_tab[ndx] = bucket;
6502 	hash->dth_nbuckets++;
6503 
6504 add:
6505 	nextp = DTRACE_HASHNEXT(hash, new);
6506 	ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
6507 	*nextp = bucket->dthb_chain;
6508 
6509 	if (bucket->dthb_chain != NULL) {
6510 		prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
6511 		ASSERT(*prevp == NULL);
6512 		*prevp = new;
6513 	}
6514 
6515 	bucket->dthb_chain = new;
6516 	bucket->dthb_len++;
6517 }
6518 
6519 static dtrace_probe_t *
6520 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
6521 {
6522 	int hashval = DTRACE_HASHSTR(hash, template);
6523 	int ndx = hashval & hash->dth_mask;
6524 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6525 
6526 	for (; bucket != NULL; bucket = bucket->dthb_next) {
6527 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6528 			return (bucket->dthb_chain);
6529 	}
6530 
6531 	return (NULL);
6532 }
6533 
6534 static int
6535 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
6536 {
6537 	int hashval = DTRACE_HASHSTR(hash, template);
6538 	int ndx = hashval & hash->dth_mask;
6539 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6540 
6541 	for (; bucket != NULL; bucket = bucket->dthb_next) {
6542 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6543 			return (bucket->dthb_len);
6544 	}
6545 
6546 	return (NULL);
6547 }
6548 
6549 static void
6550 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
6551 {
6552 	int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
6553 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6554 
6555 	dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
6556 	dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
6557 
6558 	/*
6559 	 * Find the bucket that we're removing this probe from.
6560 	 */
6561 	for (; bucket != NULL; bucket = bucket->dthb_next) {
6562 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
6563 			break;
6564 	}
6565 
6566 	ASSERT(bucket != NULL);
6567 
6568 	if (*prevp == NULL) {
6569 		if (*nextp == NULL) {
6570 			/*
6571 			 * The removed probe was the only probe on this
6572 			 * bucket; we need to remove the bucket.
6573 			 */
6574 			dtrace_hashbucket_t *b = hash->dth_tab[ndx];
6575 
6576 			ASSERT(bucket->dthb_chain == probe);
6577 			ASSERT(b != NULL);
6578 
6579 			if (b == bucket) {
6580 				hash->dth_tab[ndx] = bucket->dthb_next;
6581 			} else {
6582 				while (b->dthb_next != bucket)
6583 					b = b->dthb_next;
6584 				b->dthb_next = bucket->dthb_next;
6585 			}
6586 
6587 			ASSERT(hash->dth_nbuckets > 0);
6588 			hash->dth_nbuckets--;
6589 			kmem_free(bucket, sizeof (dtrace_hashbucket_t));
6590 			return;
6591 		}
6592 
6593 		bucket->dthb_chain = *nextp;
6594 	} else {
6595 		*(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
6596 	}
6597 
6598 	if (*nextp != NULL)
6599 		*(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
6600 }
6601 
6602 /*
6603  * DTrace Utility Functions
6604  *
6605  * These are random utility functions that are _not_ called from probe context.
6606  */
6607 static int
6608 dtrace_badattr(const dtrace_attribute_t *a)
6609 {
6610 	return (a->dtat_name > DTRACE_STABILITY_MAX ||
6611 	    a->dtat_data > DTRACE_STABILITY_MAX ||
6612 	    a->dtat_class > DTRACE_CLASS_MAX);
6613 }
6614 
6615 /*
6616  * Return a duplicate copy of a string.  If the specified string is NULL,
6617  * this function returns a zero-length string.
6618  */
6619 static char *
6620 dtrace_strdup(const char *str)
6621 {
6622 	char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
6623 
6624 	if (str != NULL)
6625 		(void) strcpy(new, str);
6626 
6627 	return (new);
6628 }
6629 
6630 #define	DTRACE_ISALPHA(c)	\
6631 	(((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
6632 
6633 static int
6634 dtrace_badname(const char *s)
6635 {
6636 	char c;
6637 
6638 	if (s == NULL || (c = *s++) == '\0')
6639 		return (0);
6640 
6641 	if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
6642 		return (1);
6643 
6644 	while ((c = *s++) != '\0') {
6645 		if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
6646 		    c != '-' && c != '_' && c != '.' && c != '`')
6647 			return (1);
6648 	}
6649 
6650 	return (0);
6651 }
6652 
6653 static void
6654 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
6655 {
6656 	uint32_t priv;
6657 
6658 	if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
6659 		/*
6660 		 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
6661 		 */
6662 		priv = DTRACE_PRIV_ALL;
6663 	} else {
6664 		*uidp = crgetuid(cr);
6665 		*zoneidp = crgetzoneid(cr);
6666 
6667 		priv = 0;
6668 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
6669 			priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
6670 		else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
6671 			priv |= DTRACE_PRIV_USER;
6672 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
6673 			priv |= DTRACE_PRIV_PROC;
6674 		if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
6675 			priv |= DTRACE_PRIV_OWNER;
6676 		if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
6677 			priv |= DTRACE_PRIV_ZONEOWNER;
6678 	}
6679 
6680 	*privp = priv;
6681 }
6682 
6683 #ifdef DTRACE_ERRDEBUG
6684 static void
6685 dtrace_errdebug(const char *str)
6686 {
6687 	int hval = dtrace_hash_str((char *)str) % DTRACE_ERRHASHSZ;
6688 	int occupied = 0;
6689 
6690 	mutex_enter(&dtrace_errlock);
6691 	dtrace_errlast = str;
6692 	dtrace_errthread = curthread;
6693 
6694 	while (occupied++ < DTRACE_ERRHASHSZ) {
6695 		if (dtrace_errhash[hval].dter_msg == str) {
6696 			dtrace_errhash[hval].dter_count++;
6697 			goto out;
6698 		}
6699 
6700 		if (dtrace_errhash[hval].dter_msg != NULL) {
6701 			hval = (hval + 1) % DTRACE_ERRHASHSZ;
6702 			continue;
6703 		}
6704 
6705 		dtrace_errhash[hval].dter_msg = str;
6706 		dtrace_errhash[hval].dter_count = 1;
6707 		goto out;
6708 	}
6709 
6710 	panic("dtrace: undersized error hash");
6711 out:
6712 	mutex_exit(&dtrace_errlock);
6713 }
6714 #endif
6715 
6716 /*
6717  * DTrace Matching Functions
6718  *
6719  * These functions are used to match groups of probes, given some elements of
6720  * a probe tuple, or some globbed expressions for elements of a probe tuple.
6721  */
6722 static int
6723 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
6724     zoneid_t zoneid)
6725 {
6726 	if (priv != DTRACE_PRIV_ALL) {
6727 		uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
6728 		uint32_t match = priv & ppriv;
6729 
6730 		/*
6731 		 * No PRIV_DTRACE_* privileges...
6732 		 */
6733 		if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
6734 		    DTRACE_PRIV_KERNEL)) == 0)
6735 			return (0);
6736 
6737 		/*
6738 		 * No matching bits, but there were bits to match...
6739 		 */
6740 		if (match == 0 && ppriv != 0)
6741 			return (0);
6742 
6743 		/*
6744 		 * Need to have permissions to the process, but don't...
6745 		 */
6746 		if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
6747 		    uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
6748 			return (0);
6749 		}
6750 
6751 		/*
6752 		 * Need to be in the same zone unless we possess the
6753 		 * privilege to examine all zones.
6754 		 */
6755 		if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
6756 		    zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
6757 			return (0);
6758 		}
6759 	}
6760 
6761 	return (1);
6762 }
6763 
6764 /*
6765  * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
6766  * consists of input pattern strings and an ops-vector to evaluate them.
6767  * This function returns >0 for match, 0 for no match, and <0 for error.
6768  */
6769 static int
6770 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
6771     uint32_t priv, uid_t uid, zoneid_t zoneid)
6772 {
6773 	dtrace_provider_t *pvp = prp->dtpr_provider;
6774 	int rv;
6775 
6776 	if (pvp->dtpv_defunct)
6777 		return (0);
6778 
6779 	if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
6780 		return (rv);
6781 
6782 	if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
6783 		return (rv);
6784 
6785 	if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
6786 		return (rv);
6787 
6788 	if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
6789 		return (rv);
6790 
6791 	if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
6792 		return (0);
6793 
6794 	return (rv);
6795 }
6796 
6797 /*
6798  * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
6799  * interface for matching a glob pattern 'p' to an input string 's'.  Unlike
6800  * libc's version, the kernel version only applies to 8-bit ASCII strings.
6801  * In addition, all of the recursion cases except for '*' matching have been
6802  * unwound.  For '*', we still implement recursive evaluation, but a depth
6803  * counter is maintained and matching is aborted if we recurse too deep.
6804  * The function returns 0 if no match, >0 if match, and <0 if recursion error.
6805  */
6806 static int
6807 dtrace_match_glob(const char *s, const char *p, int depth)
6808 {
6809 	const char *olds;
6810 	char s1, c;
6811 	int gs;
6812 
6813 	if (depth > DTRACE_PROBEKEY_MAXDEPTH)
6814 		return (-1);
6815 
6816 	if (s == NULL)
6817 		s = ""; /* treat NULL as empty string */
6818 
6819 top:
6820 	olds = s;
6821 	s1 = *s++;
6822 
6823 	if (p == NULL)
6824 		return (0);
6825 
6826 	if ((c = *p++) == '\0')
6827 		return (s1 == '\0');
6828 
6829 	switch (c) {
6830 	case '[': {
6831 		int ok = 0, notflag = 0;
6832 		char lc = '\0';
6833 
6834 		if (s1 == '\0')
6835 			return (0);
6836 
6837 		if (*p == '!') {
6838 			notflag = 1;
6839 			p++;
6840 		}
6841 
6842 		if ((c = *p++) == '\0')
6843 			return (0);
6844 
6845 		do {
6846 			if (c == '-' && lc != '\0' && *p != ']') {
6847 				if ((c = *p++) == '\0')
6848 					return (0);
6849 				if (c == '\\' && (c = *p++) == '\0')
6850 					return (0);
6851 
6852 				if (notflag) {
6853 					if (s1 < lc || s1 > c)
6854 						ok++;
6855 					else
6856 						return (0);
6857 				} else if (lc <= s1 && s1 <= c)
6858 					ok++;
6859 
6860 			} else if (c == '\\' && (c = *p++) == '\0')
6861 				return (0);
6862 
6863 			lc = c; /* save left-hand 'c' for next iteration */
6864 
6865 			if (notflag) {
6866 				if (s1 != c)
6867 					ok++;
6868 				else
6869 					return (0);
6870 			} else if (s1 == c)
6871 				ok++;
6872 
6873 			if ((c = *p++) == '\0')
6874 				return (0);
6875 
6876 		} while (c != ']');
6877 
6878 		if (ok)
6879 			goto top;
6880 
6881 		return (0);
6882 	}
6883 
6884 	case '\\':
6885 		if ((c = *p++) == '\0')
6886 			return (0);
6887 		/*FALLTHRU*/
6888 
6889 	default:
6890 		if (c != s1)
6891 			return (0);
6892 		/*FALLTHRU*/
6893 
6894 	case '?':
6895 		if (s1 != '\0')
6896 			goto top;
6897 		return (0);
6898 
6899 	case '*':
6900 		while (*p == '*')
6901 			p++; /* consecutive *'s are identical to a single one */
6902 
6903 		if (*p == '\0')
6904 			return (1);
6905 
6906 		for (s = olds; *s != '\0'; s++) {
6907 			if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
6908 				return (gs);
6909 		}
6910 
6911 		return (0);
6912 	}
6913 }
6914 
6915 /*ARGSUSED*/
6916 static int
6917 dtrace_match_string(const char *s, const char *p, int depth)
6918 {
6919 	return (s != NULL && strcmp(s, p) == 0);
6920 }
6921 
6922 /*ARGSUSED*/
6923 static int
6924 dtrace_match_nul(const char *s, const char *p, int depth)
6925 {
6926 	return (1); /* always match the empty pattern */
6927 }
6928 
6929 /*ARGSUSED*/
6930 static int
6931 dtrace_match_nonzero(const char *s, const char *p, int depth)
6932 {
6933 	return (s != NULL && s[0] != '\0');
6934 }
6935 
6936 static int
6937 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
6938     zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
6939 {
6940 	dtrace_probe_t template, *probe;
6941 	dtrace_hash_t *hash = NULL;
6942 	int len, rc, best = INT_MAX, nmatched = 0;
6943 	dtrace_id_t i;
6944 
6945 	ASSERT(MUTEX_HELD(&dtrace_lock));
6946 
6947 	/*
6948 	 * If the probe ID is specified in the key, just lookup by ID and
6949 	 * invoke the match callback once if a matching probe is found.
6950 	 */
6951 	if (pkp->dtpk_id != DTRACE_IDNONE) {
6952 		if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
6953 		    dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
6954 			if ((*matched)(probe, arg) == DTRACE_MATCH_FAIL)
6955 				return (DTRACE_MATCH_FAIL);
6956 			nmatched++;
6957 		}
6958 		return (nmatched);
6959 	}
6960 
6961 	template.dtpr_mod = (char *)pkp->dtpk_mod;
6962 	template.dtpr_func = (char *)pkp->dtpk_func;
6963 	template.dtpr_name = (char *)pkp->dtpk_name;
6964 
6965 	/*
6966 	 * We want to find the most distinct of the module name, function
6967 	 * name, and name.  So for each one that is not a glob pattern or
6968 	 * empty string, we perform a lookup in the corresponding hash and
6969 	 * use the hash table with the fewest collisions to do our search.
6970 	 */
6971 	if (pkp->dtpk_mmatch == &dtrace_match_string &&
6972 	    (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
6973 		best = len;
6974 		hash = dtrace_bymod;
6975 	}
6976 
6977 	if (pkp->dtpk_fmatch == &dtrace_match_string &&
6978 	    (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
6979 		best = len;
6980 		hash = dtrace_byfunc;
6981 	}
6982 
6983 	if (pkp->dtpk_nmatch == &dtrace_match_string &&
6984 	    (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
6985 		best = len;
6986 		hash = dtrace_byname;
6987 	}
6988 
6989 	/*
6990 	 * If we did not select a hash table, iterate over every probe and
6991 	 * invoke our callback for each one that matches our input probe key.
6992 	 */
6993 	if (hash == NULL) {
6994 		for (i = 0; i < dtrace_nprobes; i++) {
6995 			if ((probe = dtrace_probes[i]) == NULL ||
6996 			    dtrace_match_probe(probe, pkp, priv, uid,
6997 			    zoneid) <= 0)
6998 				continue;
6999 
7000 			nmatched++;
7001 
7002 			if ((rc = (*matched)(probe, arg)) !=
7003 			    DTRACE_MATCH_NEXT) {
7004 				if (rc == DTRACE_MATCH_FAIL)
7005 					return (DTRACE_MATCH_FAIL);
7006 				break;
7007 			}
7008 		}
7009 
7010 		return (nmatched);
7011 	}
7012 
7013 	/*
7014 	 * If we selected a hash table, iterate over each probe of the same key
7015 	 * name and invoke the callback for every probe that matches the other
7016 	 * attributes of our input probe key.
7017 	 */
7018 	for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
7019 	    probe = *(DTRACE_HASHNEXT(hash, probe))) {
7020 
7021 		if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
7022 			continue;
7023 
7024 		nmatched++;
7025 
7026 		if ((rc = (*matched)(probe, arg)) != DTRACE_MATCH_NEXT) {
7027 			if (rc == DTRACE_MATCH_FAIL)
7028 				return (DTRACE_MATCH_FAIL);
7029 			break;
7030 		}
7031 	}
7032 
7033 	return (nmatched);
7034 }
7035 
7036 /*
7037  * Return the function pointer dtrace_probecmp() should use to compare the
7038  * specified pattern with a string.  For NULL or empty patterns, we select
7039  * dtrace_match_nul().  For glob pattern strings, we use dtrace_match_glob().
7040  * For non-empty non-glob strings, we use dtrace_match_string().
7041  */
7042 static dtrace_probekey_f *
7043 dtrace_probekey_func(const char *p)
7044 {
7045 	char c;
7046 
7047 	if (p == NULL || *p == '\0')
7048 		return (&dtrace_match_nul);
7049 
7050 	while ((c = *p++) != '\0') {
7051 		if (c == '[' || c == '?' || c == '*' || c == '\\')
7052 			return (&dtrace_match_glob);
7053 	}
7054 
7055 	return (&dtrace_match_string);
7056 }
7057 
7058 /*
7059  * Build a probe comparison key for use with dtrace_match_probe() from the
7060  * given probe description.  By convention, a null key only matches anchored
7061  * probes: if each field is the empty string, reset dtpk_fmatch to
7062  * dtrace_match_nonzero().
7063  */
7064 static void
7065 dtrace_probekey(const dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
7066 {
7067 	pkp->dtpk_prov = pdp->dtpd_provider;
7068 	pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
7069 
7070 	pkp->dtpk_mod = pdp->dtpd_mod;
7071 	pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
7072 
7073 	pkp->dtpk_func = pdp->dtpd_func;
7074 	pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
7075 
7076 	pkp->dtpk_name = pdp->dtpd_name;
7077 	pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
7078 
7079 	pkp->dtpk_id = pdp->dtpd_id;
7080 
7081 	if (pkp->dtpk_id == DTRACE_IDNONE &&
7082 	    pkp->dtpk_pmatch == &dtrace_match_nul &&
7083 	    pkp->dtpk_mmatch == &dtrace_match_nul &&
7084 	    pkp->dtpk_fmatch == &dtrace_match_nul &&
7085 	    pkp->dtpk_nmatch == &dtrace_match_nul)
7086 		pkp->dtpk_fmatch = &dtrace_match_nonzero;
7087 }
7088 
7089 /*
7090  * DTrace Provider-to-Framework API Functions
7091  *
7092  * These functions implement much of the Provider-to-Framework API, as
7093  * described in <sys/dtrace.h>.  The parts of the API not in this section are
7094  * the functions in the API for probe management (found below), and
7095  * dtrace_probe() itself (found above).
7096  */
7097 
7098 /*
7099  * Register the calling provider with the DTrace framework.  This should
7100  * generally be called by DTrace providers in their attach(9E) entry point.
7101  */
7102 int
7103 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
7104     cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
7105 {
7106 	dtrace_provider_t *provider;
7107 
7108 	if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
7109 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7110 		    "arguments", name ? name : "<NULL>");
7111 		return (EINVAL);
7112 	}
7113 
7114 	if (name[0] == '\0' || dtrace_badname(name)) {
7115 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7116 		    "provider name", name);
7117 		return (EINVAL);
7118 	}
7119 
7120 	if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
7121 	    pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
7122 	    pops->dtps_destroy == NULL ||
7123 	    ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
7124 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7125 		    "provider ops", name);
7126 		return (EINVAL);
7127 	}
7128 
7129 	if (dtrace_badattr(&pap->dtpa_provider) ||
7130 	    dtrace_badattr(&pap->dtpa_mod) ||
7131 	    dtrace_badattr(&pap->dtpa_func) ||
7132 	    dtrace_badattr(&pap->dtpa_name) ||
7133 	    dtrace_badattr(&pap->dtpa_args)) {
7134 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7135 		    "provider attributes", name);
7136 		return (EINVAL);
7137 	}
7138 
7139 	if (priv & ~DTRACE_PRIV_ALL) {
7140 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7141 		    "privilege attributes", name);
7142 		return (EINVAL);
7143 	}
7144 
7145 	if ((priv & DTRACE_PRIV_KERNEL) &&
7146 	    (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
7147 	    pops->dtps_mode == NULL) {
7148 		cmn_err(CE_WARN, "failed to register provider '%s': need "
7149 		    "dtps_mode() op for given privilege attributes", name);
7150 		return (EINVAL);
7151 	}
7152 
7153 	provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
7154 	provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
7155 	(void) strcpy(provider->dtpv_name, name);
7156 
7157 	provider->dtpv_attr = *pap;
7158 	provider->dtpv_priv.dtpp_flags = priv;
7159 	if (cr != NULL) {
7160 		provider->dtpv_priv.dtpp_uid = crgetuid(cr);
7161 		provider->dtpv_priv.dtpp_zoneid = crgetzoneid(cr);
7162 	}
7163 	provider->dtpv_pops = *pops;
7164 
7165 	if (pops->dtps_provide == NULL) {
7166 		ASSERT(pops->dtps_provide_module != NULL);
7167 		provider->dtpv_pops.dtps_provide =
7168 		    (void (*)(void *, const dtrace_probedesc_t *))dtrace_nullop;
7169 	}
7170 
7171 	if (pops->dtps_provide_module == NULL) {
7172 		ASSERT(pops->dtps_provide != NULL);
7173 		provider->dtpv_pops.dtps_provide_module =
7174 		    (void (*)(void *, struct modctl *))dtrace_nullop;
7175 	}
7176 
7177 	if (pops->dtps_suspend == NULL) {
7178 		ASSERT(pops->dtps_resume == NULL);
7179 		provider->dtpv_pops.dtps_suspend =
7180 		    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7181 		provider->dtpv_pops.dtps_resume =
7182 		    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7183 	}
7184 
7185 	provider->dtpv_arg = arg;
7186 	*idp = (dtrace_provider_id_t)provider;
7187 
7188 	if (pops == &dtrace_provider_ops) {
7189 		ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7190 		ASSERT(MUTEX_HELD(&dtrace_lock));
7191 		ASSERT(dtrace_anon.dta_enabling == NULL);
7192 
7193 		/*
7194 		 * We make sure that the DTrace provider is at the head of
7195 		 * the provider chain.
7196 		 */
7197 		provider->dtpv_next = dtrace_provider;
7198 		dtrace_provider = provider;
7199 		return (0);
7200 	}
7201 
7202 	mutex_enter(&dtrace_provider_lock);
7203 	mutex_enter(&dtrace_lock);
7204 
7205 	/*
7206 	 * If there is at least one provider registered, we'll add this
7207 	 * provider after the first provider.
7208 	 */
7209 	if (dtrace_provider != NULL) {
7210 		provider->dtpv_next = dtrace_provider->dtpv_next;
7211 		dtrace_provider->dtpv_next = provider;
7212 	} else {
7213 		dtrace_provider = provider;
7214 	}
7215 
7216 	if (dtrace_retained != NULL) {
7217 		dtrace_enabling_provide(provider);
7218 
7219 		/*
7220 		 * Now we need to call dtrace_enabling_matchall() -- which
7221 		 * will acquire cpu_lock and dtrace_lock.  We therefore need
7222 		 * to drop all of our locks before calling into it...
7223 		 */
7224 		mutex_exit(&dtrace_lock);
7225 		mutex_exit(&dtrace_provider_lock);
7226 		dtrace_enabling_matchall();
7227 
7228 		return (0);
7229 	}
7230 
7231 	mutex_exit(&dtrace_lock);
7232 	mutex_exit(&dtrace_provider_lock);
7233 
7234 	return (0);
7235 }
7236 
7237 /*
7238  * Unregister the specified provider from the DTrace framework.  This should
7239  * generally be called by DTrace providers in their detach(9E) entry point.
7240  */
7241 int
7242 dtrace_unregister(dtrace_provider_id_t id)
7243 {
7244 	dtrace_provider_t *old = (dtrace_provider_t *)id;
7245 	dtrace_provider_t *prev = NULL;
7246 	int i, self = 0, noreap = 0;
7247 	dtrace_probe_t *probe, *first = NULL;
7248 
7249 	if (old->dtpv_pops.dtps_enable ==
7250 	    (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop) {
7251 		/*
7252 		 * If DTrace itself is the provider, we're called with locks
7253 		 * already held.
7254 		 */
7255 		ASSERT(old == dtrace_provider);
7256 		ASSERT(dtrace_devi != NULL);
7257 		ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7258 		ASSERT(MUTEX_HELD(&dtrace_lock));
7259 		self = 1;
7260 
7261 		if (dtrace_provider->dtpv_next != NULL) {
7262 			/*
7263 			 * There's another provider here; return failure.
7264 			 */
7265 			return (EBUSY);
7266 		}
7267 	} else {
7268 		mutex_enter(&dtrace_provider_lock);
7269 		mutex_enter(&mod_lock);
7270 		mutex_enter(&dtrace_lock);
7271 	}
7272 
7273 	/*
7274 	 * If anyone has /dev/dtrace open, or if there are anonymous enabled
7275 	 * probes, we refuse to let providers slither away, unless this
7276 	 * provider has already been explicitly invalidated.
7277 	 */
7278 	if (!old->dtpv_defunct &&
7279 	    (dtrace_opens || (dtrace_anon.dta_state != NULL &&
7280 	    dtrace_anon.dta_state->dts_necbs > 0))) {
7281 		if (!self) {
7282 			mutex_exit(&dtrace_lock);
7283 			mutex_exit(&mod_lock);
7284 			mutex_exit(&dtrace_provider_lock);
7285 		}
7286 		return (EBUSY);
7287 	}
7288 
7289 	/*
7290 	 * Attempt to destroy the probes associated with this provider.
7291 	 */
7292 	for (i = 0; i < dtrace_nprobes; i++) {
7293 		if ((probe = dtrace_probes[i]) == NULL)
7294 			continue;
7295 
7296 		if (probe->dtpr_provider != old)
7297 			continue;
7298 
7299 		if (probe->dtpr_ecb == NULL)
7300 			continue;
7301 
7302 		/*
7303 		 * If we are trying to unregister a defunct provider, and the
7304 		 * provider was made defunct within the interval dictated by
7305 		 * dtrace_unregister_defunct_reap, we'll (asynchronously)
7306 		 * attempt to reap our enablings.  To denote that the provider
7307 		 * should reattempt to unregister itself at some point in the
7308 		 * future, we will return a differentiable error code (EAGAIN
7309 		 * instead of EBUSY) in this case.
7310 		 */
7311 		if (dtrace_gethrtime() - old->dtpv_defunct >
7312 		    dtrace_unregister_defunct_reap)
7313 			noreap = 1;
7314 
7315 		if (!self) {
7316 			mutex_exit(&dtrace_lock);
7317 			mutex_exit(&mod_lock);
7318 			mutex_exit(&dtrace_provider_lock);
7319 		}
7320 
7321 		if (noreap)
7322 			return (EBUSY);
7323 
7324 		(void) taskq_dispatch(dtrace_taskq,
7325 		    (task_func_t *)dtrace_enabling_reap, NULL, TQ_SLEEP);
7326 
7327 		return (EAGAIN);
7328 	}
7329 
7330 	/*
7331 	 * All of the probes for this provider are disabled; we can safely
7332 	 * remove all of them from their hash chains and from the probe array.
7333 	 */
7334 	for (i = 0; i < dtrace_nprobes; i++) {
7335 		if ((probe = dtrace_probes[i]) == NULL)
7336 			continue;
7337 
7338 		if (probe->dtpr_provider != old)
7339 			continue;
7340 
7341 		dtrace_probes[i] = NULL;
7342 
7343 		dtrace_hash_remove(dtrace_bymod, probe);
7344 		dtrace_hash_remove(dtrace_byfunc, probe);
7345 		dtrace_hash_remove(dtrace_byname, probe);
7346 
7347 		if (first == NULL) {
7348 			first = probe;
7349 			probe->dtpr_nextmod = NULL;
7350 		} else {
7351 			probe->dtpr_nextmod = first;
7352 			first = probe;
7353 		}
7354 	}
7355 
7356 	/*
7357 	 * The provider's probes have been removed from the hash chains and
7358 	 * from the probe array.  Now issue a dtrace_sync() to be sure that
7359 	 * everyone has cleared out from any probe array processing.
7360 	 */
7361 	dtrace_sync();
7362 
7363 	for (probe = first; probe != NULL; probe = first) {
7364 		first = probe->dtpr_nextmod;
7365 
7366 		old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
7367 		    probe->dtpr_arg);
7368 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7369 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7370 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7371 		vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
7372 		kmem_free(probe, sizeof (dtrace_probe_t));
7373 	}
7374 
7375 	if ((prev = dtrace_provider) == old) {
7376 		ASSERT(self || dtrace_devi == NULL);
7377 		ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
7378 		dtrace_provider = old->dtpv_next;
7379 	} else {
7380 		while (prev != NULL && prev->dtpv_next != old)
7381 			prev = prev->dtpv_next;
7382 
7383 		if (prev == NULL) {
7384 			panic("attempt to unregister non-existent "
7385 			    "dtrace provider %p\n", (void *)id);
7386 		}
7387 
7388 		prev->dtpv_next = old->dtpv_next;
7389 	}
7390 
7391 	if (!self) {
7392 		mutex_exit(&dtrace_lock);
7393 		mutex_exit(&mod_lock);
7394 		mutex_exit(&dtrace_provider_lock);
7395 	}
7396 
7397 	kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
7398 	kmem_free(old, sizeof (dtrace_provider_t));
7399 
7400 	return (0);
7401 }
7402 
7403 /*
7404  * Invalidate the specified provider.  All subsequent probe lookups for the
7405  * specified provider will fail, but its probes will not be removed.
7406  */
7407 void
7408 dtrace_invalidate(dtrace_provider_id_t id)
7409 {
7410 	dtrace_provider_t *pvp = (dtrace_provider_t *)id;
7411 
7412 	ASSERT(pvp->dtpv_pops.dtps_enable !=
7413 	    (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop);
7414 
7415 	mutex_enter(&dtrace_provider_lock);
7416 	mutex_enter(&dtrace_lock);
7417 
7418 	pvp->dtpv_defunct = dtrace_gethrtime();
7419 
7420 	mutex_exit(&dtrace_lock);
7421 	mutex_exit(&dtrace_provider_lock);
7422 }
7423 
7424 /*
7425  * Indicate whether or not DTrace has attached.
7426  */
7427 int
7428 dtrace_attached(void)
7429 {
7430 	/*
7431 	 * dtrace_provider will be non-NULL iff the DTrace driver has
7432 	 * attached.  (It's non-NULL because DTrace is always itself a
7433 	 * provider.)
7434 	 */
7435 	return (dtrace_provider != NULL);
7436 }
7437 
7438 /*
7439  * Remove all the unenabled probes for the given provider.  This function is
7440  * not unlike dtrace_unregister(), except that it doesn't remove the provider
7441  * -- just as many of its associated probes as it can.
7442  */
7443 int
7444 dtrace_condense(dtrace_provider_id_t id)
7445 {
7446 	dtrace_provider_t *prov = (dtrace_provider_t *)id;
7447 	int i;
7448 	dtrace_probe_t *probe;
7449 
7450 	/*
7451 	 * Make sure this isn't the dtrace provider itself.
7452 	 */
7453 	ASSERT(prov->dtpv_pops.dtps_enable !=
7454 	    (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop);
7455 
7456 	mutex_enter(&dtrace_provider_lock);
7457 	mutex_enter(&dtrace_lock);
7458 
7459 	/*
7460 	 * Attempt to destroy the probes associated with this provider.
7461 	 */
7462 	for (i = 0; i < dtrace_nprobes; i++) {
7463 		if ((probe = dtrace_probes[i]) == NULL)
7464 			continue;
7465 
7466 		if (probe->dtpr_provider != prov)
7467 			continue;
7468 
7469 		if (probe->dtpr_ecb != NULL)
7470 			continue;
7471 
7472 		dtrace_probes[i] = NULL;
7473 
7474 		dtrace_hash_remove(dtrace_bymod, probe);
7475 		dtrace_hash_remove(dtrace_byfunc, probe);
7476 		dtrace_hash_remove(dtrace_byname, probe);
7477 
7478 		prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
7479 		    probe->dtpr_arg);
7480 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7481 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7482 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7483 		kmem_free(probe, sizeof (dtrace_probe_t));
7484 		vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
7485 	}
7486 
7487 	mutex_exit(&dtrace_lock);
7488 	mutex_exit(&dtrace_provider_lock);
7489 
7490 	return (0);
7491 }
7492 
7493 /*
7494  * DTrace Probe Management Functions
7495  *
7496  * The functions in this section perform the DTrace probe management,
7497  * including functions to create probes, look-up probes, and call into the
7498  * providers to request that probes be provided.  Some of these functions are
7499  * in the Provider-to-Framework API; these functions can be identified by the
7500  * fact that they are not declared "static".
7501  */
7502 
7503 /*
7504  * Create a probe with the specified module name, function name, and name.
7505  */
7506 dtrace_id_t
7507 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
7508     const char *func, const char *name, int aframes, void *arg)
7509 {
7510 	dtrace_probe_t *probe, **probes;
7511 	dtrace_provider_t *provider = (dtrace_provider_t *)prov;
7512 	dtrace_id_t id;
7513 
7514 	if (provider == dtrace_provider) {
7515 		ASSERT(MUTEX_HELD(&dtrace_lock));
7516 	} else {
7517 		mutex_enter(&dtrace_lock);
7518 	}
7519 
7520 	id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
7521 	    VM_BESTFIT | VM_SLEEP);
7522 	probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
7523 
7524 	probe->dtpr_id = id;
7525 	probe->dtpr_gen = dtrace_probegen++;
7526 	probe->dtpr_mod = dtrace_strdup(mod);
7527 	probe->dtpr_func = dtrace_strdup(func);
7528 	probe->dtpr_name = dtrace_strdup(name);
7529 	probe->dtpr_arg = arg;
7530 	probe->dtpr_aframes = aframes;
7531 	probe->dtpr_provider = provider;
7532 
7533 	dtrace_hash_add(dtrace_bymod, probe);
7534 	dtrace_hash_add(dtrace_byfunc, probe);
7535 	dtrace_hash_add(dtrace_byname, probe);
7536 
7537 	if (id - 1 >= dtrace_nprobes) {
7538 		size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
7539 		size_t nsize = osize << 1;
7540 
7541 		if (nsize == 0) {
7542 			ASSERT(osize == 0);
7543 			ASSERT(dtrace_probes == NULL);
7544 			nsize = sizeof (dtrace_probe_t *);
7545 		}
7546 
7547 		probes = kmem_zalloc(nsize, KM_SLEEP);
7548 
7549 		if (dtrace_probes == NULL) {
7550 			ASSERT(osize == 0);
7551 			dtrace_probes = probes;
7552 			dtrace_nprobes = 1;
7553 		} else {
7554 			dtrace_probe_t **oprobes = dtrace_probes;
7555 
7556 			bcopy(oprobes, probes, osize);
7557 			dtrace_membar_producer();
7558 			dtrace_probes = probes;
7559 
7560 			dtrace_sync();
7561 
7562 			/*
7563 			 * All CPUs are now seeing the new probes array; we can
7564 			 * safely free the old array.
7565 			 */
7566 			kmem_free(oprobes, osize);
7567 			dtrace_nprobes <<= 1;
7568 		}
7569 
7570 		ASSERT(id - 1 < dtrace_nprobes);
7571 	}
7572 
7573 	ASSERT(dtrace_probes[id - 1] == NULL);
7574 	dtrace_probes[id - 1] = probe;
7575 
7576 	if (provider != dtrace_provider)
7577 		mutex_exit(&dtrace_lock);
7578 
7579 	return (id);
7580 }
7581 
7582 static dtrace_probe_t *
7583 dtrace_probe_lookup_id(dtrace_id_t id)
7584 {
7585 	ASSERT(MUTEX_HELD(&dtrace_lock));
7586 
7587 	if (id == 0 || id > dtrace_nprobes)
7588 		return (NULL);
7589 
7590 	return (dtrace_probes[id - 1]);
7591 }
7592 
7593 static int
7594 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
7595 {
7596 	*((dtrace_id_t *)arg) = probe->dtpr_id;
7597 
7598 	return (DTRACE_MATCH_DONE);
7599 }
7600 
7601 /*
7602  * Look up a probe based on provider and one or more of module name, function
7603  * name and probe name.
7604  */
7605 dtrace_id_t
7606 dtrace_probe_lookup(dtrace_provider_id_t prid, const char *mod,
7607     const char *func, const char *name)
7608 {
7609 	dtrace_probekey_t pkey;
7610 	dtrace_id_t id;
7611 	int match;
7612 
7613 	pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
7614 	pkey.dtpk_pmatch = &dtrace_match_string;
7615 	pkey.dtpk_mod = mod;
7616 	pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
7617 	pkey.dtpk_func = func;
7618 	pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
7619 	pkey.dtpk_name = name;
7620 	pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
7621 	pkey.dtpk_id = DTRACE_IDNONE;
7622 
7623 	mutex_enter(&dtrace_lock);
7624 	match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
7625 	    dtrace_probe_lookup_match, &id);
7626 	mutex_exit(&dtrace_lock);
7627 
7628 	ASSERT(match == 1 || match == 0);
7629 	return (match ? id : 0);
7630 }
7631 
7632 /*
7633  * Returns the probe argument associated with the specified probe.
7634  */
7635 void *
7636 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
7637 {
7638 	dtrace_probe_t *probe;
7639 	void *rval = NULL;
7640 
7641 	mutex_enter(&dtrace_lock);
7642 
7643 	if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
7644 	    probe->dtpr_provider == (dtrace_provider_t *)id)
7645 		rval = probe->dtpr_arg;
7646 
7647 	mutex_exit(&dtrace_lock);
7648 
7649 	return (rval);
7650 }
7651 
7652 /*
7653  * Copy a probe into a probe description.
7654  */
7655 static void
7656 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
7657 {
7658 	bzero(pdp, sizeof (dtrace_probedesc_t));
7659 	pdp->dtpd_id = prp->dtpr_id;
7660 
7661 	(void) strncpy(pdp->dtpd_provider,
7662 	    prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
7663 
7664 	(void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
7665 	(void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
7666 	(void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
7667 }
7668 
7669 /*
7670  * Called to indicate that a probe -- or probes -- should be provided by a
7671  * specfied provider.  If the specified description is NULL, the provider will
7672  * be told to provide all of its probes.  (This is done whenever a new
7673  * consumer comes along, or whenever a retained enabling is to be matched.) If
7674  * the specified description is non-NULL, the provider is given the
7675  * opportunity to dynamically provide the specified probe, allowing providers
7676  * to support the creation of probes on-the-fly.  (So-called _autocreated_
7677  * probes.)  If the provider is NULL, the operations will be applied to all
7678  * providers; if the provider is non-NULL the operations will only be applied
7679  * to the specified provider.  The dtrace_provider_lock must be held, and the
7680  * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
7681  * will need to grab the dtrace_lock when it reenters the framework through
7682  * dtrace_probe_lookup(), dtrace_probe_create(), etc.
7683  */
7684 static void
7685 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
7686 {
7687 	struct modctl *ctl;
7688 	int all = 0;
7689 
7690 	ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7691 
7692 	if (prv == NULL) {
7693 		all = 1;
7694 		prv = dtrace_provider;
7695 	}
7696 
7697 	do {
7698 		/*
7699 		 * First, call the blanket provide operation.
7700 		 */
7701 		prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
7702 
7703 		/*
7704 		 * Now call the per-module provide operation.  We will grab
7705 		 * mod_lock to prevent the list from being modified.  Note
7706 		 * that this also prevents the mod_busy bits from changing.
7707 		 * (mod_busy can only be changed with mod_lock held.)
7708 		 */
7709 		mutex_enter(&mod_lock);
7710 
7711 		ctl = &modules;
7712 		do {
7713 			if (ctl->mod_busy || ctl->mod_mp == NULL)
7714 				continue;
7715 
7716 			prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
7717 
7718 		} while ((ctl = ctl->mod_next) != &modules);
7719 
7720 		mutex_exit(&mod_lock);
7721 	} while (all && (prv = prv->dtpv_next) != NULL);
7722 }
7723 
7724 /*
7725  * Iterate over each probe, and call the Framework-to-Provider API function
7726  * denoted by offs.
7727  */
7728 static void
7729 dtrace_probe_foreach(uintptr_t offs)
7730 {
7731 	dtrace_provider_t *prov;
7732 	void (*func)(void *, dtrace_id_t, void *);
7733 	dtrace_probe_t *probe;
7734 	dtrace_icookie_t cookie;
7735 	int i;
7736 
7737 	/*
7738 	 * We disable interrupts to walk through the probe array.  This is
7739 	 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
7740 	 * won't see stale data.
