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.
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.
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]
25 * Copyright 2008 Sun Microsystems, Inc. All rights reserved.
26 * Use is subject to license terms.
29 #pragma ident "%Z%%M% %I% %E% SMI"
32 * DTrace - Dynamic Tracing for Solaris
34 * This is the implementation of the Solaris Dynamic Tracing framework
35 * (DTrace). The user-visible interface to DTrace is described at length in
36 * the "Solaris Dynamic Tracing Guide". The interfaces between the libdtrace
37 * library, the in-kernel DTrace framework, and the DTrace providers are
38 * described in the block comments in the <sys/dtrace.h> header file. The
39 * internal architecture of DTrace is described in the block comments in the
40 * <sys/dtrace_impl.h> header file. The comments contained within the DTrace
41 * implementation very much assume mastery of all of these sources; if one has
42 * an unanswered question about the implementation, one should consult them
45 * The functions here are ordered roughly as follows:
47 * - Probe context functions
48 * - Probe hashing functions
49 * - Non-probe context utility functions
50 * - Matching functions
51 * - Provider-to-Framework API functions
52 * - Probe management functions
53 * - DIF object functions
55 * - Predicate functions
58 * - Enabling functions
60 * - Anonymous enabling functions
61 * - Consumer state functions
64 * - Driver cookbook functions
66 * Each group of functions begins with a block comment labelled the "DTrace
67 * [Group] Functions", allowing one to find each block by searching forward
68 * on capital-f functions.
70 #include <sys/errno.h>
75 #include <sys/modctl.h>
77 #include <sys/systm.h>
80 #include <sys/sunddi.h>
82 #include <sys/cpuvar.h>
85 #include <sys/strsubr.h>
87 #include <sys/sysmacros.h>
88 #include <sys/dtrace_impl.h>
89 #include <sys/atomic.h>
90 #include <sys/cmn_err.h>
92 #include <sys/mutex_impl.h>
93 #include <sys/rwlock_impl.h>
95 #include <sys/ctf_api.h>
97 #include <sys/panic.h>
98 #include <sys/priv_impl.h>
100 #include <sys/policy.h>
102 #include <sys/cred_impl.h>
103 #include <sys/procfs_isa.h>
105 #include <sys/taskq.h>
107 #include <sys/mkdev.h>
110 #include <sys/zone.h>
111 #include <sys/socket.h>
112 #include <netinet/in.h>
114 /* FreeBSD includes: */
116 #include <sys/callout.h>
117 #include <sys/ctype.h>
118 #include <sys/limits.h>
120 #include <sys/kernel.h>
121 #include <sys/malloc.h>
122 #include <sys/sysctl.h>
123 #include <sys/lock.h>
124 #include <sys/mutex.h>
125 #include <sys/rwlock.h>
127 #include <sys/dtrace_bsd.h>
128 #include <netinet/in.h>
129 #include "dtrace_cddl.h"
130 #include "dtrace_debug.c"
134 * DTrace Tunable Variables
136 * The following variables may be tuned by adding a line to /etc/system that
137 * includes both the name of the DTrace module ("dtrace") and the name of the
138 * variable. For example:
140 * set dtrace:dtrace_destructive_disallow = 1
142 * In general, the only variables that one should be tuning this way are those
143 * that affect system-wide DTrace behavior, and for which the default behavior
144 * is undesirable. Most of these variables are tunable on a per-consumer
145 * basis using DTrace options, and need not be tuned on a system-wide basis.
146 * When tuning these variables, avoid pathological values; while some attempt
147 * is made to verify the integrity of these variables, they are not considered
148 * part of the supported interface to DTrace, and they are therefore not
149 * checked comprehensively. Further, these variables should not be tuned
150 * dynamically via "mdb -kw" or other means; they should only be tuned via
153 int dtrace_destructive_disallow = 0;
154 dtrace_optval_t dtrace_nonroot_maxsize = (16 * 1024 * 1024);
155 size_t dtrace_difo_maxsize = (256 * 1024);
156 dtrace_optval_t dtrace_dof_maxsize = (256 * 1024);
157 size_t dtrace_global_maxsize = (16 * 1024);
158 size_t dtrace_actions_max = (16 * 1024);
159 size_t dtrace_retain_max = 1024;
160 dtrace_optval_t dtrace_helper_actions_max = 32;
161 dtrace_optval_t dtrace_helper_providers_max = 32;
162 dtrace_optval_t dtrace_dstate_defsize = (1 * 1024 * 1024);
163 size_t dtrace_strsize_default = 256;
164 dtrace_optval_t dtrace_cleanrate_default = 9900990; /* 101 hz */
165 dtrace_optval_t dtrace_cleanrate_min = 200000; /* 5000 hz */
166 dtrace_optval_t dtrace_cleanrate_max = (uint64_t)60 * NANOSEC; /* 1/minute */
167 dtrace_optval_t dtrace_aggrate_default = NANOSEC; /* 1 hz */
168 dtrace_optval_t dtrace_statusrate_default = NANOSEC; /* 1 hz */
169 dtrace_optval_t dtrace_statusrate_max = (hrtime_t)10 * NANOSEC; /* 6/minute */
170 dtrace_optval_t dtrace_switchrate_default = NANOSEC; /* 1 hz */
171 dtrace_optval_t dtrace_nspec_default = 1;
172 dtrace_optval_t dtrace_specsize_default = 32 * 1024;
173 dtrace_optval_t dtrace_stackframes_default = 20;
174 dtrace_optval_t dtrace_ustackframes_default = 20;
175 dtrace_optval_t dtrace_jstackframes_default = 50;
176 dtrace_optval_t dtrace_jstackstrsize_default = 512;
177 int dtrace_msgdsize_max = 128;
178 hrtime_t dtrace_chill_max = 500 * (NANOSEC / MILLISEC); /* 500 ms */
179 hrtime_t dtrace_chill_interval = NANOSEC; /* 1000 ms */
180 int dtrace_devdepth_max = 32;
181 int dtrace_err_verbose;
182 hrtime_t dtrace_deadman_interval = NANOSEC;
183 hrtime_t dtrace_deadman_timeout = (hrtime_t)10 * NANOSEC;
184 hrtime_t dtrace_deadman_user = (hrtime_t)30 * NANOSEC;
187 * DTrace External Variables
189 * As dtrace(7D) is a kernel module, any DTrace variables are obviously
190 * available to DTrace consumers via the backtick (`) syntax. One of these,
191 * dtrace_zero, is made deliberately so: it is provided as a source of
192 * well-known, zero-filled memory. While this variable is not documented,
193 * it is used by some translators as an implementation detail.
195 const char dtrace_zero[256] = { 0 }; /* zero-filled memory */
198 * DTrace Internal Variables
201 static dev_info_t *dtrace_devi; /* device info */
204 static vmem_t *dtrace_arena; /* probe ID arena */
205 static vmem_t *dtrace_minor; /* minor number arena */
206 static taskq_t *dtrace_taskq; /* task queue */
208 static struct unrhdr *dtrace_arena; /* Probe ID number. */
210 static dtrace_probe_t **dtrace_probes; /* array of all probes */
211 static int dtrace_nprobes; /* number of probes */
212 static dtrace_provider_t *dtrace_provider; /* provider list */
213 static dtrace_meta_t *dtrace_meta_pid; /* user-land meta provider */
214 static int dtrace_opens; /* number of opens */
215 static int dtrace_helpers; /* number of helpers */
217 static void *dtrace_softstate; /* softstate pointer */
219 static dtrace_hash_t *dtrace_bymod; /* probes hashed by module */
220 static dtrace_hash_t *dtrace_byfunc; /* probes hashed by function */
221 static dtrace_hash_t *dtrace_byname; /* probes hashed by name */
222 static dtrace_toxrange_t *dtrace_toxrange; /* toxic range array */
223 static int dtrace_toxranges; /* number of toxic ranges */
224 static int dtrace_toxranges_max; /* size of toxic range array */
225 static dtrace_anon_t dtrace_anon; /* anonymous enabling */
226 static kmem_cache_t *dtrace_state_cache; /* cache for dynamic state */
227 static uint64_t dtrace_vtime_references; /* number of vtimestamp refs */
228 static kthread_t *dtrace_panicked; /* panicking thread */
229 static dtrace_ecb_t *dtrace_ecb_create_cache; /* cached created ECB */
230 static dtrace_genid_t dtrace_probegen; /* current probe generation */
231 static dtrace_helpers_t *dtrace_deferred_pid; /* deferred helper list */
232 static dtrace_enabling_t *dtrace_retained; /* list of retained enablings */
233 static dtrace_dynvar_t dtrace_dynhash_sink; /* end of dynamic hash chains */
235 static struct mtx dtrace_unr_mtx;
236 MTX_SYSINIT(dtrace_unr_mtx, &dtrace_unr_mtx, "Unique resource identifier", MTX_DEF);
237 int dtrace_in_probe; /* non-zero if executing a probe */
238 #if defined(__i386__) || defined(__amd64__)
239 uintptr_t dtrace_in_probe_addr; /* Address of invop when already in probe */
245 * DTrace is protected by three (relatively coarse-grained) locks:
247 * (1) dtrace_lock is required to manipulate essentially any DTrace state,
248 * including enabling state, probes, ECBs, consumer state, helper state,
249 * etc. Importantly, dtrace_lock is _not_ required when in probe context;
250 * probe context is lock-free -- synchronization is handled via the
251 * dtrace_sync() cross call mechanism.
253 * (2) dtrace_provider_lock is required when manipulating provider state, or
254 * when provider state must be held constant.
256 * (3) dtrace_meta_lock is required when manipulating meta provider state, or
257 * when meta provider state must be held constant.
259 * The lock ordering between these three locks is dtrace_meta_lock before
260 * dtrace_provider_lock before dtrace_lock. (In particular, there are
261 * several places where dtrace_provider_lock is held by the framework as it
262 * calls into the providers -- which then call back into the framework,
263 * grabbing dtrace_lock.)
265 * There are two other locks in the mix: mod_lock and cpu_lock. With respect
266 * to dtrace_provider_lock and dtrace_lock, cpu_lock continues its historical
267 * role as a coarse-grained lock; it is acquired before both of these locks.
268 * With respect to dtrace_meta_lock, its behavior is stranger: cpu_lock must
269 * be acquired _between_ dtrace_meta_lock and any other DTrace locks.
270 * mod_lock is similar with respect to dtrace_provider_lock in that it must be
271 * acquired _between_ dtrace_provider_lock and dtrace_lock.
273 static kmutex_t dtrace_lock; /* probe state lock */
274 static kmutex_t dtrace_provider_lock; /* provider state lock */
275 static kmutex_t dtrace_meta_lock; /* meta-provider state lock */
278 /* XXX FreeBSD hacks. */
279 static kmutex_t mod_lock;
281 #define cr_suid cr_svuid
282 #define cr_sgid cr_svgid
283 #define ipaddr_t in_addr_t
284 #define mod_modname pathname
285 #define vuprintf vprintf
286 #define ttoproc(_a) ((_a)->td_proc)
287 #define crgetzoneid(_a) 0
290 #define CPU_ON_INTR(_a) 0
292 #define PRIV_EFFECTIVE (1 << 0)
293 #define PRIV_DTRACE_KERNEL (1 << 1)
294 #define PRIV_DTRACE_PROC (1 << 2)
295 #define PRIV_DTRACE_USER (1 << 3)
296 #define PRIV_PROC_OWNER (1 << 4)
297 #define PRIV_PROC_ZONE (1 << 5)
300 SYSCTL_NODE(_debug, OID_AUTO, dtrace, CTLFLAG_RD, 0, "DTrace Information");
304 #define curcpu CPU->cpu_id
309 * DTrace Provider Variables
311 * These are the variables relating to DTrace as a provider (that is, the
312 * provider of the BEGIN, END, and ERROR probes).
314 static dtrace_pattr_t dtrace_provider_attr = {
315 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
316 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
317 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
318 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
319 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
326 static dtrace_pops_t dtrace_provider_ops = {
327 (void (*)(void *, dtrace_probedesc_t *))dtrace_nullop,
328 (void (*)(void *, modctl_t *))dtrace_nullop,
329 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
330 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
331 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
332 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
336 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop
339 static dtrace_id_t dtrace_probeid_begin; /* special BEGIN probe */
340 static dtrace_id_t dtrace_probeid_end; /* special END probe */
341 dtrace_id_t dtrace_probeid_error; /* special ERROR probe */
344 * DTrace Helper Tracing Variables
346 uint32_t dtrace_helptrace_next = 0;
347 uint32_t dtrace_helptrace_nlocals;
348 char *dtrace_helptrace_buffer;
349 int dtrace_helptrace_bufsize = 512 * 1024;
352 int dtrace_helptrace_enabled = 1;
354 int dtrace_helptrace_enabled = 0;
358 * DTrace Error Hashing
360 * On DEBUG kernels, DTrace will track the errors that has seen in a hash
361 * table. This is very useful for checking coverage of tests that are
362 * expected to induce DIF or DOF processing errors, and may be useful for
363 * debugging problems in the DIF code generator or in DOF generation . The
364 * error hash may be examined with the ::dtrace_errhash MDB dcmd.
367 static dtrace_errhash_t dtrace_errhash[DTRACE_ERRHASHSZ];
368 static const char *dtrace_errlast;
369 static kthread_t *dtrace_errthread;
370 static kmutex_t dtrace_errlock;
374 * DTrace Macros and Constants
376 * These are various macros that are useful in various spots in the
377 * implementation, along with a few random constants that have no meaning
378 * outside of the implementation. There is no real structure to this cpp
379 * mishmash -- but is there ever?
381 #define DTRACE_HASHSTR(hash, probe) \
382 dtrace_hash_str(*((char **)((uintptr_t)(probe) + (hash)->dth_stroffs)))
384 #define DTRACE_HASHNEXT(hash, probe) \
385 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_nextoffs)
387 #define DTRACE_HASHPREV(hash, probe) \
388 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_prevoffs)
390 #define DTRACE_HASHEQ(hash, lhs, rhs) \
391 (strcmp(*((char **)((uintptr_t)(lhs) + (hash)->dth_stroffs)), \
392 *((char **)((uintptr_t)(rhs) + (hash)->dth_stroffs))) == 0)
394 #define DTRACE_AGGHASHSIZE_SLEW 17
396 #define DTRACE_V4MAPPED_OFFSET (sizeof (uint32_t) * 3)
399 * The key for a thread-local variable consists of the lower 61 bits of the
400 * t_did, plus the 3 bits of the highest active interrupt above LOCK_LEVEL.
401 * We add DIF_VARIABLE_MAX to t_did to assure that the thread key is never
402 * equal to a variable identifier. This is necessary (but not sufficient) to
403 * assure that global associative arrays never collide with thread-local
404 * variables. To guarantee that they cannot collide, we must also define the
405 * order for keying dynamic variables. That order is:
407 * [ key0 ] ... [ keyn ] [ variable-key ] [ tls-key ]
409 * Because the variable-key and the tls-key are in orthogonal spaces, there is
410 * no way for a global variable key signature to match a thread-local key
414 #define DTRACE_TLS_THRKEY(where) { \
416 uint_t actv = CPU->cpu_intr_actv >> (LOCK_LEVEL + 1); \
417 for (; actv; actv >>= 1) \
419 ASSERT(intr < (1 << 3)); \
420 (where) = ((curthread->t_did + DIF_VARIABLE_MAX) & \
421 (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
424 #define DTRACE_TLS_THRKEY(where) { \
425 solaris_cpu_t *_c = &solaris_cpu[curcpu]; \
427 uint_t actv = _c->cpu_intr_actv; \
428 for (; actv; actv >>= 1) \
430 ASSERT(intr < (1 << 3)); \
431 (where) = ((curthread->td_tid + DIF_VARIABLE_MAX) & \
432 (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
436 #define DT_BSWAP_8(x) ((x) & 0xff)
437 #define DT_BSWAP_16(x) ((DT_BSWAP_8(x) << 8) | DT_BSWAP_8((x) >> 8))
438 #define DT_BSWAP_32(x) ((DT_BSWAP_16(x) << 16) | DT_BSWAP_16((x) >> 16))
439 #define DT_BSWAP_64(x) ((DT_BSWAP_32(x) << 32) | DT_BSWAP_32((x) >> 32))
441 #define DT_MASK_LO 0x00000000FFFFFFFFULL
443 #define DTRACE_STORE(type, tomax, offset, what) \
444 *((type *)((uintptr_t)(tomax) + (uintptr_t)offset)) = (type)(what);
447 #define DTRACE_ALIGNCHECK(addr, size, flags) \
448 if (addr & (size - 1)) { \
449 *flags |= CPU_DTRACE_BADALIGN; \
450 cpu_core[curcpu].cpuc_dtrace_illval = addr; \
454 #define DTRACE_ALIGNCHECK(addr, size, flags)
458 * Test whether a range of memory starting at testaddr of size testsz falls
459 * within the range of memory described by addr, sz. We take care to avoid
460 * problems with overflow and underflow of the unsigned quantities, and
461 * disallow all negative sizes. Ranges of size 0 are allowed.