7741 	 */
7742 	cookie = dtrace_interrupt_disable();
7743 
7744 	for (i = 0; i < dtrace_nprobes; i++) {
7745 		if ((probe = dtrace_probes[i]) == NULL)
7746 			continue;
7747 
7748 		if (probe->dtpr_ecb == NULL) {
7749 			/*
7750 			 * This probe isn't enabled -- don't call the function.
7751 			 */
7752 			continue;
7753 		}
7754 
7755 		prov = probe->dtpr_provider;
7756 		func = *((void(**)(void *, dtrace_id_t, void *))
7757 		    ((uintptr_t)&prov->dtpv_pops + offs));
7758 
7759 		func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
7760 	}
7761 
7762 	dtrace_interrupt_enable(cookie);
7763 }
7764 
7765 static int
7766 dtrace_probe_enable(const dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
7767 {
7768 	dtrace_probekey_t pkey;
7769 	uint32_t priv;
7770 	uid_t uid;
7771 	zoneid_t zoneid;
7772 
7773 	ASSERT(MUTEX_HELD(&dtrace_lock));
7774 	dtrace_ecb_create_cache = NULL;
7775 
7776 	if (desc == NULL) {
7777 		/*
7778 		 * If we're passed a NULL description, we're being asked to
7779 		 * create an ECB with a NULL probe.
7780 		 */
7781 		(void) dtrace_ecb_create_enable(NULL, enab);
7782 		return (0);
7783 	}
7784 
7785 	dtrace_probekey(desc, &pkey);
7786 	dtrace_cred2priv(enab->dten_vstate->dtvs_state->dts_cred.dcr_cred,
7787 	    &priv, &uid, &zoneid);
7788 
7789 	return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
7790 	    enab));
7791 }
7792 
7793 /*
7794  * DTrace Helper Provider Functions
7795  */
7796 static void
7797 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
7798 {
7799 	attr->dtat_name = DOF_ATTR_NAME(dofattr);
7800 	attr->dtat_data = DOF_ATTR_DATA(dofattr);
7801 	attr->dtat_class = DOF_ATTR_CLASS(dofattr);
7802 }
7803 
7804 static void
7805 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
7806     const dof_provider_t *dofprov, char *strtab)
7807 {
7808 	hprov->dthpv_provname = strtab + dofprov->dofpv_name;
7809 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
7810 	    dofprov->dofpv_provattr);
7811 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
7812 	    dofprov->dofpv_modattr);
7813 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
7814 	    dofprov->dofpv_funcattr);
7815 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
7816 	    dofprov->dofpv_nameattr);
7817 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
7818 	    dofprov->dofpv_argsattr);
7819 }
7820 
7821 static void
7822 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
7823 {
7824 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
7825 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
7826 	dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
7827 	dof_provider_t *provider;
7828 	dof_probe_t *probe;
7829 	uint32_t *off, *enoff;
7830 	uint8_t *arg;
7831 	char *strtab;
7832 	uint_t i, nprobes;
7833 	dtrace_helper_provdesc_t dhpv;
7834 	dtrace_helper_probedesc_t dhpb;
7835 	dtrace_meta_t *meta = dtrace_meta_pid;
7836 	dtrace_mops_t *mops = &meta->dtm_mops;
7837 	void *parg;
7838 
7839 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
7840 	str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
7841 	    provider->dofpv_strtab * dof->dofh_secsize);
7842 	prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
7843 	    provider->dofpv_probes * dof->dofh_secsize);
7844 	arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
7845 	    provider->dofpv_prargs * dof->dofh_secsize);
7846 	off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
7847 	    provider->dofpv_proffs * dof->dofh_secsize);
7848 
7849 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
7850 	off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
7851 	arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
7852 	enoff = NULL;
7853 
7854 	/*
7855 	 * See dtrace_helper_provider_validate().
7856 	 */
7857 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
7858 	    provider->dofpv_prenoffs != DOF_SECT_NONE) {
7859 		enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
7860 		    provider->dofpv_prenoffs * dof->dofh_secsize);
7861 		enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
7862 	}
7863 
7864 	nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
7865 
7866 	/*
7867 	 * Create the provider.
7868 	 */
7869 	dtrace_dofprov2hprov(&dhpv, provider, strtab);
7870 
7871 	if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
7872 		return;
7873 
7874 	meta->dtm_count++;
7875 
7876 	/*
7877 	 * Create the probes.
7878 	 */
7879 	for (i = 0; i < nprobes; i++) {
7880 		probe = (dof_probe_t *)(uintptr_t)(daddr +
7881 		    prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
7882 
7883 		dhpb.dthpb_mod = dhp->dofhp_mod;
7884 		dhpb.dthpb_func = strtab + probe->dofpr_func;
7885 		dhpb.dthpb_name = strtab + probe->dofpr_name;
7886 		dhpb.dthpb_base = probe->dofpr_addr;
7887 		dhpb.dthpb_offs = off + probe->dofpr_offidx;
7888 		dhpb.dthpb_noffs = probe->dofpr_noffs;
7889 		if (enoff != NULL) {
7890 			dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
7891 			dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
7892 		} else {
7893 			dhpb.dthpb_enoffs = NULL;
7894 			dhpb.dthpb_nenoffs = 0;
7895 		}
7896 		dhpb.dthpb_args = arg + probe->dofpr_argidx;
7897 		dhpb.dthpb_nargc = probe->dofpr_nargc;
7898 		dhpb.dthpb_xargc = probe->dofpr_xargc;
7899 		dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
7900 		dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
7901 
7902 		mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
7903 	}
7904 }
7905 
7906 static void
7907 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
7908 {
7909 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
7910 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
7911 	int i;
7912 
7913 	ASSERT(MUTEX_HELD(&dtrace_meta_lock));
7914 
7915 	for (i = 0; i < dof->dofh_secnum; i++) {
7916 		dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
7917 		    dof->dofh_secoff + i * dof->dofh_secsize);
7918 
7919 		if (sec->dofs_type != DOF_SECT_PROVIDER)
7920 			continue;
7921 
7922 		dtrace_helper_provide_one(dhp, sec, pid);
7923 	}
7924 
7925 	/*
7926 	 * We may have just created probes, so we must now rematch against
7927 	 * any retained enablings.  Note that this call will acquire both
7928 	 * cpu_lock and dtrace_lock; the fact that we are holding
7929 	 * dtrace_meta_lock now is what defines the ordering with respect to
7930 	 * these three locks.
7931 	 */
7932 	dtrace_enabling_matchall();
7933 }
7934 
7935 static void
7936 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
7937 {
7938 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
7939 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
7940 	dof_sec_t *str_sec;
7941 	dof_provider_t *provider;
7942 	char *strtab;
7943 	dtrace_helper_provdesc_t dhpv;
7944 	dtrace_meta_t *meta = dtrace_meta_pid;
7945 	dtrace_mops_t *mops = &meta->dtm_mops;
7946 
7947 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
7948 	str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
7949 	    provider->dofpv_strtab * dof->dofh_secsize);
7950 
7951 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
7952 
7953 	/*
7954 	 * Create the provider.
7955 	 */
7956 	dtrace_dofprov2hprov(&dhpv, provider, strtab);
7957 
7958 	mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
7959 
7960 	meta->dtm_count--;
7961 }
7962 
7963 static void
7964 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
7965 {
7966 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
7967 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
7968 	int i;
7969 
7970 	ASSERT(MUTEX_HELD(&dtrace_meta_lock));
7971 
7972 	for (i = 0; i < dof->dofh_secnum; i++) {
7973 		dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
7974 		    dof->dofh_secoff + i * dof->dofh_secsize);
7975 
7976 		if (sec->dofs_type != DOF_SECT_PROVIDER)
7977 			continue;
7978 
7979 		dtrace_helper_provider_remove_one(dhp, sec, pid);
7980 	}
7981 }
7982 
7983 /*
7984  * DTrace Meta Provider-to-Framework API Functions
7985  *
7986  * These functions implement the Meta Provider-to-Framework API, as described
7987  * in <sys/dtrace.h>.
7988  */
7989 int
7990 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
7991     dtrace_meta_provider_id_t *idp)
7992 {
7993 	dtrace_meta_t *meta;
7994 	dtrace_helpers_t *help, *next;
7995 	int i;
7996 
7997 	*idp = DTRACE_METAPROVNONE;
7998 
7999 	/*
8000 	 * We strictly don't need the name, but we hold onto it for
8001 	 * debuggability. All hail error queues!
8002 	 */
8003 	if (name == NULL) {
8004 		cmn_err(CE_WARN, "failed to register meta-provider: "
8005 		    "invalid name");
8006 		return (EINVAL);
8007 	}
8008 
8009 	if (mops == NULL ||
8010 	    mops->dtms_create_probe == NULL ||
8011 	    mops->dtms_provide_pid == NULL ||
8012 	    mops->dtms_remove_pid == NULL) {
8013 		cmn_err(CE_WARN, "failed to register meta-register %s: "
8014 		    "invalid ops", name);
8015 		return (EINVAL);
8016 	}
8017 
8018 	meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
8019 	meta->dtm_mops = *mops;
8020 	meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
8021 	(void) strcpy(meta->dtm_name, name);
8022 	meta->dtm_arg = arg;
8023 
8024 	mutex_enter(&dtrace_meta_lock);
8025 	mutex_enter(&dtrace_lock);
8026 
8027 	if (dtrace_meta_pid != NULL) {
8028 		mutex_exit(&dtrace_lock);
8029 		mutex_exit(&dtrace_meta_lock);
8030 		cmn_err(CE_WARN, "failed to register meta-register %s: "
8031 		    "user-land meta-provider exists", name);
8032 		kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
8033 		kmem_free(meta, sizeof (dtrace_meta_t));
8034 		return (EINVAL);
8035 	}
8036 
8037 	dtrace_meta_pid = meta;
8038 	*idp = (dtrace_meta_provider_id_t)meta;
8039 
8040 	/*
8041 	 * If there are providers and probes ready to go, pass them
8042 	 * off to the new meta provider now.
8043 	 */
8044 
8045 	help = dtrace_deferred_pid;
8046 	dtrace_deferred_pid = NULL;
8047 
8048 	mutex_exit(&dtrace_lock);
8049 
8050 	while (help != NULL) {
8051 		for (i = 0; i < help->dthps_nprovs; i++) {
8052 			dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
8053 			    help->dthps_pid);
8054 		}
8055 
8056 		next = help->dthps_next;
8057 		help->dthps_next = NULL;
8058 		help->dthps_prev = NULL;
8059 		help->dthps_deferred = 0;
8060 		help = next;
8061 	}
8062 
8063 	mutex_exit(&dtrace_meta_lock);
8064 
8065 	return (0);
8066 }
8067 
8068 int
8069 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
8070 {
8071 	dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
8072 
8073 	mutex_enter(&dtrace_meta_lock);
8074 	mutex_enter(&dtrace_lock);
8075 
8076 	if (old == dtrace_meta_pid) {
8077 		pp = &dtrace_meta_pid;
8078 	} else {
8079 		panic("attempt to unregister non-existent "
8080 		    "dtrace meta-provider %p\n", (void *)old);
8081 	}
8082 
8083 	if (old->dtm_count != 0) {
8084 		mutex_exit(&dtrace_lock);
8085 		mutex_exit(&dtrace_meta_lock);
8086 		return (EBUSY);
8087 	}
8088 
8089 	*pp = NULL;
8090 
8091 	mutex_exit(&dtrace_lock);
8092 	mutex_exit(&dtrace_meta_lock);
8093 
8094 	kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
8095 	kmem_free(old, sizeof (dtrace_meta_t));
8096 
8097 	return (0);
8098 }
8099 
8100 
8101 /*
8102  * DTrace DIF Object Functions
8103  */
8104 static int
8105 dtrace_difo_err(uint_t pc, const char *format, ...)
8106 {
8107 	if (dtrace_err_verbose) {
8108 		va_list alist;
8109 
8110 		(void) uprintf("dtrace DIF object error: [%u]: ", pc);
8111 		va_start(alist, format);
8112 		(void) vuprintf(format, alist);
8113 		va_end(alist);
8114 	}
8115 
8116 #ifdef DTRACE_ERRDEBUG
8117 	dtrace_errdebug(format);
8118 #endif
8119 	return (1);
8120 }
8121 
8122 /*
8123  * Validate a DTrace DIF object by checking the IR instructions.  The following
8124  * rules are currently enforced by dtrace_difo_validate():
8125  *
8126  * 1. Each instruction must have a valid opcode
8127  * 2. Each register, string, variable, or subroutine reference must be valid
8128  * 3. No instruction can modify register %r0 (must be zero)
8129  * 4. All instruction reserved bits must be set to zero
8130  * 5. The last instruction must be a "ret" instruction
8131  * 6. All branch targets must reference a valid instruction _after_ the branch
8132  */
8133 static int
8134 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
8135     cred_t *cr)
8136 {
8137 	int err = 0, i;
8138 	int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8139 	int kcheckload;
8140 	uint_t pc;
8141 
8142 	kcheckload = cr == NULL ||
8143 	    (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
8144 
8145 	dp->dtdo_destructive = 0;
8146 
8147 	for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
8148 		dif_instr_t instr = dp->dtdo_buf[pc];
8149 
8150 		uint_t r1 = DIF_INSTR_R1(instr);
8151 		uint_t r2 = DIF_INSTR_R2(instr);
8152 		uint_t rd = DIF_INSTR_RD(instr);
8153 		uint_t rs = DIF_INSTR_RS(instr);
8154 		uint_t label = DIF_INSTR_LABEL(instr);
8155 		uint_t v = DIF_INSTR_VAR(instr);
8156 		uint_t subr = DIF_INSTR_SUBR(instr);
8157 		uint_t type = DIF_INSTR_TYPE(instr);
8158 		uint_t op = DIF_INSTR_OP(instr);
8159 
8160 		switch (op) {
8161 		case DIF_OP_OR:
8162 		case DIF_OP_XOR:
8163 		case DIF_OP_AND:
8164 		case DIF_OP_SLL:
8165 		case DIF_OP_SRL:
8166 		case DIF_OP_SRA:
8167 		case DIF_OP_SUB:
8168 		case DIF_OP_ADD:
8169 		case DIF_OP_MUL:
8170 		case DIF_OP_SDIV:
8171 		case DIF_OP_UDIV:
8172 		case DIF_OP_SREM:
8173 		case DIF_OP_UREM:
8174 		case DIF_OP_COPYS:
8175 			if (r1 >= nregs)
8176 				err += efunc(pc, "invalid register %u\n", r1);
8177 			if (r2 >= nregs)
8178 				err += efunc(pc, "invalid register %u\n", r2);
8179 			if (rd >= nregs)
8180 				err += efunc(pc, "invalid register %u\n", rd);
8181 			if (rd == 0)
8182 				err += efunc(pc, "cannot write to %r0\n");
8183 			break;
8184 		case DIF_OP_NOT:
8185 		case DIF_OP_MOV:
8186 		case DIF_OP_ALLOCS:
8187 			if (r1 >= nregs)
8188 				err += efunc(pc, "invalid register %u\n", r1);
8189 			if (r2 != 0)
8190 				err += efunc(pc, "non-zero reserved bits\n");
8191 			if (rd >= nregs)
8192 				err += efunc(pc, "invalid register %u\n", rd);
8193 			if (rd == 0)
8194 				err += efunc(pc, "cannot write to %r0\n");
8195 			break;
8196 		case DIF_OP_LDSB:
8197 		case DIF_OP_LDSH:
8198 		case DIF_OP_LDSW:
8199 		case DIF_OP_LDUB:
8200 		case DIF_OP_LDUH:
8201 		case DIF_OP_LDUW:
8202 		case DIF_OP_LDX:
8203 			if (r1 >= nregs)
8204 				err += efunc(pc, "invalid register %u\n", r1);
8205 			if (r2 != 0)
8206 				err += efunc(pc, "non-zero reserved bits\n");
8207 			if (rd >= nregs)
8208 				err += efunc(pc, "invalid register %u\n", rd);
8209 			if (rd == 0)
8210 				err += efunc(pc, "cannot write to %r0\n");
8211 			if (kcheckload)
8212 				dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
8213 				    DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
8214 			break;
8215 		case DIF_OP_RLDSB:
8216 		case DIF_OP_RLDSH:
8217 		case DIF_OP_RLDSW:
8218 		case DIF_OP_RLDUB:
8219 		case DIF_OP_RLDUH:
8220 		case DIF_OP_RLDUW:
8221 		case DIF_OP_RLDX:
8222 			if (r1 >= nregs)
8223 				err += efunc(pc, "invalid register %u\n", r1);
8224 			if (r2 != 0)
8225 				err += efunc(pc, "non-zero reserved bits\n");
8226 			if (rd >= nregs)
8227 				err += efunc(pc, "invalid register %u\n", rd);
8228 			if (rd == 0)
8229 				err += efunc(pc, "cannot write to %r0\n");
8230 			break;
8231 		case DIF_OP_ULDSB:
8232 		case DIF_OP_ULDSH:
8233 		case DIF_OP_ULDSW:
8234 		case DIF_OP_ULDUB:
8235 		case DIF_OP_ULDUH:
8236 		case DIF_OP_ULDUW:
8237 		case DIF_OP_ULDX:
8238 			if (r1 >= nregs)
8239 				err += efunc(pc, "invalid register %u\n", r1);
8240 			if (r2 != 0)
8241 				err += efunc(pc, "non-zero reserved bits\n");
8242 			if (rd >= nregs)
8243 				err += efunc(pc, "invalid register %u\n", rd);
8244 			if (rd == 0)
8245 				err += efunc(pc, "cannot write to %r0\n");
8246 			break;
8247 		case DIF_OP_STB:
8248 		case DIF_OP_STH:
8249 		case DIF_OP_STW:
8250 		case DIF_OP_STX:
8251 			if (r1 >= nregs)
8252 				err += efunc(pc, "invalid register %u\n", r1);
8253 			if (r2 != 0)
8254 				err += efunc(pc, "non-zero reserved bits\n");
8255 			if (rd >= nregs)
8256 				err += efunc(pc, "invalid register %u\n", rd);
8257 			if (rd == 0)
8258 				err += efunc(pc, "cannot write to 0 address\n");
8259 			break;
8260 		case DIF_OP_CMP:
8261 		case DIF_OP_SCMP:
8262 			if (r1 >= nregs)
8263 				err += efunc(pc, "invalid register %u\n", r1);
8264 			if (r2 >= nregs)
8265 				err += efunc(pc, "invalid register %u\n", r2);
8266 			if (rd != 0)
8267 				err += efunc(pc, "non-zero reserved bits\n");
8268 			break;
8269 		case DIF_OP_TST:
8270 			if (r1 >= nregs)
8271 				err += efunc(pc, "invalid register %u\n", r1);
8272 			if (r2 != 0 || rd != 0)
8273 				err += efunc(pc, "non-zero reserved bits\n");
8274 			break;
8275 		case DIF_OP_BA:
8276 		case DIF_OP_BE:
8277 		case DIF_OP_BNE:
8278 		case DIF_OP_BG:
8279 		case DIF_OP_BGU:
8280 		case DIF_OP_BGE:
8281 		case DIF_OP_BGEU:
8282 		case DIF_OP_BL:
8283 		case DIF_OP_BLU:
8284 		case DIF_OP_BLE:
8285 		case DIF_OP_BLEU:
8286 			if (label >= dp->dtdo_len) {
8287 				err += efunc(pc, "invalid branch target %u\n",
8288 				    label);
8289 			}
8290 			if (label <= pc) {
8291 				err += efunc(pc, "backward branch to %u\n",
8292 				    label);
8293 			}
8294 			break;
8295 		case DIF_OP_RET:
8296 			if (r1 != 0 || r2 != 0)
8297 				err += efunc(pc, "non-zero reserved bits\n");
8298 			if (rd >= nregs)
8299 				err += efunc(pc, "invalid register %u\n", rd);
8300 			break;
8301 		case DIF_OP_NOP:
8302 		case DIF_OP_POPTS:
8303 		case DIF_OP_FLUSHTS:
8304 			if (r1 != 0 || r2 != 0 || rd != 0)
8305 				err += efunc(pc, "non-zero reserved bits\n");
8306 			break;
8307 		case DIF_OP_SETX:
8308 			if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
8309 				err += efunc(pc, "invalid integer ref %u\n",
8310 				    DIF_INSTR_INTEGER(instr));
8311 			}
8312 			if (rd >= nregs)
8313 				err += efunc(pc, "invalid register %u\n", rd);
8314 			if (rd == 0)
8315 				err += efunc(pc, "cannot write to %r0\n");
8316 			break;
8317 		case DIF_OP_SETS:
8318 			if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
8319 				err += efunc(pc, "invalid string ref %u\n",
8320 				    DIF_INSTR_STRING(instr));
8321 			}
8322 			if (rd >= nregs)
8323 				err += efunc(pc, "invalid register %u\n", rd);
8324 			if (rd == 0)
8325 				err += efunc(pc, "cannot write to %r0\n");
8326 			break;
8327 		case DIF_OP_LDGA:
8328 		case DIF_OP_LDTA:
8329 			if (r1 > DIF_VAR_ARRAY_MAX)
8330 				err += efunc(pc, "invalid array %u\n", r1);
8331 			if (r2 >= nregs)
8332 				err += efunc(pc, "invalid register %u\n", r2);
8333 			if (rd >= nregs)
8334 				err += efunc(pc, "invalid register %u\n", rd);
8335 			if (rd == 0)
8336 				err += efunc(pc, "cannot write to %r0\n");
8337 			break;
8338 		case DIF_OP_LDGS:
8339 		case DIF_OP_LDTS:
8340 		case DIF_OP_LDLS:
8341 		case DIF_OP_LDGAA:
8342 		case DIF_OP_LDTAA:
8343 			if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
8344 				err += efunc(pc, "invalid variable %u\n", v);
8345 			if (rd >= nregs)
8346 				err += efunc(pc, "invalid register %u\n", rd);
8347 			if (rd == 0)
8348 				err += efunc(pc, "cannot write to %r0\n");
8349 			break;
8350 		case DIF_OP_STGS:
8351 		case DIF_OP_STTS:
8352 		case DIF_OP_STLS:
8353 		case DIF_OP_STGAA:
8354 		case DIF_OP_STTAA:
8355 			if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
8356 				err += efunc(pc, "invalid variable %u\n", v);
8357 			if (rs >= nregs)
8358 				err += efunc(pc, "invalid register %u\n", rd);
8359 			break;
8360 		case DIF_OP_CALL:
8361 			if (subr > DIF_SUBR_MAX)
8362 				err += efunc(pc, "invalid subr %u\n", subr);
8363 			if (rd >= nregs)
8364 				err += efunc(pc, "invalid register %u\n", rd);
8365 			if (rd == 0)
8366 				err += efunc(pc, "cannot write to %r0\n");
8367 
8368 			if (subr == DIF_SUBR_COPYOUT ||
8369 			    subr == DIF_SUBR_COPYOUTSTR) {
8370 				dp->dtdo_destructive = 1;
8371 			}
8372 			break;
8373 		case DIF_OP_PUSHTR:
8374 			if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
8375 				err += efunc(pc, "invalid ref type %u\n", type);
8376 			if (r2 >= nregs)
8377 				err += efunc(pc, "invalid register %u\n", r2);
8378 			if (rs >= nregs)
8379 				err += efunc(pc, "invalid register %u\n", rs);
8380 			break;
8381 		case DIF_OP_PUSHTV:
8382 			if (type != DIF_TYPE_CTF)
8383 				err += efunc(pc, "invalid val type %u\n", type);
8384 			if (r2 >= nregs)
8385 				err += efunc(pc, "invalid register %u\n", r2);
8386 			if (rs >= nregs)
8387 				err += efunc(pc, "invalid register %u\n", rs);
8388 			break;
8389 		default:
8390 			err += efunc(pc, "invalid opcode %u\n",
8391 			    DIF_INSTR_OP(instr));
8392 		}
8393 	}
8394 
8395 	if (dp->dtdo_len != 0 &&
8396 	    DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
8397 		err += efunc(dp->dtdo_len - 1,
8398 		    "expected 'ret' as last DIF instruction\n");
8399 	}
8400 
8401 	if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) {
8402 		/*
8403 		 * If we're not returning by reference, the size must be either
8404 		 * 0 or the size of one of the base types.
8405 		 */
8406 		switch (dp->dtdo_rtype.dtdt_size) {
8407 		case 0:
8408 		case sizeof (uint8_t):
8409 		case sizeof (uint16_t):
8410 		case sizeof (uint32_t):
8411 		case sizeof (uint64_t):
8412 			break;
8413 
8414 		default:
8415 			err += efunc(dp->dtdo_len - 1, "bad return size\n");
8416 		}
8417 	}
8418 
8419 	for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
8420 		dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
8421 		dtrace_diftype_t *vt, *et;
8422 		uint_t id, ndx;
8423 
8424 		if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
8425 		    v->dtdv_scope != DIFV_SCOPE_THREAD &&
8426 		    v->dtdv_scope != DIFV_SCOPE_LOCAL) {
8427 			err += efunc(i, "unrecognized variable scope %d\n",
8428 			    v->dtdv_scope);
8429 			break;
8430 		}
8431 
8432 		if (v->dtdv_kind != DIFV_KIND_ARRAY &&
8433 		    v->dtdv_kind != DIFV_KIND_SCALAR) {
8434 			err += efunc(i, "unrecognized variable type %d\n",
8435 			    v->dtdv_kind);
8436 			break;
8437 		}
8438 
8439 		if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
8440 			err += efunc(i, "%d exceeds variable id limit\n", id);
8441 			break;
8442 		}
8443 
8444 		if (id < DIF_VAR_OTHER_UBASE)
8445 			continue;
8446 
8447 		/*
8448 		 * For user-defined variables, we need to check that this
8449 		 * definition is identical to any previous definition that we
8450 		 * encountered.
8451 		 */
8452 		ndx = id - DIF_VAR_OTHER_UBASE;
8453 
8454 		switch (v->dtdv_scope) {
8455 		case DIFV_SCOPE_GLOBAL:
8456 			if (ndx < vstate->dtvs_nglobals) {
8457 				dtrace_statvar_t *svar;
8458 
8459 				if ((svar = vstate->dtvs_globals[ndx]) != NULL)
8460 					existing = &svar->dtsv_var;
8461 			}
8462 
8463 			break;
8464 
8465 		case DIFV_SCOPE_THREAD:
8466 			if (ndx < vstate->dtvs_ntlocals)
8467 				existing = &vstate->dtvs_tlocals[ndx];
8468 			break;
8469 
8470 		case DIFV_SCOPE_LOCAL:
8471 			if (ndx < vstate->dtvs_nlocals) {
8472 				dtrace_statvar_t *svar;
8473 
8474 				if ((svar = vstate->dtvs_locals[ndx]) != NULL)
8475 					existing = &svar->dtsv_var;
8476 			}
8477 
8478 			break;
8479 		}
8480 
8481 		vt = &v->dtdv_type;
8482 
8483 		if (vt->dtdt_flags & DIF_TF_BYREF) {
8484 			if (vt->dtdt_size == 0) {
8485 				err += efunc(i, "zero-sized variable\n");
8486 				break;
8487 			}
8488 
8489 			if (v->dtdv_scope == DIFV_SCOPE_GLOBAL &&
8490 			    vt->dtdt_size > dtrace_global_maxsize) {
8491 				err += efunc(i, "oversized by-ref global\n");
8492 				break;
8493 			}
8494 		}
8495 
8496 		if (existing == NULL || existing->dtdv_id == 0)
8497 			continue;
8498 
8499 		ASSERT(existing->dtdv_id == v->dtdv_id);
8500 		ASSERT(existing->dtdv_scope == v->dtdv_scope);
8501 
8502 		if (existing->dtdv_kind != v->dtdv_kind)
8503 			err += efunc(i, "%d changed variable kind\n", id);
8504 
8505 		et = &existing->dtdv_type;
8506 
8507 		if (vt->dtdt_flags != et->dtdt_flags) {
8508 			err += efunc(i, "%d changed variable type flags\n", id);
8509 			break;
8510 		}
8511 
8512 		if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
8513 			err += efunc(i, "%d changed variable type size\n", id);
8514 			break;
8515 		}
8516 	}
8517 
8518 	return (err);
8519 }
8520 
8521 /*
8522  * Validate a DTrace DIF object that it is to be used as a helper.  Helpers
8523  * are much more constrained than normal DIFOs.  Specifically, they may
8524  * not:
8525  *
8526  * 1. Make calls to subroutines other than copyin(), copyinstr() or
8527  *    miscellaneous string routines
8528  * 2. Access DTrace variables other than the args[] array, and the
8529  *    curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
8530  * 3. Have thread-local variables.
8531  * 4. Have dynamic variables.
8532  */
8533 static int
8534 dtrace_difo_validate_helper(dtrace_difo_t *dp)
8535 {
8536 	int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8537 	int err = 0;
8538 	uint_t pc;
8539 
8540 	for (pc = 0; pc < dp->dtdo_len; pc++) {
8541 		dif_instr_t instr = dp->dtdo_buf[pc];
8542 
8543 		uint_t v = DIF_INSTR_VAR(instr);
8544 		uint_t subr = DIF_INSTR_SUBR(instr);
8545 		uint_t op = DIF_INSTR_OP(instr);
8546 
8547 		switch (op) {
8548 		case DIF_OP_OR:
8549 		case DIF_OP_XOR:
8550 		case DIF_OP_AND:
8551 		case DIF_OP_SLL:
8552 		case DIF_OP_SRL:
8553 		case DIF_OP_SRA:
8554 		case DIF_OP_SUB:
8555 		case DIF_OP_ADD:
8556 		case DIF_OP_MUL:
8557 		case DIF_OP_SDIV:
8558 		case DIF_OP_UDIV:
8559 		case DIF_OP_SREM:
8560 		case DIF_OP_UREM:
8561 		case DIF_OP_COPYS:
8562 		case DIF_OP_NOT:
8563 		case DIF_OP_MOV:
8564 		case DIF_OP_RLDSB:
8565 		case DIF_OP_RLDSH:
8566 		case DIF_OP_RLDSW:
8567 		case DIF_OP_RLDUB:
8568 		case DIF_OP_RLDUH:
8569 		case DIF_OP_RLDUW:
8570 		case DIF_OP_RLDX:
8571 		case DIF_OP_ULDSB:
8572 		case DIF_OP_ULDSH:
8573 		case DIF_OP_ULDSW:
8574 		case DIF_OP_ULDUB:
8575 		case DIF_OP_ULDUH:
8576 		case DIF_OP_ULDUW:
8577 		case DIF_OP_ULDX:
8578 		case DIF_OP_STB:
8579 		case DIF_OP_STH:
8580 		case DIF_OP_STW:
8581 		case DIF_OP_STX:
8582 		case DIF_OP_ALLOCS:
8583 		case DIF_OP_CMP:
8584 		case DIF_OP_SCMP:
8585 		case DIF_OP_TST:
8586 		case DIF_OP_BA:
8587 		case DIF_OP_BE:
8588 		case DIF_OP_BNE:
8589 		case DIF_OP_BG:
8590 		case DIF_OP_BGU:
8591 		case DIF_OP_BGE:
8592 		case DIF_OP_BGEU:
8593 		case DIF_OP_BL:
8594 		case DIF_OP_BLU:
8595 		case DIF_OP_BLE:
8596 		case DIF_OP_BLEU:
8597 		case DIF_OP_RET:
8598 		case DIF_OP_NOP:
8599 		case DIF_OP_POPTS:
8600 		case DIF_OP_FLUSHTS:
8601 		case DIF_OP_SETX:
8602 		case DIF_OP_SETS:
8603 		case DIF_OP_LDGA:
8604 		case DIF_OP_LDLS:
8605 		case DIF_OP_STGS:
8606 		case DIF_OP_STLS:
8607 		case DIF_OP_PUSHTR:
8608 		case DIF_OP_PUSHTV:
8609 			break;
8610 
8611 		case DIF_OP_LDGS:
8612 			if (v >= DIF_VAR_OTHER_UBASE)
8613 				break;
8614 
8615 			if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
8616 				break;
8617 
8618 			if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
8619 			    v == DIF_VAR_PPID || v == DIF_VAR_TID ||
8620 			    v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
8621 			    v == DIF_VAR_UID || v == DIF_VAR_GID)
8622 				break;
8623 
8624 			err += efunc(pc, "illegal variable %u\n", v);
8625 			break;
8626 
8627 		case DIF_OP_LDTA:
8628 		case DIF_OP_LDTS:
8629 		case DIF_OP_LDGAA:
8630 		case DIF_OP_LDTAA:
8631 			err += efunc(pc, "illegal dynamic variable load\n");
8632 			break;
8633 
8634 		case DIF_OP_STTS:
8635 		case DIF_OP_STGAA:
8636 		case DIF_OP_STTAA:
8637 			err += efunc(pc, "illegal dynamic variable store\n");
8638 			break;
8639 
8640 		case DIF_OP_CALL:
8641 			if (subr == DIF_SUBR_ALLOCA ||
8642 			    subr == DIF_SUBR_BCOPY ||
8643 			    subr == DIF_SUBR_COPYIN ||
8644 			    subr == DIF_SUBR_COPYINTO ||
8645 			    subr == DIF_SUBR_COPYINSTR ||
8646 			    subr == DIF_SUBR_INDEX ||
8647 			    subr == DIF_SUBR_INET_NTOA ||
8648 			    subr == DIF_SUBR_INET_NTOA6 ||
8649 			    subr == DIF_SUBR_INET_NTOP ||
8650 			    subr == DIF_SUBR_LLTOSTR ||
8651 			    subr == DIF_SUBR_RINDEX ||
8652 			    subr == DIF_SUBR_STRCHR ||
8653 			    subr == DIF_SUBR_STRJOIN ||
8654 			    subr == DIF_SUBR_STRRCHR ||
8655 			    subr == DIF_SUBR_STRSTR ||
8656 			    subr == DIF_SUBR_HTONS ||
8657 			    subr == DIF_SUBR_HTONL ||
8658 			    subr == DIF_SUBR_HTONLL ||
8659 			    subr == DIF_SUBR_NTOHS ||
8660 			    subr == DIF_SUBR_NTOHL ||
8661 			    subr == DIF_SUBR_NTOHLL)
8662 				break;
8663 
8664 			err += efunc(pc, "invalid subr %u\n", subr);
8665 			break;
8666 
8667 		default:
8668 			err += efunc(pc, "invalid opcode %u\n",
8669 			    DIF_INSTR_OP(instr));
8670 		}
8671 	}
8672 
8673 	return (err);
8674 }
8675 
8676 /*
8677  * Returns 1 if the expression in the DIF object can be cached on a per-thread
8678  * basis; 0 if not.
8679  */
8680 static int
8681 dtrace_difo_cacheable(dtrace_difo_t *dp)
8682 {
8683 	int i;
8684 
8685 	if (dp == NULL)
8686 		return (0);
8687 
8688 	for (i = 0; i < dp->dtdo_varlen; i++) {
8689 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
8690 
8691 		if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
8692 			continue;
8693 
8694 		switch (v->dtdv_id) {
8695 		case DIF_VAR_CURTHREAD:
8696 		case DIF_VAR_PID:
8697 		case DIF_VAR_TID:
8698 		case DIF_VAR_EXECNAME:
8699 		case DIF_VAR_ZONENAME:
8700 			break;
8701 
8702 		default:
8703 			return (0);
8704 		}
8705 	}
8706 
8707 	/*
8708 	 * This DIF object may be cacheable.  Now we need to look for any
8709 	 * array loading instructions, any memory loading instructions, or
8710 	 * any stores to thread-local variables.
8711 	 */
8712 	for (i = 0; i < dp->dtdo_len; i++) {
8713 		uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
8714 
8715 		if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
8716 		    (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
8717 		    (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
8718 		    op == DIF_OP_LDGA || op == DIF_OP_STTS)
8719 			return (0);
8720 	}
8721 
8722 	return (1);
8723 }
8724 
8725 static void
8726 dtrace_difo_hold(dtrace_difo_t *dp)
8727 {
8728 	int i;
8729 
8730 	ASSERT(MUTEX_HELD(&dtrace_lock));
8731 
8732 	dp->dtdo_refcnt++;
8733 	ASSERT(dp->dtdo_refcnt != 0);
8734 
8735 	/*
8736 	 * We need to check this DIF object for references to the variable
8737 	 * DIF_VAR_VTIMESTAMP.
8738 	 */
8739 	for (i = 0; i < dp->dtdo_varlen; i++) {
8740 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
8741 
8742 		if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
8743 			continue;
8744 
8745 		if (dtrace_vtime_references++ == 0)
8746 			dtrace_vtime_enable();
8747 	}
8748 }
8749 
8750 /*
8751  * This routine calculates the dynamic variable chunksize for a given DIF
8752  * object.  The calculation is not fool-proof, and can probably be tricked by
8753  * malicious DIF -- but it works for all compiler-generated DIF.  Because this
8754  * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
8755  * if a dynamic variable size exceeds the chunksize.
8756  */
8757 static void
8758 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
8759 {
8760 	uint64_t sval;
8761 	dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
8762 	const dif_instr_t *text = dp->dtdo_buf;
8763 	uint_t pc, srd = 0;
8764 	uint_t ttop = 0;
8765 	size_t size, ksize;
8766 	uint_t id, i;
8767 
8768 	for (pc = 0; pc < dp->dtdo_len; pc++) {
8769 		dif_instr_t instr = text[pc];
8770 		uint_t op = DIF_INSTR_OP(instr);
8771 		uint_t rd = DIF_INSTR_RD(instr);
8772 		uint_t r1 = DIF_INSTR_R1(instr);
8773 		uint_t nkeys = 0;
8774 		uchar_t scope;
8775 
8776 		dtrace_key_t *key = tupregs;
8777 
8778 		switch (op) {
8779 		case DIF_OP_SETX:
8780 			sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
8781 			srd = rd;
8782 			continue;
8783 
8784 		case DIF_OP_STTS:
8785 			key = &tupregs[DIF_DTR_NREGS];
8786 			key[0].dttk_size = 0;
8787 			key[1].dttk_size = 0;
8788 			nkeys = 2;
8789 			scope = DIFV_SCOPE_THREAD;
8790 			break;
8791 
8792 		case DIF_OP_STGAA:
8793 		case DIF_OP_STTAA:
8794 			nkeys = ttop;
8795 
8796 			if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
8797 				key[nkeys++].dttk_size = 0;
8798 
8799 			key[nkeys++].dttk_size = 0;
8800 
8801 			if (op == DIF_OP_STTAA) {
8802 				scope = DIFV_SCOPE_THREAD;
8803 			} else {
8804 				scope = DIFV_SCOPE_GLOBAL;
8805 			}
8806 
8807 			break;
8808 
8809 		case DIF_OP_PUSHTR:
8810 			if (ttop == DIF_DTR_NREGS)
8811 				return;
8812 
8813 			if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
8814 				/*
8815 				 * If the register for the size of the "pushtr"
8816 				 * is %r0 (or the value is 0) and the type is
8817 				 * a string, we'll use the system-wide default
8818 				 * string size.
8819 				 */
8820 				tupregs[ttop++].dttk_size =
8821 				    dtrace_strsize_default;
8822 			} else {
8823 				if (srd == 0)
8824 					return;
8825 
8826 				tupregs[ttop++].dttk_size = sval;
8827 			}
8828 
8829 			break;
8830 
8831 		case DIF_OP_PUSHTV:
8832 			if (ttop == DIF_DTR_NREGS)
8833 				return;
8834 
8835 			tupregs[ttop++].dttk_size = 0;
8836 			break;
8837 
8838 		case DIF_OP_FLUSHTS:
8839 			ttop = 0;
8840 			break;
8841 
8842 		case DIF_OP_POPTS:
8843 			if (ttop != 0)
8844 				ttop--;
8845 			break;
8846 		}
8847 
8848 		sval = 0;
8849 		srd = 0;
8850 
8851 		if (nkeys == 0)
8852 			continue;
8853 
8854 		/*
8855 		 * We have a dynamic variable allocation; calculate its size.