463 #define DTRACE_INRANGE(testaddr, testsz, baseaddr, basesz) \
464 ((testaddr) - (baseaddr) < (basesz) && \
465 (testaddr) + (testsz) - (baseaddr) <= (basesz) && \
466 (testaddr) + (testsz) >= (testaddr))
469 * Test whether alloc_sz bytes will fit in the scratch region. We isolate
470 * alloc_sz on the righthand side of the comparison in order to avoid overflow
471 * or underflow in the comparison with it. This is simpler than the INRANGE
472 * check above, because we know that the dtms_scratch_ptr is valid in the
473 * range. Allocations of size zero are allowed.
475 #define DTRACE_INSCRATCH(mstate, alloc_sz) \
476 ((mstate)->dtms_scratch_base + (mstate)->dtms_scratch_size - \
477 (mstate)->dtms_scratch_ptr >= (alloc_sz))
479 #define DTRACE_LOADFUNC(bits) \
482 dtrace_load##bits(uintptr_t addr) \
484 size_t size = bits / NBBY; \
486 uint##bits##_t rval; \
488 volatile uint16_t *flags = (volatile uint16_t *) \
489 &cpu_core[curcpu].cpuc_dtrace_flags; \
491 DTRACE_ALIGNCHECK(addr, size, flags); \
493 for (i = 0; i < dtrace_toxranges; i++) { \
494 if (addr >= dtrace_toxrange[i].dtt_limit) \
497 if (addr + size <= dtrace_toxrange[i].dtt_base) \
501 * This address falls within a toxic region; return 0. \
503 *flags |= CPU_DTRACE_BADADDR; \
504 cpu_core[curcpu].cpuc_dtrace_illval = addr; \
508 *flags |= CPU_DTRACE_NOFAULT; \
510 rval = *((volatile uint##bits##_t *)addr); \
511 *flags &= ~CPU_DTRACE_NOFAULT; \
513 return (!(*flags & CPU_DTRACE_FAULT) ? rval : 0); \
517 #define dtrace_loadptr dtrace_load64
519 #define dtrace_loadptr dtrace_load32
522 #define DTRACE_DYNHASH_FREE 0
523 #define DTRACE_DYNHASH_SINK 1
524 #define DTRACE_DYNHASH_VALID 2
526 #define DTRACE_MATCH_NEXT 0
527 #define DTRACE_MATCH_DONE 1
528 #define DTRACE_ANCHORED(probe) ((probe)->dtpr_func[0] != '\0')
529 #define DTRACE_STATE_ALIGN 64
531 #define DTRACE_FLAGS2FLT(flags) \
532 (((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR : \
533 ((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP : \
534 ((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO : \
535 ((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV : \
536 ((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV : \
537 ((flags) & CPU_DTRACE_TUPOFLOW) ? DTRACEFLT_TUPOFLOW : \
538 ((flags) & CPU_DTRACE_BADALIGN) ? DTRACEFLT_BADALIGN : \
539 ((flags) & CPU_DTRACE_NOSCRATCH) ? DTRACEFLT_NOSCRATCH : \
540 ((flags) & CPU_DTRACE_BADSTACK) ? DTRACEFLT_BADSTACK : \
543 #define DTRACEACT_ISSTRING(act) \
544 ((act)->dta_kind == DTRACEACT_DIFEXPR && \
545 (act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING)
547 /* Function prototype definitions: */
548 static size_t dtrace_strlen(const char *, size_t);
549 static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id);
550 static void dtrace_enabling_provide(dtrace_provider_t *);
551 static int dtrace_enabling_match(dtrace_enabling_t *, int *);
552 static void dtrace_enabling_matchall(void);
553 static dtrace_state_t *dtrace_anon_grab(void);
554 static uint64_t dtrace_helper(int, dtrace_mstate_t *,
555 dtrace_state_t *, uint64_t, uint64_t);
556 static dtrace_helpers_t *dtrace_helpers_create(proc_t *);
557 static void dtrace_buffer_drop(dtrace_buffer_t *);
558 static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t,
559 dtrace_state_t *, dtrace_mstate_t *);
560 static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t,
562 static int dtrace_ecb_create_enable(dtrace_probe_t *, void *);
563 static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);
564 uint16_t dtrace_load16(uintptr_t);
565 uint32_t dtrace_load32(uintptr_t);
566 uint64_t dtrace_load64(uintptr_t);
567 uint8_t dtrace_load8(uintptr_t);
568 void dtrace_dynvar_clean(dtrace_dstate_t *);
569 dtrace_dynvar_t *dtrace_dynvar(dtrace_dstate_t *, uint_t, dtrace_key_t *,
570 size_t, dtrace_dynvar_op_t, dtrace_mstate_t *, dtrace_vstate_t *);
571 uintptr_t dtrace_dif_varstr(uintptr_t, dtrace_state_t *, dtrace_mstate_t *);
574 * DTrace Probe Context Functions
576 * These functions are called from probe context. Because probe context is
577 * any context in which C may be called, arbitrarily locks may be held,
578 * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
579 * As a result, functions called from probe context may only call other DTrace
580 * support functions -- they may not interact at all with the system at large.
581 * (Note that the ASSERT macro is made probe-context safe by redefining it in
582 * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
583 * loads are to be performed from probe context, they _must_ be in terms of
584 * the safe dtrace_load*() variants.
586 * Some functions in this block are not actually called from probe context;
587 * for these functions, there will be a comment above the function reading
588 * "Note: not called from probe context."
591 dtrace_panic(const char *format, ...)
595 va_start(alist, format);
596 dtrace_vpanic(format, alist);
601 dtrace_assfail(const char *a, const char *f, int l)
603 dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);
606 * We just need something here that even the most clever compiler
607 * cannot optimize away.
609 return (a[(uintptr_t)f]);
613 * Atomically increment a specified error counter from probe context.
616 dtrace_error(uint32_t *counter)
619 * Most counters stored to in probe context are per-CPU counters.
620 * However, there are some error conditions that are sufficiently
621 * arcane that they don't merit per-CPU storage. If these counters
622 * are incremented concurrently on different CPUs, scalability will be
623 * adversely affected -- but we don't expect them to be white-hot in a
624 * correctly constructed enabling...
631 if ((nval = oval + 1) == 0) {
633 * If the counter would wrap, set it to 1 -- assuring
634 * that the counter is never zero when we have seen
635 * errors. (The counter must be 32-bits because we
636 * aren't guaranteed a 64-bit compare&swap operation.)
637 * To save this code both the infamy of being fingered
638 * by a priggish news story and the indignity of being
639 * the target of a neo-puritan witch trial, we're
640 * carefully avoiding any colorful description of the
641 * likelihood of this condition -- but suffice it to
642 * say that it is only slightly more likely than the
643 * overflow of predicate cache IDs, as discussed in
644 * dtrace_predicate_create().
648 } while (dtrace_cas32(counter, oval, nval) != oval);
652 * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
653 * uint8_t, a uint16_t, a uint32_t and a uint64_t.
661 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
663 if (dest < mstate->dtms_scratch_base)
666 if (dest + size < dest)
669 if (dest + size > mstate->dtms_scratch_ptr)
676 dtrace_canstore_statvar(uint64_t addr, size_t sz,
677 dtrace_statvar_t **svars, int nsvars)
681 for (i = 0; i < nsvars; i++) {
682 dtrace_statvar_t *svar = svars[i];
684 if (svar == NULL || svar->dtsv_size == 0)
687 if (DTRACE_INRANGE(addr, sz, svar->dtsv_data, svar->dtsv_size))
695 * Check to see if the address is within a memory region to which a store may
696 * be issued. This includes the DTrace scratch areas, and any DTrace variable
697 * region. The caller of dtrace_canstore() is responsible for performing any
698 * alignment checks that are needed before stores are actually executed.
701 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
702 dtrace_vstate_t *vstate)
705 * First, check to see if the address is in scratch space...
707 if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base,
708 mstate->dtms_scratch_size))
712 * Now check to see if it's a dynamic variable. This check will pick
713 * up both thread-local variables and any global dynamically-allocated
716 if (DTRACE_INRANGE(addr, sz, (uintptr_t)vstate->dtvs_dynvars.dtds_base,
717 vstate->dtvs_dynvars.dtds_size)) {
718 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
719 uintptr_t base = (uintptr_t)dstate->dtds_base +
720 (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t));
724 * Before we assume that we can store here, we need to make
725 * sure that it isn't in our metadata -- storing to our
726 * dynamic variable metadata would corrupt our state. For
727 * the range to not include any dynamic variable metadata,
730 * (1) Start above the hash table that is at the base of
731 * the dynamic variable space
733 * (2) Have a starting chunk offset that is beyond the
734 * dtrace_dynvar_t that is at the base of every chunk
736 * (3) Not span a chunk boundary
742 chunkoffs = (addr - base) % dstate->dtds_chunksize;
744 if (chunkoffs < sizeof (dtrace_dynvar_t))
747 if (chunkoffs + sz > dstate->dtds_chunksize)
754 * Finally, check the static local and global variables. These checks
755 * take the longest, so we perform them last.
757 if (dtrace_canstore_statvar(addr, sz,
758 vstate->dtvs_locals, vstate->dtvs_nlocals))
761 if (dtrace_canstore_statvar(addr, sz,
762 vstate->dtvs_globals, vstate->dtvs_nglobals))
770 * Convenience routine to check to see if the address is within a memory
771 * region in which a load may be issued given the user's privilege level;
772 * if not, it sets the appropriate error flags and loads 'addr' into the
773 * illegal value slot.
775 * DTrace subroutines (DIF_SUBR_*) should use this helper to implement
776 * appropriate memory access protection.
779 dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
780 dtrace_vstate_t *vstate)
782 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
785 * If we hold the privilege to read from kernel memory, then
786 * everything is readable.
788 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
792 * You can obviously read that which you can store.
794 if (dtrace_canstore(addr, sz, mstate, vstate))
798 * We're allowed to read from our own string table.
800 if (DTRACE_INRANGE(addr, sz, (uintptr_t)mstate->dtms_difo->dtdo_strtab,
801 mstate->dtms_difo->dtdo_strlen))
804 DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV);
810 * Convenience routine to check to see if a given string is within a memory
811 * region in which a load may be issued given the user's privilege level;
812 * this exists so that we don't need to issue unnecessary dtrace_strlen()
813 * calls in the event that the user has all privileges.
816 dtrace_strcanload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
817 dtrace_vstate_t *vstate)
822 * If we hold the privilege to read from kernel memory, then
823 * everything is readable.
825 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
828 strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr, sz);
829 if (dtrace_canload(addr, strsz, mstate, vstate))
836 * Convenience routine to check to see if a given variable is within a memory
837 * region in which a load may be issued given the user's privilege level.
840 dtrace_vcanload(void *src, dtrace_diftype_t *type, dtrace_mstate_t *mstate,
841 dtrace_vstate_t *vstate)
844 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
847 * If we hold the privilege to read from kernel memory, then
848 * everything is readable.
850 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
853 if (type->dtdt_kind == DIF_TYPE_STRING)
854 sz = dtrace_strlen(src,
855 vstate->dtvs_state->dts_options[DTRACEOPT_STRSIZE]) + 1;
857 sz = type->dtdt_size;
859 return (dtrace_canload((uintptr_t)src, sz, mstate, vstate));
863 * Compare two strings using safe loads.
866 dtrace_strncmp(char *s1, char *s2, size_t limit)
869 volatile uint16_t *flags;
871 if (s1 == s2 || limit == 0)
874 flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
880 c1 = dtrace_load8((uintptr_t)s1++);
886 c2 = dtrace_load8((uintptr_t)s2++);
891 } while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));
897 * Compute strlen(s) for a string using safe memory accesses. The additional
898 * len parameter is used to specify a maximum length to ensure completion.
901 dtrace_strlen(const char *s, size_t lim)
905 for (len = 0; len != lim; len++) {
906 if (dtrace_load8((uintptr_t)s++) == '\0')
914 * Check if an address falls within a toxic region.
917 dtrace_istoxic(uintptr_t kaddr, size_t size)
919 uintptr_t taddr, tsize;
922 for (i = 0; i < dtrace_toxranges; i++) {
923 taddr = dtrace_toxrange[i].dtt_base;
924 tsize = dtrace_toxrange[i].dtt_limit - taddr;
926 if (kaddr - taddr < tsize) {
927 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
928 cpu_core[curcpu].cpuc_dtrace_illval = kaddr;
932 if (taddr - kaddr < size) {
933 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
934 cpu_core[curcpu].cpuc_dtrace_illval = taddr;
943 * Copy src to dst using safe memory accesses. The src is assumed to be unsafe
944 * memory specified by the DIF program. The dst is assumed to be safe memory
945 * that we can store to directly because it is managed by DTrace. As with
946 * standard bcopy, overlapping copies are handled properly.
949 dtrace_bcopy(const void *src, void *dst, size_t len)
953 const uint8_t *s2 = src;
957 *s1++ = dtrace_load8((uintptr_t)s2++);
958 } while (--len != 0);
964 *--s1 = dtrace_load8((uintptr_t)--s2);
965 } while (--len != 0);
971 * Copy src to dst using safe memory accesses, up to either the specified
972 * length, or the point that a nul byte is encountered. The src is assumed to
973 * be unsafe memory specified by the DIF program. The dst is assumed to be
974 * safe memory that we can store to directly because it is managed by DTrace.
975 * Unlike dtrace_bcopy(), overlapping regions are not handled.
978 dtrace_strcpy(const void *src, void *dst, size_t len)
981 uint8_t *s1 = dst, c;
982 const uint8_t *s2 = src;
985 *s1++ = c = dtrace_load8((uintptr_t)s2++);
986 } while (--len != 0 && c != '\0');
991 * Copy src to dst, deriving the size and type from the specified (BYREF)
992 * variable type. The src is assumed to be unsafe memory specified by the DIF
993 * program. The dst is assumed to be DTrace variable memory that is of the
994 * specified type; we assume that we can store to directly.
997 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type)
999 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
1001 if (type->dtdt_kind == DIF_TYPE_STRING) {
1002 dtrace_strcpy(src, dst, type->dtdt_size);
1004 dtrace_bcopy(src, dst, type->dtdt_size);
1009 * Compare s1 to s2 using safe memory accesses. The s1 data is assumed to be
1010 * unsafe memory specified by the DIF program. The s2 data is assumed to be
1011 * safe memory that we can access directly because it is managed by DTrace.
1014 dtrace_bcmp(const void *s1, const void *s2, size_t len)
1016 volatile uint16_t *flags;
1018 flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
1023 if (s1 == NULL || s2 == NULL)
1026 if (s1 != s2 && len != 0) {
1027 const uint8_t *ps1 = s1;
1028 const uint8_t *ps2 = s2;
1031 if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
1033 } while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
1039 * Zero the specified region using a simple byte-by-byte loop. Note that this
1040 * is for safe DTrace-managed memory only.
1043 dtrace_bzero(void *dst, size_t len)
1047 for (cp = dst; len != 0; len--)
1052 dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum)
1056 result[0] = addend1[0] + addend2[0];
1057 result[1] = addend1[1] + addend2[1] +
1058 (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0);
1065 * Shift the 128-bit value in a by b. If b is positive, shift left.
1066 * If b is negative, shift right.