8856 		 */
8857 		for (ksize = 0, i = 0; i < nkeys; i++)
8858 			ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
8859 
8860 		size = sizeof (dtrace_dynvar_t);
8861 		size += sizeof (dtrace_key_t) * (nkeys - 1);
8862 		size += ksize;
8863 
8864 		/*
8865 		 * Now we need to determine the size of the stored data.
8866 		 */
8867 		id = DIF_INSTR_VAR(instr);
8868 
8869 		for (i = 0; i < dp->dtdo_varlen; i++) {
8870 			dtrace_difv_t *v = &dp->dtdo_vartab[i];
8871 
8872 			if (v->dtdv_id == id && v->dtdv_scope == scope) {
8873 				size += v->dtdv_type.dtdt_size;
8874 				break;
8875 			}
8876 		}
8877 
8878 		if (i == dp->dtdo_varlen)
8879 			return;
8880 
8881 		/*
8882 		 * We have the size.  If this is larger than the chunk size
8883 		 * for our dynamic variable state, reset the chunk size.
8884 		 */
8885 		size = P2ROUNDUP(size, sizeof (uint64_t));
8886 
8887 		if (size > vstate->dtvs_dynvars.dtds_chunksize)
8888 			vstate->dtvs_dynvars.dtds_chunksize = size;
8889 	}
8890 }
8891 
8892 static void
8893 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
8894 {
8895 	int i, oldsvars, osz, nsz, otlocals, ntlocals;
8896 	uint_t id;
8897 
8898 	ASSERT(MUTEX_HELD(&dtrace_lock));
8899 	ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
8900 
8901 	for (i = 0; i < dp->dtdo_varlen; i++) {
8902 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
8903 		dtrace_statvar_t *svar, ***svarp;
8904 		size_t dsize = 0;
8905 		uint8_t scope = v->dtdv_scope;
8906 		int *np;
8907 
8908 		if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
8909 			continue;
8910 
8911 		id -= DIF_VAR_OTHER_UBASE;
8912 
8913 		switch (scope) {
8914 		case DIFV_SCOPE_THREAD:
8915 			while (id >= (otlocals = vstate->dtvs_ntlocals)) {
8916 				dtrace_difv_t *tlocals;
8917 
8918 				if ((ntlocals = (otlocals << 1)) == 0)
8919 					ntlocals = 1;
8920 
8921 				osz = otlocals * sizeof (dtrace_difv_t);
8922 				nsz = ntlocals * sizeof (dtrace_difv_t);
8923 
8924 				tlocals = kmem_zalloc(nsz, KM_SLEEP);
8925 
8926 				if (osz != 0) {
8927 					bcopy(vstate->dtvs_tlocals,
8928 					    tlocals, osz);
8929 					kmem_free(vstate->dtvs_tlocals, osz);
8930 				}
8931 
8932 				vstate->dtvs_tlocals = tlocals;
8933 				vstate->dtvs_ntlocals = ntlocals;
8934 			}
8935 
8936 			vstate->dtvs_tlocals[id] = *v;
8937 			continue;
8938 
8939 		case DIFV_SCOPE_LOCAL:
8940 			np = &vstate->dtvs_nlocals;
8941 			svarp = &vstate->dtvs_locals;
8942 
8943 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
8944 				dsize = NCPU * (v->dtdv_type.dtdt_size +
8945 				    sizeof (uint64_t));
8946 			else
8947 				dsize = NCPU * sizeof (uint64_t);
8948 
8949 			break;
8950 
8951 		case DIFV_SCOPE_GLOBAL:
8952 			np = &vstate->dtvs_nglobals;
8953 			svarp = &vstate->dtvs_globals;
8954 
8955 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
8956 				dsize = v->dtdv_type.dtdt_size +
8957 				    sizeof (uint64_t);
8958 
8959 			break;
8960 
8961 		default:
8962 			ASSERT(0);
8963 		}
8964 
8965 		while (id >= (oldsvars = *np)) {
8966 			dtrace_statvar_t **statics;
8967 			int newsvars, oldsize, newsize;
8968 
8969 			if ((newsvars = (oldsvars << 1)) == 0)
8970 				newsvars = 1;
8971 
8972 			oldsize = oldsvars * sizeof (dtrace_statvar_t *);
8973 			newsize = newsvars * sizeof (dtrace_statvar_t *);
8974 
8975 			statics = kmem_zalloc(newsize, KM_SLEEP);
8976 
8977 			if (oldsize != 0) {
8978 				bcopy(*svarp, statics, oldsize);
8979 				kmem_free(*svarp, oldsize);
8980 			}
8981 
8982 			*svarp = statics;
8983 			*np = newsvars;
8984 		}
8985 
8986 		if ((svar = (*svarp)[id]) == NULL) {
8987 			svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
8988 			svar->dtsv_var = *v;
8989 
8990 			if ((svar->dtsv_size = dsize) != 0) {
8991 				svar->dtsv_data = (uint64_t)(uintptr_t)
8992 				    kmem_zalloc(dsize, KM_SLEEP);
8993 			}
8994 
8995 			(*svarp)[id] = svar;
8996 		}
8997 
8998 		svar->dtsv_refcnt++;
8999 	}
9000 
9001 	dtrace_difo_chunksize(dp, vstate);
9002 	dtrace_difo_hold(dp);
9003 }
9004 
9005 static dtrace_difo_t *
9006 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9007 {
9008 	dtrace_difo_t *new;
9009 	size_t sz;
9010 
9011 	ASSERT(dp->dtdo_buf != NULL);
9012 	ASSERT(dp->dtdo_refcnt != 0);
9013 
9014 	new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
9015 
9016 	ASSERT(dp->dtdo_buf != NULL);
9017 	sz = dp->dtdo_len * sizeof (dif_instr_t);
9018 	new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
9019 	bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
9020 	new->dtdo_len = dp->dtdo_len;
9021 
9022 	if (dp->dtdo_strtab != NULL) {
9023 		ASSERT(dp->dtdo_strlen != 0);
9024 		new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
9025 		bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
9026 		new->dtdo_strlen = dp->dtdo_strlen;
9027 	}
9028 
9029 	if (dp->dtdo_inttab != NULL) {
9030 		ASSERT(dp->dtdo_intlen != 0);
9031 		sz = dp->dtdo_intlen * sizeof (uint64_t);
9032 		new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
9033 		bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
9034 		new->dtdo_intlen = dp->dtdo_intlen;
9035 	}
9036 
9037 	if (dp->dtdo_vartab != NULL) {
9038 		ASSERT(dp->dtdo_varlen != 0);
9039 		sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
9040 		new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
9041 		bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
9042 		new->dtdo_varlen = dp->dtdo_varlen;
9043 	}
9044 
9045 	dtrace_difo_init(new, vstate);
9046 	return (new);
9047 }
9048 
9049 static void
9050 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9051 {
9052 	int i;
9053 
9054 	ASSERT(dp->dtdo_refcnt == 0);
9055 
9056 	for (i = 0; i < dp->dtdo_varlen; i++) {
9057 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
9058 		dtrace_statvar_t *svar, **svarp;
9059 		uint_t id;
9060 		uint8_t scope = v->dtdv_scope;
9061 		int *np;
9062 
9063 		switch (scope) {
9064 		case DIFV_SCOPE_THREAD:
9065 			continue;
9066 
9067 		case DIFV_SCOPE_LOCAL:
9068 			np = &vstate->dtvs_nlocals;
9069 			svarp = vstate->dtvs_locals;
9070 			break;
9071 
9072 		case DIFV_SCOPE_GLOBAL:
9073 			np = &vstate->dtvs_nglobals;
9074 			svarp = vstate->dtvs_globals;
9075 			break;
9076 
9077 		default:
9078 			ASSERT(0);
9079 		}
9080 
9081 		if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9082 			continue;
9083 
9084 		id -= DIF_VAR_OTHER_UBASE;
9085 		ASSERT(id < *np);
9086 
9087 		svar = svarp[id];
9088 		ASSERT(svar != NULL);
9089 		ASSERT(svar->dtsv_refcnt > 0);
9090 
9091 		if (--svar->dtsv_refcnt > 0)
9092 			continue;
9093 
9094 		if (svar->dtsv_size != 0) {
9095 			ASSERT(svar->dtsv_data != NULL);
9096 			kmem_free((void *)(uintptr_t)svar->dtsv_data,
9097 			    svar->dtsv_size);
9098 		}
9099 
9100 		kmem_free(svar, sizeof (dtrace_statvar_t));
9101 		svarp[id] = NULL;
9102 	}
9103 
9104 	kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
9105 	kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
9106 	kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
9107 	kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
9108 
9109 	kmem_free(dp, sizeof (dtrace_difo_t));
9110 }
9111 
9112 static void
9113 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9114 {
9115 	int i;
9116 
9117 	ASSERT(MUTEX_HELD(&dtrace_lock));
9118 	ASSERT(dp->dtdo_refcnt != 0);
9119 
9120 	for (i = 0; i < dp->dtdo_varlen; i++) {
9121 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
9122 
9123 		if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9124 			continue;
9125 
9126 		ASSERT(dtrace_vtime_references > 0);
9127 		if (--dtrace_vtime_references == 0)
9128 			dtrace_vtime_disable();
9129 	}
9130 
9131 	if (--dp->dtdo_refcnt == 0)
9132 		dtrace_difo_destroy(dp, vstate);
9133 }
9134 
9135 /*
9136  * DTrace Format Functions
9137  */
9138 static uint16_t
9139 dtrace_format_add(dtrace_state_t *state, char *str)
9140 {
9141 	char *fmt, **new;
9142 	uint16_t ndx, len = strlen(str) + 1;
9143 
9144 	fmt = kmem_zalloc(len, KM_SLEEP);
9145 	bcopy(str, fmt, len);
9146 
9147 	for (ndx = 0; ndx < state->dts_nformats; ndx++) {
9148 		if (state->dts_formats[ndx] == NULL) {
9149 			state->dts_formats[ndx] = fmt;
9150 			return (ndx + 1);
9151 		}
9152 	}
9153 
9154 	if (state->dts_nformats == USHRT_MAX) {
9155 		/*
9156 		 * This is only likely if a denial-of-service attack is being
9157 		 * attempted.  As such, it's okay to fail silently here.
9158 		 */
9159 		kmem_free(fmt, len);
9160 		return (0);
9161 	}
9162 
9163 	/*
9164 	 * For simplicity, we always resize the formats array to be exactly the
9165 	 * number of formats.
9166 	 */
9167 	ndx = state->dts_nformats++;
9168 	new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
9169 
9170 	if (state->dts_formats != NULL) {
9171 		ASSERT(ndx != 0);
9172 		bcopy(state->dts_formats, new, ndx * sizeof (char *));
9173 		kmem_free(state->dts_formats, ndx * sizeof (char *));
9174 	}
9175 
9176 	state->dts_formats = new;
9177 	state->dts_formats[ndx] = fmt;
9178 
9179 	return (ndx + 1);
9180 }
9181 
9182 static void
9183 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
9184 {
9185 	char *fmt;
9186 
9187 	ASSERT(state->dts_formats != NULL);
9188 	ASSERT(format <= state->dts_nformats);
9189 	ASSERT(state->dts_formats[format - 1] != NULL);
9190 
9191 	fmt = state->dts_formats[format - 1];
9192 	kmem_free(fmt, strlen(fmt) + 1);
9193 	state->dts_formats[format - 1] = NULL;
9194 }
9195 
9196 static void
9197 dtrace_format_destroy(dtrace_state_t *state)
9198 {
9199 	int i;
9200 
9201 	if (state->dts_nformats == 0) {
9202 		ASSERT(state->dts_formats == NULL);
9203 		return;
9204 	}
9205 
9206 	ASSERT(state->dts_formats != NULL);
9207 
9208 	for (i = 0; i < state->dts_nformats; i++) {
9209 		char *fmt = state->dts_formats[i];
9210 
9211 		if (fmt == NULL)
9212 			continue;
9213 
9214 		kmem_free(fmt, strlen(fmt) + 1);
9215 	}
9216 
9217 	kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
9218 	state->dts_nformats = 0;
9219 	state->dts_formats = NULL;
9220 }
9221 
9222 /*
9223  * DTrace Predicate Functions
9224  */
9225 static dtrace_predicate_t *
9226 dtrace_predicate_create(dtrace_difo_t *dp)
9227 {
9228 	dtrace_predicate_t *pred;
9229 
9230 	ASSERT(MUTEX_HELD(&dtrace_lock));
9231 	ASSERT(dp->dtdo_refcnt != 0);
9232 
9233 	pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
9234 	pred->dtp_difo = dp;
9235 	pred->dtp_refcnt = 1;
9236 
9237 	if (!dtrace_difo_cacheable(dp))
9238 		return (pred);
9239 
9240 	if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
9241 		/*
9242 		 * This is only theoretically possible -- we have had 2^32
9243 		 * cacheable predicates on this machine.  We cannot allow any
9244 		 * more predicates to become cacheable:  as unlikely as it is,
9245 		 * there may be a thread caching a (now stale) predicate cache
9246 		 * ID. (N.B.: the temptation is being successfully resisted to
9247 		 * have this cmn_err() "Holy shit -- we executed this code!")
9248 		 */
9249 		return (pred);
9250 	}
9251 
9252 	pred->dtp_cacheid = dtrace_predcache_id++;
9253 
9254 	return (pred);
9255 }
9256 
9257 static void
9258 dtrace_predicate_hold(dtrace_predicate_t *pred)
9259 {
9260 	ASSERT(MUTEX_HELD(&dtrace_lock));
9261 	ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
9262 	ASSERT(pred->dtp_refcnt > 0);
9263 
9264 	pred->dtp_refcnt++;
9265 }
9266 
9267 static void
9268 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
9269 {
9270 	dtrace_difo_t *dp = pred->dtp_difo;
9271 
9272 	ASSERT(MUTEX_HELD(&dtrace_lock));
9273 	ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
9274 	ASSERT(pred->dtp_refcnt > 0);
9275 
9276 	if (--pred->dtp_refcnt == 0) {
9277 		dtrace_difo_release(pred->dtp_difo, vstate);
9278 		kmem_free(pred, sizeof (dtrace_predicate_t));
9279 	}
9280 }
9281 
9282 /*
9283  * DTrace Action Description Functions
9284  */
9285 static dtrace_actdesc_t *
9286 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
9287     uint64_t uarg, uint64_t arg)
9288 {
9289 	dtrace_actdesc_t *act;
9290 
9291 	ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
9292 	    arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
9293 
9294 	act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
9295 	act->dtad_kind = kind;
9296 	act->dtad_ntuple = ntuple;
9297 	act->dtad_uarg = uarg;
9298 	act->dtad_arg = arg;
9299 	act->dtad_refcnt = 1;
9300 
9301 	return (act);
9302 }
9303 
9304 static void
9305 dtrace_actdesc_hold(dtrace_actdesc_t *act)
9306 {
9307 	ASSERT(act->dtad_refcnt >= 1);
9308 	act->dtad_refcnt++;
9309 }
9310 
9311 static void
9312 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
9313 {
9314 	dtrace_actkind_t kind = act->dtad_kind;
9315 	dtrace_difo_t *dp;
9316 
9317 	ASSERT(act->dtad_refcnt >= 1);
9318 
9319 	if (--act->dtad_refcnt != 0)
9320 		return;
9321 
9322 	if ((dp = act->dtad_difo) != NULL)
9323 		dtrace_difo_release(dp, vstate);
9324 
9325 	if (DTRACEACT_ISPRINTFLIKE(kind)) {
9326 		char *str = (char *)(uintptr_t)act->dtad_arg;
9327 
9328 		ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
9329 		    (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
9330 
9331 		if (str != NULL)
9332 			kmem_free(str, strlen(str) + 1);
9333 	}
9334 
9335 	kmem_free(act, sizeof (dtrace_actdesc_t));
9336 }
9337 
9338 /*
9339  * DTrace ECB Functions
9340  */
9341 static dtrace_ecb_t *
9342 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
9343 {
9344 	dtrace_ecb_t *ecb;
9345 	dtrace_epid_t epid;
9346 
9347 	ASSERT(MUTEX_HELD(&dtrace_lock));
9348 
9349 	ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
9350 	ecb->dte_predicate = NULL;
9351 	ecb->dte_probe = probe;
9352 
9353 	/*
9354 	 * The default size is the size of the default action: recording
9355 	 * the epid.
9356 	 */
9357 	ecb->dte_size = ecb->dte_needed = sizeof (dtrace_epid_t);
9358 	ecb->dte_alignment = sizeof (dtrace_epid_t);
9359 
9360 	epid = state->dts_epid++;
9361 
9362 	if (epid - 1 >= state->dts_necbs) {
9363 		dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
9364 		int necbs = state->dts_necbs << 1;
9365 
9366 		ASSERT(epid == state->dts_necbs + 1);
9367 
9368 		if (necbs == 0) {
9369 			ASSERT(oecbs == NULL);
9370 			necbs = 1;
9371 		}
9372 
9373 		ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
9374 
9375 		if (oecbs != NULL)
9376 			bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
9377 
9378 		dtrace_membar_producer();
9379 		state->dts_ecbs = ecbs;
9380 
9381 		if (oecbs != NULL) {
9382 			/*
9383 			 * If this state is active, we must dtrace_sync()
9384 			 * before we can free the old dts_ecbs array:  we're
9385 			 * coming in hot, and there may be active ring
9386 			 * buffer processing (which indexes into the dts_ecbs
9387 			 * array) on another CPU.
9388 			 */
9389 			if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
9390 				dtrace_sync();
9391 
9392 			kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
9393 		}
9394 
9395 		dtrace_membar_producer();
9396 		state->dts_necbs = necbs;
9397 	}
9398 
9399 	ecb->dte_state = state;
9400 
9401 	ASSERT(state->dts_ecbs[epid - 1] == NULL);
9402 	dtrace_membar_producer();
9403 	state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
9404 
9405 	return (ecb);
9406 }
9407 
9408 static int
9409 dtrace_ecb_enable(dtrace_ecb_t *ecb)
9410 {
9411 	dtrace_probe_t *probe = ecb->dte_probe;
9412 
9413 	ASSERT(MUTEX_HELD(&cpu_lock));
9414 	ASSERT(MUTEX_HELD(&dtrace_lock));
9415 	ASSERT(ecb->dte_next == NULL);
9416 
9417 	if (probe == NULL) {
9418 		/*
9419 		 * This is the NULL probe -- there's nothing to do.
9420 		 */
9421 		return (0);
9422 	}
9423 
9424 	if (probe->dtpr_ecb == NULL) {
9425 		dtrace_provider_t *prov = probe->dtpr_provider;
9426 
9427 		/*
9428 		 * We're the first ECB on this probe.
9429 		 */
9430 		probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
9431 
9432 		if (ecb->dte_predicate != NULL)
9433 			probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
9434 
9435 		return (prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
9436 		    probe->dtpr_id, probe->dtpr_arg));
9437 	} else {
9438 		/*
9439 		 * This probe is already active.  Swing the last pointer to
9440 		 * point to the new ECB, and issue a dtrace_sync() to assure
9441 		 * that all CPUs have seen the change.
9442 		 */
9443 		ASSERT(probe->dtpr_ecb_last != NULL);
9444 		probe->dtpr_ecb_last->dte_next = ecb;
9445 		probe->dtpr_ecb_last = ecb;
9446 		probe->dtpr_predcache = 0;
9447 
9448 		dtrace_sync();
9449 		return (0);
9450 	}
9451 }
9452 
9453 static void
9454 dtrace_ecb_resize(dtrace_ecb_t *ecb)
9455 {
9456 	uint32_t maxalign = sizeof (dtrace_epid_t);
9457 	uint32_t align = sizeof (uint8_t), offs, diff;
9458 	dtrace_action_t *act;
9459 	int wastuple = 0;
9460 	uint32_t aggbase = UINT32_MAX;
9461 	dtrace_state_t *state = ecb->dte_state;
9462 
9463 	/*
9464 	 * If we record anything, we always record the epid.  (And we always
9465 	 * record it first.)
9466 	 */
9467 	offs = sizeof (dtrace_epid_t);
9468 	ecb->dte_size = ecb->dte_needed = sizeof (dtrace_epid_t);
9469 
9470 	for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
9471 		dtrace_recdesc_t *rec = &act->dta_rec;
9472 
9473 		if ((align = rec->dtrd_alignment) > maxalign)
9474 			maxalign = align;
9475 
9476 		if (!wastuple && act->dta_intuple) {
9477 			/*
9478 			 * This is the first record in a tuple.  Align the
9479 			 * offset to be at offset 4 in an 8-byte aligned
9480 			 * block.
9481 			 */
9482 			diff = offs + sizeof (dtrace_aggid_t);
9483 
9484 			if (diff = (diff & (sizeof (uint64_t) - 1)))
9485 				offs += sizeof (uint64_t) - diff;
9486 
9487 			aggbase = offs - sizeof (dtrace_aggid_t);
9488 			ASSERT(!(aggbase & (sizeof (uint64_t) - 1)));
9489 		}
9490 
9491 		/*LINTED*/
9492 		if (rec->dtrd_size != 0 && (diff = (offs & (align - 1)))) {
9493 			/*
9494 			 * The current offset is not properly aligned; align it.
9495 			 */
9496 			offs += align - diff;
9497 		}
9498 
9499 		rec->dtrd_offset = offs;
9500 
9501 		if (offs + rec->dtrd_size > ecb->dte_needed) {
9502 			ecb->dte_needed = offs + rec->dtrd_size;
9503 
9504 			if (ecb->dte_needed > state->dts_needed)
9505 				state->dts_needed = ecb->dte_needed;
9506 		}
9507 
9508 		if (DTRACEACT_ISAGG(act->dta_kind)) {
9509 			dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
9510 			dtrace_action_t *first = agg->dtag_first, *prev;
9511 
9512 			ASSERT(rec->dtrd_size != 0 && first != NULL);
9513 			ASSERT(wastuple);
9514 			ASSERT(aggbase != UINT32_MAX);
9515 
9516 			agg->dtag_base = aggbase;
9517 
9518 			while ((prev = first->dta_prev) != NULL &&
9519 			    DTRACEACT_ISAGG(prev->dta_kind)) {
9520 				agg = (dtrace_aggregation_t *)prev;
9521 				first = agg->dtag_first;
9522 			}
9523 
9524 			if (prev != NULL) {
9525 				offs = prev->dta_rec.dtrd_offset +
9526 				    prev->dta_rec.dtrd_size;
9527 			} else {
9528 				offs = sizeof (dtrace_epid_t);
9529 			}
9530 			wastuple = 0;
9531 		} else {
9532 			if (!act->dta_intuple)
9533 				ecb->dte_size = offs + rec->dtrd_size;
9534 
9535 			offs += rec->dtrd_size;
9536 		}
9537 
9538 		wastuple = act->dta_intuple;
9539 	}
9540 
9541 	if ((act = ecb->dte_action) != NULL &&
9542 	    !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
9543 	    ecb->dte_size == sizeof (dtrace_epid_t)) {
9544 		/*
9545 		 * If the size is still sizeof (dtrace_epid_t), then all
9546 		 * actions store no data; set the size to 0.
9547 		 */
9548 		ecb->dte_alignment = maxalign;
9549 		ecb->dte_size = 0;
9550 
9551 		/*
9552 		 * If the needed space is still sizeof (dtrace_epid_t), then
9553 		 * all actions need no additional space; set the needed
9554 		 * size to 0.
9555 		 */
9556 		if (ecb->dte_needed == sizeof (dtrace_epid_t))
9557 			ecb->dte_needed = 0;
9558 
9559 		return;
9560 	}
9561 
9562 	/*
9563 	 * Set our alignment, and make sure that the dte_size and dte_needed
9564 	 * are aligned to the size of an EPID.
9565 	 */
9566 	ecb->dte_alignment = maxalign;
9567 	ecb->dte_size = (ecb->dte_size + (sizeof (dtrace_epid_t) - 1)) &
9568 	    ~(sizeof (dtrace_epid_t) - 1);
9569 	ecb->dte_needed = (ecb->dte_needed + (sizeof (dtrace_epid_t) - 1)) &
9570 	    ~(sizeof (dtrace_epid_t) - 1);
9571 	ASSERT(ecb->dte_size <= ecb->dte_needed);
9572 }
9573 
9574 static dtrace_action_t *
9575 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
9576 {
9577 	dtrace_aggregation_t *agg;
9578 	size_t size = sizeof (uint64_t);
9579 	int ntuple = desc->dtad_ntuple;
9580 	dtrace_action_t *act;
9581 	dtrace_recdesc_t *frec;
9582 	dtrace_aggid_t aggid;
9583 	dtrace_state_t *state = ecb->dte_state;
9584 
9585 	agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
9586 	agg->dtag_ecb = ecb;
9587 
9588 	ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
9589 
9590 	switch (desc->dtad_kind) {
9591 	case DTRACEAGG_MIN:
9592 		agg->dtag_initial = INT64_MAX;
9593 		agg->dtag_aggregate = dtrace_aggregate_min;
9594 		break;
9595 
9596 	case DTRACEAGG_MAX:
9597 		agg->dtag_initial = INT64_MIN;
9598 		agg->dtag_aggregate = dtrace_aggregate_max;
9599 		break;
9600 
9601 	case DTRACEAGG_COUNT:
9602 		agg->dtag_aggregate = dtrace_aggregate_count;
9603 		break;
9604 
9605 	case DTRACEAGG_QUANTIZE:
9606 		agg->dtag_aggregate = dtrace_aggregate_quantize;
9607 		size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
9608 		    sizeof (uint64_t);
9609 		break;
9610 
9611 	case DTRACEAGG_LQUANTIZE: {
9612 		uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
9613 		uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
9614 
9615 		agg->dtag_initial = desc->dtad_arg;
9616 		agg->dtag_aggregate = dtrace_aggregate_lquantize;
9617 
9618 		if (step == 0 || levels == 0)
9619 			goto err;
9620 
9621 		size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
9622 		break;
9623 	}
9624 
9625 	case DTRACEAGG_LLQUANTIZE: {
9626 		uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(desc->dtad_arg);
9627 		uint16_t low = DTRACE_LLQUANTIZE_LOW(desc->dtad_arg);
9628 		uint16_t high = DTRACE_LLQUANTIZE_HIGH(desc->dtad_arg);
9629 		uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(desc->dtad_arg);
9630 		int64_t v;
9631 
9632 		agg->dtag_initial = desc->dtad_arg;
9633 		agg->dtag_aggregate = dtrace_aggregate_llquantize;
9634 
9635 		if (factor < 2 || low >= high || nsteps < factor)
9636 			goto err;
9637 
9638 		/*
9639 		 * Now check that the number of steps evenly divides a power
9640 		 * of the factor.  (This assures both integer bucket size and
9641 		 * linearity within each magnitude.)
9642 		 */
9643 		for (v = factor; v < nsteps; v *= factor)
9644 			continue;
9645 
9646 		if ((v % nsteps) || (nsteps % factor))
9647 			goto err;
9648 
9649 		size = (dtrace_aggregate_llquantize_bucket(factor,
9650 		    low, high, nsteps, INT64_MAX) + 2) * sizeof (uint64_t);
9651 		break;
9652 	}
9653 
9654 	case DTRACEAGG_AVG:
9655 		agg->dtag_aggregate = dtrace_aggregate_avg;
9656 		size = sizeof (uint64_t) * 2;
9657 		break;
9658 
9659 	case DTRACEAGG_STDDEV:
9660 		agg->dtag_aggregate = dtrace_aggregate_stddev;
9661 		size = sizeof (uint64_t) * 4;
9662 		break;
9663 
9664 	case DTRACEAGG_SUM:
9665 		agg->dtag_aggregate = dtrace_aggregate_sum;
9666 		break;
9667 
9668 	default:
9669 		goto err;
9670 	}
9671 
9672 	agg->dtag_action.dta_rec.dtrd_size = size;
9673 
9674 	if (ntuple == 0)
9675 		goto err;
9676 
9677 	/*
9678 	 * We must make sure that we have enough actions for the n-tuple.
9679 	 */
9680 	for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
9681 		if (DTRACEACT_ISAGG(act->dta_kind))
9682 			break;
9683 
9684 		if (--ntuple == 0) {
9685 			/*
9686 			 * This is the action with which our n-tuple begins.
9687 			 */
9688 			agg->dtag_first = act;
9689 			goto success;
9690 		}
9691 	}
9692 
9693 	/*
9694 	 * This n-tuple is short by ntuple elements.  Return failure.
9695 	 */
9696 	ASSERT(ntuple != 0);
9697 err:
9698 	kmem_free(agg, sizeof (dtrace_aggregation_t));
9699 	return (NULL);
9700 
9701 success:
9702 	/*
9703 	 * If the last action in the tuple has a size of zero, it's actually
9704 	 * an expression argument for the aggregating action.
9705 	 */
9706 	ASSERT(ecb->dte_action_last != NULL);
9707 	act = ecb->dte_action_last;
9708 
9709 	if (act->dta_kind == DTRACEACT_DIFEXPR) {
9710 		ASSERT(act->dta_difo != NULL);
9711 
9712 		if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
9713 			agg->dtag_hasarg = 1;
9714 	}
9715 
9716 	/*
9717 	 * We need to allocate an id for this aggregation.
9718 	 */
9719 	aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
9720 	    VM_BESTFIT | VM_SLEEP);
9721 
9722 	if (aggid - 1 >= state->dts_naggregations) {
9723 		dtrace_aggregation_t **oaggs = state->dts_aggregations;
9724 		dtrace_aggregation_t **aggs;
9725 		int naggs = state->dts_naggregations << 1;
9726 		int onaggs = state->dts_naggregations;
9727 
9728 		ASSERT(aggid == state->dts_naggregations + 1);
9729 
9730 		if (naggs == 0) {
9731 			ASSERT(oaggs == NULL);
9732 			naggs = 1;
9733 		}
9734 
9735 		aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
9736 
9737 		if (oaggs != NULL) {
9738 			bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
9739 			kmem_free(oaggs, onaggs * sizeof (*aggs));
9740 		}
9741 
9742 		state->dts_aggregations = aggs;
9743 		state->dts_naggregations = naggs;
9744 	}
9745 
9746 	ASSERT(state->dts_aggregations[aggid - 1] == NULL);
9747 	state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
9748 
9749 	frec = &agg->dtag_first->dta_rec;
9750 	if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
9751 		frec->dtrd_alignment = sizeof (dtrace_aggid_t);
9752 
9753 	for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
9754 		ASSERT(!act->dta_intuple);
9755 		act->dta_intuple = 1;
9756 	}
9757 
9758 	return (&agg->dtag_action);
9759 }
9760 
9761 static void
9762 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
9763 {
9764 	dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
9765 	dtrace_state_t *state = ecb->dte_state;
9766 	dtrace_aggid_t aggid = agg->dtag_id;
9767 
9768 	ASSERT(DTRACEACT_ISAGG(act->dta_kind));
9769 	vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
9770 
9771 	ASSERT(state->dts_aggregations[aggid - 1] == agg);
9772 	state->dts_aggregations[aggid - 1] = NULL;
9773 
9774 	kmem_free(agg, sizeof (dtrace_aggregation_t));
9775 }
9776 
9777 static int
9778 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
9779 {
9780 	dtrace_action_t *action, *last;
9781 	dtrace_difo_t *dp = desc->dtad_difo;
9782 	uint32_t size = 0, align = sizeof (uint8_t), mask;
9783 	uint16_t format = 0;
9784 	dtrace_recdesc_t *rec;
9785 	dtrace_state_t *state = ecb->dte_state;
9786 	dtrace_optval_t *opt = state->dts_options, nframes, strsize;
9787 	uint64_t arg = desc->dtad_arg;
9788 
9789 	ASSERT(MUTEX_HELD(&dtrace_lock));
9790 	ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
9791 
9792 	if (DTRACEACT_ISAGG(desc->dtad_kind)) {
9793 		/*
9794 		 * If this is an aggregating action, there must be neither
9795 		 * a speculate nor a commit on the action chain.
9796 		 */
9797 		dtrace_action_t *act;
9798 
9799 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
9800 			if (act->dta_kind == DTRACEACT_COMMIT)
9801 				return (EINVAL);
9802 
9803 			if (act->dta_kind == DTRACEACT_SPECULATE)
9804 				return (EINVAL);
9805 		}
9806 
9807 		action = dtrace_ecb_aggregation_create(ecb, desc);
9808 
9809 		if (action == NULL)
9810 			return (EINVAL);
9811 	} else {
9812 		if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
9813 		    (desc->dtad_kind == DTRACEACT_DIFEXPR &&
9814 		    dp != NULL && dp->dtdo_destructive)) {
9815 			state->dts_destructive = 1;
9816 		}
9817 
9818 		switch (desc->dtad_kind) {
9819 		case DTRACEACT_PRINTF:
9820 		case DTRACEACT_PRINTA:
9821 		case DTRACEACT_SYSTEM:
9822 		case DTRACEACT_FREOPEN:
9823 		case DTRACEACT_DIFEXPR:
9824 			/*
9825 			 * We know that our arg is a string -- turn it into a
9826 			 * format.
9827 			 */
9828 			if (arg == NULL) {
9829 				ASSERT(desc->dtad_kind == DTRACEACT_PRINTA ||
9830 				    desc->dtad_kind == DTRACEACT_DIFEXPR);
9831 				format = 0;
9832 			} else {
9833 				ASSERT(arg != NULL);
9834 				ASSERT(arg > KERNELBASE);
9835 				format = dtrace_format_add(state,
9836 				    (char *)(uintptr_t)arg);
9837 			}
9838 
9839 			/*FALLTHROUGH*/
9840 		case DTRACEACT_LIBACT:
9841 		case DTRACEACT_TRACEMEM:
9842 		case DTRACEACT_TRACEMEM_DYNSIZE:
9843 			if (dp == NULL)
9844 				return (EINVAL);
9845 
9846 			if ((size = dp->dtdo_rtype.dtdt_size) != 0)
9847 				break;
9848 
9849 			if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
9850 				if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
9851 					return (EINVAL);
9852 
9853 				size = opt[DTRACEOPT_STRSIZE];
9854 			}
9855 
9856 			break;
9857 
9858 		case DTRACEACT_STACK:
9859 			if ((nframes = arg) == 0) {
9860 				nframes = opt[DTRACEOPT_STACKFRAMES];
9861 				ASSERT(nframes > 0);
9862 				arg = nframes;
9863 			}
9864 
9865 			size = nframes * sizeof (pc_t);
9866 			break;
9867 
9868 		case DTRACEACT_JSTACK:
9869 			if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
9870 				strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
9871 
9872 			if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
9873 				nframes = opt[DTRACEOPT_JSTACKFRAMES];
9874 
9875 			arg = DTRACE_USTACK_ARG(nframes, strsize);
9876 
9877 			/*FALLTHROUGH*/
9878 		case DTRACEACT_USTACK:
9879 			if (desc->dtad_kind != DTRACEACT_JSTACK &&
9880 			    (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
9881 				strsize = DTRACE_USTACK_STRSIZE(arg);
9882 				nframes = opt[DTRACEOPT_USTACKFRAMES];
9883 				ASSERT(nframes > 0);
9884 				arg = DTRACE_USTACK_ARG(nframes, strsize);
9885 			}
9886 
9887 			/*
9888 			 * Save a slot for the pid.
9889 			 */
9890 			size = (nframes + 1) * sizeof (uint64_t);
9891 			size += DTRACE_USTACK_STRSIZE(arg);
9892 			size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
9893 
9894 			break;
9895 
9896 		case DTRACEACT_SYM:
9897 		case DTRACEACT_MOD:
9898 			if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
9899 			    sizeof (uint64_t)) ||
9900 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
9901 				return (EINVAL);
9902 			break;
9903 
9904 		case DTRACEACT_USYM:
9905 		case DTRACEACT_UMOD:
9906 		case DTRACEACT_UADDR:
9907 			if (dp == NULL ||
9908 			    (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
9909 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
9910 				return (EINVAL);
9911 
9912 			/*
9913 			 * We have a slot for the pid, plus a slot for the
9914 			 * argument.  To keep things simple (aligned with
9915 			 * bitness-neutral sizing), we store each as a 64-bit
9916 			 * quantity.
9917 			 */
9918 			size = 2 * sizeof (uint64_t);
9919 			break;
9920 
9921 		case DTRACEACT_STOP:
9922 		case DTRACEACT_BREAKPOINT:
9923 		case DTRACEACT_PANIC:
9924 			break;
9925 
9926 		case DTRACEACT_CHILL:
9927 		case DTRACEACT_DISCARD:
9928 		case DTRACEACT_RAISE:
9929 			if (dp == NULL)
9930 				return (EINVAL);
9931 			break;
9932 
9933 		case DTRACEACT_EXIT:
9934 			if (dp == NULL ||
9935 			    (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
9936 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
9937 				return (EINVAL);
9938 			break;
9939 
9940 		case DTRACEACT_SPECULATE:
9941 			if (ecb->dte_size > sizeof (dtrace_epid_t))
9942 				return (EINVAL);
9943 
9944 			if (dp == NULL)
9945 				return (EINVAL);
9946 
9947 			state->dts_speculates = 1;
9948 			break;
9949 
9950 		case DTRACEACT_COMMIT: {
9951 			dtrace_action_t *act = ecb->dte_action;
9952 
9953 			for (; act != NULL; act = act->dta_next) {
9954 				if (act->dta_kind == DTRACEACT_COMMIT)
9955 					return (EINVAL);
9956 			}
9957 
9958 			if (dp == NULL)
9959 				return (EINVAL);
9960 			break;
9961 		}
9962 
9963 		default:
9964 			return (EINVAL);
9965 		}
9966 
9967 		if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
9968 			/*
9969 			 * If this is a data-storing action or a speculate,
9970 			 * we must be sure that there isn't a commit on the
9971 			 * action chain.
9972 			 */
9973 			dtrace_action_t *act = ecb->dte_action;
9974 
9975 			for (; act != NULL; act = act->dta_next) {
9976 				if (act->dta_kind == DTRACEACT_COMMIT)
9977 					return (EINVAL);
9978 			}
9979 		}
9980 
9981 		action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
9982 		action->dta_rec.dtrd_size = size;
9983 	}
9984 
9985 	action->dta_refcnt = 1;
9986 	rec = &action->dta_rec;
9987 	size = rec->dtrd_size;
9988 
9989 	for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
9990 		if (!(size & mask)) {
9991 			align = mask + 1;
9992 			break;
9993 		}
9994 	}
9995 
9996 	action->dta_kind = desc->dtad_kind;
9997 
9998 	if ((action->dta_difo = dp) != NULL)
9999 		dtrace_difo_hold(dp);
10000 
10001 	rec->dtrd_action = action->dta_kind;
10002 	rec->dtrd_arg = arg;
10003 	rec->dtrd_uarg = desc->dtad_uarg;
10004 	rec->dtrd_alignment = (uint16_t)align;
10005 	rec->dtrd_format = format;
10006 
10007 	if ((last = ecb->dte_action_last) != NULL) {
10008 		ASSERT(ecb->dte_action != NULL);
10009 		action->dta_prev = last;
10010 		last->dta_next = action;
10011 	} else {
10012 		ASSERT(ecb->dte_action == NULL);
10013 		ecb->dte_action = action;
10014 	}
10015 
10016 	ecb->dte_action_last = action;
10017 
10018 	return (0);
10019 }
10020 
10021 static void
10022 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
10023 {
10024 	dtrace_action_t *act = ecb->dte_action, *next;
10025 	dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
10026 	dtrace_difo_t *dp;
10027 	uint16_t format;
10028 
10029 	if (act != NULL && act->dta_refcnt > 1) {
10030 		ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
10031 		act->dta_refcnt--;
10032 	} else {
10033 		for (; act != NULL; act = next) {
10034 			next = act->dta_next;
10035 			ASSERT(next != NULL || act == ecb->dte_action_last);
10036 			ASSERT(act->dta_refcnt == 1);
10037 
10038 			if ((format = act->dta_rec.dtrd_format) != 0)
10039 				dtrace_format_remove(ecb->dte_state, format);
10040 
10041 			if ((dp = act->dta_difo) != NULL)
10042 				dtrace_difo_release(dp, vstate);
10043 
10044 			if (DTRACEACT_ISAGG(act->dta_kind)) {
10045 				dtrace_ecb_aggregation_destroy(ecb, act);
10046 			} else {
10047 				kmem_free(act, sizeof (dtrace_action_t));
10048 			}
10049 		}
10050 	}
10051 
10052 	ecb->dte_action = NULL;
10053 	ecb->dte_action_last = NULL;
10054 	ecb->dte_size = sizeof (dtrace_epid_t);
10055 }
10056 
10057 static void
10058 dtrace_ecb_disable(dtrace_ecb_t *ecb)
10059 {
10060 	/*
10061 	 * We disable the ECB by removing it from its probe.