1069 dtrace_shift_128(uint64_t *a, int b)
1079 a[0] = a[1] >> (b - 64);
1083 mask = 1LL << (64 - b);
1085 a[0] |= ((a[1] & mask) << (64 - b));
1090 a[1] = a[0] << (b - 64);
1094 mask = a[0] >> (64 - b);
1102 * The basic idea is to break the 2 64-bit values into 4 32-bit values,
1103 * use native multiplication on those, and then re-combine into the
1104 * resulting 128-bit value.
1106 * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) =
1113 dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product)
1115 uint64_t hi1, hi2, lo1, lo2;
1118 hi1 = factor1 >> 32;
1119 hi2 = factor2 >> 32;
1121 lo1 = factor1 & DT_MASK_LO;
1122 lo2 = factor2 & DT_MASK_LO;
1124 product[0] = lo1 * lo2;
1125 product[1] = hi1 * hi2;
1129 dtrace_shift_128(tmp, 32);
1130 dtrace_add_128(product, tmp, product);
1134 dtrace_shift_128(tmp, 32);
1135 dtrace_add_128(product, tmp, product);
1139 * This privilege check should be used by actions and subroutines to
1140 * verify that the user credentials of the process that enabled the
1141 * invoking ECB match the target credentials
1144 dtrace_priv_proc_common_user(dtrace_state_t *state)
1146 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1149 * We should always have a non-NULL state cred here, since if cred
1150 * is null (anonymous tracing), we fast-path bypass this routine.
1152 ASSERT(s_cr != NULL);
1154 if ((cr = CRED()) != NULL &&
1155 s_cr->cr_uid == cr->cr_uid &&
1156 s_cr->cr_uid == cr->cr_ruid &&
1157 s_cr->cr_uid == cr->cr_suid &&
1158 s_cr->cr_gid == cr->cr_gid &&
1159 s_cr->cr_gid == cr->cr_rgid &&
1160 s_cr->cr_gid == cr->cr_sgid)
1167 * This privilege check should be used by actions and subroutines to
1168 * verify that the zone of the process that enabled the invoking ECB
1169 * matches the target credentials
1172 dtrace_priv_proc_common_zone(dtrace_state_t *state)
1175 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1178 * We should always have a non-NULL state cred here, since if cred
1179 * is null (anonymous tracing), we fast-path bypass this routine.
1181 ASSERT(s_cr != NULL);
1183 if ((cr = CRED()) != NULL &&
1184 s_cr->cr_zone == cr->cr_zone)
1194 * This privilege check should be used by actions and subroutines to
1195 * verify that the process has not setuid or changed credentials.
1198 dtrace_priv_proc_common_nocd(void)
1202 if ((proc = ttoproc(curthread)) != NULL &&
1203 !(proc->p_flag & SNOCD))
1210 dtrace_priv_proc_destructive(dtrace_state_t *state)
1212 int action = state->dts_cred.dcr_action;
1214 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
1215 dtrace_priv_proc_common_zone(state) == 0)
1218 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
1219 dtrace_priv_proc_common_user(state) == 0)
1222 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
1223 dtrace_priv_proc_common_nocd() == 0)
1229 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1235 dtrace_priv_proc_control(dtrace_state_t *state)
1237 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
1240 if (dtrace_priv_proc_common_zone(state) &&
1241 dtrace_priv_proc_common_user(state) &&
1242 dtrace_priv_proc_common_nocd())
1245 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1251 dtrace_priv_proc(dtrace_state_t *state)
1253 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC)
1256 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1262 dtrace_priv_kernel(dtrace_state_t *state)
1264 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
1267 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1273 dtrace_priv_kernel_destructive(dtrace_state_t *state)
1275 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
1278 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1284 * Note: not called from probe context. This function is called
1285 * asynchronously (and at a regular interval) from outside of probe context to
1286 * clean the dirty dynamic variable lists on all CPUs. Dynamic variable
1287 * cleaning is explained in detail in <sys/dtrace_impl.h>.
1290 dtrace_dynvar_clean(dtrace_dstate_t *dstate)
1292 dtrace_dynvar_t *dirty;
1293 dtrace_dstate_percpu_t *dcpu;
1296 for (i = 0; i < NCPU; i++) {
1297 dcpu = &dstate->dtds_percpu[i];
1299 ASSERT(dcpu->dtdsc_rinsing == NULL);
1302 * If the dirty list is NULL, there is no dirty work to do.
1304 if (dcpu->dtdsc_dirty == NULL)
1308 * If the clean list is non-NULL, then we're not going to do
1309 * any work for this CPU -- it means that there has not been
1310 * a dtrace_dynvar() allocation on this CPU (or from this CPU)
1311 * since the last time we cleaned house.
1313 if (dcpu->dtdsc_clean != NULL)
1319 * Atomically move the dirty list aside.
1322 dirty = dcpu->dtdsc_dirty;
1325 * Before we zap the dirty list, set the rinsing list.
1326 * (This allows for a potential assertion in
1327 * dtrace_dynvar(): if a free dynamic variable appears
1328 * on a hash chain, either the dirty list or the
1329 * rinsing list for some CPU must be non-NULL.)
1331 dcpu->dtdsc_rinsing = dirty;
1332 dtrace_membar_producer();
1333 } while (dtrace_casptr(&dcpu->dtdsc_dirty,
1334 dirty, NULL) != dirty);
1339 * We have no work to do; we can simply return.
1346 for (i = 0; i < NCPU; i++) {
1347 dcpu = &dstate->dtds_percpu[i];
1349 if (dcpu->dtdsc_rinsing == NULL)
1353 * We are now guaranteed that no hash chain contains a pointer
1354 * into this dirty list; we can make it clean.
1356 ASSERT(dcpu->dtdsc_clean == NULL);
1357 dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
1358 dcpu->dtdsc_rinsing = NULL;
1362 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
1363 * sure that all CPUs have seen all of the dtdsc_clean pointers.
1364 * This prevents a race whereby a CPU incorrectly decides that
1365 * the state should be something other than DTRACE_DSTATE_CLEAN
1366 * after dtrace_dynvar_clean() has completed.
1370 dstate->dtds_state = DTRACE_DSTATE_CLEAN;
1374 * Depending on the value of the op parameter, this function looks-up,
1375 * allocates or deallocates an arbitrarily-keyed dynamic variable. If an
1376 * allocation is requested, this function will return a pointer to a
1377 * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
1378 * variable can be allocated. If NULL is returned, the appropriate counter
1379 * will be incremented.
1382 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
1383 dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op,
1384 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1386 uint64_t hashval = DTRACE_DYNHASH_VALID;
1387 dtrace_dynhash_t *hash = dstate->dtds_hash;
1388 dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
1389 processorid_t me = curcpu, cpu = me;
1390 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
1391 size_t bucket, ksize;
1392 size_t chunksize = dstate->dtds_chunksize;
1393 uintptr_t kdata, lock, nstate;
1399 * Hash the key. As with aggregations, we use Jenkins' "One-at-a-time"
1400 * algorithm. For the by-value portions, we perform the algorithm in
1401 * 16-bit chunks (as opposed to 8-bit chunks). This speeds things up a
1402 * bit, and seems to have only a minute effect on distribution. For
1403 * the by-reference data, we perform "One-at-a-time" iterating (safely)
1404 * over each referenced byte. It's painful to do this, but it's much
1405 * better than pathological hash distribution. The efficacy of the
1406 * hashing algorithm (and a comparison with other algorithms) may be
1407 * found by running the ::dtrace_dynstat MDB dcmd.
1409 for (i = 0; i < nkeys; i++) {
1410 if (key[i].dttk_size == 0) {
1411 uint64_t val = key[i].dttk_value;
1413 hashval += (val >> 48) & 0xffff;
1414 hashval += (hashval << 10);
1415 hashval ^= (hashval >> 6);
1417 hashval += (val >> 32) & 0xffff;
1418 hashval += (hashval << 10);
1419 hashval ^= (hashval >> 6);
1421 hashval += (val >> 16) & 0xffff;
1422 hashval += (hashval << 10);
1423 hashval ^= (hashval >> 6);
1425 hashval += val & 0xffff;
1426 hashval += (hashval << 10);
1427 hashval ^= (hashval >> 6);
1430 * This is incredibly painful, but it beats the hell
1431 * out of the alternative.
1433 uint64_t j, size = key[i].dttk_size;
1434 uintptr_t base = (uintptr_t)key[i].dttk_value;
1436 if (!dtrace_canload(base, size, mstate, vstate))
1439 for (j = 0; j < size; j++) {
1440 hashval += dtrace_load8(base + j);
1441 hashval += (hashval << 10);
1442 hashval ^= (hashval >> 6);
1447 if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT))
1450 hashval += (hashval << 3);
1451 hashval ^= (hashval >> 11);
1452 hashval += (hashval << 15);
1455 * There is a remote chance (ideally, 1 in 2^31) that our hashval
1456 * comes out to be one of our two sentinel hash values. If this
1457 * actually happens, we set the hashval to be a value known to be a
1458 * non-sentinel value.
1460 if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
1461 hashval = DTRACE_DYNHASH_VALID;
1464 * Yes, it's painful to do a divide here. If the cycle count becomes
1465 * important here, tricks can be pulled to reduce it. (However, it's
1466 * critical that hash collisions be kept to an absolute minimum;
1467 * they're much more painful than a divide.) It's better to have a
1468 * solution that generates few collisions and still keeps things
1469 * relatively simple.
1471 bucket = hashval % dstate->dtds_hashsize;
1473 if (op == DTRACE_DYNVAR_DEALLOC) {
1474 volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;
1477 while ((lock = *lockp) & 1)
1480 if (dtrace_casptr((volatile void *)lockp,
1481 (volatile void *)lock, (volatile void *)(lock + 1)) == (void *)lock)
1485 dtrace_membar_producer();
1490 lock = hash[bucket].dtdh_lock;
1492 dtrace_membar_consumer();
1494 start = hash[bucket].dtdh_chain;
1495 ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
1496 start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
1497 op != DTRACE_DYNVAR_DEALLOC));
1499 for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
1500 dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
1501 dtrace_key_t *dkey = &dtuple->dtt_key[0];
1503 if (dvar->dtdv_hashval != hashval) {
1504 if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
1506 * We've reached the sink, and therefore the
1507 * end of the hash chain; we can kick out of
1508 * the loop knowing that we have seen a valid
1509 * snapshot of state.
1511 ASSERT(dvar->dtdv_next == NULL);
1512 ASSERT(dvar == &dtrace_dynhash_sink);
1516 if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
1518 * We've gone off the rails: somewhere along
1519 * the line, one of the members of this hash
1520 * chain was deleted. Note that we could also
1521 * detect this by simply letting this loop run
1522 * to completion, as we would eventually hit
1523 * the end of the dirty list. However, we
1524 * want to avoid running the length of the
1525 * dirty list unnecessarily (it might be quite
1526 * long), so we catch this as early as
1527 * possible by detecting the hash marker. In
1528 * this case, we simply set dvar to NULL and
1529 * break; the conditional after the loop will
1530 * send us back to top.
1539 if (dtuple->dtt_nkeys != nkeys)
1542 for (i = 0; i < nkeys; i++, dkey++) {
1543 if (dkey->dttk_size != key[i].dttk_size)
1544 goto next; /* size or type mismatch */
1546 if (dkey->dttk_size != 0) {
1548 (void *)(uintptr_t)key[i].dttk_value,
1549 (void *)(uintptr_t)dkey->dttk_value,
1553 if (dkey->dttk_value != key[i].dttk_value)
1558 if (op != DTRACE_DYNVAR_DEALLOC)
1561 ASSERT(dvar->dtdv_next == NULL ||
1562 dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);
1565 ASSERT(hash[bucket].dtdh_chain != dvar);
1566 ASSERT(start != dvar);
1567 ASSERT(prev->dtdv_next == dvar);
1568 prev->dtdv_next = dvar->dtdv_next;
1570 if (dtrace_casptr(&hash[bucket].dtdh_chain,
1571 start, dvar->dtdv_next) != start) {
1573 * We have failed to atomically swing the
1574 * hash table head pointer, presumably because
1575 * of a conflicting allocation on another CPU.
1576 * We need to reread the hash chain and try
1583 dtrace_membar_producer();
1586 * Now set the hash value to indicate that it's free.
1588 ASSERT(hash[bucket].dtdh_chain != dvar);
1589 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1591 dtrace_membar_producer();
1594 * Set the next pointer to point at the dirty list, and
1595 * atomically swing the dirty pointer to the newly freed dvar.
1598 next = dcpu->dtdsc_dirty;
1599 dvar->dtdv_next = next;
1600 } while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);
1603 * Finally, unlock this hash bucket.
1605 ASSERT(hash[bucket].dtdh_lock == lock);
1607 hash[bucket].dtdh_lock++;
1617 * If dvar is NULL, it is because we went off the rails:
1618 * one of the elements that we traversed in the hash chain
1619 * was deleted while we were traversing it. In this case,
1620 * we assert that we aren't doing a dealloc (deallocs lock
1621 * the hash bucket to prevent themselves from racing with
1622 * one another), and retry the hash chain traversal.
1624 ASSERT(op != DTRACE_DYNVAR_DEALLOC);
1628 if (op != DTRACE_DYNVAR_ALLOC) {
1630 * If we are not to allocate a new variable, we want to
1631 * return NULL now. Before we return, check that the value
1632 * of the lock word hasn't changed. If it has, we may have
1633 * seen an inconsistent snapshot.
1635 if (op == DTRACE_DYNVAR_NOALLOC) {
1636 if (hash[bucket].dtdh_lock != lock)
1639 ASSERT(op == DTRACE_DYNVAR_DEALLOC);
1640 ASSERT(hash[bucket].dtdh_lock == lock);
1642 hash[bucket].dtdh_lock++;
1649 * We need to allocate a new dynamic variable. The size we need is the
1650 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
1651 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
1652 * the size of any referred-to data (dsize). We then round the final
1653 * size up to the chunksize for allocation.
1655 for (ksize = 0, i = 0; i < nkeys; i++)
1656 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
1659 * This should be pretty much impossible, but could happen if, say,
1660 * strange DIF specified the tuple. Ideally, this should be an
1661 * assertion and not an error condition -- but that requires that the
1662 * chunksize calculation in dtrace_difo_chunksize() be absolutely
1663 * bullet-proof. (That is, it must not be able to be fooled by
1664 * malicious DIF.) Given the lack of backwards branches in DIF,
1665 * solving this would presumably not amount to solving the Halting
1666 * Problem -- but it still seems awfully hard.
1668 if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
1669 ksize + dsize > chunksize) {
1670 dcpu->dtdsc_drops++;
1674 nstate = DTRACE_DSTATE_EMPTY;
1678 free = dcpu->dtdsc_free;
1681 dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
1684 if (clean == NULL) {
1686 * We're out of dynamic variable space on
1687 * this CPU. Unless we have tried all CPUs,
1688 * we'll try to allocate from a different
1691 switch (dstate->dtds_state) {
1692 case DTRACE_DSTATE_CLEAN: {
1693 void *sp = &dstate->dtds_state;
1698 if (dcpu->dtdsc_dirty != NULL &&
1699 nstate == DTRACE_DSTATE_EMPTY)
1700 nstate = DTRACE_DSTATE_DIRTY;
1702 if (dcpu->dtdsc_rinsing != NULL)
1703 nstate = DTRACE_DSTATE_RINSING;
1705 dcpu = &dstate->dtds_percpu[cpu];
1710 (void) dtrace_cas32(sp,
1711 DTRACE_DSTATE_CLEAN, nstate);
1714 * To increment the correct bean
1715 * counter, take another lap.
1720 case DTRACE_DSTATE_DIRTY:
1721 dcpu->dtdsc_dirty_drops++;
1724 case DTRACE_DSTATE_RINSING:
1725 dcpu->dtdsc_rinsing_drops++;
1728 case DTRACE_DSTATE_EMPTY:
1729 dcpu->dtdsc_drops++;
1733 DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
1738 * The clean list appears to be non-empty. We want to
1739 * move the clean list to the free list; we start by
1740 * moving the clean pointer aside.