10062 	 */
10063 	dtrace_ecb_t *pecb, *prev = NULL;
10064 	dtrace_probe_t *probe = ecb->dte_probe;
10065 
10066 	ASSERT(MUTEX_HELD(&dtrace_lock));
10067 
10068 	if (probe == NULL) {
10069 		/*
10070 		 * This is the NULL probe; there is nothing to disable.
10071 		 */
10072 		return;
10073 	}
10074 
10075 	for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
10076 		if (pecb == ecb)
10077 			break;
10078 		prev = pecb;
10079 	}
10080 
10081 	ASSERT(pecb != NULL);
10082 
10083 	if (prev == NULL) {
10084 		probe->dtpr_ecb = ecb->dte_next;
10085 	} else {
10086 		prev->dte_next = ecb->dte_next;
10087 	}
10088 
10089 	if (ecb == probe->dtpr_ecb_last) {
10090 		ASSERT(ecb->dte_next == NULL);
10091 		probe->dtpr_ecb_last = prev;
10092 	}
10093 
10094 	/*
10095 	 * The ECB has been disconnected from the probe; now sync to assure
10096 	 * that all CPUs have seen the change before returning.
10097 	 */
10098 	dtrace_sync();
10099 
10100 	if (probe->dtpr_ecb == NULL) {
10101 		/*
10102 		 * That was the last ECB on the probe; clear the predicate
10103 		 * cache ID for the probe, disable it and sync one more time
10104 		 * to assure that we'll never hit it again.
10105 		 */
10106 		dtrace_provider_t *prov = probe->dtpr_provider;
10107 
10108 		ASSERT(ecb->dte_next == NULL);
10109 		ASSERT(probe->dtpr_ecb_last == NULL);
10110 		probe->dtpr_predcache = DTRACE_CACHEIDNONE;
10111 		prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
10112 		    probe->dtpr_id, probe->dtpr_arg);
10113 		dtrace_sync();
10114 	} else {
10115 		/*
10116 		 * There is at least one ECB remaining on the probe.  If there
10117 		 * is _exactly_ one, set the probe's predicate cache ID to be
10118 		 * the predicate cache ID of the remaining ECB.
10119 		 */
10120 		ASSERT(probe->dtpr_ecb_last != NULL);
10121 		ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
10122 
10123 		if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
10124 			dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
10125 
10126 			ASSERT(probe->dtpr_ecb->dte_next == NULL);
10127 
10128 			if (p != NULL)
10129 				probe->dtpr_predcache = p->dtp_cacheid;
10130 		}
10131 
10132 		ecb->dte_next = NULL;
10133 	}
10134 }
10135 
10136 static void
10137 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
10138 {
10139 	dtrace_state_t *state = ecb->dte_state;
10140 	dtrace_vstate_t *vstate = &state->dts_vstate;
10141 	dtrace_predicate_t *pred;
10142 	dtrace_epid_t epid = ecb->dte_epid;
10143 
10144 	ASSERT(MUTEX_HELD(&dtrace_lock));
10145 	ASSERT(ecb->dte_next == NULL);
10146 	ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
10147 
10148 	if ((pred = ecb->dte_predicate) != NULL)
10149 		dtrace_predicate_release(pred, vstate);
10150 
10151 	dtrace_ecb_action_remove(ecb);
10152 
10153 	ASSERT(state->dts_ecbs[epid - 1] == ecb);
10154 	state->dts_ecbs[epid - 1] = NULL;
10155 
10156 	kmem_free(ecb, sizeof (dtrace_ecb_t));
10157 }
10158 
10159 static dtrace_ecb_t *
10160 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
10161     dtrace_enabling_t *enab)
10162 {
10163 	dtrace_ecb_t *ecb;
10164 	dtrace_predicate_t *pred;
10165 	dtrace_actdesc_t *act;
10166 	dtrace_provider_t *prov;
10167 	dtrace_ecbdesc_t *desc = enab->dten_current;
10168 
10169 	ASSERT(MUTEX_HELD(&dtrace_lock));
10170 	ASSERT(state != NULL);
10171 
10172 	ecb = dtrace_ecb_add(state, probe);
10173 	ecb->dte_uarg = desc->dted_uarg;
10174 
10175 	if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
10176 		dtrace_predicate_hold(pred);
10177 		ecb->dte_predicate = pred;
10178 	}
10179 
10180 	if (probe != NULL) {
10181 		/*
10182 		 * If the provider shows more leg than the consumer is old
10183 		 * enough to see, we need to enable the appropriate implicit
10184 		 * predicate bits to prevent the ecb from activating at
10185 		 * revealing times.
10186 		 *
10187 		 * Providers specifying DTRACE_PRIV_USER at register time
10188 		 * are stating that they need the /proc-style privilege
10189 		 * model to be enforced, and this is what DTRACE_COND_OWNER
10190 		 * and DTRACE_COND_ZONEOWNER will then do at probe time.
10191 		 */
10192 		prov = probe->dtpr_provider;
10193 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
10194 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10195 			ecb->dte_cond |= DTRACE_COND_OWNER;
10196 
10197 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
10198 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10199 			ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
10200 
10201 		/*
10202 		 * If the provider shows us kernel innards and the user
10203 		 * is lacking sufficient privilege, enable the
10204 		 * DTRACE_COND_USERMODE implicit predicate.
10205 		 */
10206 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
10207 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
10208 			ecb->dte_cond |= DTRACE_COND_USERMODE;
10209 	}
10210 
10211 	if (dtrace_ecb_create_cache != NULL) {
10212 		/*
10213 		 * If we have a cached ecb, we'll use its action list instead
10214 		 * of creating our own (saving both time and space).
10215 		 */
10216 		dtrace_ecb_t *cached = dtrace_ecb_create_cache;
10217 		dtrace_action_t *act = cached->dte_action;
10218 
10219 		if (act != NULL) {
10220 			ASSERT(act->dta_refcnt > 0);
10221 			act->dta_refcnt++;
10222 			ecb->dte_action = act;
10223 			ecb->dte_action_last = cached->dte_action_last;
10224 			ecb->dte_needed = cached->dte_needed;
10225 			ecb->dte_size = cached->dte_size;
10226 			ecb->dte_alignment = cached->dte_alignment;
10227 		}
10228 
10229 		return (ecb);
10230 	}
10231 
10232 	for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
10233 		if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
10234 			dtrace_ecb_destroy(ecb);
10235 			return (NULL);
10236 		}
10237 	}
10238 
10239 	dtrace_ecb_resize(ecb);
10240 
10241 	return (dtrace_ecb_create_cache = ecb);
10242 }
10243 
10244 static int
10245 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
10246 {
10247 	dtrace_ecb_t *ecb;
10248 	dtrace_enabling_t *enab = arg;
10249 	dtrace_state_t *state = enab->dten_vstate->dtvs_state;
10250 
10251 	ASSERT(state != NULL);
10252 
10253 	if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
10254 		/*
10255 		 * This probe was created in a generation for which this
10256 		 * enabling has previously created ECBs; we don't want to
10257 		 * enable it again, so just kick out.
10258 		 */
10259 		return (DTRACE_MATCH_NEXT);
10260 	}
10261 
10262 	if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
10263 		return (DTRACE_MATCH_DONE);
10264 
10265 	if (dtrace_ecb_enable(ecb) < 0)
10266 		return (DTRACE_MATCH_FAIL);
10267 
10268 	return (DTRACE_MATCH_NEXT);
10269 }
10270 
10271 static dtrace_ecb_t *
10272 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
10273 {
10274 	dtrace_ecb_t *ecb;
10275 
10276 	ASSERT(MUTEX_HELD(&dtrace_lock));
10277 
10278 	if (id == 0 || id > state->dts_necbs)
10279 		return (NULL);
10280 
10281 	ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
10282 	ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
10283 
10284 	return (state->dts_ecbs[id - 1]);
10285 }
10286 
10287 static dtrace_aggregation_t *
10288 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
10289 {
10290 	dtrace_aggregation_t *agg;
10291 
10292 	ASSERT(MUTEX_HELD(&dtrace_lock));
10293 
10294 	if (id == 0 || id > state->dts_naggregations)
10295 		return (NULL);
10296 
10297 	ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
10298 	ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
10299 	    agg->dtag_id == id);
10300 
10301 	return (state->dts_aggregations[id - 1]);
10302 }
10303 
10304 /*
10305  * DTrace Buffer Functions
10306  *
10307  * The following functions manipulate DTrace buffers.  Most of these functions
10308  * are called in the context of establishing or processing consumer state;
10309  * exceptions are explicitly noted.
10310  */
10311 
10312 /*
10313  * Note:  called from cross call context.  This function switches the two
10314  * buffers on a given CPU.  The atomicity of this operation is assured by
10315  * disabling interrupts while the actual switch takes place; the disabling of
10316  * interrupts serializes the execution with any execution of dtrace_probe() on
10317  * the same CPU.
10318  */
10319 static void
10320 dtrace_buffer_switch(dtrace_buffer_t *buf)
10321 {
10322 	caddr_t tomax = buf->dtb_tomax;
10323 	caddr_t xamot = buf->dtb_xamot;
10324 	dtrace_icookie_t cookie;
10325 	hrtime_t now = dtrace_gethrtime();
10326 
10327 	ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
10328 	ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
10329 
10330 	cookie = dtrace_interrupt_disable();
10331 	buf->dtb_tomax = xamot;
10332 	buf->dtb_xamot = tomax;
10333 	buf->dtb_xamot_drops = buf->dtb_drops;
10334 	buf->dtb_xamot_offset = buf->dtb_offset;
10335 	buf->dtb_xamot_errors = buf->dtb_errors;
10336 	buf->dtb_xamot_flags = buf->dtb_flags;
10337 	buf->dtb_offset = 0;
10338 	buf->dtb_drops = 0;
10339 	buf->dtb_errors = 0;
10340 	buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
10341 	buf->dtb_interval = now - buf->dtb_switched;
10342 	buf->dtb_switched = now;
10343 	dtrace_interrupt_enable(cookie);
10344 }
10345 
10346 /*
10347  * Note:  called from cross call context.  This function activates a buffer
10348  * on a CPU.  As with dtrace_buffer_switch(), the atomicity of the operation
10349  * is guaranteed by the disabling of interrupts.
10350  */
10351 static void
10352 dtrace_buffer_activate(dtrace_state_t *state)
10353 {
10354 	dtrace_buffer_t *buf;
10355 	dtrace_icookie_t cookie = dtrace_interrupt_disable();
10356 
10357 	buf = &state->dts_buffer[CPU->cpu_id];
10358 
10359 	if (buf->dtb_tomax != NULL) {
10360 		/*
10361 		 * We might like to assert that the buffer is marked inactive,
10362 		 * but this isn't necessarily true:  the buffer for the CPU
10363 		 * that processes the BEGIN probe has its buffer activated
10364 		 * manually.  In this case, we take the (harmless) action
10365 		 * re-clearing the bit INACTIVE bit.
10366 		 */
10367 		buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
10368 	}
10369 
10370 	dtrace_interrupt_enable(cookie);
10371 }
10372 
10373 static int
10374 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
10375     processorid_t cpu, int *factor)
10376 {
10377 	cpu_t *cp;
10378 	dtrace_buffer_t *buf;
10379 	int allocated = 0, desired = 0;
10380 
10381 	ASSERT(MUTEX_HELD(&cpu_lock));
10382 	ASSERT(MUTEX_HELD(&dtrace_lock));
10383 
10384 	*factor = 1;
10385 
10386 	if (size > dtrace_nonroot_maxsize &&
10387 	    !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
10388 		return (EFBIG);
10389 
10390 	cp = cpu_list;
10391 
10392 	do {
10393 		if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10394 			continue;
10395 
10396 		buf = &bufs[cp->cpu_id];
10397 
10398 		/*
10399 		 * If there is already a buffer allocated for this CPU, it
10400 		 * is only possible that this is a DR event.  In this case,
10401 		 * the buffer size must match our specified size.
10402 		 */
10403 		if (buf->dtb_tomax != NULL) {
10404 			ASSERT(buf->dtb_size == size);
10405 			continue;
10406 		}
10407 
10408 		ASSERT(buf->dtb_xamot == NULL);
10409 
10410 		if ((buf->dtb_tomax = kmem_zalloc(size,
10411 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL)
10412 			goto err;
10413 
10414 		buf->dtb_size = size;
10415 		buf->dtb_flags = flags;
10416 		buf->dtb_offset = 0;
10417 		buf->dtb_drops = 0;
10418 
10419 		if (flags & DTRACEBUF_NOSWITCH)
10420 			continue;
10421 
10422 		if ((buf->dtb_xamot = kmem_zalloc(size,
10423 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL)
10424 			goto err;
10425 	} while ((cp = cp->cpu_next) != cpu_list);
10426 
10427 	return (0);
10428 
10429 err:
10430 	cp = cpu_list;
10431 
10432 	do {
10433 		if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10434 			continue;
10435 
10436 		buf = &bufs[cp->cpu_id];
10437 		desired += 2;
10438 
10439 		if (buf->dtb_xamot != NULL) {
10440 			ASSERT(buf->dtb_tomax != NULL);
10441 			ASSERT(buf->dtb_size == size);
10442 			kmem_free(buf->dtb_xamot, size);
10443 			allocated++;
10444 		}
10445 
10446 		if (buf->dtb_tomax != NULL) {
10447 			ASSERT(buf->dtb_size == size);
10448 			kmem_free(buf->dtb_tomax, size);
10449 			allocated++;
10450 		}
10451 
10452 		buf->dtb_tomax = NULL;
10453 		buf->dtb_xamot = NULL;
10454 		buf->dtb_size = 0;
10455 	} while ((cp = cp->cpu_next) != cpu_list);
10456 
10457 	*factor = desired / (allocated > 0 ? allocated : 1);
10458 
10459 	return (ENOMEM);
10460 }
10461 
10462 /*
10463  * Note:  called from probe context.  This function just increments the drop
10464  * count on a buffer.  It has been made a function to allow for the
10465  * possibility of understanding the source of mysterious drop counts.  (A
10466  * problem for which one may be particularly disappointed that DTrace cannot
10467  * be used to understand DTrace.)
10468  */
10469 static void
10470 dtrace_buffer_drop(dtrace_buffer_t *buf)
10471 {
10472 	buf->dtb_drops++;
10473 }
10474 
10475 /*
10476  * Note:  called from probe context.  This function is called to reserve space
10477  * in a buffer.  If mstate is non-NULL, sets the scratch base and size in the
10478  * mstate.  Returns the new offset in the buffer, or a negative value if an
10479  * error has occurred.
10480  */
10481 static intptr_t
10482 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
10483     dtrace_state_t *state, dtrace_mstate_t *mstate)
10484 {
10485 	intptr_t offs = buf->dtb_offset, soffs;
10486 	intptr_t woffs;
10487 	caddr_t tomax;
10488 	size_t total;
10489 
10490 	if (buf->dtb_flags & DTRACEBUF_INACTIVE)
10491 		return (-1);
10492 
10493 	if ((tomax = buf->dtb_tomax) == NULL) {
10494 		dtrace_buffer_drop(buf);
10495 		return (-1);
10496 	}
10497 
10498 	if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
10499 		while (offs & (align - 1)) {
10500 			/*
10501 			 * Assert that our alignment is off by a number which
10502 			 * is itself sizeof (uint32_t) aligned.
10503 			 */
10504 			ASSERT(!((align - (offs & (align - 1))) &
10505 			    (sizeof (uint32_t) - 1)));
10506 			DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
10507 			offs += sizeof (uint32_t);
10508 		}
10509 
10510 		if ((soffs = offs + needed) > buf->dtb_size) {
10511 			dtrace_buffer_drop(buf);
10512 			return (-1);
10513 		}
10514 
10515 		if (mstate == NULL)
10516 			return (offs);
10517 
10518 		mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
10519 		mstate->dtms_scratch_size = buf->dtb_size - soffs;
10520 		mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
10521 
10522 		return (offs);
10523 	}
10524 
10525 	if (buf->dtb_flags & DTRACEBUF_FILL) {
10526 		if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
10527 		    (buf->dtb_flags & DTRACEBUF_FULL))
10528 			return (-1);
10529 		goto out;
10530 	}
10531 
10532 	total = needed + (offs & (align - 1));
10533 
10534 	/*
10535 	 * For a ring buffer, life is quite a bit more complicated.  Before
10536 	 * we can store any padding, we need to adjust our wrapping offset.
10537 	 * (If we've never before wrapped or we're not about to, no adjustment
10538 	 * is required.)
10539 	 */
10540 	if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
10541 	    offs + total > buf->dtb_size) {
10542 		woffs = buf->dtb_xamot_offset;
10543 
10544 		if (offs + total > buf->dtb_size) {
10545 			/*
10546 			 * We can't fit in the end of the buffer.  First, a
10547 			 * sanity check that we can fit in the buffer at all.
10548 			 */
10549 			if (total > buf->dtb_size) {
10550 				dtrace_buffer_drop(buf);
10551 				return (-1);
10552 			}
10553 
10554 			/*
10555 			 * We're going to be storing at the top of the buffer,
10556 			 * so now we need to deal with the wrapped offset.  We
10557 			 * only reset our wrapped offset to 0 if it is
10558 			 * currently greater than the current offset.  If it
10559 			 * is less than the current offset, it is because a
10560 			 * previous allocation induced a wrap -- but the
10561 			 * allocation didn't subsequently take the space due
10562 			 * to an error or false predicate evaluation.  In this
10563 			 * case, we'll just leave the wrapped offset alone: if
10564 			 * the wrapped offset hasn't been advanced far enough
10565 			 * for this allocation, it will be adjusted in the
10566 			 * lower loop.
10567 			 */
10568 			if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
10569 				if (woffs >= offs)
10570 					woffs = 0;
10571 			} else {
10572 				woffs = 0;
10573 			}
10574 
10575 			/*
10576 			 * Now we know that we're going to be storing to the
10577 			 * top of the buffer and that there is room for us
10578 			 * there.  We need to clear the buffer from the current
10579 			 * offset to the end (there may be old gunk there).
10580 			 */
10581 			while (offs < buf->dtb_size)
10582 				tomax[offs++] = 0;
10583 
10584 			/*
10585 			 * We need to set our offset to zero.  And because we
10586 			 * are wrapping, we need to set the bit indicating as
10587 			 * much.  We can also adjust our needed space back
10588 			 * down to the space required by the ECB -- we know
10589 			 * that the top of the buffer is aligned.
10590 			 */
10591 			offs = 0;
10592 			total = needed;
10593 			buf->dtb_flags |= DTRACEBUF_WRAPPED;
10594 		} else {
10595 			/*
10596 			 * There is room for us in the buffer, so we simply
10597 			 * need to check the wrapped offset.
10598 			 */
10599 			if (woffs < offs) {
10600 				/*
10601 				 * The wrapped offset is less than the offset.
10602 				 * This can happen if we allocated buffer space
10603 				 * that induced a wrap, but then we didn't
10604 				 * subsequently take the space due to an error
10605 				 * or false predicate evaluation.  This is
10606 				 * okay; we know that _this_ allocation isn't
10607 				 * going to induce a wrap.  We still can't
10608 				 * reset the wrapped offset to be zero,
10609 				 * however: the space may have been trashed in
10610 				 * the previous failed probe attempt.  But at
10611 				 * least the wrapped offset doesn't need to
10612 				 * be adjusted at all...
10613 				 */
10614 				goto out;
10615 			}
10616 		}
10617 
10618 		while (offs + total > woffs) {
10619 			dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
10620 			size_t size;
10621 
10622 			if (epid == DTRACE_EPIDNONE) {
10623 				size = sizeof (uint32_t);
10624 			} else {
10625 				ASSERT(epid <= state->dts_necbs);
10626 				ASSERT(state->dts_ecbs[epid - 1] != NULL);
10627 
10628 				size = state->dts_ecbs[epid - 1]->dte_size;
10629 			}
10630 
10631 			ASSERT(woffs + size <= buf->dtb_size);
10632 			ASSERT(size != 0);
10633 
10634 			if (woffs + size == buf->dtb_size) {
10635 				/*
10636 				 * We've reached the end of the buffer; we want
10637 				 * to set the wrapped offset to 0 and break
10638 				 * out.  However, if the offs is 0, then we're
10639 				 * in a strange edge-condition:  the amount of
10640 				 * space that we want to reserve plus the size
10641 				 * of the record that we're overwriting is
10642 				 * greater than the size of the buffer.  This
10643 				 * is problematic because if we reserve the
10644 				 * space but subsequently don't consume it (due
10645 				 * to a failed predicate or error) the wrapped
10646 				 * offset will be 0 -- yet the EPID at offset 0
10647 				 * will not be committed.  This situation is
10648 				 * relatively easy to deal with:  if we're in
10649 				 * this case, the buffer is indistinguishable
10650 				 * from one that hasn't wrapped; we need only
10651 				 * finish the job by clearing the wrapped bit,
10652 				 * explicitly setting the offset to be 0, and
10653 				 * zero'ing out the old data in the buffer.
10654 				 */
10655 				if (offs == 0) {
10656 					buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
10657 					buf->dtb_offset = 0;
10658 					woffs = total;
10659 
10660 					while (woffs < buf->dtb_size)
10661 						tomax[woffs++] = 0;
10662 				}
10663 
10664 				woffs = 0;
10665 				break;
10666 			}
10667 
10668 			woffs += size;
10669 		}
10670 
10671 		/*
10672 		 * We have a wrapped offset.  It may be that the wrapped offset
10673 		 * has become zero -- that's okay.
10674 		 */
10675 		buf->dtb_xamot_offset = woffs;
10676 	}
10677 
10678 out:
10679 	/*
10680 	 * Now we can plow the buffer with any necessary padding.
10681 	 */
10682 	while (offs & (align - 1)) {
10683 		/*
10684 		 * Assert that our alignment is off by a number which
10685 		 * is itself sizeof (uint32_t) aligned.
10686 		 */
10687 		ASSERT(!((align - (offs & (align - 1))) &
10688 		    (sizeof (uint32_t) - 1)));
10689 		DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
10690 		offs += sizeof (uint32_t);
10691 	}
10692 
10693 	if (buf->dtb_flags & DTRACEBUF_FILL) {
10694 		if (offs + needed > buf->dtb_size - state->dts_reserve) {
10695 			buf->dtb_flags |= DTRACEBUF_FULL;
10696 			return (-1);
10697 		}
10698 	}
10699 
10700 	if (mstate == NULL)
10701 		return (offs);
10702 
10703 	/*
10704 	 * For ring buffers and fill buffers, the scratch space is always
10705 	 * the inactive buffer.
10706 	 */
10707 	mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
10708 	mstate->dtms_scratch_size = buf->dtb_size;
10709 	mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
10710 
10711 	return (offs);
10712 }
10713 
10714 static void
10715 dtrace_buffer_polish(dtrace_buffer_t *buf)
10716 {
10717 	ASSERT(buf->dtb_flags & DTRACEBUF_RING);
10718 	ASSERT(MUTEX_HELD(&dtrace_lock));
10719 
10720 	if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
10721 		return;
10722 
10723 	/*
10724 	 * We need to polish the ring buffer.  There are three cases:
10725 	 *
10726 	 * - The first (and presumably most common) is that there is no gap
10727 	 *   between the buffer offset and the wrapped offset.  In this case,
10728 	 *   there is nothing in the buffer that isn't valid data; we can
10729 	 *   mark the buffer as polished and return.
10730 	 *
10731 	 * - The second (less common than the first but still more common
10732 	 *   than the third) is that there is a gap between the buffer offset
10733 	 *   and the wrapped offset, and the wrapped offset is larger than the
10734 	 *   buffer offset.  This can happen because of an alignment issue, or
10735 	 *   can happen because of a call to dtrace_buffer_reserve() that
10736 	 *   didn't subsequently consume the buffer space.  In this case,
10737 	 *   we need to zero the data from the buffer offset to the wrapped
10738 	 *   offset.
10739 	 *
10740 	 * - The third (and least common) is that there is a gap between the
10741 	 *   buffer offset and the wrapped offset, but the wrapped offset is
10742 	 *   _less_ than the buffer offset.  This can only happen because a
10743 	 *   call to dtrace_buffer_reserve() induced a wrap, but the space
10744 	 *   was not subsequently consumed.  In this case, we need to zero the
10745 	 *   space from the offset to the end of the buffer _and_ from the
10746 	 *   top of the buffer to the wrapped offset.
10747 	 */
10748 	if (buf->dtb_offset < buf->dtb_xamot_offset) {
10749 		bzero(buf->dtb_tomax + buf->dtb_offset,
10750 		    buf->dtb_xamot_offset - buf->dtb_offset);
10751 	}
10752 
10753 	if (buf->dtb_offset > buf->dtb_xamot_offset) {
10754 		bzero(buf->dtb_tomax + buf->dtb_offset,
10755 		    buf->dtb_size - buf->dtb_offset);
10756 		bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
10757 	}
10758 }
10759 
10760 /*
10761  * This routine determines if data generated at the specified time has likely
10762  * been entirely consumed at user-level.  This routine is called to determine
10763  * if an ECB on a defunct probe (but for an active enabling) can be safely
10764  * disabled and destroyed.
10765  */
10766 static int
10767 dtrace_buffer_consumed(dtrace_buffer_t *bufs, hrtime_t when)
10768 {
10769 	int i;
10770 
10771 	for (i = 0; i < NCPU; i++) {
10772 		dtrace_buffer_t *buf = &bufs[i];
10773 
10774 		if (buf->dtb_size == 0)
10775 			continue;
10776 
10777 		if (buf->dtb_flags & DTRACEBUF_RING)
10778 			return (0);
10779 
10780 		if (!buf->dtb_switched && buf->dtb_offset != 0)
10781 			return (0);
10782 
10783 		if (buf->dtb_switched - buf->dtb_interval < when)
10784 			return (0);
10785 	}
10786 
10787 	return (1);
10788 }
10789 
10790 static void
10791 dtrace_buffer_free(dtrace_buffer_t *bufs)
10792 {
10793 	int i;
10794 
10795 	for (i = 0; i < NCPU; i++) {
10796 		dtrace_buffer_t *buf = &bufs[i];
10797 
10798 		if (buf->dtb_tomax == NULL) {
10799 			ASSERT(buf->dtb_xamot == NULL);
10800 			ASSERT(buf->dtb_size == 0);
10801 			continue;
10802 		}
10803 
10804 		if (buf->dtb_xamot != NULL) {
10805 			ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
10806 			kmem_free(buf->dtb_xamot, buf->dtb_size);
10807 		}
10808 
10809 		kmem_free(buf->dtb_tomax, buf->dtb_size);
10810 		buf->dtb_size = 0;
10811 		buf->dtb_tomax = NULL;
10812 		buf->dtb_xamot = NULL;
10813 	}
10814 }
10815 
10816 /*
10817  * DTrace Enabling Functions
10818  */
10819 static dtrace_enabling_t *
10820 dtrace_enabling_create(dtrace_vstate_t *vstate)
10821 {
10822 	dtrace_enabling_t *enab;
10823 
10824 	enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
10825 	enab->dten_vstate = vstate;
10826 
10827 	return (enab);
10828 }
10829 
10830 static void
10831 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
10832 {
10833 	dtrace_ecbdesc_t **ndesc;
10834 	size_t osize, nsize;
10835 
10836 	/*
10837 	 * We can't add to enablings after we've enabled them, or after we've
10838 	 * retained them.
10839 	 */
10840 	ASSERT(enab->dten_probegen == 0);
10841 	ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
10842 
10843 	if (enab->dten_ndesc < enab->dten_maxdesc) {
10844 		enab->dten_desc[enab->dten_ndesc++] = ecb;
10845 		return;
10846 	}
10847 
10848 	osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
10849 
10850 	if (enab->dten_maxdesc == 0) {
10851 		enab->dten_maxdesc = 1;
10852 	} else {
10853 		enab->dten_maxdesc <<= 1;
10854 	}
10855 
10856 	ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
10857 
10858 	nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
10859 	ndesc = kmem_zalloc(nsize, KM_SLEEP);
10860 	bcopy(enab->dten_desc, ndesc, osize);
10861 	kmem_free(enab->dten_desc, osize);
10862 
10863 	enab->dten_desc = ndesc;
10864 	enab->dten_desc[enab->dten_ndesc++] = ecb;
10865 }
10866 
10867 static void
10868 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
10869     dtrace_probedesc_t *pd)
10870 {
10871 	dtrace_ecbdesc_t *new;
10872 	dtrace_predicate_t *pred;
10873 	dtrace_actdesc_t *act;
10874 
10875 	/*
10876 	 * We're going to create a new ECB description that matches the
10877 	 * specified ECB in every way, but has the specified probe description.
10878 	 */
10879 	new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
10880 
10881 	if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
10882 		dtrace_predicate_hold(pred);
10883 
10884 	for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
10885 		dtrace_actdesc_hold(act);
10886 
10887 	new->dted_action = ecb->dted_action;
10888 	new->dted_pred = ecb->dted_pred;
10889 	new->dted_probe = *pd;
10890 	new->dted_uarg = ecb->dted_uarg;
10891 
10892 	dtrace_enabling_add(enab, new);
10893 }
10894 
10895 static void
10896 dtrace_enabling_dump(dtrace_enabling_t *enab)
10897 {
10898 	int i;
10899 
10900 	for (i = 0; i < enab->dten_ndesc; i++) {
10901 		dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
10902 
10903 		cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
10904 		    desc->dtpd_provider, desc->dtpd_mod,
10905 		    desc->dtpd_func, desc->dtpd_name);
10906 	}
10907 }
10908 
10909 static void
10910 dtrace_enabling_destroy(dtrace_enabling_t *enab)
10911 {
10912 	int i;
10913 	dtrace_ecbdesc_t *ep;
10914 	dtrace_vstate_t *vstate = enab->dten_vstate;
10915 
10916 	ASSERT(MUTEX_HELD(&dtrace_lock));
10917 
10918 	for (i = 0; i < enab->dten_ndesc; i++) {
10919 		dtrace_actdesc_t *act, *next;
10920 		dtrace_predicate_t *pred;
10921 
10922 		ep = enab->dten_desc[i];
10923 
10924 		if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
10925 			dtrace_predicate_release(pred, vstate);
10926 
10927 		for (act = ep->dted_action; act != NULL; act = next) {
10928 			next = act->dtad_next;
10929 			dtrace_actdesc_release(act, vstate);
10930 		}
10931 
10932 		kmem_free(ep, sizeof (dtrace_ecbdesc_t));
10933 	}
10934 
10935 	kmem_free(enab->dten_desc,
10936 	    enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
10937 
10938 	/*
10939 	 * If this was a retained enabling, decrement the dts_nretained count
10940 	 * and take it off of the dtrace_retained list.
10941 	 */
10942 	if (enab->dten_prev != NULL || enab->dten_next != NULL ||
10943 	    dtrace_retained == enab) {
10944 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
10945 		ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
10946 		enab->dten_vstate->dtvs_state->dts_nretained--;
10947 		dtrace_retained_gen++;
10948 	}
10949 
10950 	if (enab->dten_prev == NULL) {
10951 		if (dtrace_retained == enab) {
10952 			dtrace_retained = enab->dten_next;
10953 
10954 			if (dtrace_retained != NULL)
10955 				dtrace_retained->dten_prev = NULL;
10956 		}
10957 	} else {
10958 		ASSERT(enab != dtrace_retained);
10959 		ASSERT(dtrace_retained != NULL);
10960 		enab->dten_prev->dten_next = enab->dten_next;
10961 	}
10962 
10963 	if (enab->dten_next != NULL) {
10964 		ASSERT(dtrace_retained != NULL);
10965 		enab->dten_next->dten_prev = enab->dten_prev;
10966 	}
10967 
10968 	kmem_free(enab, sizeof (dtrace_enabling_t));
10969 }
10970 
10971 static int
10972 dtrace_enabling_retain(dtrace_enabling_t *enab)
10973 {
10974 	dtrace_state_t *state;
10975 
10976 	ASSERT(MUTEX_HELD(&dtrace_lock));
10977 	ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
10978 	ASSERT(enab->dten_vstate != NULL);
10979 
10980 	state = enab->dten_vstate->dtvs_state;
10981 	ASSERT(state != NULL);
10982 
10983 	/*
10984 	 * We only allow each state to retain dtrace_retain_max enablings.
10985 	 */
10986 	if (state->dts_nretained >= dtrace_retain_max)
10987 		return (ENOSPC);
10988 
10989 	state->dts_nretained++;
10990 	dtrace_retained_gen++;
10991 
10992 	if (dtrace_retained == NULL) {
10993 		dtrace_retained = enab;
10994 		return (0);
10995 	}
10996 
10997 	enab->dten_next = dtrace_retained;
10998 	dtrace_retained->dten_prev = enab;
10999 	dtrace_retained = enab;
11000 
11001 	return (0);
11002 }
11003 
11004 static int
11005 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
11006     dtrace_probedesc_t *create)
11007 {
11008 	dtrace_enabling_t *new, *enab;
11009 	int found = 0, err = ENOENT;
11010 
11011 	ASSERT(MUTEX_HELD(&dtrace_lock));
11012 	ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
11013 	ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
11014 	ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
11015 	ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
11016 
11017 	new = dtrace_enabling_create(&state->dts_vstate);
11018 
11019 	/*
11020 	 * Iterate over all retained enablings, looking for enablings that
11021 	 * match the specified state.
11022 	 */
11023 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11024 		int i;
11025 
11026 		/*
11027 		 * dtvs_state can only be NULL for helper enablings -- and
11028 		 * helper enablings can't be retained.
11029 		 */
11030 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
11031 
11032 		if (enab->dten_vstate->dtvs_state != state)
11033 			continue;
11034 
11035 		/*
11036 		 * Now iterate over each probe description; we're looking for
11037 		 * an exact match to the specified probe description.
11038 		 */
11039 		for (i = 0; i < enab->dten_ndesc; i++) {
11040 			dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11041 			dtrace_probedesc_t *pd = &ep->dted_probe;
11042 
11043 			if (strcmp(pd->dtpd_provider, match->dtpd_provider))
11044 				continue;
11045 
11046 			if (strcmp(pd->dtpd_mod, match->dtpd_mod))
11047 				continue;
11048 
11049 			if (strcmp(pd->dtpd_func, match->dtpd_func))
11050 				continue;
11051 
11052 			if (strcmp(pd->dtpd_name, match->dtpd_name))
11053 				continue;
11054 
11055 			/*
11056 			 * We have a winning probe!  Add it to our growing
11057 			 * enabling.
11058 			 */
11059 			found = 1;
11060 			dtrace_enabling_addlike(new, ep, create);
11061 		}
11062 	}
11063 
11064 	if (!found || (err = dtrace_enabling_retain(new)) != 0) {
11065 		dtrace_enabling_destroy(new);
11066 		return (err);
11067 	}
11068 
11069 	return (0);
11070 }
11071 
11072 static void
11073 dtrace_enabling_retract(dtrace_state_t *state)
11074 {
11075 	dtrace_enabling_t *enab, *next;
11076 
11077 	ASSERT(MUTEX_HELD(&dtrace_lock));
11078 
11079 	/*
11080 	 * Iterate over all retained enablings, destroy the enablings retained
11081 	 * for the specified state.
11082 	 */
11083 	for (enab = dtrace_retained; enab != NULL; enab = next) {
11084 		next = enab->dten_next;
11085 
11086 		/*
11087 		 * dtvs_state can only be NULL for helper enablings -- and
11088 		 * helper enablings can't be retained.
11089 		 */
11090 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
11091 
11092 		if (enab->dten_vstate->dtvs_state == state) {
11093 			ASSERT(state->dts_nretained > 0);
11094 			dtrace_enabling_destroy(enab);
11095 		}
11096 	}
11097 
11098 	ASSERT(state->dts_nretained == 0);
11099 }
11100 
11101 static int
11102 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
11103 {
11104 	int i = 0;
11105 	int total_matched = 0, matched = 0;
11106 
11107 	ASSERT(MUTEX_HELD(&cpu_lock));
11108 	ASSERT(MUTEX_HELD(&dtrace_lock));
11109 
11110 	for (i = 0; i < enab->dten_ndesc; i++) {
11111 		dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11112 
11113 		enab->dten_current = ep;
11114 		enab->dten_error = 0;
11115 
11116 		/*
11117 		 * If a provider failed to enable a probe then get out and
11118 		 * let the consumer know we failed.
11119 		 */
11120 		if ((matched = dtrace_probe_enable(&ep->dted_probe, enab)) < 0)
11121 			return (EBUSY);
11122 
11123 		total_matched += matched;
11124 
11125 		if (enab->dten_error != 0) {
11126 			/*
11127 			 * If we get an error half-way through enabling the
11128 			 * probes, we kick out -- perhaps with some number of
11129 			 * them enabled.  Leaving enabled probes enabled may
11130 			 * be slightly confusing for user-level, but we expect
11131 			 * that no one will attempt to actually drive on in
11132 			 * the face of such errors.  If this is an anonymous
11133 			 * enabling (indicated with a NULL nmatched pointer),
11134 			 * we cmn_err() a message.  We aren't expecting to
11135 			 * get such an error -- such as it can exist at all,
11136 			 * it would be a result of corrupted DOF in the driver
11137 			 * properties.
11138 			 */
11139 			if (nmatched == NULL) {
11140 				cmn_err(CE_WARN, "dtrace_enabling_match() "
11141 				    "error on %p: %d", (void *)ep,
11142 				    enab->dten_error);
11143 			}
11144 
11145 			return (enab->dten_error);
11146 		}
11147 	}
11148 
11149 	enab->dten_probegen = dtrace_probegen;
11150 	if (nmatched != NULL)
11151 		*nmatched = total_matched;
11152 
11153 	return (0);
11154 }
11155 
11156 static void
11157 dtrace_enabling_matchall(void)
11158 {
11159 	dtrace_enabling_t *enab;
11160 
11161 	mutex_enter(&cpu_lock);
11162 	mutex_enter(&dtrace_lock);
11163 
11164 	/*
11165 	 * Iterate over all retained enablings to see if any probes match
11166 	 * against them.  We only perform this operation on enablings for which
11167 	 * we have sufficient permissions by virtue of being in the global zone
11168 	 * or in the same zone as the DTrace client.  Because we can be called
11169 	 * after dtrace_detach() has been called, we cannot assert that there
11170 	 * are retained enablings.  We can safely load from dtrace_retained,
11171 	 * however:  the taskq_destroy() at the end of dtrace_detach() will
11172 	 * block pending our completion.
11173 	 */
11174 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11175 		dtrace_cred_t *dcr = &enab->dten_vstate->dtvs_state->dts_cred;
11176 		cred_t *cr = dcr->dcr_cred;
11177 		zoneid_t zone = cr != NULL ? crgetzoneid(cr) : 0;
11178 
11179 		if ((dcr->dcr_visible & DTRACE_CRV_ALLZONE) || (cr != NULL &&
11180 		    (zone == GLOBAL_ZONEID || getzoneid() == zone)))
11181 			(void) dtrace_enabling_match(enab, NULL);
11182 	}
11183 
11184 	mutex_exit(&dtrace_lock);
11185 	mutex_exit(&cpu_lock);
11186 }
11187 
11188 /*
11189  * If an enabling is to be enabled without having matched probes (that is, if
11190  * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
11191  * enabling must be _primed_ by creating an ECB for every ECB description.