1742 if (dtrace_casptr(&dcpu->dtdsc_clean,
1743 clean, NULL) != clean) {
1745 * We are in one of two situations:
1747 * (a) The clean list was switched to the
1748 * free list by another CPU.
1750 * (b) The clean list was added to by the
1753 * In either of these situations, we can
1754 * just reattempt the free list allocation.
1759 ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);
1762 * Now we'll move the clean list to the free list.
1763 * It's impossible for this to fail: the only way
1764 * the free list can be updated is through this
1765 * code path, and only one CPU can own the clean list.
1766 * Thus, it would only be possible for this to fail if
1767 * this code were racing with dtrace_dynvar_clean().
1768 * (That is, if dtrace_dynvar_clean() updated the clean
1769 * list, and we ended up racing to update the free
1770 * list.) This race is prevented by the dtrace_sync()
1771 * in dtrace_dynvar_clean() -- which flushes the
1772 * owners of the clean lists out before resetting
1775 rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
1776 ASSERT(rval == NULL);
1781 new_free = dvar->dtdv_next;
1782 } while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);
1785 * We have now allocated a new chunk. We copy the tuple keys into the
1786 * tuple array and copy any referenced key data into the data space
1787 * following the tuple array. As we do this, we relocate dttk_value
1788 * in the final tuple to point to the key data address in the chunk.
1790 kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
1791 dvar->dtdv_data = (void *)(kdata + ksize);
1792 dvar->dtdv_tuple.dtt_nkeys = nkeys;
1794 for (i = 0; i < nkeys; i++) {
1795 dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
1796 size_t kesize = key[i].dttk_size;
1800 (const void *)(uintptr_t)key[i].dttk_value,
1801 (void *)kdata, kesize);
1802 dkey->dttk_value = kdata;
1803 kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
1805 dkey->dttk_value = key[i].dttk_value;
1808 dkey->dttk_size = kesize;
1811 ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
1812 dvar->dtdv_hashval = hashval;
1813 dvar->dtdv_next = start;
1815 if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
1819 * The cas has failed. Either another CPU is adding an element to
1820 * this hash chain, or another CPU is deleting an element from this
1821 * hash chain. The simplest way to deal with both of these cases
1822 * (though not necessarily the most efficient) is to free our
1823 * allocated block and tail-call ourselves. Note that the free is
1824 * to the dirty list and _not_ to the free list. This is to prevent
1825 * races with allocators, above.
1827 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1829 dtrace_membar_producer();
1832 free = dcpu->dtdsc_dirty;
1833 dvar->dtdv_next = free;
1834 } while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);
1836 return (dtrace_dynvar(dstate, nkeys, key, dsize, op, mstate, vstate));
1841 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
1843 if ((int64_t)nval < (int64_t)*oval)
1849 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
1851 if ((int64_t)nval > (int64_t)*oval)
1856 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
1858 int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
1859 int64_t val = (int64_t)nval;
1862 for (i = 0; i < zero; i++) {
1863 if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
1869 for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
1870 if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
1871 quanta[i - 1] += incr;
1876 quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
1884 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
1886 uint64_t arg = *lquanta++;
1887 int32_t base = DTRACE_LQUANTIZE_BASE(arg);
1888 uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
1889 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
1890 int32_t val = (int32_t)nval, level;
1893 ASSERT(levels != 0);
1897 * This is an underflow.
1903 level = (val - base) / step;
1905 if (level < levels) {
1906 lquanta[level + 1] += incr;
1911 * This is an overflow.
1913 lquanta[levels + 1] += incr;
1918 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
1926 dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg)
1928 int64_t snval = (int64_t)nval;
1935 * What we want to say here is:
1937 * data[2] += nval * nval;
1939 * But given that nval is 64-bit, we could easily overflow, so
1940 * we do this as 128-bit arithmetic.
1945 dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp);
1946 dtrace_add_128(data + 2, tmp, data + 2);
1951 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
1958 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
1964 * Aggregate given the tuple in the principal data buffer, and the aggregating
1965 * action denoted by the specified dtrace_aggregation_t. The aggregation
1966 * buffer is specified as the buf parameter. This routine does not return
1967 * failure; if there is no space in the aggregation buffer, the data will be
1968 * dropped, and a corresponding counter incremented.
1971 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
1972 intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
1974 dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
1975 uint32_t i, ndx, size, fsize;
1976 uint32_t align = sizeof (uint64_t) - 1;
1977 dtrace_aggbuffer_t *agb;
1978 dtrace_aggkey_t *key;
1979 uint32_t hashval = 0, limit, isstr;
1980 caddr_t tomax, data, kdata;
1981 dtrace_actkind_t action;
1982 dtrace_action_t *act;
1988 if (!agg->dtag_hasarg) {
1990 * Currently, only quantize() and lquantize() take additional
1991 * arguments, and they have the same semantics: an increment
1992 * value that defaults to 1 when not present. If additional
1993 * aggregating actions take arguments, the setting of the
1994 * default argument value will presumably have to become more
2000 action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
2001 size = rec->dtrd_offset - agg->dtag_base;
2002 fsize = size + rec->dtrd_size;
2004 ASSERT(dbuf->dtb_tomax != NULL);
2005 data = dbuf->dtb_tomax + offset + agg->dtag_base;
2007 if ((tomax = buf->dtb_tomax) == NULL) {
2008 dtrace_buffer_drop(buf);
2013 * The metastructure is always at the bottom of the buffer.
2015 agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
2016 sizeof (dtrace_aggbuffer_t));
2018 if (buf->dtb_offset == 0) {
2020 * We just kludge up approximately 1/8th of the size to be
2021 * buckets. If this guess ends up being routinely
2022 * off-the-mark, we may need to dynamically readjust this
2023 * based on past performance.
2025 uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);
2027 if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
2028 (uintptr_t)tomax || hashsize == 0) {
2030 * We've been given a ludicrously small buffer;
2031 * increment our drop count and leave.
2033 dtrace_buffer_drop(buf);
2038 * And now, a pathetic attempt to try to get a an odd (or
2039 * perchance, a prime) hash size for better hash distribution.
2041 if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
2042 hashsize -= DTRACE_AGGHASHSIZE_SLEW;
2044 agb->dtagb_hashsize = hashsize;
2045 agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
2046 agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
2047 agb->dtagb_free = (uintptr_t)agb->dtagb_hash;
2049 for (i = 0; i < agb->dtagb_hashsize; i++)
2050 agb->dtagb_hash[i] = NULL;
2053 ASSERT(agg->dtag_first != NULL);
2054 ASSERT(agg->dtag_first->dta_intuple);
2057 * Calculate the hash value based on the key. Note that we _don't_
2058 * include the aggid in the hashing (but we will store it as part of
2059 * the key). The hashing algorithm is Bob Jenkins' "One-at-a-time"
2060 * algorithm: a simple, quick algorithm that has no known funnels, and
2061 * gets good distribution in practice. The efficacy of the hashing
2062 * algorithm (and a comparison with other algorithms) may be found by
2063 * running the ::dtrace_aggstat MDB dcmd.
2065 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2066 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2067 limit = i + act->dta_rec.dtrd_size;
2068 ASSERT(limit <= size);
2069 isstr = DTRACEACT_ISSTRING(act);
2071 for (; i < limit; i++) {
2073 hashval += (hashval << 10);
2074 hashval ^= (hashval >> 6);
2076 if (isstr && data[i] == '\0')
2081 hashval += (hashval << 3);
2082 hashval ^= (hashval >> 11);
2083 hashval += (hashval << 15);
2086 * Yes, the divide here is expensive -- but it's generally the least
2087 * of the performance issues given the amount of data that we iterate
2088 * over to compute hash values, compare data, etc.
2090 ndx = hashval % agb->dtagb_hashsize;
2092 for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
2093 ASSERT((caddr_t)key >= tomax);
2094 ASSERT((caddr_t)key < tomax + buf->dtb_size);
2096 if (hashval != key->dtak_hashval || key->dtak_size != size)
2099 kdata = key->dtak_data;
2100 ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);
2102 for (act = agg->dtag_first; act->dta_intuple;
2103 act = act->dta_next) {
2104 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2105 limit = i + act->dta_rec.dtrd_size;
2106 ASSERT(limit <= size);
2107 isstr = DTRACEACT_ISSTRING(act);
2109 for (; i < limit; i++) {
2110 if (kdata[i] != data[i])
2113 if (isstr && data[i] == '\0')
2118 if (action != key->dtak_action) {
2120 * We are aggregating on the same value in the same
2121 * aggregation with two different aggregating actions.