11192  * This must be done to assure that we know the number of speculations, the
11193  * number of aggregations, the minimum buffer size needed, etc. before we
11194  * transition out of DTRACE_ACTIVITY_INACTIVE.  To do this without actually
11195  * enabling any probes, we create ECBs for every ECB decription, but with a
11196  * NULL probe -- which is exactly what this function does.
11197  */
11198 static void
11199 dtrace_enabling_prime(dtrace_state_t *state)
11200 {
11201 	dtrace_enabling_t *enab;
11202 	int i;
11203 
11204 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11205 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
11206 
11207 		if (enab->dten_vstate->dtvs_state != state)
11208 			continue;
11209 
11210 		/*
11211 		 * We don't want to prime an enabling more than once, lest
11212 		 * we allow a malicious user to induce resource exhaustion.
11213 		 * (The ECBs that result from priming an enabling aren't
11214 		 * leaked -- but they also aren't deallocated until the
11215 		 * consumer state is destroyed.)
11216 		 */
11217 		if (enab->dten_primed)
11218 			continue;
11219 
11220 		for (i = 0; i < enab->dten_ndesc; i++) {
11221 			enab->dten_current = enab->dten_desc[i];
11222 			(void) dtrace_probe_enable(NULL, enab);
11223 		}
11224 
11225 		enab->dten_primed = 1;
11226 	}
11227 }
11228 
11229 /*
11230  * Called to indicate that probes should be provided due to retained
11231  * enablings.  This is implemented in terms of dtrace_probe_provide(), but it
11232  * must take an initial lap through the enabling calling the dtps_provide()
11233  * entry point explicitly to allow for autocreated probes.
11234  */
11235 static void
11236 dtrace_enabling_provide(dtrace_provider_t *prv)
11237 {
11238 	int i, all = 0;
11239 	dtrace_probedesc_t desc;
11240 	dtrace_genid_t gen;
11241 
11242 	ASSERT(MUTEX_HELD(&dtrace_lock));
11243 	ASSERT(MUTEX_HELD(&dtrace_provider_lock));
11244 
11245 	if (prv == NULL) {
11246 		all = 1;
11247 		prv = dtrace_provider;
11248 	}
11249 
11250 	do {
11251 		dtrace_enabling_t *enab;
11252 		void *parg = prv->dtpv_arg;
11253 
11254 retry:
11255 		gen = dtrace_retained_gen;
11256 		for (enab = dtrace_retained; enab != NULL;
11257 		    enab = enab->dten_next) {
11258 			for (i = 0; i < enab->dten_ndesc; i++) {
11259 				desc = enab->dten_desc[i]->dted_probe;
11260 				mutex_exit(&dtrace_lock);
11261 				prv->dtpv_pops.dtps_provide(parg, &desc);
11262 				mutex_enter(&dtrace_lock);
11263 				/*
11264 				 * Process the retained enablings again if
11265 				 * they have changed while we weren't holding
11266 				 * dtrace_lock.
11267 				 */
11268 				if (gen != dtrace_retained_gen)
11269 					goto retry;
11270 			}
11271 		}
11272 	} while (all && (prv = prv->dtpv_next) != NULL);
11273 
11274 	mutex_exit(&dtrace_lock);
11275 	dtrace_probe_provide(NULL, all ? NULL : prv);
11276 	mutex_enter(&dtrace_lock);
11277 }
11278 
11279 /*
11280  * Called to reap ECBs that are attached to probes from defunct providers.
11281  */
11282 static void
11283 dtrace_enabling_reap(void)
11284 {
11285 	dtrace_provider_t *prov;
11286 	dtrace_probe_t *probe;
11287 	dtrace_ecb_t *ecb;
11288 	hrtime_t when;
11289 	int i;
11290 
11291 	mutex_enter(&cpu_lock);
11292 	mutex_enter(&dtrace_lock);
11293 
11294 	for (i = 0; i < dtrace_nprobes; i++) {
11295 		if ((probe = dtrace_probes[i]) == NULL)
11296 			continue;
11297 
11298 		if (probe->dtpr_ecb == NULL)
11299 			continue;
11300 
11301 		prov = probe->dtpr_provider;
11302 
11303 		if ((when = prov->dtpv_defunct) == 0)
11304 			continue;
11305 
11306 		/*
11307 		 * We have ECBs on a defunct provider:  we want to reap these
11308 		 * ECBs to allow the provider to unregister.  The destruction
11309 		 * of these ECBs must be done carefully:  if we destroy the ECB
11310 		 * and the consumer later wishes to consume an EPID that
11311 		 * corresponds to the destroyed ECB (and if the EPID metadata
11312 		 * has not been previously consumed), the consumer will abort
11313 		 * processing on the unknown EPID.  To reduce (but not, sadly,
11314 		 * eliminate) the possibility of this, we will only destroy an
11315 		 * ECB for a defunct provider if, for the state that
11316 		 * corresponds to the ECB:
11317 		 *
11318 		 *  (a)	There is no speculative tracing (which can effectively
11319 		 *	cache an EPID for an arbitrary amount of time).
11320 		 *
11321 		 *  (b)	The principal buffers have been switched twice since the
11322 		 *	provider became defunct.
11323 		 *
11324 		 *  (c)	The aggregation buffers are of zero size or have been
11325 		 *	switched twice since the provider became defunct.
11326 		 *
11327 		 * We use dts_speculates to determine (a) and call a function
11328 		 * (dtrace_buffer_consumed()) to determine (b) and (c).  Note
11329 		 * that as soon as we've been unable to destroy one of the ECBs
11330 		 * associated with the probe, we quit trying -- reaping is only
11331 		 * fruitful in as much as we can destroy all ECBs associated
11332 		 * with the defunct provider's probes.
11333 		 */
11334 		while ((ecb = probe->dtpr_ecb) != NULL) {
11335 			dtrace_state_t *state = ecb->dte_state;
11336 			dtrace_buffer_t *buf = state->dts_buffer;
11337 			dtrace_buffer_t *aggbuf = state->dts_aggbuffer;
11338 
11339 			if (state->dts_speculates)
11340 				break;
11341 
11342 			if (!dtrace_buffer_consumed(buf, when))
11343 				break;
11344 
11345 			if (!dtrace_buffer_consumed(aggbuf, when))
11346 				break;
11347 
11348 			dtrace_ecb_disable(ecb);
11349 			ASSERT(probe->dtpr_ecb != ecb);
11350 			dtrace_ecb_destroy(ecb);
11351 		}
11352 	}
11353 
11354 	mutex_exit(&dtrace_lock);
11355 	mutex_exit(&cpu_lock);
11356 }
11357 
11358 /*
11359  * DTrace DOF Functions
11360  */
11361 /*ARGSUSED*/
11362 static void
11363 dtrace_dof_error(dof_hdr_t *dof, const char *str)
11364 {
11365 	if (dtrace_err_verbose)
11366 		cmn_err(CE_WARN, "failed to process DOF: %s", str);
11367 
11368 #ifdef DTRACE_ERRDEBUG
11369 	dtrace_errdebug(str);
11370 #endif
11371 }
11372 
11373 /*
11374  * Create DOF out of a currently enabled state.  Right now, we only create
11375  * DOF containing the run-time options -- but this could be expanded to create
11376  * complete DOF representing the enabled state.
11377  */
11378 static dof_hdr_t *
11379 dtrace_dof_create(dtrace_state_t *state)
11380 {
11381 	dof_hdr_t *dof;
11382 	dof_sec_t *sec;
11383 	dof_optdesc_t *opt;
11384 	int i, len = sizeof (dof_hdr_t) +
11385 	    roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
11386 	    sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11387 
11388 	ASSERT(MUTEX_HELD(&dtrace_lock));
11389 
11390 	dof = kmem_zalloc(len, KM_SLEEP);
11391 	dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
11392 	dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
11393 	dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
11394 	dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
11395 
11396 	dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
11397 	dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
11398 	dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
11399 	dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
11400 	dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
11401 	dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
11402 
11403 	dof->dofh_flags = 0;
11404 	dof->dofh_hdrsize = sizeof (dof_hdr_t);
11405 	dof->dofh_secsize = sizeof (dof_sec_t);
11406 	dof->dofh_secnum = 1;	/* only DOF_SECT_OPTDESC */
11407 	dof->dofh_secoff = sizeof (dof_hdr_t);
11408 	dof->dofh_loadsz = len;
11409 	dof->dofh_filesz = len;
11410 	dof->dofh_pad = 0;
11411 
11412 	/*
11413 	 * Fill in the option section header...
11414 	 */
11415 	sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
11416 	sec->dofs_type = DOF_SECT_OPTDESC;
11417 	sec->dofs_align = sizeof (uint64_t);
11418 	sec->dofs_flags = DOF_SECF_LOAD;
11419 	sec->dofs_entsize = sizeof (dof_optdesc_t);
11420 
11421 	opt = (dof_optdesc_t *)((uintptr_t)sec +
11422 	    roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
11423 
11424 	sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
11425 	sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11426 
11427 	for (i = 0; i < DTRACEOPT_MAX; i++) {
11428 		opt[i].dofo_option = i;
11429 		opt[i].dofo_strtab = DOF_SECIDX_NONE;
11430 		opt[i].dofo_value = state->dts_options[i];
11431 	}
11432 
11433 	return (dof);
11434 }
11435 
11436 static dof_hdr_t *
11437 dtrace_dof_copyin(uintptr_t uarg, int *errp)
11438 {
11439 	dof_hdr_t hdr, *dof;
11440 
11441 	ASSERT(!MUTEX_HELD(&dtrace_lock));
11442 
11443 	/*
11444 	 * First, we're going to copyin() the sizeof (dof_hdr_t).
11445 	 */
11446 	if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
11447 		dtrace_dof_error(NULL, "failed to copyin DOF header");
11448 		*errp = EFAULT;
11449 		return (NULL);
11450 	}
11451 
11452 	/*
11453 	 * Now we'll allocate the entire DOF and copy it in -- provided
11454 	 * that the length isn't outrageous.
11455 	 */
11456 	if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
11457 		dtrace_dof_error(&hdr, "load size exceeds maximum");
11458 		*errp = E2BIG;
11459 		return (NULL);
11460 	}
11461 
11462 	if (hdr.dofh_loadsz < sizeof (hdr)) {
11463 		dtrace_dof_error(&hdr, "invalid load size");
11464 		*errp = EINVAL;
11465 		return (NULL);
11466 	}
11467 
11468 	dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
11469 
11470 	if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0 ||
11471 	    dof->dofh_loadsz != hdr.dofh_loadsz) {
11472 		kmem_free(dof, hdr.dofh_loadsz);
11473 		*errp = EFAULT;
11474 		return (NULL);
11475 	}
11476 
11477 	return (dof);
11478 }
11479 
11480 static dof_hdr_t *
11481 dtrace_dof_property(const char *name)
11482 {
11483 	uchar_t *buf;
11484 	uint64_t loadsz;
11485 	unsigned int len, i;
11486 	dof_hdr_t *dof;
11487 
11488 	/*
11489 	 * Unfortunately, array of values in .conf files are always (and
11490 	 * only) interpreted to be integer arrays.  We must read our DOF
11491 	 * as an integer array, and then squeeze it into a byte array.
11492 	 */
11493 	if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
11494 	    (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
11495 		return (NULL);
11496 
11497 	for (i = 0; i < len; i++)
11498 		buf[i] = (uchar_t)(((int *)buf)[i]);
11499 
11500 	if (len < sizeof (dof_hdr_t)) {
11501 		ddi_prop_free(buf);
11502 		dtrace_dof_error(NULL, "truncated header");
11503 		return (NULL);
11504 	}
11505 
11506 	if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
11507 		ddi_prop_free(buf);
11508 		dtrace_dof_error(NULL, "truncated DOF");
11509 		return (NULL);
11510 	}
11511 
11512 	if (loadsz >= dtrace_dof_maxsize) {
11513 		ddi_prop_free(buf);
11514 		dtrace_dof_error(NULL, "oversized DOF");
11515 		return (NULL);
11516 	}
11517 
11518 	dof = kmem_alloc(loadsz, KM_SLEEP);
11519 	bcopy(buf, dof, loadsz);
11520 	ddi_prop_free(buf);
11521 
11522 	return (dof);
11523 }
11524 
11525 static void
11526 dtrace_dof_destroy(dof_hdr_t *dof)
11527 {
11528 	kmem_free(dof, dof->dofh_loadsz);
11529 }
11530 
11531 /*
11532  * Return the dof_sec_t pointer corresponding to a given section index.  If the
11533  * index is not valid, dtrace_dof_error() is called and NULL is returned.  If
11534  * a type other than DOF_SECT_NONE is specified, the header is checked against
11535  * this type and NULL is returned if the types do not match.
11536  */
11537 static dof_sec_t *
11538 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
11539 {
11540 	dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
11541 	    ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
11542 
11543 	if (i >= dof->dofh_secnum) {
11544 		dtrace_dof_error(dof, "referenced section index is invalid");
11545 		return (NULL);
11546 	}
11547 
11548 	if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
11549 		dtrace_dof_error(dof, "referenced section is not loadable");
11550 		return (NULL);
11551 	}
11552 
11553 	if (type != DOF_SECT_NONE && type != sec->dofs_type) {
11554 		dtrace_dof_error(dof, "referenced section is the wrong type");
11555 		return (NULL);
11556 	}
11557 
11558 	return (sec);
11559 }
11560 
11561 static dtrace_probedesc_t *
11562 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
11563 {
11564 	dof_probedesc_t *probe;
11565 	dof_sec_t *strtab;
11566 	uintptr_t daddr = (uintptr_t)dof;
11567 	uintptr_t str;
11568 	size_t size;
11569 
11570 	if (sec->dofs_type != DOF_SECT_PROBEDESC) {
11571 		dtrace_dof_error(dof, "invalid probe section");
11572 		return (NULL);
11573 	}
11574 
11575 	if (sec->dofs_align != sizeof (dof_secidx_t)) {
11576 		dtrace_dof_error(dof, "bad alignment in probe description");
11577 		return (NULL);
11578 	}
11579 
11580 	if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
11581 		dtrace_dof_error(dof, "truncated probe description");
11582 		return (NULL);
11583 	}
11584 
11585 	probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
11586 	strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
11587 
11588 	if (strtab == NULL)
11589 		return (NULL);
11590 
11591 	str = daddr + strtab->dofs_offset;
11592 	size = strtab->dofs_size;
11593 
11594 	if (probe->dofp_provider >= strtab->dofs_size) {
11595 		dtrace_dof_error(dof, "corrupt probe provider");
11596 		return (NULL);
11597 	}
11598 
11599 	(void) strncpy(desc->dtpd_provider,
11600 	    (char *)(str + probe->dofp_provider),
11601 	    MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
11602 
11603 	if (probe->dofp_mod >= strtab->dofs_size) {
11604 		dtrace_dof_error(dof, "corrupt probe module");
11605 		return (NULL);
11606 	}
11607 
11608 	(void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
11609 	    MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
11610 
11611 	if (probe->dofp_func >= strtab->dofs_size) {
11612 		dtrace_dof_error(dof, "corrupt probe function");
11613 		return (NULL);
11614 	}
11615 
11616 	(void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
11617 	    MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
11618 
11619 	if (probe->dofp_name >= strtab->dofs_size) {
11620 		dtrace_dof_error(dof, "corrupt probe name");
11621 		return (NULL);
11622 	}
11623 
11624 	(void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
11625 	    MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
11626 
11627 	return (desc);
11628 }
11629 
11630 static dtrace_difo_t *
11631 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
11632     cred_t *cr)
11633 {
11634 	dtrace_difo_t *dp;
11635 	size_t ttl = 0;
11636 	dof_difohdr_t *dofd;
11637 	uintptr_t daddr = (uintptr_t)dof;
11638 	size_t max = dtrace_difo_maxsize;
11639 	int i, l, n;
11640 
11641 	static const struct {
11642 		int section;
11643 		int bufoffs;
11644 		int lenoffs;
11645 		int entsize;
11646 		int align;
11647 		const char *msg;
11648 	} difo[] = {
11649 		{ DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
11650 		offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
11651 		sizeof (dif_instr_t), "multiple DIF sections" },
11652 
11653 		{ DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
11654 		offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
11655 		sizeof (uint64_t), "multiple integer tables" },
11656 
11657 		{ DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
11658 		offsetof(dtrace_difo_t, dtdo_strlen), 0,
11659 		sizeof (char), "multiple string tables" },
11660 
11661 		{ DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
11662 		offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
11663 		sizeof (uint_t), "multiple variable tables" },
11664 
11665 		{ DOF_SECT_NONE, 0, 0, 0, NULL }
11666 	};
11667 
11668 	if (sec->dofs_type != DOF_SECT_DIFOHDR) {
11669 		dtrace_dof_error(dof, "invalid DIFO header section");
11670 		return (NULL);
11671 	}
11672 
11673 	if (sec->dofs_align != sizeof (dof_secidx_t)) {
11674 		dtrace_dof_error(dof, "bad alignment in DIFO header");
11675 		return (NULL);
11676 	}
11677 
11678 	if (sec->dofs_size < sizeof (dof_difohdr_t) ||
11679 	    sec->dofs_size % sizeof (dof_secidx_t)) {
11680 		dtrace_dof_error(dof, "bad size in DIFO header");
11681 		return (NULL);
11682 	}
11683 
11684 	dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
11685 	n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
11686 
11687 	dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
11688 	dp->dtdo_rtype = dofd->dofd_rtype;
11689 
11690 	for (l = 0; l < n; l++) {
11691 		dof_sec_t *subsec;
11692 		void **bufp;
11693 		uint32_t *lenp;
11694 
11695 		if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
11696 		    dofd->dofd_links[l])) == NULL)
11697 			goto err; /* invalid section link */
11698 
11699 		if (ttl + subsec->dofs_size > max) {
11700 			dtrace_dof_error(dof, "exceeds maximum size");
11701 			goto err;
11702 		}
11703 
11704 		ttl += subsec->dofs_size;
11705 
11706 		for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
11707 			if (subsec->dofs_type != difo[i].section)
11708 				continue;
11709 
11710 			if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
11711 				dtrace_dof_error(dof, "section not loaded");
11712 				goto err;
11713 			}
11714 
11715 			if (subsec->dofs_align != difo[i].align) {
11716 				dtrace_dof_error(dof, "bad alignment");
11717 				goto err;
11718 			}
11719 
11720 			bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
11721 			lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
11722 
11723 			if (*bufp != NULL) {
11724 				dtrace_dof_error(dof, difo[i].msg);
11725 				goto err;
11726 			}
11727 
11728 			if (difo[i].entsize != subsec->dofs_entsize) {
11729 				dtrace_dof_error(dof, "entry size mismatch");
11730 				goto err;
11731 			}
11732 
11733 			if (subsec->dofs_entsize != 0 &&
11734 			    (subsec->dofs_size % subsec->dofs_entsize) != 0) {
11735 				dtrace_dof_error(dof, "corrupt entry size");
11736 				goto err;
11737 			}
11738 
11739 			*lenp = subsec->dofs_size;
11740 			*bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
11741 			bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
11742 			    *bufp, subsec->dofs_size);
11743 
11744 			if (subsec->dofs_entsize != 0)
11745 				*lenp /= subsec->dofs_entsize;
11746 
11747 			break;
11748 		}
11749 
11750 		/*
11751 		 * If we encounter a loadable DIFO sub-section that is not
11752 		 * known to us, assume this is a broken program and fail.
11753 		 */
11754 		if (difo[i].section == DOF_SECT_NONE &&
11755 		    (subsec->dofs_flags & DOF_SECF_LOAD)) {
11756 			dtrace_dof_error(dof, "unrecognized DIFO subsection");
11757 			goto err;
11758 		}
11759 	}
11760 
11761 	if (dp->dtdo_buf == NULL) {
11762 		/*
11763 		 * We can't have a DIF object without DIF text.
11764 		 */
11765 		dtrace_dof_error(dof, "missing DIF text");
11766 		goto err;
11767 	}
11768 
11769 	/*
11770 	 * Before we validate the DIF object, run through the variable table
11771 	 * looking for the strings -- if any of their size are under, we'll set
11772 	 * their size to be the system-wide default string size.  Note that
11773 	 * this should _not_ happen if the "strsize" option has been set --
11774 	 * in this case, the compiler should have set the size to reflect the
11775 	 * setting of the option.
11776 	 */
11777 	for (i = 0; i < dp->dtdo_varlen; i++) {
11778 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
11779 		dtrace_diftype_t *t = &v->dtdv_type;
11780 
11781 		if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
11782 			continue;
11783 
11784 		if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
11785 			t->dtdt_size = dtrace_strsize_default;
11786 	}
11787 
11788 	if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
11789 		goto err;
11790 
11791 	dtrace_difo_init(dp, vstate);
11792 	return (dp);
11793 
11794 err:
11795 	kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
11796 	kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
11797 	kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
11798 	kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
11799 
11800 	kmem_free(dp, sizeof (dtrace_difo_t));
11801 	return (NULL);
11802 }
11803 
11804 static dtrace_predicate_t *
11805 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
11806     cred_t *cr)
11807 {
11808 	dtrace_difo_t *dp;
11809 
11810 	if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
11811 		return (NULL);
11812 
11813 	return (dtrace_predicate_create(dp));
11814 }
11815 
11816 static dtrace_actdesc_t *
11817 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
11818     cred_t *cr)
11819 {
11820 	dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
11821 	dof_actdesc_t *desc;
11822 	dof_sec_t *difosec;
11823 	size_t offs;
11824 	uintptr_t daddr = (uintptr_t)dof;
11825 	uint64_t arg;
11826 	dtrace_actkind_t kind;
11827 
11828 	if (sec->dofs_type != DOF_SECT_ACTDESC) {
11829 		dtrace_dof_error(dof, "invalid action section");
11830 		return (NULL);
11831 	}
11832 
11833 	if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
11834 		dtrace_dof_error(dof, "truncated action description");
11835 		return (NULL);
11836 	}
11837 
11838 	if (sec->dofs_align != sizeof (uint64_t)) {
11839 		dtrace_dof_error(dof, "bad alignment in action description");
11840 		return (NULL);
11841 	}
11842 
11843 	if (sec->dofs_size < sec->dofs_entsize) {
11844 		dtrace_dof_error(dof, "section entry size exceeds total size");
11845 		return (NULL);
11846 	}
11847 
11848 	if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
11849 		dtrace_dof_error(dof, "bad entry size in action description");
11850 		return (NULL);
11851 	}
11852 
11853 	if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
11854 		dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
11855 		return (NULL);
11856 	}
11857 
11858 	for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
11859 		desc = (dof_actdesc_t *)(daddr +
11860 		    (uintptr_t)sec->dofs_offset + offs);
11861 		kind = (dtrace_actkind_t)desc->dofa_kind;
11862 
11863 		if ((DTRACEACT_ISPRINTFLIKE(kind) &&
11864 		    (kind != DTRACEACT_PRINTA ||
11865 		    desc->dofa_strtab != DOF_SECIDX_NONE)) ||
11866 		    (kind == DTRACEACT_DIFEXPR &&
11867 		    desc->dofa_strtab != DOF_SECIDX_NONE)) {
11868 			dof_sec_t *strtab;
11869 			char *str, *fmt;
11870 			uint64_t i;
11871 
11872 			/*
11873 			 * The argument to these actions is an index into the
11874 			 * DOF string table.  For printf()-like actions, this
11875 			 * is the format string.  For print(), this is the
11876 			 * CTF type of the expression result.
11877 			 */
11878 			if ((strtab = dtrace_dof_sect(dof,
11879 			    DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
11880 				goto err;
11881 
11882 			str = (char *)((uintptr_t)dof +
11883 			    (uintptr_t)strtab->dofs_offset);
11884 
11885 			for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
11886 				if (str[i] == '\0')
11887 					break;
11888 			}
11889 
11890 			if (i >= strtab->dofs_size) {
11891 				dtrace_dof_error(dof, "bogus format string");
11892 				goto err;
11893 			}
11894 
11895 			if (i == desc->dofa_arg) {
11896 				dtrace_dof_error(dof, "empty format string");
11897 				goto err;
11898 			}
11899 
11900 			i -= desc->dofa_arg;
11901 			fmt = kmem_alloc(i + 1, KM_SLEEP);
11902 			bcopy(&str[desc->dofa_arg], fmt, i + 1);
11903 			arg = (uint64_t)(uintptr_t)fmt;
11904 		} else {
11905 			if (kind == DTRACEACT_PRINTA) {
11906 				ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
11907 				arg = 0;
11908 			} else {
11909 				arg = desc->dofa_arg;
11910 			}
11911 		}
11912 
11913 		act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
11914 		    desc->dofa_uarg, arg);
11915 
11916 		if (last != NULL) {
11917 			last->dtad_next = act;
11918 		} else {
11919 			first = act;
11920 		}
11921 
11922 		last = act;
11923 
11924 		if (desc->dofa_difo == DOF_SECIDX_NONE)
11925 			continue;
11926 
11927 		if ((difosec = dtrace_dof_sect(dof,
11928 		    DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
11929 			goto err;
11930 
11931 		act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
11932 
11933 		if (act->dtad_difo == NULL)
11934 			goto err;
11935 	}
11936 
11937 	ASSERT(first != NULL);
11938 	return (first);
11939 
11940 err:
11941 	for (act = first; act != NULL; act = next) {
11942 		next = act->dtad_next;
11943 		dtrace_actdesc_release(act, vstate);
11944 	}
11945 
11946 	return (NULL);
11947 }
11948 
11949 static dtrace_ecbdesc_t *
11950 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
11951     cred_t *cr)
11952 {
11953 	dtrace_ecbdesc_t *ep;
11954 	dof_ecbdesc_t *ecb;
11955 	dtrace_probedesc_t *desc;
11956 	dtrace_predicate_t *pred = NULL;
11957 
11958 	if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
11959 		dtrace_dof_error(dof, "truncated ECB description");
11960 		return (NULL);
11961 	}
11962 
11963 	if (sec->dofs_align != sizeof (uint64_t)) {
11964 		dtrace_dof_error(dof, "bad alignment in ECB description");
11965 		return (NULL);
11966 	}
11967 
11968 	ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
11969 	sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
11970 
11971 	if (sec == NULL)
11972 		return (NULL);
11973 
11974 	ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
11975 	ep->dted_uarg = ecb->dofe_uarg;
11976 	desc = &ep->dted_probe;
11977 
11978 	if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
11979 		goto err;
11980 
11981 	if (ecb->dofe_pred != DOF_SECIDX_NONE) {
11982 		if ((sec = dtrace_dof_sect(dof,
11983 		    DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
11984 			goto err;
11985 
11986 		if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
11987 			goto err;
11988 
11989 		ep->dted_pred.dtpdd_predicate = pred;
11990 	}
11991 
11992 	if (ecb->dofe_actions != DOF_SECIDX_NONE) {
11993 		if ((sec = dtrace_dof_sect(dof,
11994 		    DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
11995 			goto err;
11996 
11997 		ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
11998 
11999 		if (ep->dted_action == NULL)
12000 			goto err;
12001 	}
12002 
12003 	return (ep);
12004 
12005 err:
12006 	if (pred != NULL)
12007 		dtrace_predicate_release(pred, vstate);
12008 	kmem_free(ep, sizeof (dtrace_ecbdesc_t));
12009 	return (NULL);
12010 }
12011 
12012 /*
12013  * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
12014  * specified DOF.  At present, this amounts to simply adding 'ubase' to the
12015  * site of any user SETX relocations to account for load object base address.
12016  * In the future, if we need other relocations, this function can be extended.
12017  */
12018 static int
12019 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase)
12020 {
12021 	uintptr_t daddr = (uintptr_t)dof;
12022 	dof_relohdr_t *dofr =
12023 	    (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12024 	dof_sec_t *ss, *rs, *ts;
12025 	dof_relodesc_t *r;
12026 	uint_t i, n;
12027 
12028 	if (sec->dofs_size < sizeof (dof_relohdr_t) ||
12029 	    sec->dofs_align != sizeof (dof_secidx_t)) {
12030 		dtrace_dof_error(dof, "invalid relocation header");
12031 		return (-1);
12032 	}
12033 
12034 	ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
12035 	rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
12036 	ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
12037 
12038 	if (ss == NULL || rs == NULL || ts == NULL)
12039 		return (-1); /* dtrace_dof_error() has been called already */
12040 
12041 	if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
12042 	    rs->dofs_align != sizeof (uint64_t)) {
12043 		dtrace_dof_error(dof, "invalid relocation section");
12044 		return (-1);
12045 	}
12046 
12047 	r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
12048 	n = rs->dofs_size / rs->dofs_entsize;
12049 
12050 	for (i = 0; i < n; i++) {
12051 		uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
12052 
12053 		switch (r->dofr_type) {
12054 		case DOF_RELO_NONE:
12055 			break;
12056 		case DOF_RELO_SETX:
12057 			if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
12058 			    sizeof (uint64_t) > ts->dofs_size) {
12059 				dtrace_dof_error(dof, "bad relocation offset");
12060 				return (-1);
12061 			}
12062 
12063 			if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
12064 				dtrace_dof_error(dof, "misaligned setx relo");
12065 				return (-1);
12066 			}
12067 
12068 			*(uint64_t *)taddr += ubase;
12069 			break;
12070 		default:
12071 			dtrace_dof_error(dof, "invalid relocation type");
12072 			return (-1);
12073 		}
12074 
12075 		r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
12076 	}
12077 
12078 	return (0);
12079 }
12080 
12081 /*
12082  * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
12083  * header:  it should be at the front of a memory region that is at least
12084  * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
12085  * size.  It need not be validated in any other way.
12086  */
12087 static int
12088 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
12089     dtrace_enabling_t **enabp, uint64_t ubase, int noprobes)
12090 {
12091 	uint64_t len = dof->dofh_loadsz, seclen;
12092 	uintptr_t daddr = (uintptr_t)dof;
12093 	dtrace_ecbdesc_t *ep;
12094 	dtrace_enabling_t *enab;
12095 	uint_t i;
12096 
12097 	ASSERT(MUTEX_HELD(&dtrace_lock));
12098 	ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
12099 
12100 	/*
12101 	 * Check the DOF header identification bytes.  In addition to checking
12102 	 * valid settings, we also verify that unused bits/bytes are zeroed so
12103 	 * we can use them later without fear of regressing existing binaries.
12104 	 */
12105 	if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
12106 	    DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
12107 		dtrace_dof_error(dof, "DOF magic string mismatch");
12108 		return (-1);
12109 	}
12110 
12111 	if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
12112 	    dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
12113 		dtrace_dof_error(dof, "DOF has invalid data model");
12114 		return (-1);
12115 	}
12116 
12117 	if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
12118 		dtrace_dof_error(dof, "DOF encoding mismatch");
12119 		return (-1);
12120 	}
12121 
12122 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
12123 	    dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
12124 		dtrace_dof_error(dof, "DOF version mismatch");
12125 		return (-1);
12126 	}
12127 
12128 	if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
12129 		dtrace_dof_error(dof, "DOF uses unsupported instruction set");
12130 		return (-1);
12131 	}
12132 
12133 	if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
12134 		dtrace_dof_error(dof, "DOF uses too many integer registers");
12135 		return (-1);
12136 	}
12137 
12138 	if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
12139 		dtrace_dof_error(dof, "DOF uses too many tuple registers");
12140 		return (-1);
12141 	}
12142 
12143 	for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
12144 		if (dof->dofh_ident[i] != 0) {
12145 			dtrace_dof_error(dof, "DOF has invalid ident byte set");
12146 			return (-1);
12147 		}
12148 	}
12149 
12150 	if (dof->dofh_flags & ~DOF_FL_VALID) {
12151 		dtrace_dof_error(dof, "DOF has invalid flag bits set");
12152 		return (-1);
12153 	}
12154 
12155 	if (dof->dofh_secsize == 0) {
12156 		dtrace_dof_error(dof, "zero section header size");
12157 		return (-1);
12158 	}
12159 
12160 	/*
12161 	 * Check that the section headers don't exceed the amount of DOF
12162 	 * data.  Note that we cast the section size and number of sections
12163 	 * to uint64_t's to prevent possible overflow in the multiplication.
12164 	 */
12165 	seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
12166 
12167 	if (dof->dofh_secoff > len || seclen > len ||
12168 	    dof->dofh_secoff + seclen > len) {
12169 		dtrace_dof_error(dof, "truncated section headers");
12170 		return (-1);
12171 	}
12172 
12173 	if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
12174 		dtrace_dof_error(dof, "misaligned section headers");
12175 		return (-1);
12176 	}
12177 
12178 	if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
12179 		dtrace_dof_error(dof, "misaligned section size");
12180 		return (-1);
12181 	}
12182 
12183 	/*
12184 	 * Take an initial pass through the section headers to be sure that
12185 	 * the headers don't have stray offsets.  If the 'noprobes' flag is
12186 	 * set, do not permit sections relating to providers, probes, or args.
12187 	 */
12188 	for (i = 0; i < dof->dofh_secnum; i++) {
12189 		dof_sec_t *sec = (dof_sec_t *)(daddr +
12190 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12191 
12192 		if (noprobes) {
12193 			switch (sec->dofs_type) {
12194 			case DOF_SECT_PROVIDER:
12195 			case DOF_SECT_PROBES:
12196 			case DOF_SECT_PRARGS:
12197 			case DOF_SECT_PROFFS:
12198 				dtrace_dof_error(dof, "illegal sections "
12199 				    "for enabling");
12200 				return (-1);
12201 			}
12202 		}
12203 
12204 		if (DOF_SEC_ISLOADABLE(sec->dofs_type) &&
12205 		    !(sec->dofs_flags & DOF_SECF_LOAD)) {
12206 			dtrace_dof_error(dof, "loadable section with load "
12207 			    "flag unset");
12208 			return (-1);
12209 		}
12210 
12211 		if (!(sec->dofs_flags & DOF_SECF_LOAD))
12212 			continue; /* just ignore non-loadable sections */
12213 
12214 		if (sec->dofs_align & (sec->dofs_align - 1)) {
12215 			dtrace_dof_error(dof, "bad section alignment");
12216 			return (-1);
12217 		}
12218 
12219 		if (sec->dofs_offset & (sec->dofs_align - 1)) {
12220 			dtrace_dof_error(dof, "misaligned section");
12221 			return (-1);
12222 		}
12223 
12224 		if (sec->dofs_offset > len || sec->dofs_size > len ||
12225 		    sec->dofs_offset + sec->dofs_size > len) {
12226 			dtrace_dof_error(dof, "corrupt section header");
12227 			return (-1);
12228 		}
12229 
12230 		if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
12231 		    sec->dofs_offset + sec->dofs_size - 1) != '\0') {
12232 			dtrace_dof_error(dof, "non-terminating string table");
12233 			return (-1);
12234 		}
12235 	}
12236 
12237 	/*
12238 	 * Take a second pass through the sections and locate and perform any
12239 	 * relocations that are present.  We do this after the first pass to
12240 	 * be sure that all sections have had their headers validated.
12241 	 */
12242 	for (i = 0; i < dof->dofh_secnum; i++) {
12243 		dof_sec_t *sec = (dof_sec_t *)(daddr +
12244 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12245 
12246 		if (!(sec->dofs_flags & DOF_SECF_LOAD))
12247 			continue; /* skip sections that are not loadable */
12248 
12249 		switch (sec->dofs_type) {
12250 		case DOF_SECT_URELHDR:
12251 			if (dtrace_dof_relocate(dof, sec, ubase) != 0)
12252 				return (-1);
12253 			break;
12254 		}
12255 	}
12256 
12257 	if ((enab = *enabp) == NULL)
12258 		enab = *enabp = dtrace_enabling_create(vstate);
12259 
12260 	for (i = 0; i < dof->dofh_secnum; i++) {
12261 		dof_sec_t *sec = (dof_sec_t *)(daddr +
12262 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12263 
12264 		if (sec->dofs_type != DOF_SECT_ECBDESC)
12265 			continue;
12266 
12267 		if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
12268 			dtrace_enabling_destroy(enab);
12269 			*enabp = NULL;
12270 			return (-1);
12271 		}
12272 
12273 		dtrace_enabling_add(enab, ep);
12274 	}
12275 
12276 	return (0);
12277 }
12278 
12279 /*
12280  * Process DOF for any options.  This routine assumes that the DOF has been
12281  * at least processed by dtrace_dof_slurp().
12282  */
12283 static int
12284 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
12285 {
12286 	int i, rval;
12287 	uint32_t entsize;
12288 	size_t offs;
12289 	dof_optdesc_t *desc;
12290 
12291 	for (i = 0; i < dof->dofh_secnum; i++) {
12292 		dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
12293 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12294 
12295 		if (sec->dofs_type != DOF_SECT_OPTDESC)
12296 			continue;
12297 
12298 		if (sec->dofs_align != sizeof (uint64_t)) {
12299 			dtrace_dof_error(dof, "bad alignment in "
12300 			    "option description");
12301 			return (EINVAL);
12302 		}
12303 
12304 		if ((entsize = sec->dofs_entsize) == 0) {
12305 			dtrace_dof_error(dof, "zeroed option entry size");
12306 			return (EINVAL);
12307 		}
12308 
12309 		if (entsize < sizeof (dof_optdesc_t)) {
12310 			dtrace_dof_error(dof, "bad option entry size");
12311 			return (EINVAL);
12312 		}
12313 
12314 		for (offs = 0; offs < sec->dofs_size; offs += entsize) {
12315 			desc = (dof_optdesc_t *)((uintptr_t)dof +
12316 			    (uintptr_t)sec->dofs_offset + offs);
12317 
12318 			if (desc->dofo_strtab != DOF_SECIDX_NONE) {
12319 				dtrace_dof_error(dof, "non-zero option string");
12320 				return (EINVAL);
12321 			}
12322 
12323 			if (desc->dofo_value == DTRACEOPT_UNSET) {
12324 				dtrace_dof_error(dof, "unset option");
12325 				return (EINVAL);
12326 			}
12327 
12328 			if ((rval = dtrace_state_option(state,
12329 			    desc->dofo_option, desc->dofo_value)) != 0) {
12330 				dtrace_dof_error(dof, "rejected option");
12331 				return (rval);
12332 			}
12333 		}
12334 	}
12335 
12336 	return (0);
12337 }
12338 
12339 /*
12340  * DTrace Consumer State Functions
12341  */
12342 int
12343 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
12344 {
12345 	size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
12346 	void *base;
12347 	uintptr_t limit;
12348 	dtrace_dynvar_t *dvar, *next, *start;
12349 	int i;
12350 
12351 	ASSERT(MUTEX_HELD(&dtrace_lock));
12352 	ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
12353 
12354 	bzero(dstate, sizeof (dtrace_dstate_t));
12355 
12356 	if ((dstate->dtds_chunksize = chunksize) == 0)
12357 		dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
12358 
12359 	if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
12360 		size = min;
12361 
12362 	if ((base = kmem_zalloc(size, KM_NOSLEEP | KM_NORMALPRI)) == NULL)
12363 		return (ENOMEM);
12364 
12365 	dstate->dtds_size = size;
12366 	dstate->dtds_base = base;
12367 	dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
12368 	bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
12369 
12370 	hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
12371 
12372 	if (hashsize != 1 && (hashsize & 1))
12373 		hashsize--;
12374 
12375 	dstate->dtds_hashsize = hashsize;
12376 	dstate->dtds_hash = dstate->dtds_base;
12377 
12378 	/*
12379 	 * Set all of our hash buckets to point to the single sink, and (if
12380 	 * it hasn't already been set), set the sink's hash value to be the
12381 	 * sink sentinel value.  The sink is needed for dynamic variable
12382 	 * lookups to know that they have iterated over an entire, valid hash
12383 	 * chain.
12384 	 */
12385 	for (i = 0; i < hashsize; i++)
12386 		dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
12387 
12388 	if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
12389 		dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
12390 
12391 	/*
12392 	 * Determine number of active CPUs.  Divide free list evenly among
12393 	 * active CPUs.
12394 	 */
12395 	start = (dtrace_dynvar_t *)
12396 	    ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
12397 	limit = (uintptr_t)base + size;
12398 
12399 	maxper = (limit - (uintptr_t)start) / NCPU;
12400 	maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
12401 
12402 	for (i = 0; i < NCPU; i++) {
12403 		dstate->dtds_percpu[i].dtdsc_free = dvar = start;
12404 
12405 		/*
12406 		 * If we don't even have enough chunks to make it once through
12407 		 * NCPUs, we're just going to allocate everything to the first
12408 		 * CPU.  And if we're on the last CPU, we're going to allocate
12409 		 * whatever is left over.  In either case, we set the limit to
12410 		 * be the limit of the dynamic variable space.
12411 		 */
12412 		if (maxper == 0 || i == NCPU - 1) {
12413 			limit = (uintptr_t)base + size;
12414 			start = NULL;
12415 		} else {
12416 			limit = (uintptr_t)start + maxper;
12417 			start = (dtrace_dynvar_t *)limit;
12418 		}
12419 
12420 		ASSERT(limit <= (uintptr_t)base + size);
12421 
12422 		for (;;) {
12423 			next = (dtrace_dynvar_t *)((uintptr_t)dvar +
12424 			    dstate->dtds_chunksize);
12425 
12426 			if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
12427 				break;
12428 
12429 			dvar->dtdv_next = next;
12430 			dvar = next;
12431 		}
12432 
12433 		if (maxper == 0)
12434 			break;
12435 	}
12436 
12437 	return (0);
12438 }
12439 
12440 void
12441 dtrace_dstate_fini(dtrace_dstate_t *dstate)
12442 {
12443 	ASSERT(MUTEX_HELD(&cpu_lock));
12444 
12445 	if (dstate->dtds_base == NULL)
12446 		return;
12447 
12448 	kmem_free(dstate->dtds_base, dstate->dtds_size);
12449 	kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
12450 }
12451 
12452 static void
12453 dtrace_vstate_fini(dtrace_vstate_t *vstate)
12454 {
12455 	/*
12456 	 * Logical XOR, where are you?
12457 	 */
12458 	ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
12459 
12460 	if (vstate->dtvs_nglobals > 0) {
12461 		kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
12462 		    sizeof (dtrace_statvar_t *));
12463 	}
12464 
12465 	if (vstate->dtvs_ntlocals > 0) {
12466 		kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
12467 		    sizeof (dtrace_difv_t));
12468 	}
12469 
12470 	ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
12471 
12472 	if (vstate->dtvs_nlocals > 0) {
12473 		kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
12474 		    sizeof (dtrace_statvar_t *));
12475 	}
12476 }
12477 
12478 static void
12479 dtrace_state_clean(dtrace_state_t *state)
12480 {
12481 	if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
12482 		return;
12483 
12484 	dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
12485 	dtrace_speculation_clean(state);
12486 }
12487 
12488 static void
12489 dtrace_state_deadman(dtrace_state_t *state)
12490 {
12491 	hrtime_t now;
12492 
12493 	dtrace_sync();
12494 
12495 	now = dtrace_gethrtime();
12496 
12497 	if (state != dtrace_anon.dta_state &&
12498 	    now - state->dts_laststatus >= dtrace_deadman_user)
12499 		return;
12500 
12501 	/*
12502 	 * We must be sure that dts_alive never appears to be less than the
12503 	 * value upon entry to dtrace_state_deadman(), and because we lack a
12504 	 * dtrace_cas64(), we cannot store to it atomically.  We thus instead
12505 	 * store INT64_MAX to it, followed by a memory barrier, followed by
12506 	 * the new value.  This assures that dts_alive never appears to be
12507 	 * less than its true value, regardless of the order in which the
12508 	 * stores to the underlying storage are issued.
12509 	 */
12510 	state->dts_alive = INT64_MAX;
12511 	dtrace_membar_producer();
12512 	state->dts_alive = now;
12513 }
12514 
12515 dtrace_state_t *
12516 dtrace_state_create(dev_t *devp, cred_t *cr)
12517 {
12518 	minor_t minor;
12519 	major_t major;
12520 	char c[30];
12521 	dtrace_state_t *state;
12522 	dtrace_optval_t *opt;
12523 	int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
12524 
12525 	ASSERT(MUTEX_HELD(&dtrace_lock));
12526 	ASSERT(MUTEX_HELD(&cpu_lock));
12527 
12528 	minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
12529 	    VM_BESTFIT | VM_SLEEP);
12530 
12531 	if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
12532 		vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
12533 		return (NULL);
12534 	}
12535 
12536 	state = ddi_get_soft_state(dtrace_softstate, minor);
12537 	state->dts_epid = DTRACE_EPIDNONE + 1;
12538 
12539 	(void) snprintf(c, sizeof (c), "dtrace_aggid_%d", minor);
12540 	state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
12541 	    NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
12542 
12543 	if (devp != NULL) {
12544 		major = getemajor(*devp);
12545 	} else {
12546 		major = ddi_driver_major(dtrace_devi);
12547 	}
12548 
12549 	state->dts_dev = makedevice(major, minor);
12550 
12551 	if (devp != NULL)
12552 		*devp = state->dts_dev;
12553 
12554 	/*
12555 	 * We allocate NCPU buffers.  On the one hand, this can be quite
12556 	 * a bit of memory per instance (nearly 36K on a Starcat).  On the
12557 	 * other hand, it saves an additional memory reference in the probe
12558 	 * path.
12559 	 */
12560 	state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
12561 	state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
12562 	state->dts_cleaner = CYCLIC_NONE;
12563 	state->dts_deadman = CYCLIC_NONE;
12564 	state->dts_vstate.dtvs_state = state;
12565 
12566 	for (i = 0; i < DTRACEOPT_MAX; i++)
12567 		state->dts_options[i] = DTRACEOPT_UNSET;
12568 
12569 	/*
12570 	 * Set the default options.
12571 	 */
12572 	opt = state->dts_options;
12573 	opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
12574 	opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
12575 	opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
12576 	opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
12577 	opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
12578 	opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
12579 	opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
12580 	opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
12581 	opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
12582 	opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
12583 	opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
12584 	opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
12585 	opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
12586 	opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
12587 
12588 	state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
12589 
12590 	/*
12591 	 * Depending on the user credentials, we set flag bits which alter probe
12592 	 * visibility or the amount of destructiveness allowed.  In the case of
12593 	 * actual anonymous tracing, or the possession of all privileges, all of
12594 	 * the normal checks are bypassed.
12595 	 */
12596 	if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
12597 		state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
12598 		state->dts_cred.dcr_action = DTRACE_CRA_ALL;
12599 	} else {
12600 		/*
12601 		 * Set up the credentials for this instantiation.  We take a
12602 		 * hold on the credential to prevent it from disappearing on
12603 		 * us; this in turn prevents the zone_t referenced by this
12604 		 * credential from disappearing.  This means that we can
12605 		 * examine the credential and the zone from probe context.
12606 		 */
12607 		crhold(cr);
12608 		state->dts_cred.dcr_cred = cr;
12609 
12610 		/*
12611 		 * CRA_PROC means "we have *some* privilege for dtrace" and
12612 		 * unlocks the use of variables like pid, zonename, etc.
12613 		 */
12614 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
12615 		    PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
12616 			state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
12617 		}
12618 
12619 		/*
12620 		 * dtrace_user allows use of syscall and profile providers.
12621 		 * If the user also has proc_owner and/or proc_zone, we
12622 		 * extend the scope to include additional visibility and
12623 		 * destructive power.
12624 		 */
12625 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
12626 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
12627 				state->dts_cred.dcr_visible |=
12628 				    DTRACE_CRV_ALLPROC;
12629 
12630 				state->dts_cred.dcr_action |=
12631 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
12632 			}
12633 
12634 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
12635 				state->dts_cred.dcr_visible |=
12636 				    DTRACE_CRV_ALLZONE;
12637 
12638 				state->dts_cred.dcr_action |=
12639 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
12640 			}
12641 
12642 			/*
12643 			 * If we have all privs in whatever zone this is,
12644 			 * we can do destructive things to processes which
12645 			 * have altered credentials.
12646 			 */
12647 			if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
12648 			    cr->cr_zone->zone_privset)) {
12649 				state->dts_cred.dcr_action |=
12650 				    DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
12651 			}
12652 		}
12653 
12654 		/*
12655 		 * Holding the dtrace_kernel privilege also implies that
12656 		 * the user has the dtrace_user privilege from a visibility
12657 		 * perspective.  But without further privileges, some
12658 		 * destructive actions are not available.
12659 		 */
12660 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
12661 			/*
12662 			 * Make all probes in all zones visible.  However,
12663 			 * this doesn't mean that all actions become available
12664 			 * to all zones.
12665 			 */
12666 			state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
12667 			    DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
12668 
12669 			state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
12670 			    DTRACE_CRA_PROC;
12671 			/*
12672 			 * Holding proc_owner means that destructive actions
12673 			 * for *this* zone are allowed.
12674 			 */
12675 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
12676 				state->dts_cred.dcr_action |=
12677 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
12678 
12679 			/*
12680 			 * Holding proc_zone means that destructive actions
12681 			 * for this user/group ID in all zones is allowed.
12682 			 */
12683 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
12684 				state->dts_cred.dcr_action |=
12685 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
12686 
12687 			/*
12688 			 * If we have all privs in whatever zone this is,
12689 			 * we can do destructive things to processes which
12690 			 * have altered credentials.
12691 			 */
12692 			if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
12693 			    cr->cr_zone->zone_privset)) {
12694 				state->dts_cred.dcr_action |=
12695 				    DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
12696 			}
12697 		}
12698 
12699 		/*
12700 		 * Holding the dtrace_proc privilege gives control over fasttrap
12701 		 * and pid providers.  We need to grant wider destructive
12702 		 * privileges in the event that the user has proc_owner and/or
12703 		 * proc_zone.
12704 		 */
12705 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
12706 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
12707 				state->dts_cred.dcr_action |=
12708 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
12709 
12710 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
12711 				state->dts_cred.dcr_action |=
12712 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
12713 		}
12714 	}
12715 
12716 	return (state);
12717 }
12718 
12719 static int
12720 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
12721 {
12722 	dtrace_optval_t *opt = state->dts_options, size;
12723 	processorid_t cpu;
12724 	int flags = 0, rval, factor, divisor = 1;
12725 
12726 	ASSERT(MUTEX_HELD(&dtrace_lock));
12727 	ASSERT(MUTEX_HELD(&cpu_lock));
12728 	ASSERT(which < DTRACEOPT_MAX);
12729 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
12730 	    (state == dtrace_anon.dta_state &&
12731 	    state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
12732 
12733 	if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
12734 		return (0);
12735 
12736 	if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
12737 		cpu = opt[DTRACEOPT_CPU];
12738 
12739 	if (which == DTRACEOPT_SPECSIZE)
12740 		flags |= DTRACEBUF_NOSWITCH;
12741 
12742 	if (which == DTRACEOPT_BUFSIZE) {
12743 		if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
12744 			flags |= DTRACEBUF_RING;
12745 
12746 		if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
12747 			flags |= DTRACEBUF_FILL;
12748 
12749 		if (state != dtrace_anon.dta_state ||
12750 		    state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
12751 			flags |= DTRACEBUF_INACTIVE;
12752 	}
12753 
12754 	for (size = opt[which]; size >= sizeof (uint64_t); size /= divisor) {
12755 		/*
12756 		 * The size must be 8-byte aligned.  If the size is not 8-byte
12757 		 * aligned, drop it down by the difference.
12758 		 */
12759 		if (size & (sizeof (uint64_t) - 1))
12760 			size -= size & (sizeof (uint64_t) - 1);
12761 
12762 		if (size < state->dts_reserve) {
12763 			/*
12764 			 * Buffers always must be large enough to accommodate
12765 			 * their prereserved space.  We return E2BIG instead
12766 			 * of ENOMEM in this case to allow for user-level
12767 			 * software to differentiate the cases.
12768 			 */
12769 			return (E2BIG);
12770 		}
12771 
12772 		rval = dtrace_buffer_alloc(buf, size, flags, cpu, &factor);
12773 
12774 		if (rval != ENOMEM) {
12775 			opt[which] = size;
12776 			return (rval);
12777 		}
12778 
12779 		if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
12780 			return (rval);
12781 
12782 		for (divisor = 2; divisor < factor; divisor <<= 1)
12783 			continue;
12784 	}
12785 
12786 	return (ENOMEM);
12787 }
12788 
12789 static int
12790 dtrace_state_buffers(dtrace_state_t *state)
12791 {
12792 	dtrace_speculation_t *spec = state->dts_speculations;
12793 	int rval, i;
12794 
12795 	if ((rval = dtrace_state_buffer(state, state->dts_buffer,
12796 	    DTRACEOPT_BUFSIZE)) != 0)
12797 		return (rval);
12798 
12799 	if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
12800 	    DTRACEOPT_AGGSIZE)) != 0)
12801 		return (rval);
12802 
12803 	for (i = 0; i < state->dts_nspeculations; i++) {
12804 		if ((rval = dtrace_state_buffer(state,
12805 		    spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
12806 			return (rval);
12807 	}
12808 
12809 	return (0);
12810 }
12811 
12812 static void
12813 dtrace_state_prereserve(dtrace_state_t *state)
12814 {
12815 	dtrace_ecb_t *ecb;
12816 	dtrace_probe_t *probe;
12817 
12818 	state->dts_reserve = 0;
12819 
12820 	if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
12821 		return;
12822 
12823 	/*
12824 	 * If our buffer policy is a "fill" buffer policy, we need to set the
12825 	 * prereserved space to be the space required by the END probes.
12826 	 */
12827 	probe = dtrace_probes[dtrace_probeid_end - 1];
12828 	ASSERT(probe != NULL);
12829 
12830 	for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
12831 		if (ecb->dte_state != state)
12832 			continue;
12833 
12834 		state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
12835 	}
12836 }
12837 
12838 static int
12839 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
12840 {
12841 	dtrace_optval_t *opt = state->dts_options, sz, nspec;
12842 	dtrace_speculation_t *spec;
12843 	dtrace_buffer_t *buf;
12844 	cyc_handler_t hdlr;
12845 	cyc_time_t when;
12846 	int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
12847 	dtrace_icookie_t cookie;
12848 
12849 	mutex_enter(&cpu_lock);
12850 	mutex_enter(&dtrace_lock);
12851 
12852 	if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
12853 		rval = EBUSY;
12854 		goto out;
12855 	}
12856 
12857 	/*
12858 	 * Before we can perform any checks, we must prime all of the
12859 	 * retained enablings that correspond to this state.
12860 	 */
12861 	dtrace_enabling_prime(state);
12862 
12863 	if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
12864 		rval = EACCES;
12865 		goto out;
12866 	}
12867 
12868 	dtrace_state_prereserve(state);
12869 
12870 	/*
12871 	 * Now we want to do is try to allocate our speculations.
12872 	 * We do not automatically resize the number of speculations; if
12873 	 * this fails, we will fail the operation.
12874 	 */
12875 	nspec = opt[DTRACEOPT_NSPEC];
12876 	ASSERT(nspec != DTRACEOPT_UNSET);
12877 
12878 	if (nspec > INT_MAX) {
12879 		rval = ENOMEM;
12880 		goto out;
12881 	}
12882 
12883 	spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t),
12884 	    KM_NOSLEEP | KM_NORMALPRI);
12885 
12886 	if (spec == NULL) {
12887 		rval = ENOMEM;
12888 		goto out;
12889 	}
12890 
12891 	state->dts_speculations = spec;
12892 	state->dts_nspeculations = (int)nspec;
12893 
12894 	for (i = 0; i < nspec; i++) {
12895 		if ((buf = kmem_zalloc(bufsize,
12896 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL) {
12897 			rval = ENOMEM;
12898 			goto err;
12899 		}
12900 
12901 		spec[i].dtsp_buffer = buf;
12902 	}
12903 
12904 	if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
12905 		if (dtrace_anon.dta_state == NULL) {
12906 			rval = ENOENT;
12907 			goto out;
12908 		}
12909 
12910 		if (state->dts_necbs != 0) {
12911 			rval = EALREADY;
12912 			goto out;
12913 		}
12914 
12915 		state->dts_anon = dtrace_anon_grab();
12916 		ASSERT(state->dts_anon != NULL);
12917 		state = state->dts_anon;
12918 
12919 		/*
12920 		 * We want "grabanon" to be set in the grabbed state, so we'll
12921 		 * copy that option value from the grabbing state into the
12922 		 * grabbed state.
12923 		 */
12924 		state->dts_options[DTRACEOPT_GRABANON] =
12925 		    opt[DTRACEOPT_GRABANON];
12926 
12927 		*cpu = dtrace_anon.dta_beganon;
12928 
12929 		/*
12930 		 * If the anonymous state is active (as it almost certainly
12931 		 * is if the anonymous enabling ultimately matched anything),
12932 		 * we don't allow any further option processing -- but we
12933 		 * don't return failure.
12934 		 */
12935 		if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
12936 			goto out;
12937 	}
12938 
12939 	if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
12940 	    opt[DTRACEOPT_AGGSIZE] != 0) {
12941 		if (state->dts_aggregations == NULL) {
12942 			/*
12943 			 * We're not going to create an aggregation buffer
12944 			 * because we don't have any ECBs that contain
12945 			 * aggregations -- set this option to 0.
12946 			 */
12947 			opt[DTRACEOPT_AGGSIZE] = 0;
12948 		} else {
12949 			/*
12950 			 * If we have an aggregation buffer, we must also have
12951 			 * a buffer to use as scratch.
12952 			 */
12953 			if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
12954 			    opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
12955 				opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
12956 			}
12957 		}
12958 	}
12959 
12960 	if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
12961 	    opt[DTRACEOPT_SPECSIZE] != 0) {
12962 		if (!state->dts_speculates) {
12963 			/*
12964 			 * We're not going to create speculation buffers
12965 			 * because we don't have any ECBs that actually
12966 			 * speculate -- set the speculation size to 0.
12967 			 */
12968 			opt[DTRACEOPT_SPECSIZE] = 0;
12969 		}
12970 	}
12971 
12972 	/*
12973 	 * The bare minimum size for any buffer that we're actually going to
12974 	 * do anything to is sizeof (uint64_t).
12975 	 */
12976 	sz = sizeof (uint64_t);
12977 
12978 	if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
12979 	    (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
12980 	    (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
12981 		/*
12982 		 * A buffer size has been explicitly set to 0 (or to a size
12983 		 * that will be adjusted to 0) and we need the space -- we
12984 		 * need to return failure.  We return ENOSPC to differentiate
12985 		 * it from failing to allocate a buffer due to failure to meet
12986 		 * the reserve (for which we return E2BIG).
12987 		 */
12988 		rval = ENOSPC;
12989 		goto out;
12990 	}
12991 
12992 	if ((rval = dtrace_state_buffers(state)) != 0)
12993 		goto err;
12994 
12995 	if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
12996 		sz = dtrace_dstate_defsize;
12997 
12998 	do {
12999 		rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
13000 
13001 		if (rval == 0)
13002 			break;
13003 
13004 		if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13005 			goto err;
13006 	} while (sz >>= 1);
13007 
13008 	opt[DTRACEOPT_DYNVARSIZE] = sz;
13009 
13010 	if (rval != 0)
13011 		goto err;
13012 
13013 	if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
13014 		opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
13015 
13016 	if (opt[DTRACEOPT_CLEANRATE] == 0)
13017 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13018 
13019 	if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
13020 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
13021 
13022 	if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
13023 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13024 
13025 	hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
13026 	hdlr.cyh_arg = state;
13027 	hdlr.cyh_level = CY_LOW_LEVEL;
13028 
13029 	when.cyt_when = 0;
13030 	when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
13031 
13032 	state->dts_cleaner = cyclic_add(&hdlr, &when);
13033 
13034 	hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
13035 	hdlr.cyh_arg = state;
13036 	hdlr.cyh_level = CY_LOW_LEVEL;
13037 
13038 	when.cyt_when = 0;
13039 	when.cyt_interval = dtrace_deadman_interval;
13040 
13041 	state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
13042 	state->dts_deadman = cyclic_add(&hdlr, &when);
13043 
13044 	state->dts_activity = DTRACE_ACTIVITY_WARMUP;
13045 
13046 	/*
13047 	 * Now it's time to actually fire the BEGIN probe.  We need to disable
13048 	 * interrupts here both to record the CPU on which we fired the BEGIN
13049 	 * probe (the data from this CPU will be processed first at user
13050 	 * level) and to manually activate the buffer for this CPU.
13051 	 */
13052 	cookie = dtrace_interrupt_disable();
13053 	*cpu = CPU->cpu_id;
13054 	ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
13055 	state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
13056 
13057 	dtrace_probe(dtrace_probeid_begin,
13058 	    (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13059 	dtrace_interrupt_enable(cookie);
13060 	/*
13061 	 * We may have had an exit action from a BEGIN probe; only change our
13062 	 * state to ACTIVE if we're still in WARMUP.
13063 	 */
13064 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
13065 	    state->dts_activity == DTRACE_ACTIVITY_DRAINING);
13066 
13067 	if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
13068 		state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
13069 
13070 	/*
13071 	 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
13072 	 * want each CPU to transition its principal buffer out of the
13073 	 * INACTIVE state.  Doing this assures that no CPU will suddenly begin
13074 	 * processing an ECB halfway down a probe's ECB chain; all CPUs will
13075 	 * atomically transition from processing none of a state's ECBs to
13076 	 * processing all of them.
13077 	 */
13078 	dtrace_xcall(DTRACE_CPUALL,
13079 	    (dtrace_xcall_t)dtrace_buffer_activate, state);
13080 	goto out;
13081 
13082 err:
13083 	dtrace_buffer_free(state->dts_buffer);
13084 	dtrace_buffer_free(state->dts_aggbuffer);
13085 
13086 	if ((nspec = state->dts_nspeculations) == 0) {
13087 		ASSERT(state->dts_speculations == NULL);
13088 		goto out;
13089 	}
13090 
13091 	spec = state->dts_speculations;
13092 	ASSERT(spec != NULL);
13093 
13094 	for (i = 0; i < state->dts_nspeculations; i++) {
13095 		if ((buf = spec[i].dtsp_buffer) == NULL)
13096 			break;
13097 
13098 		dtrace_buffer_free(buf);
13099 		kmem_free(buf, bufsize);
13100 	}
13101 
13102 	kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13103 	state->dts_nspeculations = 0;
13104 	state->dts_speculations = NULL;
13105 
13106 out:
13107 	mutex_exit(&dtrace_lock);
13108 	mutex_exit(&cpu_lock);
13109 
13110 	return (rval);
13111 }
13112 
13113 static int
13114 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
13115 {
13116 	dtrace_icookie_t cookie;
13117 
13118 	ASSERT(MUTEX_HELD(&dtrace_lock));
13119 
13120 	if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
13121 	    state->dts_activity != DTRACE_ACTIVITY_DRAINING)
13122 		return (EINVAL);
13123 
13124 	/*
13125 	 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
13126 	 * to be sure that every CPU has seen it.  See below for the details
13127 	 * on why this is done.
13128 	 */
13129 	state->dts_activity = DTRACE_ACTIVITY_DRAINING;
13130 	dtrace_sync();
13131 
13132 	/*
13133 	 * By this point, it is impossible for any CPU to be still processing
13134 	 * with DTRACE_ACTIVITY_ACTIVE.  We can thus set our activity to
13135 	 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
13136 	 * other CPU in dtrace_buffer_reserve().  This allows dtrace_probe()
13137 	 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
13138 	 * iff we're in the END probe.
13139 	 */
13140 	state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
13141 	dtrace_sync();
13142 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
13143 
13144 	/*
13145 	 * Finally, we can release the reserve and call the END probe.  We
13146 	 * disable interrupts across calling the END probe to allow us to
13147 	 * return the CPU on which we actually called the END probe.  This
13148 	 * allows user-land to be sure that this CPU's principal buffer is
13149 	 * processed last.
13150 	 */
13151 	state->dts_reserve = 0;
13152 
13153 	cookie = dtrace_interrupt_disable();
13154 	*cpu = CPU->cpu_id;
13155 	dtrace_probe(dtrace_probeid_end,
13156 	    (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13157 	dtrace_interrupt_enable(cookie);
13158 
13159 	state->dts_activity = DTRACE_ACTIVITY_STOPPED;
13160 	dtrace_sync();
13161 
13162 	return (0);
13163 }
13164 
13165 static int
13166 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
13167     dtrace_optval_t val)
13168 {
13169 	ASSERT(MUTEX_HELD(&dtrace_lock));
13170 
13171 	if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13172 		return (EBUSY);
13173 
13174 	if (option >= DTRACEOPT_MAX)
13175 		return (EINVAL);
13176 
13177 	if (option != DTRACEOPT_CPU && val < 0)
13178 		return (EINVAL);
13179 
13180 	switch (option) {
13181 	case DTRACEOPT_DESTRUCTIVE:
13182 		if (dtrace_destructive_disallow)
13183 			return (EACCES);
13184 
13185 		state->dts_cred.dcr_destructive = 1;
13186 		break;
13187 
13188 	case DTRACEOPT_BUFSIZE:
13189 	case DTRACEOPT_DYNVARSIZE:
13190 	case DTRACEOPT_AGGSIZE:
13191 	case DTRACEOPT_SPECSIZE:
13192 	case DTRACEOPT_STRSIZE:
13193 		if (val < 0)
13194 			return (EINVAL);
13195 
13196 		if (val >= LONG_MAX) {
13197 			/*
13198 			 * If this is an otherwise negative value, set it to
13199 			 * the highest multiple of 128m less than LONG_MAX.
13200 			 * Technically, we're adjusting the size without
13201 			 * regard to the buffer resizing policy, but in fact,
13202 			 * this has no effect -- if we set the buffer size to
13203 			 * ~LONG_MAX and the buffer policy is ultimately set to
13204 			 * be "manual", the buffer allocation is guaranteed to
13205 			 * fail, if only because the allocation requires two
13206 			 * buffers.  (We set the the size to the highest
13207 			 * multiple of 128m because it ensures that the size
13208 			 * will remain a multiple of a megabyte when
13209 			 * repeatedly halved -- all the way down to 15m.)
13210 			 */
13211 			val = LONG_MAX - (1 << 27) + 1;
13212 		}
13213 	}
13214 
13215 	state->dts_options[option] = val;
13216 
13217 	return (0);
13218 }
13219 
13220 static void
13221 dtrace_state_destroy(dtrace_state_t *state)
13222 {
13223 	dtrace_ecb_t *ecb;
13224 	dtrace_vstate_t *vstate = &state->dts_vstate;
13225 	minor_t minor = getminor(state->dts_dev);
13226 	int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13227 	dtrace_speculation_t *spec = state->dts_speculations;
13228 	int nspec = state->dts_nspeculations;
13229 	uint32_t match;
13230 
13231 	ASSERT(MUTEX_HELD(&dtrace_lock));
13232 	ASSERT(MUTEX_HELD(&cpu_lock));
13233 
13234 	/*
13235 	 * First, retract any retained enablings for this state.
13236 	 */
13237 	dtrace_enabling_retract(state);
13238 	ASSERT(state->dts_nretained == 0);
13239 
13240 	if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
13241 	    state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
13242 		/*
13243 		 * We have managed to come into dtrace_state_destroy() on a
13244 		 * hot enabling -- almost certainly because of a disorderly
13245 		 * shutdown of a consumer.  (That is, a consumer that is
13246 		 * exiting without having called dtrace_stop().) In this case,
13247 		 * we're going to set our activity to be KILLED, and then
13248 		 * issue a sync to be sure that everyone is out of probe
13249 		 * context before we start blowing away ECBs.
13250 		 */
13251 		state->dts_activity = DTRACE_ACTIVITY_KILLED;
13252 		dtrace_sync();
13253 	}
13254 
13255 	/*
13256 	 * Release the credential hold we took in dtrace_state_create().
13257 	 */
13258 	if (state->dts_cred.dcr_cred != NULL)
13259 		crfree(state->dts_cred.dcr_cred);
13260 
13261 	/*
13262 	 * Now we can safely disable and destroy any enabled probes.  Because
13263 	 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
13264 	 * (especially if they're all enabled), we take two passes through the
13265 	 * ECBs:  in the first, we disable just DTRACE_PRIV_KERNEL probes, and
13266 	 * in the second we disable whatever is left over.
13267 	 */
13268 	for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
13269 		for (i = 0; i < state->dts_necbs; i++) {
13270 			if ((ecb = state->dts_ecbs[i]) == NULL)
13271 				continue;
13272 
13273 			if (match && ecb->dte_probe != NULL) {
13274 				dtrace_probe_t *probe = ecb->dte_probe;
13275 				dtrace_provider_t *prov = probe->dtpr_provider;
13276 
13277 				if (!(prov->dtpv_priv.dtpp_flags & match))
13278 					continue;
13279 			}
13280 
13281 			dtrace_ecb_disable(ecb);
13282 			dtrace_ecb_destroy(ecb);
13283 		}
13284 
13285 		if (!match)
13286 			break;
13287 	}
13288 
13289 	/*
13290 	 * Before we free the buffers, perform one more sync to assure that
13291 	 * every CPU is out of probe context.
13292 	 */
13293 	dtrace_sync();
13294 
13295 	dtrace_buffer_free(state->dts_buffer);
13296 	dtrace_buffer_free(state->dts_aggbuffer);
13297 
13298 	for (i = 0; i < nspec; i++)
13299 		dtrace_buffer_free(spec[i].dtsp_buffer);
13300 
13301 	if (state->dts_cleaner != CYCLIC_NONE)
13302 		cyclic_remove(state->dts_cleaner);
13303 
13304 	if (state->dts_deadman != CYCLIC_NONE)
13305 		cyclic_remove(state->dts_deadman);
13306 
13307 	dtrace_dstate_fini(&vstate->dtvs_dynvars);
13308 	dtrace_vstate_fini(vstate);
13309 	kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
13310 
13311 	if (state->dts_aggregations != NULL) {
13312 #ifdef DEBUG
13313 		for (i = 0; i < state->dts_naggregations; i++)
13314 			ASSERT(state->dts_aggregations[i] == NULL);
13315 #endif
13316 		ASSERT(state->dts_naggregations > 0);
13317 		kmem_free(state->dts_aggregations,
13318 		    state->dts_naggregations * sizeof (dtrace_aggregation_t *));
13319 	}
13320 
13321 	kmem_free(state->dts_buffer, bufsize);
13322 	kmem_free(state->dts_aggbuffer, bufsize);
13323 
13324 	for (i = 0; i < nspec; i++)
13325 		kmem_free(spec[i].dtsp_buffer, bufsize);
13326 
13327 	kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13328 
13329 	dtrace_format_destroy(state);
13330 
13331 	vmem_destroy(state->dts_aggid_arena);
13332 	ddi_soft_state_free(dtrace_softstate, minor);
13333 	vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
13334 }
13335 
13336 /*
13337  * DTrace Anonymous Enabling Functions
13338  */
13339 static dtrace_state_t *
13340 dtrace_anon_grab(void)
13341 {
13342 	dtrace_state_t *state;
13343 
13344 	ASSERT(MUTEX_HELD(&dtrace_lock));
13345 
13346 	if ((state = dtrace_anon.dta_state) == NULL) {
13347 		ASSERT(dtrace_anon.dta_enabling == NULL);
13348 		return (NULL);
13349 	}
13350 
13351 	ASSERT(dtrace_anon.dta_enabling != NULL);
13352 	ASSERT(dtrace_retained != NULL);
13353 
13354 	dtrace_enabling_destroy(dtrace_anon.dta_enabling);
13355 	dtrace_anon.dta_enabling = NULL;
13356 	dtrace_anon.dta_state = NULL;
13357 
13358 	return (state);
13359 }
13360 
13361 static void
13362 dtrace_anon_property(void)
13363 {
13364 	int i, rv;
13365 	dtrace_state_t *state;
13366 	dof_hdr_t *dof;
13367 	char c[32];		/* enough for "dof-data-" + digits */
13368 
13369 	ASSERT(MUTEX_HELD(&dtrace_lock));
13370 	ASSERT(MUTEX_HELD(&cpu_lock));
13371 
13372 	for (i = 0; ; i++) {
13373 		(void) snprintf(c, sizeof (c), "dof-data-%d", i);
13374 
13375 		dtrace_err_verbose = 1;
13376 
13377 		if ((dof = dtrace_dof_property(c)) == NULL) {
13378 			dtrace_err_verbose = 0;
13379 			break;
13380 		}
13381 
13382 		/*
13383 		 * We want to create anonymous state, so we need to transition
13384 		 * the kernel debugger to indicate that DTrace is active.  If
13385 		 * this fails (e.g. because the debugger has modified text in
13386 		 * some way), we won't continue with the processing.
13387 		 */
13388 		if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
13389 			cmn_err(CE_NOTE, "kernel debugger active; anonymous "
13390 			    "enabling ignored.");
13391 			dtrace_dof_destroy(dof);
13392 			break;
13393 		}
13394 
13395 		/*
13396 		 * If we haven't allocated an anonymous state, we'll do so now.
13397 		 */
13398 		if ((state = dtrace_anon.dta_state) == NULL) {
13399 			state = dtrace_state_create(NULL, NULL);
13400 			dtrace_anon.dta_state = state;
13401 
13402 			if (state == NULL) {
13403 				/*
13404 				 * This basically shouldn't happen:  the only
13405 				 * failure mode from dtrace_state_create() is a
13406 				 * failure of ddi_soft_state_zalloc() that
13407 				 * itself should never happen.  Still, the
13408 				 * interface allows for a failure mode, and
13409 				 * we want to fail as gracefully as possible:
13410 				 * we'll emit an error message and cease
13411 				 * processing anonymous state in this case.
13412 				 */
13413 				cmn_err(CE_WARN, "failed to create "
13414 				    "anonymous state");
13415 				dtrace_dof_destroy(dof);
13416 				break;
13417 			}
13418 		}
13419 
13420 		rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
13421 		    &dtrace_anon.dta_enabling, 0, B_TRUE);
13422 
13423 		if (rv == 0)
13424 			rv = dtrace_dof_options(dof, state);
13425 
13426 		dtrace_err_verbose = 0;
13427 		dtrace_dof_destroy(dof);
13428 
13429 		if (rv != 0) {
13430 			/*
13431 			 * This is malformed DOF; chuck any anonymous state
13432 			 * that we created.
13433 			 */
13434 			ASSERT(dtrace_anon.dta_enabling == NULL);
13435 			dtrace_state_destroy(state);
13436 			dtrace_anon.dta_state = NULL;
13437 			break;
13438 		}
13439 
13440 		ASSERT(dtrace_anon.dta_enabling != NULL);
13441 	}
13442 
13443 	if (dtrace_anon.dta_enabling != NULL) {
13444 		int rval;
13445 
13446 		/*
13447 		 * dtrace_enabling_retain() can only fail because we are
13448 		 * trying to retain more enablings than are allowed -- but
13449 		 * we only have one anonymous enabling, and we are guaranteed
13450 		 * to be allowed at least one retained enabling; we assert
13451 		 * that dtrace_enabling_retain() returns success.
13452 		 */
13453 		rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
13454 		ASSERT(rval == 0);
13455 
13456 		dtrace_enabling_dump(dtrace_anon.dta_enabling);
13457 	}
13458 }
13459 
13460 /*
13461  * DTrace Helper Functions
13462  */
13463 static void
13464 dtrace_helper_trace(dtrace_helper_action_t *helper,
13465     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
13466 {
13467 	uint32_t size, next, nnext, i;
13468 	dtrace_helptrace_t *ent;
13469 	uint16_t flags = cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
13470 
13471 	if (!dtrace_helptrace_enabled)
13472 		return;
13473 
13474 	ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
13475 
13476 	/*
13477 	 * What would a tracing framework be without its own tracing
13478 	 * framework?  (Well, a hell of a lot simpler, for starters...)
13479 	 */
13480 	size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
13481 	    sizeof (uint64_t) - sizeof (uint64_t);
13482 
13483 	/*
13484 	 * Iterate until we can allocate a slot in the trace buffer.
13485 	 */
13486 	do {
13487 		next = dtrace_helptrace_next;
13488 
13489 		if (next + size < dtrace_helptrace_bufsize) {
13490 			nnext = next + size;
13491 		} else {
13492 			nnext = size;
13493 		}
13494 	} while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
13495 
13496 	/*
13497 	 * We have our slot; fill it in.
13498 	 */
13499 	if (nnext == size)
13500 		next = 0;
13501 
13502 	ent = (dtrace_helptrace_t *)&dtrace_helptrace_buffer[next];
13503 	ent->dtht_helper = helper;
13504 	ent->dtht_where = where;
13505 	ent->dtht_nlocals = vstate->dtvs_nlocals;
13506 
13507 	ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
13508 	    mstate->dtms_fltoffs : -1;
13509 	ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
13510 	ent->dtht_illval = cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
13511 
13512 	for (i = 0; i < vstate->dtvs_nlocals; i++) {
13513 		dtrace_statvar_t *svar;
13514 
13515 		if ((svar = vstate->dtvs_locals[i]) == NULL)
13516 			continue;
13517 
13518 		ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
13519 		ent->dtht_locals[i] =
13520 		    ((uint64_t *)(uintptr_t)svar->dtsv_data)[CPU->cpu_id];
13521 	}
13522 }
13523 
13524 static uint64_t
13525 dtrace_helper(int which, dtrace_mstate_t *mstate,
13526     dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
13527 {
13528 	uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
13529 	uint64_t sarg0 = mstate->dtms_arg[0];
13530 	uint64_t sarg1 = mstate->dtms_arg[1];
13531 	uint64_t rval;
13532 	dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
13533 	dtrace_helper_action_t *helper;
13534 	dtrace_vstate_t *vstate;
13535 	dtrace_difo_t *pred;
13536 	int i, trace = dtrace_helptrace_enabled;
13537 
13538 	ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
13539 
13540 	if (helpers == NULL)
13541 		return (0);
13542 
13543 	if ((helper = helpers->dthps_actions[which]) == NULL)
13544 		return (0);
13545 
13546 	vstate = &helpers->dthps_vstate;
13547 	mstate->dtms_arg[0] = arg0;
13548 	mstate->dtms_arg[1] = arg1;
13549 
13550 	/*
13551 	 * Now iterate over each helper.  If its predicate evaluates to 'true',
13552 	 * we'll call the corresponding actions.  Note that the below calls
13553 	 * to dtrace_dif_emulate() may set faults in machine state.  This is
13554 	 * okay:  our caller (the outer dtrace_dif_emulate()) will simply plow
13555 	 * the stored DIF offset with its own (which is the desired behavior).
13556 	 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
13557 	 * from machine state; this is okay, too.
13558 	 */
13559 	for (; helper != NULL; helper = helper->dtha_next) {
13560 		if ((pred = helper->dtha_predicate) != NULL) {
13561 			if (trace)
13562 				dtrace_helper_trace(helper, mstate, vstate, 0);
13563 
13564 			if (!dtrace_dif_emulate(pred, mstate, vstate, state))
13565 				goto next;
13566 
13567 			if (*flags & CPU_DTRACE_FAULT)
13568 				goto err;
13569 		}
13570 
13571 		for (i = 0; i < helper->dtha_nactions; i++) {
13572 			if (trace)
13573 				dtrace_helper_trace(helper,
13574 				    mstate, vstate, i + 1);
13575 
13576 			rval = dtrace_dif_emulate(helper->dtha_actions[i],
13577 			    mstate, vstate, state);
13578 
13579 			if (*flags & CPU_DTRACE_FAULT)
13580 				goto err;
13581 		}
13582 
13583 next:
13584 		if (trace)
13585 			dtrace_helper_trace(helper, mstate, vstate,
13586 			    DTRACE_HELPTRACE_NEXT);
13587 	}
13588 
13589 	if (trace)
13590 		dtrace_helper_trace(helper, mstate, vstate,
13591 		    DTRACE_HELPTRACE_DONE);
13592 
13593 	/*
13594 	 * Restore the arg0 that we saved upon entry.
13595 	 */
13596 	mstate->dtms_arg[0] = sarg0;
13597 	mstate->dtms_arg[1] = sarg1;
13598 
13599 	return (rval);
13600 
13601 err:
13602 	if (trace)
13603 		dtrace_helper_trace(helper, mstate, vstate,
13604 		    DTRACE_HELPTRACE_ERR);
13605 
13606 	/*
13607 	 * Restore the arg0 that we saved upon entry.
13608 	 */
13609 	mstate->dtms_arg[0] = sarg0;
13610 	mstate->dtms_arg[1] = sarg1;
13611 
13612 	return (NULL);
13613 }
13614 
13615 static void
13616 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
13617     dtrace_vstate_t *vstate)
13618 {
13619 	int i;
13620 
13621 	if (helper->dtha_predicate != NULL)
13622 		dtrace_difo_release(helper->dtha_predicate, vstate);
13623 
13624 	for (i = 0; i < helper->dtha_nactions; i++) {
13625 		ASSERT(helper->dtha_actions[i] != NULL);
13626 		dtrace_difo_release(helper->dtha_actions[i], vstate);
13627 	}
13628 
13629 	kmem_free(helper->dtha_actions,
13630 	    helper->dtha_nactions * sizeof (dtrace_difo_t *));
13631 	kmem_free(helper, sizeof (dtrace_helper_action_t));
13632 }
13633 
13634 static int
13635 dtrace_helper_destroygen(int gen)
13636 {
13637 	proc_t *p = curproc;
13638 	dtrace_helpers_t *help = p->p_dtrace_helpers;
13639 	dtrace_vstate_t *vstate;
13640 	int i;
13641 
13642 	ASSERT(MUTEX_HELD(&dtrace_lock));
13643 
13644 	if (help == NULL || gen > help->dthps_generation)
13645 		return (EINVAL);
13646 
13647 	vstate = &help->dthps_vstate;
13648 
13649 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
13650 		dtrace_helper_action_t *last = NULL, *h, *next;
13651 
13652 		for (h = help->dthps_actions[i]; h != NULL; h = next) {
13653 			next = h->dtha_next;
13654 
13655 			if (h->dtha_generation == gen) {
13656 				if (last != NULL) {
13657 					last->dtha_next = next;
13658 				} else {
13659 					help->dthps_actions[i] = next;
13660 				}
13661 
13662 				dtrace_helper_action_destroy(h, vstate);
13663 			} else {
13664 				last = h;
13665 			}
13666 		}
13667 	}
13668 
13669 	/*
13670 	 * Interate until we've cleared out all helper providers with the
13671 	 * given generation number.
13672 	 */
13673 	for (;;) {
13674 		dtrace_helper_provider_t *prov;
13675 
13676 		/*
13677 		 * Look for a helper provider with the right generation. We
13678 		 * have to start back at the beginning of the list each time
13679 		 * because we drop dtrace_lock. It's unlikely that we'll make
13680 		 * more than two passes.
13681 		 */
13682 		for (i = 0; i < help->dthps_nprovs; i++) {
13683 			prov = help->dthps_provs[i];
13684 
13685 			if (prov->dthp_generation == gen)
13686 				break;
13687 		}
13688 
13689 		/*
13690 		 * If there were no matches, we're done.
13691 		 */
13692 		if (i == help->dthps_nprovs)
13693 			break;
13694 
13695 		/*
13696 		 * Move the last helper provider into this slot.
13697 		 */
13698 		help->dthps_nprovs--;
13699 		help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
13700 		help->dthps_provs[help->dthps_nprovs] = NULL;
13701 
13702 		mutex_exit(&dtrace_lock);
13703 
13704 		/*
13705 		 * If we have a meta provider, remove this helper provider.
13706 		 */
13707 		mutex_enter(&dtrace_meta_lock);
13708 		if (dtrace_meta_pid != NULL) {
13709 			ASSERT(dtrace_deferred_pid == NULL);
13710 			dtrace_helper_provider_remove(&prov->dthp_prov,
13711 			    p->p_pid);
13712 		}
13713 		mutex_exit(&dtrace_meta_lock);
13714 
13715 		dtrace_helper_provider_destroy(prov);
13716 
13717 		mutex_enter(&dtrace_lock);
13718 	}
13719 
13720 	return (0);
13721 }
13722 
13723 static int
13724 dtrace_helper_validate(dtrace_helper_action_t *helper)
13725 {
13726 	int err = 0, i;
13727 	dtrace_difo_t *dp;
13728 
13729 	if ((dp = helper->dtha_predicate) != NULL)
13730 		err += dtrace_difo_validate_helper(dp);
13731 
13732 	for (i = 0; i < helper->dtha_nactions; i++)
13733 		err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
13734 
13735 	return (err == 0);
13736 }
13737 
13738 static int
13739 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep)
13740 {
13741 	dtrace_helpers_t *help;
13742 	dtrace_helper_action_t *helper, *last;
13743 	dtrace_actdesc_t *act;
13744 	dtrace_vstate_t *vstate;
13745 	dtrace_predicate_t *pred;
13746 	int count = 0, nactions = 0, i;
13747 
13748 	if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
13749 		return (EINVAL);
13750 
13751 	help = curproc->p_dtrace_helpers;
13752 	last = help->dthps_actions[which];
13753 	vstate = &help->dthps_vstate;
13754 
13755 	for (count = 0; last != NULL; last = last->dtha_next) {
13756 		count++;
13757 		if (last->dtha_next == NULL)
13758 			break;
13759 	}
13760 
13761 	/*
13762 	 * If we already have dtrace_helper_actions_max helper actions for this
13763 	 * helper action type, we'll refuse to add a new one.
13764 	 */
13765 	if (count >= dtrace_helper_actions_max)
13766 		return (ENOSPC);
13767 
13768 	helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
13769 	helper->dtha_generation = help->dthps_generation;
13770 
13771 	if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
13772 		ASSERT(pred->dtp_difo != NULL);
13773 		dtrace_difo_hold(pred->dtp_difo);
13774 		helper->dtha_predicate = pred->dtp_difo;
13775 	}
13776 
13777 	for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
13778 		if (act->dtad_kind != DTRACEACT_DIFEXPR)
13779 			goto err;
13780 
13781 		if (act->dtad_difo == NULL)
13782 			goto err;
13783 
13784 		nactions++;
13785 	}
13786 
13787 	helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
13788 	    (helper->dtha_nactions = nactions), KM_SLEEP);
13789 
13790 	for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
13791 		dtrace_difo_hold(act->dtad_difo);
13792 		helper->dtha_actions[i++] = act->dtad_difo;
13793 	}
13794 
13795 	if (!dtrace_helper_validate(helper))
13796 		goto err;
13797 
13798 	if (last == NULL) {
13799 		help->dthps_actions[which] = helper;
13800 	} else {
13801 		last->dtha_next = helper;
13802 	}
13803 
13804 	if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
13805 		dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
13806 		dtrace_helptrace_next = 0;
13807 	}
13808 
13809 	return (0);
13810 err:
13811 	dtrace_helper_action_destroy(helper, vstate);
13812 	return (EINVAL);
13813 }
13814 
13815 static void
13816 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
13817     dof_helper_t *dofhp)
13818 {
13819 	ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
13820 
13821 	mutex_enter(&dtrace_meta_lock);
13822 	mutex_enter(&dtrace_lock);
13823 
13824 	if (!dtrace_attached() || dtrace_meta_pid == NULL) {
13825 		/*
13826 		 * If the dtrace module is loaded but not attached, or if
13827 		 * there aren't isn't a meta provider registered to deal with
13828 		 * these provider descriptions, we need to postpone creating
13829 		 * the actual providers until later.
13830 		 */
13831 
13832 		if (help->dthps_next == NULL && help->dthps_prev == NULL &&
13833 		    dtrace_deferred_pid != help) {
13834 			help->dthps_deferred = 1;
13835 			help->dthps_pid = p->p_pid;
13836 			help->dthps_next = dtrace_deferred_pid;
13837 			help->dthps_prev = NULL;
13838 			if (dtrace_deferred_pid != NULL)
13839 				dtrace_deferred_pid->dthps_prev = help;
13840 			dtrace_deferred_pid = help;
13841 		}
13842 
13843 		mutex_exit(&dtrace_lock);
13844 
13845 	} else if (dofhp != NULL) {
13846 		/*
13847 		 * If the dtrace module is loaded and we have a particular
13848 		 * helper provider description, pass that off to the
13849 		 * meta provider.
13850 		 */
13851 
13852 		mutex_exit(&dtrace_lock);
13853 
13854 		dtrace_helper_provide(dofhp, p->p_pid);
13855 
13856 	} else {
13857 		/*
13858 		 * Otherwise, just pass all the helper provider descriptions
13859 		 * off to the meta provider.
13860 		 */
13861 
13862 		int i;
13863 		mutex_exit(&dtrace_lock);
13864 
13865 		for (i = 0; i < help->dthps_nprovs; i++) {
13866 			dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
13867 			    p->p_pid);
13868 		}
13869 	}
13870 
13871 	mutex_exit(&dtrace_meta_lock);
13872 }
13873 
13874 static int
13875 dtrace_helper_provider_add(dof_helper_t *dofhp, int gen)
13876 {
13877 	dtrace_helpers_t *help;
13878 	dtrace_helper_provider_t *hprov, **tmp_provs;
13879 	uint_t tmp_maxprovs, i;
13880 
13881 	ASSERT(MUTEX_HELD(&dtrace_lock));
13882 
13883 	help = curproc->p_dtrace_helpers;
13884 	ASSERT(help != NULL);
13885 
13886 	/*
13887 	 * If we already have dtrace_helper_providers_max helper providers,
13888 	 * we're refuse to add a new one.
13889 	 */
13890 	if (help->dthps_nprovs >= dtrace_helper_providers_max)
13891 		return (ENOSPC);
13892 
13893 	/*
13894 	 * Check to make sure this isn't a duplicate.
13895 	 */
13896 	for (i = 0; i < help->dthps_nprovs; i++) {
13897 		if (dofhp->dofhp_addr ==
13898 		    help->dthps_provs[i]->dthp_prov.dofhp_addr)
13899 			return (EALREADY);
13900 	}
13901 
13902 	hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
13903 	hprov->dthp_prov = *dofhp;
13904 	hprov->dthp_ref = 1;
13905 	hprov->dthp_generation = gen;
13906 
13907 	/*
13908 	 * Allocate a bigger table for helper providers if it's already full.
13909 	 */
13910 	if (help->dthps_maxprovs == help->dthps_nprovs) {
13911 		tmp_maxprovs = help->dthps_maxprovs;
13912 		tmp_provs = help->dthps_provs;
13913 
13914 		if (help->dthps_maxprovs == 0)
13915 			help->dthps_maxprovs = 2;
13916 		else
13917 			help->dthps_maxprovs *= 2;
13918 		if (help->dthps_maxprovs > dtrace_helper_providers_max)
13919 			help->dthps_maxprovs = dtrace_helper_providers_max;
13920 
13921 		ASSERT(tmp_maxprovs < help->dthps_maxprovs);
13922 
13923 		help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
13924 		    sizeof (dtrace_helper_provider_t *), KM_SLEEP);
13925 
13926 		if (tmp_provs != NULL) {
13927 			bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
13928 			    sizeof (dtrace_helper_provider_t *));
13929 			kmem_free(tmp_provs, tmp_maxprovs *
13930 			    sizeof (dtrace_helper_provider_t *));
13931 		}
13932 	}
13933 
13934 	help->dthps_provs[help->dthps_nprovs] = hprov;
13935 	help->dthps_nprovs++;
13936 
13937 	return (0);
13938 }
13939 
13940 static void
13941 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
13942 {
13943 	mutex_enter(&dtrace_lock);
13944 
13945 	if (--hprov->dthp_ref == 0) {
13946 		dof_hdr_t *dof;
13947 		mutex_exit(&dtrace_lock);
13948 		dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
13949 		dtrace_dof_destroy(dof);
13950 		kmem_free(hprov, sizeof (dtrace_helper_provider_t));
13951 	} else {
13952 		mutex_exit(&dtrace_lock);
13953 	}
13954 }
13955 
13956 static int
13957 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
13958 {
13959 	uintptr_t daddr = (uintptr_t)dof;
13960 	dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
13961 	dof_provider_t *provider;
13962 	dof_probe_t *probe;
13963 	uint8_t *arg;
13964 	char *strtab, *typestr;
13965 	dof_stridx_t typeidx;
13966 	size_t typesz;
13967 	uint_t nprobes, j, k;
13968 
13969 	ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
13970 
13971 	if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
13972 		dtrace_dof_error(dof, "misaligned section offset");
13973 		return (-1);
13974 	}
13975 
13976 	/*
13977 	 * The section needs to be large enough to contain the DOF provider
13978 	 * structure appropriate for the given version.
13979 	 */
13980 	if (sec->dofs_size <
13981 	    ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
13982 	    offsetof(dof_provider_t, dofpv_prenoffs) :
13983 	    sizeof (dof_provider_t))) {
13984 		dtrace_dof_error(dof, "provider section too small");
13985 		return (-1);
13986 	}
13987 
13988 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
13989 	str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
13990 	prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
13991 	arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
13992 	off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
13993 
13994 	if (str_sec == NULL || prb_sec == NULL ||
13995 	    arg_sec == NULL || off_sec == NULL)
13996 		return (-1);
13997 
13998 	enoff_sec = NULL;
13999 
14000 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
14001 	    provider->dofpv_prenoffs != DOF_SECT_NONE &&
14002 	    (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
14003 	    provider->dofpv_prenoffs)) == NULL)
14004 		return (-1);
14005 
14006 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
14007 
14008 	if (provider->dofpv_name >= str_sec->dofs_size ||
14009 	    strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
14010 		dtrace_dof_error(dof, "invalid provider name");
14011 		return (-1);
14012 	}
14013 
14014 	if (prb_sec->dofs_entsize == 0 ||
14015 	    prb_sec->dofs_entsize > prb_sec->dofs_size) {
14016 		dtrace_dof_error(dof, "invalid entry size");
14017 		return (-1);
14018 	}
14019 
14020 	if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
14021 		dtrace_dof_error(dof, "misaligned entry size");
14022 		return (-1);
14023 	}
14024 
14025 	if (off_sec->dofs_entsize != sizeof (uint32_t)) {
14026 		dtrace_dof_error(dof, "invalid entry size");
14027 		return (-1);
14028 	}
14029 
14030 	if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
14031 		dtrace_dof_error(dof, "misaligned section offset");
14032 		return (-1);
14033 	}
14034 
14035 	if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
14036 		dtrace_dof_error(dof, "invalid entry size");
14037 		return (-1);
14038 	}
14039 
14040 	arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
14041 
14042 	nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
14043 
14044 	/*
14045 	 * Take a pass through the probes to check for errors.
14046 	 */
14047 	for (j = 0; j < nprobes; j++) {
14048 		probe = (dof_probe_t *)(uintptr_t)(daddr +
14049 		    prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
14050 
14051 		if (probe->dofpr_func >= str_sec->dofs_size) {
14052 			dtrace_dof_error(dof, "invalid function name");
14053 			return (-1);
14054 		}
14055 
14056 		if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
14057 			dtrace_dof_error(dof, "function name too long");
14058 			return (-1);
14059 		}
14060 
14061 		if (probe->dofpr_name >= str_sec->dofs_size ||
14062 		    strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
14063 			dtrace_dof_error(dof, "invalid probe name");
14064 			return (-1);
14065 		}
14066 
14067 		/*
14068 		 * The offset count must not wrap the index, and the offsets
14069 		 * must also not overflow the section's data.
14070 		 */
14071 		if (probe->dofpr_offidx + probe->dofpr_noffs <
14072 		    probe->dofpr_offidx ||
14073 		    (probe->dofpr_offidx + probe->dofpr_noffs) *
14074 		    off_sec->dofs_entsize > off_sec->dofs_size) {
14075 			dtrace_dof_error(dof, "invalid probe offset");
14076 			return (-1);
14077 		}
14078 
14079 		if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
14080 			/*
14081 			 * If there's no is-enabled offset section, make sure
14082 			 * there aren't any is-enabled offsets. Otherwise
14083 			 * perform the same checks as for probe offsets
14084 			 * (immediately above).
14085 			 */
14086 			if (enoff_sec == NULL) {
14087 				if (probe->dofpr_enoffidx != 0 ||
14088 				    probe->dofpr_nenoffs != 0) {
14089 					dtrace_dof_error(dof, "is-enabled "
14090 					    "offsets with null section");
14091 					return (-1);
14092 				}
14093 			} else if (probe->dofpr_enoffidx +
14094 			    probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
14095 			    (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
14096 			    enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
14097 				dtrace_dof_error(dof, "invalid is-enabled "
14098 				    "offset");
14099 				return (-1);
14100 			}
14101 
14102 			if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
14103 				dtrace_dof_error(dof, "zero probe and "
14104 				    "is-enabled offsets");
14105 				return (-1);
14106 			}
14107 		} else if (probe->dofpr_noffs == 0) {
14108 			dtrace_dof_error(dof, "zero probe offsets");
14109 			return (-1);
14110 		}
14111 
14112 		if (probe->dofpr_argidx + probe->dofpr_xargc <
14113 		    probe->dofpr_argidx ||
14114 		    (probe->dofpr_argidx + probe->dofpr_xargc) *
14115 		    arg_sec->dofs_entsize > arg_sec->dofs_size) {
14116 			dtrace_dof_error(dof, "invalid args");
14117 			return (-1);
14118 		}
14119 
14120 		typeidx = probe->dofpr_nargv;
14121 		typestr = strtab + probe->dofpr_nargv;
14122 		for (k = 0; k < probe->dofpr_nargc; k++) {
14123 			if (typeidx >= str_sec->dofs_size) {
14124 				dtrace_dof_error(dof, "bad "
14125 				    "native argument type");
14126 				return (-1);
14127 			}
14128 
14129 			typesz = strlen(typestr) + 1;
14130 			if (typesz > DTRACE_ARGTYPELEN) {
14131 				dtrace_dof_error(dof, "native "
14132 				    "argument type too long");
14133 				return (-1);
14134 			}
14135 			typeidx += typesz;
14136 			typestr += typesz;
14137 		}
14138 
14139 		typeidx = probe->dofpr_xargv;
14140 		typestr = strtab + probe->dofpr_xargv;
14141 		for (k = 0; k < probe->dofpr_xargc; k++) {
14142 			if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
14143 				dtrace_dof_error(dof, "bad "
14144 				    "native argument index");
14145 				return (-1);
14146 			}
14147 
14148 			if (typeidx >= str_sec->dofs_size) {
14149 				dtrace_dof_error(dof, "bad "
14150 				    "translated argument type");
14151 				return (-1);
14152 			}
14153 
14154 			typesz = strlen(typestr) + 1;
14155 			if (typesz > DTRACE_ARGTYPELEN) {
14156 				dtrace_dof_error(dof, "translated argument "
14157 				    "type too long");
14158 				return (-1);
14159 			}
14160 
14161 			typeidx += typesz;
14162 			typestr += typesz;
14163 		}
14164 	}
14165 
14166 	return (0);
14167 }
14168 
14169 static int
14170 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp)
14171 {
14172 	dtrace_helpers_t *help;
14173 	dtrace_vstate_t *vstate;
14174 	dtrace_enabling_t *enab = NULL;
14175 	int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
14176 	uintptr_t daddr = (uintptr_t)dof;
14177 
14178 	ASSERT(MUTEX_HELD(&dtrace_lock));
14179 
14180 	if ((help = curproc->p_dtrace_helpers) == NULL)
14181 		help = dtrace_helpers_create(curproc);
14182 
14183 	vstate = &help->dthps_vstate;
14184 
14185 	if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab,
14186 	    dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) {
14187 		dtrace_dof_destroy(dof);
14188 		return (rv);
14189 	}
14190 
14191 	/*
14192 	 * Look for helper providers and validate their descriptions.
14193 	 */
14194 	if (dhp != NULL) {
14195 		for (i = 0; i < dof->dofh_secnum; i++) {
14196 			dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
14197 			    dof->dofh_secoff + i * dof->dofh_secsize);
14198 
14199 			if (sec->dofs_type != DOF_SECT_PROVIDER)
14200 				continue;
14201 
14202 			if (dtrace_helper_provider_validate(dof, sec) != 0) {
14203 				dtrace_enabling_destroy(enab);
14204 				dtrace_dof_destroy(dof);
14205 				return (-1);
14206 			}
14207 
14208 			nprovs++;
14209 		}
14210 	}
14211 
14212 	/*
14213 	 * Now we need to walk through the ECB descriptions in the enabling.
14214 	 */
14215 	for (i = 0; i < enab->dten_ndesc; i++) {
14216 		dtrace_ecbdesc_t *ep = enab->dten_desc[i];
14217 		dtrace_probedesc_t *desc = &ep->dted_probe;
14218 
14219 		if (strcmp(desc->dtpd_provider, "dtrace") != 0)
14220 			continue;
14221 
14222 		if (strcmp(desc->dtpd_mod, "helper") != 0)
14223 			continue;
14224 
14225 		if (strcmp(desc->dtpd_func, "ustack") != 0)
14226 			continue;
14227 
14228 		if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
14229 		    ep)) != 0) {
14230 			/*
14231 			 * Adding this helper action failed -- we are now going
14232 			 * to rip out the entire generation and return failure.
14233 			 */
14234 			(void) dtrace_helper_destroygen(help->dthps_generation);
14235 			dtrace_enabling_destroy(enab);
14236 			dtrace_dof_destroy(dof);
14237 			return (-1);
14238 		}
14239 
14240 		nhelpers++;
14241 	}
14242 
14243 	if (nhelpers < enab->dten_ndesc)
14244 		dtrace_dof_error(dof, "unmatched helpers");
14245 
14246 	gen = help->dthps_generation++;
14247 	dtrace_enabling_destroy(enab);
14248 
14249 	if (dhp != NULL && nprovs > 0) {
14250 		dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
14251 		if (dtrace_helper_provider_add(dhp, gen) == 0) {
14252 			mutex_exit(&dtrace_lock);
14253 			dtrace_helper_provider_register(curproc, help, dhp);
14254 			mutex_enter(&dtrace_lock);
14255 
14256 			destroy = 0;
14257 		}
14258 	}
14259 
14260 	if (destroy)
14261 		dtrace_dof_destroy(dof);
14262 
14263 	return (gen);
14264 }
14265 
14266 static dtrace_helpers_t *
14267 dtrace_helpers_create(proc_t *p)
14268 {
14269 	dtrace_helpers_t *help;
14270 
14271 	ASSERT(MUTEX_HELD(&dtrace_lock));
14272 	ASSERT(p->p_dtrace_helpers == NULL);
14273 
14274 	help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
14275 	help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
14276 	    DTRACE_NHELPER_ACTIONS, KM_SLEEP);
14277 
14278 	p->p_dtrace_helpers = help;
14279 	dtrace_helpers++;
14280 
14281 	return (help);
14282 }
14283 
14284 static void
14285 dtrace_helpers_destroy(void)
14286 {
14287 	dtrace_helpers_t *help;
14288 	dtrace_vstate_t *vstate;
14289 	proc_t *p = curproc;
14290 	int i;
14291 
14292 	mutex_enter(&dtrace_lock);
14293 
14294 	ASSERT(p->p_dtrace_helpers != NULL);
14295 	ASSERT(dtrace_helpers > 0);
14296 
14297 	help = p->p_dtrace_helpers;
14298 	vstate = &help->dthps_vstate;
14299 
14300 	/*
14301 	 * We're now going to lose the help from this process.
14302 	 */
14303 	p->p_dtrace_helpers = NULL;
14304 	dtrace_sync();
14305 
14306 	/*
14307 	 * Destory the helper actions.
14308 	 */
14309 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14310 		dtrace_helper_action_t *h, *next;
14311 
14312 		for (h = help->dthps_actions[i]; h != NULL; h = next) {
14313 			next = h->dtha_next;
14314 			dtrace_helper_action_destroy(h, vstate);
14315 			h = next;
14316 		}
14317 	}
14318 
14319 	mutex_exit(&dtrace_lock);
14320 
14321 	/*
14322 	 * Destroy the helper providers.
14323 	 */
14324 	if (help->dthps_maxprovs > 0) {
14325 		mutex_enter(&dtrace_meta_lock);
14326 		if (dtrace_meta_pid != NULL) {
14327 			ASSERT(dtrace_deferred_pid == NULL);
14328 
14329 			for (i = 0; i < help->dthps_nprovs; i++) {
14330 				dtrace_helper_provider_remove(
14331 				    &help->dthps_provs[i]->dthp_prov, p->p_pid);
14332 			}
14333 		} else {
14334 			mutex_enter(&dtrace_lock);
14335 			ASSERT(help->dthps_deferred == 0 ||
14336 			    help->dthps_next != NULL ||
14337 			    help->dthps_prev != NULL ||
14338 			    help == dtrace_deferred_pid);
14339 
14340 			/*
14341 			 * Remove the helper from the deferred list.
14342 			 */
14343 			if (help->dthps_next != NULL)
14344 				help->dthps_next->dthps_prev = help->dthps_prev;
14345 			if (help->dthps_prev != NULL)
14346 				help->dthps_prev->dthps_next = help->dthps_next;
14347 			if (dtrace_deferred_pid == help) {
14348 				dtrace_deferred_pid = help->dthps_next;
14349 				ASSERT(help->dthps_prev == NULL);
14350 			}
14351 
14352 			mutex_exit(&dtrace_lock);
14353 		}
14354 
14355 		mutex_exit(&dtrace_meta_lock);
14356 
14357 		for (i = 0; i < help->dthps_nprovs; i++) {
14358 			dtrace_helper_provider_destroy(help->dthps_provs[i]);
14359 		}
14360 
14361 		kmem_free(help->dthps_provs, help->dthps_maxprovs *
14362 		    sizeof (dtrace_helper_provider_t *));
14363 	}
14364 
14365 	mutex_enter(&dtrace_lock);
14366 
14367 	dtrace_vstate_fini(&help->dthps_vstate);
14368 	kmem_free(help->dthps_actions,
14369 	    sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
14370 	kmem_free(help, sizeof (dtrace_helpers_t));
14371 
14372 	--dtrace_helpers;
14373 	mutex_exit(&dtrace_lock);
14374 }
14375 
14376 static void
14377 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
14378 {
14379 	dtrace_helpers_t *help, *newhelp;
14380 	dtrace_helper_action_t *helper, *new, *last;
14381 	dtrace_difo_t *dp;
14382 	dtrace_vstate_t *vstate;
14383 	int i, j, sz, hasprovs = 0;
14384 
14385 	mutex_enter(&dtrace_lock);
14386 	ASSERT(from->p_dtrace_helpers != NULL);
14387 	ASSERT(dtrace_helpers > 0);
14388 
14389 	help = from->p_dtrace_helpers;
14390 	newhelp = dtrace_helpers_create(to);
14391 	ASSERT(to->p_dtrace_helpers != NULL);
14392 
14393 	newhelp->dthps_generation = help->dthps_generation;
14394 	vstate = &newhelp->dthps_vstate;
14395 
14396 	/*
14397 	 * Duplicate the helper actions.
14398 	 */
14399 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14400 		if ((helper = help->dthps_actions[i]) == NULL)
14401 			continue;
14402 
14403 		for (last = NULL; helper != NULL; helper = helper->dtha_next) {
14404 			new = kmem_zalloc(sizeof (dtrace_helper_action_t),
14405 			    KM_SLEEP);
14406 			new->dtha_generation = helper->dtha_generation;
14407 
14408 			if ((dp = helper->dtha_predicate) != NULL) {
14409 				dp = dtrace_difo_duplicate(dp, vstate);
14410 				new->dtha_predicate = dp;
14411 			}
14412 
14413 			new->dtha_nactions = helper->dtha_nactions;
14414 			sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
14415 			new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
14416 
14417 			for (j = 0; j < new->dtha_nactions; j++) {
14418 				dtrace_difo_t *dp = helper->dtha_actions[j];
14419 
14420 				ASSERT(dp != NULL);
14421 				dp = dtrace_difo_duplicate(dp, vstate);
14422 				new->dtha_actions[j] = dp;
14423 			}
14424 
14425 			if (last != NULL) {
14426 				last->dtha_next = new;
14427 			} else {
14428 				newhelp->dthps_actions[i] = new;
14429 			}
14430 
14431 			last = new;
14432 		}
14433 	}
14434 
14435 	/*
14436 	 * Duplicate the helper providers and register them with the
14437 	 * DTrace framework.
14438 	 */
14439 	if (help->dthps_nprovs > 0) {
14440 		newhelp->dthps_nprovs = help->dthps_nprovs;
14441 		newhelp->dthps_maxprovs = help->dthps_nprovs;
14442 		newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
14443 		    sizeof (dtrace_helper_provider_t *), KM_SLEEP);
14444 		for (i = 0; i < newhelp->dthps_nprovs; i++) {
14445 			newhelp->dthps_provs[i] = help->dthps_provs[i];
14446 			newhelp->dthps_provs[i]->dthp_ref++;
14447 		}
14448 
14449 		hasprovs = 1;
14450 	}
14451 
14452 	mutex_exit(&dtrace_lock);
14453 
14454 	if (hasprovs)
14455 		dtrace_helper_provider_register(to, newhelp, NULL);
14456 }
14457 
14458 /*
14459  * DTrace Hook Functions
14460  */
14461 static void
14462 dtrace_module_loaded(struct modctl *ctl)
14463 {
14464 	dtrace_provider_t *prv;
14465 
14466 	mutex_enter(&dtrace_provider_lock);
14467 	mutex_enter(&mod_lock);
14468 
14469 	ASSERT(ctl->mod_busy);
14470 
14471 	/*
14472 	 * We're going to call each providers per-module provide operation
14473 	 * specifying only this module.
14474 	 */
14475 	for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
14476 		prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
14477 
14478 	mutex_exit(&mod_lock);
14479 	mutex_exit(&dtrace_provider_lock);
14480 
14481 	/*
14482 	 * If we have any retained enablings, we need to match against them.
14483 	 * Enabling probes requires that cpu_lock be held, and we cannot hold
14484 	 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
14485 	 * module.  (In particular, this happens when loading scheduling
14486 	 * classes.)  So if we have any retained enablings, we need to dispatch
14487 	 * our task queue to do the match for us.
14488 	 */
14489 	mutex_enter(&dtrace_lock);
14490 
14491 	if (dtrace_retained == NULL) {
14492 		mutex_exit(&dtrace_lock);
14493 		return;
14494 	}
14495 
14496 	(void) taskq_dispatch(dtrace_taskq,
14497 	    (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
14498 
14499 	mutex_exit(&dtrace_lock);
14500 
14501 	/*
14502 	 * And now, for a little heuristic sleaze:  in general, we want to
14503 	 * match modules as soon as they load.  However, we cannot guarantee
14504 	 * this, because it would lead us to the lock ordering violation
14505 	 * outlined above.  The common case, of course, is that cpu_lock is
14506 	 * _not_ held -- so we delay here for a clock tick, hoping that that's
14507 	 * long enough for the task queue to do its work.  If it's not, it's
14508 	 * not a serious problem -- it just means that the module that we
14509 	 * just loaded may not be immediately instrumentable.
14510 	 */
14511 	delay(1);
14512 }
14513 
14514 static void
14515 dtrace_module_unloaded(struct modctl *ctl)
14516 {
14517 	dtrace_probe_t template, *probe, *first, *next;
14518 	dtrace_provider_t *prov;
14519 
14520 	template.dtpr_mod = ctl->mod_modname;
14521 
14522 	mutex_enter(&dtrace_provider_lock);
14523 	mutex_enter(&mod_lock);
14524 	mutex_enter(&dtrace_lock);
14525 
14526 	if (dtrace_bymod == NULL) {
14527 		/*
14528 		 * The DTrace module is loaded (obviously) but not attached;
14529 		 * we don't have any work to do.
14530 		 */
14531 		mutex_exit(&dtrace_provider_lock);
14532 		mutex_exit(&mod_lock);
14533 		mutex_exit(&dtrace_lock);
14534 		return;
14535 	}
14536 
14537 	for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
14538 	    probe != NULL; probe = probe->dtpr_nextmod) {
14539 		if (probe->dtpr_ecb != NULL) {
14540 			mutex_exit(&dtrace_provider_lock);
14541 			mutex_exit(&mod_lock);
14542 			mutex_exit(&dtrace_lock);
14543 
14544 			/*
14545 			 * This shouldn't _actually_ be possible -- we're
14546 			 * unloading a module that has an enabled probe in it.
14547 			 * (It's normally up to the provider to make sure that
14548 			 * this can't happen.)  However, because dtps_enable()
14549 			 * doesn't have a failure mode, there can be an
14550 			 * enable/unload race.  Upshot:  we don't want to
14551 			 * assert, but we're not going to disable the
14552 			 * probe, either.
14553 			 */
14554 			if (dtrace_err_verbose) {
14555 				cmn_err(CE_WARN, "unloaded module '%s' had "
14556 				    "enabled probes", ctl->mod_modname);
14557 			}
14558 
14559 			return;
14560 		}
14561 	}
14562 
14563 	probe = first;
14564 
14565 	for (first = NULL; probe != NULL; probe = next) {
14566 		ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
14567 
14568 		dtrace_probes[probe->dtpr_id - 1] = NULL;
14569 
14570 		next = probe->dtpr_nextmod;
14571 		dtrace_hash_remove(dtrace_bymod, probe);
14572 		dtrace_hash_remove(dtrace_byfunc, probe);
14573 		dtrace_hash_remove(dtrace_byname, probe);
14574 
14575 		if (first == NULL) {
14576 			first = probe;
14577 			probe->dtpr_nextmod = NULL;
14578 		} else {
14579 			probe->dtpr_nextmod = first;
14580 			first = probe;
14581 		}
14582 	}
14583 
14584 	/*
14585 	 * We've removed all of the module's probes from the hash chains and
14586 	 * from the probe array.  Now issue a dtrace_sync() to be sure that
14587 	 * everyone has cleared out from any probe array processing.
14588 	 */
14589 	dtrace_sync();
14590 
14591 	for (probe = first; probe != NULL; probe = first) {
14592 		first = probe->dtpr_nextmod;
14593 		prov = probe->dtpr_provider;
14594 		prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
14595 		    probe->dtpr_arg);
14596 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
14597 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
14598 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
14599 		vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
14600 		kmem_free(probe, sizeof (dtrace_probe_t));
14601 	}
14602 
14603 	mutex_exit(&dtrace_lock);
14604 	mutex_exit(&mod_lock);
14605 	mutex_exit(&dtrace_provider_lock);
14606 }
14607 
14608 void
14609 dtrace_suspend(void)
14610 {
14611 	dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
14612 }
14613 
14614 void
14615 dtrace_resume(void)
14616 {
14617 	dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
14618 }
14619 
14620 static int
14621 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
14622 {
14623 	ASSERT(MUTEX_HELD(&cpu_lock));
14624 	mutex_enter(&dtrace_lock);
14625 
14626 	switch (what) {
14627 	case CPU_CONFIG: {
14628 		dtrace_state_t *state;
14629 		dtrace_optval_t *opt, rs, c;
14630 
14631 		/*
14632 		 * For now, we only allocate a new buffer for anonymous state.
14633 		 */
14634 		if ((state = dtrace_anon.dta_state) == NULL)
14635 			break;
14636 
14637 		if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
14638 			break;
14639 
14640 		opt = state->dts_options;
14641 		c = opt[DTRACEOPT_CPU];
14642 
14643 		if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
14644 			break;
14645 
14646 		/*
14647 		 * Regardless of what the actual policy is, we're going to
14648 		 * temporarily set our resize policy to be manual.  We're
14649 		 * also going to temporarily set our CPU option to denote
14650 		 * the newly configured CPU.
14651 		 */
14652 		rs = opt[DTRACEOPT_BUFRESIZE];
14653 		opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
14654 		opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
14655 
14656 		(void) dtrace_state_buffers(state);
14657 
14658 		opt[DTRACEOPT_BUFRESIZE] = rs;
14659 		opt[DTRACEOPT_CPU] = c;
14660 
14661 		break;
14662 	}
14663 
14664 	case CPU_UNCONFIG:
14665 		/*
14666 		 * We don't free the buffer in the CPU_UNCONFIG case.  (The
14667 		 * buffer will be freed when the consumer exits.)
14668 		 */
14669 		break;
14670 
14671 	default:
14672 		break;
14673 	}
14674 
14675 	mutex_exit(&dtrace_lock);
14676 	return (0);
14677 }
14678 
14679 static void
14680 dtrace_cpu_setup_initial(processorid_t cpu)
14681 {
14682 	(void) dtrace_cpu_setup(CPU_CONFIG, cpu);
14683 }
14684 
14685 static void
14686 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
14687 {
14688 	if (dtrace_toxranges >= dtrace_toxranges_max) {
14689 		int osize, nsize;
14690 		dtrace_toxrange_t *range;
14691 
14692 		osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
14693 
14694 		if (osize == 0) {
14695 			ASSERT(dtrace_toxrange == NULL);
14696 			ASSERT(dtrace_toxranges_max == 0);
14697 			dtrace_toxranges_max = 1;
14698 		} else {
14699 			dtrace_toxranges_max <<= 1;
14700 		}
14701 
14702 		nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
14703 		range = kmem_zalloc(nsize, KM_SLEEP);
14704 
14705 		if (dtrace_toxrange != NULL) {
14706 			ASSERT(osize != 0);
14707 			bcopy(dtrace_toxrange, range, osize);
14708 			kmem_free(dtrace_toxrange, osize);
14709 		}
14710 
14711 		dtrace_toxrange = range;
14712 	}
14713 
14714 	ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == NULL);
14715 	ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == NULL);
14716 
14717 	dtrace_toxrange[dtrace_toxranges].dtt_base = base;
14718 	dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
14719 	dtrace_toxranges++;
14720 }
14721 
14722 /*
14723  * DTrace Driver Cookbook Functions
14724  */
14725 /*ARGSUSED*/
14726 static int
14727 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
14728 {
14729 	dtrace_provider_id_t id;
14730 	dtrace_state_t *state = NULL;
14731 	dtrace_enabling_t *enab;
14732 
14733 	mutex_enter(&cpu_lock);
14734 	mutex_enter(&dtrace_provider_lock);
14735 	mutex_enter(&dtrace_lock);
14736 
14737 	if (ddi_soft_state_init(&dtrace_softstate,
14738 	    sizeof (dtrace_state_t), 0) != 0) {
14739 		cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
14740 		mutex_exit(&cpu_lock);
14741 		mutex_exit(&dtrace_provider_lock);
14742 		mutex_exit(&dtrace_lock);
14743 		return (DDI_FAILURE);
14744 	}
14745 
14746 	if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
14747 	    DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
14748 	    ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
14749 	    DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
14750 		cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
14751 		ddi_remove_minor_node(devi, NULL);
14752 		ddi_soft_state_fini(&dtrace_softstate);
14753 		mutex_exit(&cpu_lock);
14754 		mutex_exit(&dtrace_provider_lock);
14755 		mutex_exit(&dtrace_lock);
14756 		return (DDI_FAILURE);
14757 	}
14758 
14759 	ddi_report_dev(devi);
14760 	dtrace_devi = devi;
14761 
14762 	dtrace_modload = dtrace_module_loaded;
14763 	dtrace_modunload = dtrace_module_unloaded;
14764 	dtrace_cpu_init = dtrace_cpu_setup_initial;
14765 	dtrace_helpers_cleanup = dtrace_helpers_destroy;
14766 	dtrace_helpers_fork = dtrace_helpers_duplicate;
14767 	dtrace_cpustart_init = dtrace_suspend;
14768 	dtrace_cpustart_fini = dtrace_resume;
14769 	dtrace_debugger_init = dtrace_suspend;
14770 	dtrace_debugger_fini = dtrace_resume;
14771 
14772 	register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
14773 
14774 	ASSERT(MUTEX_HELD(&cpu_lock));
14775 
14776 	dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
14777 	    NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
14778 	dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
14779 	    UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
14780 	    VM_SLEEP | VMC_IDENTIFIER);
14781 	dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
14782 	    1, INT_MAX, 0);
14783 
14784 	dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
14785 	    sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
14786 	    NULL, NULL, NULL, NULL, NULL, 0);
14787 
14788 	ASSERT(MUTEX_HELD(&cpu_lock));
14789 	dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
14790 	    offsetof(dtrace_probe_t, dtpr_nextmod),
14791 	    offsetof(dtrace_probe_t, dtpr_prevmod));
14792 
14793 	dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
14794 	    offsetof(dtrace_probe_t, dtpr_nextfunc),
14795 	    offsetof(dtrace_probe_t, dtpr_prevfunc));
14796 
14797 	dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
14798 	    offsetof(dtrace_probe_t, dtpr_nextname),
14799 	    offsetof(dtrace_probe_t, dtpr_prevname));
14800 
14801 	if (dtrace_retain_max < 1) {
14802 		cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; "
14803 		    "setting to 1", dtrace_retain_max);
14804 		dtrace_retain_max = 1;
14805 	}
14806 
14807 	/*
14808 	 * Now discover our toxic ranges.
14809 	 */
14810 	dtrace_toxic_ranges(dtrace_toxrange_add);
14811 
14812 	/*
14813 	 * Before we register ourselves as a provider to our own framework,
14814 	 * we would like to assert that dtrace_provider is NULL -- but that's
14815 	 * not true if we were loaded as a dependency of a DTrace provider.
14816 	 * Once we've registered, we can assert that dtrace_provider is our
14817 	 * pseudo provider.
14818 	 */
14819 	(void) dtrace_register("dtrace", &dtrace_provider_attr,
14820 	    DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
14821 
14822 	ASSERT(dtrace_provider != NULL);
14823 	ASSERT((dtrace_provider_id_t)dtrace_provider == id);
14824 
14825 	dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
14826 	    dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
14827 	dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
14828 	    dtrace_provider, NULL, NULL, "END", 0, NULL);
14829 	dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
14830 	    dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
14831 
14832 	dtrace_anon_property();
14833 	mutex_exit(&cpu_lock);
14834 
14835 	/*
14836 	 * If DTrace helper tracing is enabled, we need to allocate the
14837 	 * trace buffer and initialize the values.
14838 	 */
14839 	if (dtrace_helptrace_enabled) {
14840 		ASSERT(dtrace_helptrace_buffer == NULL);
14841 		dtrace_helptrace_buffer =
14842 		    kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
14843 		dtrace_helptrace_next = 0;
14844 	}
14845 
14846 	/*
14847 	 * If there are already providers, we must ask them to provide their
14848 	 * probes, and then match any anonymous enabling against them.  Note
14849 	 * that there should be no other retained enablings at this time:
14850 	 * the only retained enablings at this time should be the anonymous
14851 	 * enabling.
14852 	 */
14853 	if (dtrace_anon.dta_enabling != NULL) {
14854 		ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
14855 
14856 		dtrace_enabling_provide(NULL);
14857 		state = dtrace_anon.dta_state;
14858 
14859 		/*
14860 		 * We couldn't hold cpu_lock across the above call to
14861 		 * dtrace_enabling_provide(), but we must hold it to actually
14862 		 * enable the probes.  We have to drop all of our locks, pick
14863 		 * up cpu_lock, and regain our locks before matching the
14864 		 * retained anonymous enabling.
14865 		 */
14866 		mutex_exit(&dtrace_lock);
14867 		mutex_exit(&dtrace_provider_lock);
14868 
14869 		mutex_enter(&cpu_lock);
14870 		mutex_enter(&dtrace_provider_lock);
14871 		mutex_enter(&dtrace_lock);
14872 
14873 		if ((enab = dtrace_anon.dta_enabling) != NULL)
14874 			(void) dtrace_enabling_match(enab, NULL);
14875 
14876 		mutex_exit(&cpu_lock);
14877 	}
14878 
14879 	mutex_exit(&dtrace_lock);
14880 	mutex_exit(&dtrace_provider_lock);
14881 
14882 	if (state != NULL) {
14883 		/*
14884 		 * If we created any anonymous state, set it going now.
14885 		 */
14886 		(void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
14887 	}
14888 
14889 	return (DDI_SUCCESS);
14890 }
14891 
14892 /*ARGSUSED*/
14893 static int
14894 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
14895 {
14896 	dtrace_state_t *state;
14897 	uint32_t priv;
14898 	uid_t uid;
14899 	zoneid_t zoneid;
14900 
14901 	if (getminor(*devp) == DTRACEMNRN_HELPER)
14902 		return (0);
14903 
14904 	/*
14905 	 * If this wasn't an open with the "helper" minor, then it must be
14906 	 * the "dtrace" minor.
14907 	 */
14908 	if (getminor(*devp) != DTRACEMNRN_DTRACE)
14909 		return (ENXIO);
14910 
14911 	/*
14912 	 * If no DTRACE_PRIV_* bits are set in the credential, then the
14913 	 * caller lacks sufficient permission to do anything with DTrace.
14914 	 */
14915 	dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
14916 	if (priv == DTRACE_PRIV_NONE)
14917 		return (EACCES);
14918 
14919 	/*
14920 	 * Ask all providers to provide all their probes.
14921 	 */
14922 	mutex_enter(&dtrace_provider_lock);
14923 	dtrace_probe_provide(NULL, NULL);
14924 	mutex_exit(&dtrace_provider_lock);
14925 
14926 	mutex_enter(&cpu_lock);
14927 	mutex_enter(&dtrace_lock);
14928 	dtrace_opens++;
14929 	dtrace_membar_producer();
14930 
14931 	/*
14932 	 * If the kernel debugger is active (that is, if the kernel debugger
14933 	 * modified text in some way), we won't allow the open.
14934 	 */
14935 	if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
14936 		dtrace_opens--;
14937 		mutex_exit(&cpu_lock);
14938 		mutex_exit(&dtrace_lock);
14939 		return (EBUSY);
14940 	}
14941 
14942 	state = dtrace_state_create(devp, cred_p);
14943 	mutex_exit(&cpu_lock);
14944 
14945 	if (state == NULL) {
14946 		if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
14947 			(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
14948 		mutex_exit(&dtrace_lock);
14949 		return (EAGAIN);
14950 	}
14951 
14952 	mutex_exit(&dtrace_lock);
14953 
14954 	return (0);
14955 }
14956 
14957 /*ARGSUSED*/
14958 static int
14959 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
14960 {
14961 	minor_t minor = getminor(dev);
14962 	dtrace_state_t *state;
14963 
14964 	if (minor == DTRACEMNRN_HELPER)
14965 		return (0);
14966 
14967 	state = ddi_get_soft_state(dtrace_softstate, minor);
14968 
14969 	mutex_enter(&cpu_lock);
14970 	mutex_enter(&dtrace_lock);
14971 
14972 	if (state->dts_anon) {
14973 		/*
14974 		 * There is anonymous state. Destroy that first.
14975 		 */
14976 		ASSERT(dtrace_anon.dta_state == NULL);
14977 		dtrace_state_destroy(state->dts_anon);
14978 	}
14979 
14980 	dtrace_state_destroy(state);
14981 	ASSERT(dtrace_opens > 0);
14982 
14983 	/*
14984 	 * Only relinquish control of the kernel debugger interface when there
14985 	 * are no consumers and no anonymous enablings.
14986 	 */
14987 	if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
14988 		(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
14989 
14990 	mutex_exit(&dtrace_lock);
14991 	mutex_exit(&cpu_lock);
14992 
14993 	return (0);
14994 }
14995 
14996 /*ARGSUSED*/
14997 static int
14998 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
14999 {
15000 	int rval;
15001 	dof_helper_t help, *dhp = NULL;
15002 
15003 	switch (cmd) {
15004 	case DTRACEHIOC_ADDDOF:
15005 		if (copyin((void *)arg, &help, sizeof (help)) != 0) {
15006 			dtrace_dof_error(NULL, "failed to copyin DOF helper");
15007 			return (EFAULT);
15008 		}
15009 
15010 		dhp = &help;
15011 		arg = (intptr_t)help.dofhp_dof;
15012 		/*FALLTHROUGH*/
15013 
15014 	case DTRACEHIOC_ADD: {
15015 		dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
15016 
15017 		if (dof == NULL)
15018 			return (rval);
15019 
15020 		mutex_enter(&dtrace_lock);
15021 
15022 		/*
15023 		 * dtrace_helper_slurp() takes responsibility for the dof --
15024 		 * it may free it now or it may save it and free it later.
15025 		 */
15026 		if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
15027 			*rv = rval;
15028 			rval = 0;
15029 		} else {
15030 			rval = EINVAL;
15031 		}
15032 
15033 		mutex_exit(&dtrace_lock);
15034 		return (rval);
15035 	}
15036 
15037 	case DTRACEHIOC_REMOVE: {
15038 		mutex_enter(&dtrace_lock);
15039 		rval = dtrace_helper_destroygen(arg);
15040 		mutex_exit(&dtrace_lock);
15041 
15042 		return (rval);
15043 	}
15044 
15045 	default:
15046 		break;
15047 	}
15048 
15049 	return (ENOTTY);
15050 }
15051 
15052 /*ARGSUSED*/
15053 static int
15054 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
15055 {
15056 	minor_t minor = getminor(dev);
15057 	dtrace_state_t *state;
15058 	int rval;
15059 
15060 	if (minor == DTRACEMNRN_HELPER)
15061 		return (dtrace_ioctl_helper(cmd, arg, rv));
15062 
15063 	state = ddi_get_soft_state(dtrace_softstate, minor);
15064 
15065 	if (state->dts_anon) {
15066 		ASSERT(dtrace_anon.dta_state == NULL);
15067 		state = state->dts_anon;
15068 	}
15069 
15070 	switch (cmd) {
15071 	case DTRACEIOC_PROVIDER: {
15072 		dtrace_providerdesc_t pvd;
15073 		dtrace_provider_t *pvp;
15074 
15075 		if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
15076 			return (EFAULT);
15077 
15078 		pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
15079 		mutex_enter(&dtrace_provider_lock);
15080 
15081 		for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
15082 			if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
15083 				break;
15084 		}
15085 
15086 		mutex_exit(&dtrace_provider_lock);
15087 
15088 		if (pvp == NULL)
15089 			return (ESRCH);
15090 
15091 		bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
15092 		bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
15093 		if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
15094 			return (EFAULT);
15095 
15096 		return (0);
15097 	}
15098 
15099 	case DTRACEIOC_EPROBE: {
15100 		dtrace_eprobedesc_t epdesc;
15101 		dtrace_ecb_t *ecb;
15102 		dtrace_action_t *act;
15103 		void *buf;
15104 		size_t size;
15105 		uintptr_t dest;
15106 		int nrecs;
15107 
15108 		if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
15109 			return (EFAULT);
15110 
15111 		mutex_enter(&dtrace_lock);
15112 
15113 		if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
15114 			mutex_exit(&dtrace_lock);
15115 			return (EINVAL);
15116 		}
15117 
15118 		if (ecb->dte_probe == NULL) {
15119 			mutex_exit(&dtrace_lock);
15120 			return (EINVAL);
15121 		}
15122 
15123 		epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
15124 		epdesc.dtepd_uarg = ecb->dte_uarg;
15125 		epdesc.dtepd_size = ecb->dte_size;
15126 
15127 		nrecs = epdesc.dtepd_nrecs;
15128 		epdesc.dtepd_nrecs = 0;
15129 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15130 			if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15131 				continue;
15132 
15133 			epdesc.dtepd_nrecs++;
15134 		}
15135 
15136 		/*
15137 		 * Now that we have the size, we need to allocate a temporary
15138 		 * buffer in which to store the complete description.  We need
15139 		 * the temporary buffer to be able to drop dtrace_lock()
15140 		 * across the copyout(), below.
15141 		 */
15142 		size = sizeof (dtrace_eprobedesc_t) +
15143 		    (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
15144 
15145 		buf = kmem_alloc(size, KM_SLEEP);
15146 		dest = (uintptr_t)buf;
15147 
15148 		bcopy(&epdesc, (void *)dest, sizeof (epdesc));
15149 		dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
15150 
15151 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15152 			if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15153 				continue;
15154 
15155 			if (nrecs-- == 0)
15156 				break;
15157 
15158 			bcopy(&act->dta_rec, (void *)dest,
15159 			    sizeof (dtrace_recdesc_t));
15160 			dest += sizeof (dtrace_recdesc_t);
15161 		}
15162 
15163 		mutex_exit(&dtrace_lock);
15164 
15165 		if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
15166 			kmem_free(buf, size);
15167 			return (EFAULT);
15168 		}
15169 
15170 		kmem_free(buf, size);
15171 		return (0);
15172 	}
15173 
15174 	case DTRACEIOC_AGGDESC: {
15175 		dtrace_aggdesc_t aggdesc;
15176 		dtrace_action_t *act;
15177 		dtrace_aggregation_t *agg;
15178 		int nrecs;
15179 		uint32_t offs;
15180 		dtrace_recdesc_t *lrec;
15181 		void *buf;
15182 		size_t size;
15183 		uintptr_t dest;
15184 
15185 		if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
15186 			return (EFAULT);
15187 
15188 		mutex_enter(&dtrace_lock);
15189 
15190 		if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
15191 			mutex_exit(&dtrace_lock);
15192 			return (EINVAL);
15193 		}
15194 
15195 		aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
15196 
15197 		nrecs = aggdesc.dtagd_nrecs;
15198 		aggdesc.dtagd_nrecs = 0;
15199 
15200 		offs = agg->dtag_base;
15201 		lrec = &agg->dtag_action.dta_rec;
15202 		aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
15203 
15204 		for (act = agg->dtag_first; ; act = act->dta_next) {
15205 			ASSERT(act->dta_intuple ||
15206 			    DTRACEACT_ISAGG(act->dta_kind));
15207 
15208 			/*
15209 			 * If this action has a record size of zero, it
15210 			 * denotes an argument to the aggregating action.
15211 			 * Because the presence of this record doesn't (or
15212 			 * shouldn't) affect the way the data is interpreted,
15213 			 * we don't copy it out to save user-level the
15214 			 * confusion of dealing with a zero-length record.
15215 			 */
15216 			if (act->dta_rec.dtrd_size == 0) {
15217 				ASSERT(agg->dtag_hasarg);
15218 				continue;
15219 			}
15220 
15221 			aggdesc.dtagd_nrecs++;
15222 
15223 			if (act == &agg->dtag_action)
15224 				break;
15225 		}
15226 
15227 		/*
15228 		 * Now that we have the size, we need to allocate a temporary
15229 		 * buffer in which to store the complete description.  We need
15230 		 * the temporary buffer to be able to drop dtrace_lock()
15231 		 * across the copyout(), below.
15232 		 */
15233 		size = sizeof (dtrace_aggdesc_t) +
15234 		    (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
15235 
15236 		buf = kmem_alloc(size, KM_SLEEP);
15237 		dest = (uintptr_t)buf;
15238 
15239 		bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
15240 		dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
15241 
15242 		for (act = agg->dtag_first; ; act = act->dta_next) {
15243 			dtrace_recdesc_t rec = act->dta_rec;
15244 
15245 			/*
15246 			 * See the comment in the above loop for why we pass
15247 			 * over zero-length records.
15248 			 */
15249 			if (rec.dtrd_size == 0) {
15250 				ASSERT(agg->dtag_hasarg);
15251 				continue;
15252 			}
15253 
15254 			if (nrecs-- == 0)
15255 				break;
15256 
15257 			rec.dtrd_offset -= offs;
15258 			bcopy(&rec, (void *)dest, sizeof (rec));
15259 			dest += sizeof (dtrace_recdesc_t);
15260 
15261 			if (act == &agg->dtag_action)
15262 				break;
15263 		}
15264 
15265 		mutex_exit(&dtrace_lock);
15266 
15267 		if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
15268 			kmem_free(buf, size);
15269 			return (EFAULT);
15270 		}
15271 
15272 		kmem_free(buf, size);
15273 		return (0);
15274 	}
15275 
15276 	case DTRACEIOC_ENABLE: {
15277 		dof_hdr_t *dof;
15278 		dtrace_enabling_t *enab = NULL;
15279 		dtrace_vstate_t *vstate;
15280 		int err = 0;
15281 
15282 		*rv = 0;
15283 
15284 		/*
15285 		 * If a NULL argument has been passed, we take this as our
15286 		 * cue to reevaluate our enablings.
15287 		 */
15288 		if (arg == NULL) {
15289 			dtrace_enabling_matchall();
15290 
15291 			return (0);
15292 		}
15293 
15294 		if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
15295 			return (rval);
15296 
15297 		mutex_enter(&cpu_lock);
15298 		mutex_enter(&dtrace_lock);
15299 		vstate = &state->dts_vstate;
15300 
15301 		if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
15302 			mutex_exit(&dtrace_lock);
15303 			mutex_exit(&cpu_lock);
15304 			dtrace_dof_destroy(dof);
15305 			return (EBUSY);
15306 		}
15307 
15308 		if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
15309 			mutex_exit(&dtrace_lock);
15310 			mutex_exit(&cpu_lock);
15311 			dtrace_dof_destroy(dof);
15312 			return (EINVAL);
15313 		}
15314 
15315 		if ((rval = dtrace_dof_options(dof, state)) != 0) {
15316 			dtrace_enabling_destroy(enab);
15317 			mutex_exit(&dtrace_lock);
15318 			mutex_exit(&cpu_lock);
15319 			dtrace_dof_destroy(dof);
15320 			return (rval);
15321 		}
15322 
15323 		if ((err = dtrace_enabling_match(enab, rv)) == 0) {
15324 			err = dtrace_enabling_retain(enab);
15325 		} else {
15326 			dtrace_enabling_destroy(enab);
15327 		}
15328 
15329 		mutex_exit(&cpu_lock);
15330 		mutex_exit(&dtrace_lock);
15331 		dtrace_dof_destroy(dof);
15332 
15333 		return (err);
15334 	}
15335 
15336 	case DTRACEIOC_REPLICATE: {
15337 		dtrace_repldesc_t desc;
15338 		dtrace_probedesc_t *match = &desc.dtrpd_match;
15339 		dtrace_probedesc_t *create = &desc.dtrpd_create;
15340 		int err;
15341 
15342 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15343 			return (EFAULT);
15344 
15345 		match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
15346 		match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
15347 		match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
15348 		match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
15349 
15350 		create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
15351 		create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
15352 		create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
15353 		create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
15354 
15355 		mutex_enter(&dtrace_lock);
15356 		err = dtrace_enabling_replicate(state, match, create);
15357 		mutex_exit(&dtrace_lock);
15358 
15359 		return (err);
15360 	}
15361 
15362 	case DTRACEIOC_PROBEMATCH:
15363 	case DTRACEIOC_PROBES: {
15364 		dtrace_probe_t *probe = NULL;
15365 		dtrace_probedesc_t desc;
15366 		dtrace_probekey_t pkey;
15367 		dtrace_id_t i;
15368 		int m = 0;
15369 		uint32_t priv;
15370 		uid_t uid;
15371 		zoneid_t zoneid;
15372 
15373 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15374 			return (EFAULT);
15375 
15376 		desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
15377 		desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
15378 		desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
15379 		desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
15380 
15381 		/*
15382 		 * Before we attempt to match this probe, we want to give
15383 		 * all providers the opportunity to provide it.
15384 		 */
15385 		if (desc.dtpd_id == DTRACE_IDNONE) {
15386 			mutex_enter(&dtrace_provider_lock);
15387 			dtrace_probe_provide(&desc, NULL);
15388 			mutex_exit(&dtrace_provider_lock);
15389 			desc.dtpd_id++;
15390 		}
15391 
15392 		if (cmd == DTRACEIOC_PROBEMATCH)  {
15393 			dtrace_probekey(&desc, &pkey);
15394 			pkey.dtpk_id = DTRACE_IDNONE;
15395 		}
15396 
15397 		dtrace_cred2priv(cr, &priv, &uid, &zoneid);
15398 
15399 		mutex_enter(&dtrace_lock);
15400 
15401 		if (cmd == DTRACEIOC_PROBEMATCH) {
15402 			for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
15403 				if ((probe = dtrace_probes[i - 1]) != NULL &&
15404 				    (m = dtrace_match_probe(probe, &pkey,
15405 				    priv, uid, zoneid)) != 0)
15406 					break;
15407 			}
15408 
15409 			if (m < 0) {
15410 				mutex_exit(&dtrace_lock);
15411 				return (EINVAL);
15412 			}
15413 
15414 		} else {
15415 			for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
15416 				if ((probe = dtrace_probes[i - 1]) != NULL &&
15417 				    dtrace_match_priv(probe, priv, uid, zoneid))
15418 					break;
15419 			}
15420 		}
15421 
15422 		if (probe == NULL) {
15423 			mutex_exit(&dtrace_lock);
15424 			return (ESRCH);
15425 		}
15426 
15427 		dtrace_probe_description(probe, &desc);
15428 		mutex_exit(&dtrace_lock);
15429 
15430 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15431 			return (EFAULT);
15432 
15433 		return (0);
15434 	}
15435 
15436 	case DTRACEIOC_PROBEARG: {
15437 		dtrace_argdesc_t desc;
15438 		dtrace_probe_t *probe;
15439 		dtrace_provider_t *prov;
15440 
15441 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15442 			return (EFAULT);
15443 
15444 		if (desc.dtargd_id == DTRACE_IDNONE)
15445 			return (EINVAL);
15446 
15447 		if (desc.dtargd_ndx == DTRACE_ARGNONE)
15448 			return (EINVAL);
15449 
15450 		mutex_enter(&dtrace_provider_lock);
15451 		mutex_enter(&mod_lock);
15452 		mutex_enter(&dtrace_lock);
15453 
15454 		if (desc.dtargd_id > dtrace_nprobes) {
15455 			mutex_exit(&dtrace_lock);
15456 			mutex_exit(&mod_lock);
15457 			mutex_exit(&dtrace_provider_lock);
15458 			return (EINVAL);
15459 		}
15460 
15461 		if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
15462 			mutex_exit(&dtrace_lock);
15463 			mutex_exit(&mod_lock);
15464 			mutex_exit(&dtrace_provider_lock);
15465 			return (EINVAL);
15466 		}
15467 
15468 		mutex_exit(&dtrace_lock);
15469 
15470 		prov = probe->dtpr_provider;
15471 
15472 		if (prov->dtpv_pops.dtps_getargdesc == NULL) {
15473 			/*
15474 			 * There isn't any typed information for this probe.
15475 			 * Set the argument number to DTRACE_ARGNONE.
15476 			 */
15477 			desc.dtargd_ndx = DTRACE_ARGNONE;
15478 		} else {
15479 			desc.dtargd_native[0] = '\0';
15480 			desc.dtargd_xlate[0] = '\0';
15481 			desc.dtargd_mapping = desc.dtargd_ndx;
15482 
15483 			prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
15484 			    probe->dtpr_id, probe->dtpr_arg, &desc);
15485 		}
15486 
15487 		mutex_exit(&mod_lock);
15488 		mutex_exit(&dtrace_provider_lock);
15489 
15490 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15491 			return (EFAULT);
15492 
15493 		return (0);
15494 	}
15495 
15496 	case DTRACEIOC_GO: {
15497 		processorid_t cpuid;
15498 		rval = dtrace_state_go(state, &cpuid);
15499 
15500 		if (rval != 0)
15501 			return (rval);
15502 
15503 		if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
15504 			return (EFAULT);
15505 
15506 		return (0);
15507 	}
15508 
15509 	case DTRACEIOC_STOP: {
15510 		processorid_t cpuid;
15511 
15512 		mutex_enter(&dtrace_lock);
15513 		rval = dtrace_state_stop(state, &cpuid);
15514 		mutex_exit(&dtrace_lock);
15515 
15516 		if (rval != 0)
15517 			return (rval);
15518 
15519 		if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
15520 			return (EFAULT);
15521 
15522 		return (0);
15523 	}
15524 
15525 	case DTRACEIOC_DOFGET: {
15526 		dof_hdr_t hdr, *dof;
15527 		uint64_t len;
15528 
15529 		if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
15530 			return (EFAULT);
15531 
15532 		mutex_enter(&dtrace_lock);
15533 		dof = dtrace_dof_create(state);
15534 		mutex_exit(&dtrace_lock);
15535 
15536 		len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
15537 		rval = copyout(dof, (void *)arg, len);
15538 		dtrace_dof_destroy(dof);
15539 
15540 		return (rval == 0 ? 0 : EFAULT);
15541 	}
15542 
15543 	case DTRACEIOC_AGGSNAP:
15544 	case DTRACEIOC_BUFSNAP: {
15545 		dtrace_bufdesc_t desc;
15546 		caddr_t cached;
15547 		dtrace_buffer_t *buf;
15548 
15549 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15550 			return (EFAULT);
15551 
15552 		if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
15553 			return (EINVAL);
15554 
15555 		mutex_enter(&dtrace_lock);
15556 
15557 		if (cmd == DTRACEIOC_BUFSNAP) {
15558 			buf = &state->dts_buffer[desc.dtbd_cpu];
15559 		} else {
15560 			buf = &state->dts_aggbuffer[desc.dtbd_cpu];
15561 		}
15562 
15563 		if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
15564 			size_t sz = buf->dtb_offset;
15565 
15566 			if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
15567 				mutex_exit(&dtrace_lock);
15568 				return (EBUSY);
15569 			}
15570 
15571 			/*
15572 			 * If this buffer has already been consumed, we're
15573 			 * going to indicate that there's nothing left here
15574 			 * to consume.
15575 			 */
15576 			if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
15577 				mutex_exit(&dtrace_lock);
15578 
15579 				desc.dtbd_size = 0;
15580 				desc.dtbd_drops = 0;
15581 				desc.dtbd_errors = 0;
15582 				desc.dtbd_oldest = 0;
15583 				sz = sizeof (desc);
15584 
15585 				if (copyout(&desc, (void *)arg, sz) != 0)
15586 					return (EFAULT);
15587 
15588 				return (0);
15589 			}
15590 
15591 			/*
15592 			 * If this is a ring buffer that has wrapped, we want
15593 			 * to copy the whole thing out.
15594 			 */
15595 			if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
15596 				dtrace_buffer_polish(buf);
15597 				sz = buf->dtb_size;
15598 			}
15599 
15600 			if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
15601 				mutex_exit(&dtrace_lock);
15602 				return (EFAULT);
15603 			}
15604 
15605 			desc.dtbd_size = sz;
15606 			desc.dtbd_drops = buf->dtb_drops;
15607 			desc.dtbd_errors = buf->dtb_errors;
15608 			desc.dtbd_oldest = buf->dtb_xamot_offset;
15609 
15610 			mutex_exit(&dtrace_lock);
15611 
15612 			if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15613 				return (EFAULT);
15614 
15615 			buf->dtb_flags |= DTRACEBUF_CONSUMED;
15616 
15617 			return (0);
15618 		}
15619 
15620 		if (buf->dtb_tomax == NULL) {
15621 			ASSERT(buf->dtb_xamot == NULL);
15622 			mutex_exit(&dtrace_lock);
15623 			return (ENOENT);
15624 		}
15625 
15626 		cached = buf->dtb_tomax;
15627 		ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
15628 
15629 		dtrace_xcall(desc.dtbd_cpu,
15630 		    (dtrace_xcall_t)dtrace_buffer_switch, buf);
15631 
15632 		state->dts_errors += buf->dtb_xamot_errors;
15633 
15634 		/*
15635 		 * If the buffers did not actually switch, then the cross call
15636 		 * did not take place -- presumably because the given CPU is
15637 		 * not in the ready set.  If this is the case, we'll return
15638 		 * ENOENT.
15639 		 */
15640 		if (buf->dtb_tomax == cached) {
15641 			ASSERT(buf->dtb_xamot != cached);
15642 			mutex_exit(&dtrace_lock);
15643 			return (ENOENT);
15644 		}
15645 
15646 		ASSERT(cached == buf->dtb_xamot);
15647 
15648 		/*
15649 		 * We have our snapshot; now copy it out.
15650 		 */
15651 		if (copyout(buf->dtb_xamot, desc.dtbd_data,
15652 		    buf->dtb_xamot_offset) != 0) {
15653 			mutex_exit(&dtrace_lock);
15654 			return (EFAULT);
15655 		}
15656 
15657 		desc.dtbd_size = buf->dtb_xamot_offset;
15658 		desc.dtbd_drops = buf->dtb_xamot_drops;
15659 		desc.dtbd_errors = buf->dtb_xamot_errors;
15660 		desc.dtbd_oldest = 0;
15661 
15662 		mutex_exit(&dtrace_lock);
15663 
15664 		/*
15665 		 * Finally, copy out the buffer description.
15666 		 */
15667 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15668 			return (EFAULT);
15669 
15670 		return (0);
15671 	}
15672 
15673 	case DTRACEIOC_CONF: {
15674 		dtrace_conf_t conf;
15675 
15676 		bzero(&conf, sizeof (conf));
15677 		conf.dtc_difversion = DIF_VERSION;
15678 		conf.dtc_difintregs = DIF_DIR_NREGS;
15679 		conf.dtc_diftupregs = DIF_DTR_NREGS;
15680 		conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
15681 
15682 		if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
15683 			return (EFAULT);
15684 
15685 		return (0);
15686 	}
15687 
15688 	case DTRACEIOC_STATUS: {
15689 		dtrace_status_t stat;
15690 		dtrace_dstate_t *dstate;
15691 		int i, j;
15692 		uint64_t nerrs;
15693 
15694 		/*
15695 		 * See the comment in dtrace_state_deadman() for the reason
15696 		 * for setting dts_laststatus to INT64_MAX before setting
15697 		 * it to the correct value.
15698 		 */
15699 		state->dts_laststatus = INT64_MAX;
15700 		dtrace_membar_producer();
15701 		state->dts_laststatus = dtrace_gethrtime();
15702 
15703 		bzero(&stat, sizeof (stat));
15704 
15705 		mutex_enter(&dtrace_lock);
15706 
15707 		if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
15708 			mutex_exit(&dtrace_lock);
15709 			return (ENOENT);
15710 		}
15711 
15712 		if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
15713 			stat.dtst_exiting = 1;
15714 
15715 		nerrs = state->dts_errors;
15716 		dstate = &state->dts_vstate.dtvs_dynvars;
15717 
15718 		for (i = 0; i < NCPU; i++) {
15719 			dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
15720 
15721 			stat.dtst_dyndrops += dcpu->dtdsc_drops;
15722 			stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
15723 			stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
15724 
15725 			if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
15726 				stat.dtst_filled++;
15727 
15728 			nerrs += state->dts_buffer[i].dtb_errors;
15729 
15730 			for (j = 0; j < state->dts_nspeculations; j++) {
15731 				dtrace_speculation_t *spec;
15732 				dtrace_buffer_t *buf;
15733 
15734 				spec = &state->dts_speculations[j];
15735 				buf = &spec->dtsp_buffer[i];
15736 				stat.dtst_specdrops += buf->dtb_xamot_drops;
15737 			}
15738 		}
15739 
15740 		stat.dtst_specdrops_busy = state->dts_speculations_busy;
15741 		stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
15742 		stat.dtst_stkstroverflows = state->dts_stkstroverflows;
15743 		stat.dtst_dblerrors = state->dts_dblerrors;
15744 		stat.dtst_killed =
15745 		    (state->dts_activity == DTRACE_ACTIVITY_KILLED);
15746 		stat.dtst_errors = nerrs;
15747 
15748 		mutex_exit(&dtrace_lock);
15749 
15750 		if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
15751 			return (EFAULT);
15752 
15753 		return (0);
15754 	}
15755 
15756 	case DTRACEIOC_FORMAT: {
15757 		dtrace_fmtdesc_t fmt;
15758 		char *str;
15759 		int len;
15760 
15761 		if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
15762 			return (EFAULT);
15763 
15764 		mutex_enter(&dtrace_lock);
15765 
15766 		if (fmt.dtfd_format == 0 ||
15767 		    fmt.dtfd_format > state->dts_nformats) {
15768 			mutex_exit(&dtrace_lock);
15769 			return (EINVAL);
15770 		}
15771 
15772 		/*
15773 		 * Format strings are allocated contiguously and they are
15774 		 * never freed; if a format index is less than the number
15775 		 * of formats, we can assert that the format map is non-NULL
15776 		 * and that the format for the specified index is non-NULL.
15777 		 */
15778 		ASSERT(state->dts_formats != NULL);
15779 		str = state->dts_formats[fmt.dtfd_format - 1];
15780 		ASSERT(str != NULL);
15781 
15782 		len = strlen(str) + 1;
15783 
15784 		if (len > fmt.dtfd_length) {
15785 			fmt.dtfd_length = len;
15786 
15787 			if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
15788 				mutex_exit(&dtrace_lock);
15789 				return (EINVAL);
15790 			}
15791 		} else {
15792 			if (copyout(str, fmt.dtfd_string, len) != 0) {
15793 				mutex_exit(&dtrace_lock);
15794 				return (EINVAL);
15795 			}
15796 		}
15797 
15798 		mutex_exit(&dtrace_lock);
15799 		return (0);
15800 	}
15801 
15802 	default:
15803 		break;
15804 	}
15805 
15806 	return (ENOTTY);
15807 }
15808 
15809 /*ARGSUSED*/
15810 static int
15811 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
15812 {
15813 	dtrace_state_t *state;
15814 
15815 	switch (cmd) {
15816 	case DDI_DETACH:
15817 		break;
15818 
15819 	case DDI_SUSPEND:
15820 		return (DDI_SUCCESS);
15821 
15822 	default:
15823 		return (DDI_FAILURE);
15824 	}
15825 
15826 	mutex_enter(&cpu_lock);
15827 	mutex_enter(&dtrace_provider_lock);
15828 	mutex_enter(&dtrace_lock);
15829 
15830 	ASSERT(dtrace_opens == 0);
15831 
15832 	if (dtrace_helpers > 0) {
15833 		mutex_exit(&dtrace_provider_lock);
15834 		mutex_exit(&dtrace_lock);
15835 		mutex_exit(&cpu_lock);
15836 		return (DDI_FAILURE);
15837 	}
15838 
15839 	if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
15840 		mutex_exit(&dtrace_provider_lock);
15841 		mutex_exit(&dtrace_lock);
15842 		mutex_exit(&cpu_lock);
15843 		return (DDI_FAILURE);
15844 	}
15845 
15846 	dtrace_provider = NULL;
15847 
15848 	if ((state = dtrace_anon_grab()) != NULL) {
15849 		/*
15850 		 * If there were ECBs on this state, the provider should
15851 		 * have not been allowed to detach; assert that there is
15852 		 * none.
15853 		 */
15854 		ASSERT(state->dts_necbs == 0);
15855 		dtrace_state_destroy(state);
15856 
15857 		/*
15858 		 * If we're being detached with anonymous state, we need to
15859 		 * indicate to the kernel debugger that DTrace is now inactive.
15860 		 */
15861 		(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15862 	}
15863 
15864 	bzero(&dtrace_anon, sizeof (dtrace_anon_t));
15865 	unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
15866 	dtrace_cpu_init = NULL;
15867 	dtrace_helpers_cleanup = NULL;
15868 	dtrace_helpers_fork = NULL;
15869 	dtrace_cpustart_init = NULL;
15870 	dtrace_cpustart_fini = NULL;
15871 	dtrace_debugger_init = NULL;
15872 	dtrace_debugger_fini = NULL;
15873 	dtrace_modload = NULL;
15874 	dtrace_modunload = NULL;
15875 
15876 	mutex_exit(&cpu_lock);
15877 
15878 	if (dtrace_helptrace_enabled) {
15879 		kmem_free(dtrace_helptrace_buffer, dtrace_helptrace_bufsize);
15880 		dtrace_helptrace_buffer = NULL;
15881 	}
15882 
15883 	kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
15884 	dtrace_probes = NULL;
15885 	dtrace_nprobes = 0;
15886 
15887 	dtrace_hash_destroy(dtrace_bymod);
15888 	dtrace_hash_destroy(dtrace_byfunc);
15889 	dtrace_hash_destroy(dtrace_byname);
15890 	dtrace_bymod = NULL;
15891 	dtrace_byfunc = NULL;
15892 	dtrace_byname = NULL;
15893 
15894 	kmem_cache_destroy(dtrace_state_cache);
15895 	vmem_destroy(dtrace_minor);
15896 	vmem_destroy(dtrace_arena);
15897 
15898 	if (dtrace_toxrange != NULL) {
15899 		kmem_free(dtrace_toxrange,
15900 		    dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
15901 		dtrace_toxrange = NULL;
15902 		dtrace_toxranges = 0;
15903 		dtrace_toxranges_max = 0;
15904 	}
15905 
15906 	ddi_remove_minor_node(dtrace_devi, NULL);
15907 	dtrace_devi = NULL;
15908 
15909 	ddi_soft_state_fini(&dtrace_softstate);
15910 
15911 	ASSERT(dtrace_vtime_references == 0);
15912 	ASSERT(dtrace_opens == 0);
15913 	ASSERT(dtrace_retained == NULL);
15914 
15915 	mutex_exit(&dtrace_lock);
15916 	mutex_exit(&dtrace_provider_lock);
15917 
15918 	/*
15919 	 * We don't destroy the task queue until after we have dropped our
15920 	 * locks (taskq_destroy() may block on running tasks).  To prevent
15921 	 * attempting to do work after we have effectively detached but before
15922 	 * the task queue has been destroyed, all tasks dispatched via the
15923 	 * task queue must check that DTrace is still attached before
15924 	 * performing any operation.
15925 	 */
15926 	taskq_destroy(dtrace_taskq);
15927 	dtrace_taskq = NULL;
15928 
15929 	return (DDI_SUCCESS);
15930 }
15931 
15932 /*ARGSUSED*/
15933 static int
15934 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
15935 {
15936 	int error;
15937 
15938 	switch (infocmd) {
15939 	case DDI_INFO_DEVT2DEVINFO:
15940 		*result = (void *)dtrace_devi;
15941 		error = DDI_SUCCESS;
15942 		break;
15943 	case DDI_INFO_DEVT2INSTANCE:
15944 		*result = (void *)0;
15945 		error = DDI_SUCCESS;
15946 		break;
15947 	default:
15948 		error = DDI_FAILURE;
15949 	}
15950 	return (error);
15951 }
15952 
15953 static struct cb_ops dtrace_cb_ops = {
15954 	dtrace_open,		/* open */
15955 	dtrace_close,		/* close */
15956 	nulldev,		/* strategy */
15957 	nulldev,		/* print */
15958 	nodev,			/* dump */
15959 	nodev,			/* read */
15960 	nodev,			/* write */
15961 	dtrace_ioctl,		/* ioctl */
15962 	nodev,			/* devmap */
15963 	nodev,			/* mmap */
15964 	nodev,			/* segmap */
15965 	nochpoll,		/* poll */
15966 	ddi_prop_op,		/* cb_prop_op */
15967 	0,			/* streamtab  */
15968 	D_NEW | D_MP		/* Driver compatibility flag */
15969 };
15970 
15971 static struct dev_ops dtrace_ops = {
15972 	DEVO_REV,		/* devo_rev */
15973 	0,			/* refcnt */
15974 	dtrace_info,		/* get_dev_info */
15975 	nulldev,		/* identify */
15976 	nulldev,		/* probe */
15977 	dtrace_attach,		/* attach */
15978 	dtrace_detach,		/* detach */
15979 	nodev,			/* reset */
15980 	&dtrace_cb_ops,		/* driver operations */
15981 	NULL,			/* bus operations */
15982 	nodev,			/* dev power */
15983 	ddi_quiesce_not_needed,		/* quiesce */
15984 };
15985 
15986 static struct modldrv modldrv = {
15987 	&mod_driverops,		/* module type (this is a pseudo driver) */
15988 	"Dynamic Tracing",	/* name of module */
15989 	&dtrace_ops,		/* driver ops */
15990 };
15991 
15992 static struct modlinkage modlinkage = {
15993 	MODREV_1,
15994 	(void *)&modldrv,
15995 	NULL
15996 };
15997 
15998 int
15999 _init(void)
16000 {
16001 	return (mod_install(&modlinkage));
16002 }
16003 
16004 int
16005 _info(struct modinfo *modinfop)
16006 {
16007 	return (mod_info(&modlinkage, modinfop));
16008 }
16009 
16010 int
16011 _fini(void)
16012 {
16013 	return (mod_remove(&modlinkage));
16014 }
16015