2122 * (This should have been picked up in the compiler,
2123 * so we may be dealing with errant or devious DIF.)
2124 * This is an error condition; we indicate as much,
2127 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2132 * This is a hit: we need to apply the aggregator to
2133 * the value at this key.
2135 agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
2142 * We didn't find it. We need to allocate some zero-filled space,
2143 * link it into the hash table appropriately, and apply the aggregator
2144 * to the (zero-filled) value.
2146 offs = buf->dtb_offset;
2147 while (offs & (align - 1))
2148 offs += sizeof (uint32_t);
2151 * If we don't have enough room to both allocate a new key _and_
2152 * its associated data, increment the drop count and return.
2154 if ((uintptr_t)tomax + offs + fsize >
2155 agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
2156 dtrace_buffer_drop(buf);
2161 ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
2162 key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
2163 agb->dtagb_free -= sizeof (dtrace_aggkey_t);
2165 key->dtak_data = kdata = tomax + offs;
2166 buf->dtb_offset = offs + fsize;
2169 * Now copy the data across.
2171 *((dtrace_aggid_t *)kdata) = agg->dtag_id;
2173 for (i = sizeof (dtrace_aggid_t); i < size; i++)
2177 * Because strings are not zeroed out by default, we need to iterate
2178 * looking for actions that store strings, and we need to explicitly
2179 * pad these strings out with zeroes.
2181 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2184 if (!DTRACEACT_ISSTRING(act))
2187 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2188 limit = i + act->dta_rec.dtrd_size;
2189 ASSERT(limit <= size);
2191 for (nul = 0; i < limit; i++) {
2197 if (data[i] != '\0')
2204 for (i = size; i < fsize; i++)
2207 key->dtak_hashval = hashval;
2208 key->dtak_size = size;
2209 key->dtak_action = action;
2210 key->dtak_next = agb->dtagb_hash[ndx];
2211 agb->dtagb_hash[ndx] = key;
2214 * Finally, apply the aggregator.
2216 *((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
2217 agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
2221 * Given consumer state, this routine finds a speculation in the INACTIVE
2222 * state and transitions it into the ACTIVE state. If there is no speculation
2223 * in the INACTIVE state, 0 is returned. In this case, no error counter is
2224 * incremented -- it is up to the caller to take appropriate action.
2227 dtrace_speculation(dtrace_state_t *state)
2230 dtrace_speculation_state_t current;
2231 uint32_t *stat = &state->dts_speculations_unavail, count;
2233 while (i < state->dts_nspeculations) {
2234 dtrace_speculation_t *spec = &state->dts_speculations[i];
2236 current = spec->dtsp_state;
2238 if (current != DTRACESPEC_INACTIVE) {
2239 if (current == DTRACESPEC_COMMITTINGMANY ||
2240 current == DTRACESPEC_COMMITTING ||
2241 current == DTRACESPEC_DISCARDING)
2242 stat = &state->dts_speculations_busy;
2247 if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2248 current, DTRACESPEC_ACTIVE) == current)
2253 * We couldn't find a speculation. If we found as much as a single
2254 * busy speculation buffer, we'll attribute this failure as "busy"
2255 * instead of "unavail".
2259 } while (dtrace_cas32(stat, count, count + 1) != count);
2265 * This routine commits an active speculation. If the specified speculation
2266 * is not in a valid state to perform a commit(), this routine will silently do
2267 * nothing. The state of the specified speculation is transitioned according
2268 * to the state transition diagram outlined in <sys/dtrace_impl.h>
2271 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
2272 dtrace_specid_t which)
2274 dtrace_speculation_t *spec;
2275 dtrace_buffer_t *src, *dest;
2276 uintptr_t daddr, saddr, dlimit;
2277 dtrace_speculation_state_t current, new = 0;
2283 if (which > state->dts_nspeculations) {
2284 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2288 spec = &state->dts_speculations[which - 1];
2289 src = &spec->dtsp_buffer[cpu];
2290 dest = &state->dts_buffer[cpu];
2293 current = spec->dtsp_state;
2295 if (current == DTRACESPEC_COMMITTINGMANY)
2299 case DTRACESPEC_INACTIVE:
2300 case DTRACESPEC_DISCARDING:
2303 case DTRACESPEC_COMMITTING:
2305 * This is only possible if we are (a) commit()'ing
2306 * without having done a prior speculate() on this CPU
2307 * and (b) racing with another commit() on a different
2308 * CPU. There's nothing to do -- we just assert that
2311 ASSERT(src->dtb_offset == 0);
2314 case DTRACESPEC_ACTIVE:
2315 new = DTRACESPEC_COMMITTING;
2318 case DTRACESPEC_ACTIVEONE:
2320 * This speculation is active on one CPU. If our
2321 * buffer offset is non-zero, we know that the one CPU
2322 * must be us. Otherwise, we are committing on a
2323 * different CPU from the speculate(), and we must
2324 * rely on being asynchronously cleaned.
2326 if (src->dtb_offset != 0) {
2327 new = DTRACESPEC_COMMITTING;
2332 case DTRACESPEC_ACTIVEMANY:
2333 new = DTRACESPEC_COMMITTINGMANY;
2339 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2340 current, new) != current);
2343 * We have set the state to indicate that we are committing this
2344 * speculation. Now reserve the necessary space in the destination
2347 if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
2348 sizeof (uint64_t), state, NULL)) < 0) {
2349 dtrace_buffer_drop(dest);
2354 * We have the space; copy the buffer across. (Note that this is a
2355 * highly subobtimal bcopy(); in the unlikely event that this becomes
2356 * a serious performance issue, a high-performance DTrace-specific
2357 * bcopy() should obviously be invented.)
2359 daddr = (uintptr_t)dest->dtb_tomax + offs;
2360 dlimit = daddr + src->dtb_offset;
2361 saddr = (uintptr_t)src->dtb_tomax;
2364 * First, the aligned portion.
2366 while (dlimit - daddr >= sizeof (uint64_t)) {
2367 *((uint64_t *)daddr) = *((uint64_t *)saddr);
2369 daddr += sizeof (uint64_t);
2370 saddr += sizeof (uint64_t);
2374 * Now any left-over bit...
2376 while (dlimit - daddr)
2377 *((uint8_t *)daddr++) = *((uint8_t *)saddr++);
2380 * Finally, commit the reserved space in the destination buffer.
2382 dest->dtb_offset = offs + src->dtb_offset;
2386 * If we're lucky enough to be the only active CPU on this speculation
2387 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
2389 if (current == DTRACESPEC_ACTIVE ||
2390 (current == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
2391 uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
2392 DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);
2394 ASSERT(rval == DTRACESPEC_COMMITTING);
2397 src->dtb_offset = 0;
2398 src->dtb_xamot_drops += src->dtb_drops;
2403 * This routine discards an active speculation. If the specified speculation
2404 * is not in a valid state to perform a discard(), this routine will silently
2405 * do nothing. The state of the specified speculation is transitioned
2406 * according to the state transition diagram outlined in <sys/dtrace_impl.h>
2409 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
2410 dtrace_specid_t which)
2412 dtrace_speculation_t *spec;
2413 dtrace_speculation_state_t current, new = 0;
2414 dtrace_buffer_t *buf;
2419 if (which > state->dts_nspeculations) {
2420 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2424 spec = &state->dts_speculations[which - 1];
2425 buf = &spec->dtsp_buffer[cpu];
2428 current = spec->dtsp_state;
2431 case DTRACESPEC_INACTIVE:
2432 case DTRACESPEC_COMMITTINGMANY:
2433 case DTRACESPEC_COMMITTING:
2434 case DTRACESPEC_DISCARDING:
2437 case DTRACESPEC_ACTIVE:
2438 case DTRACESPEC_ACTIVEMANY:
2439 new = DTRACESPEC_DISCARDING;
2442 case DTRACESPEC_ACTIVEONE:
2443 if (buf->dtb_offset != 0) {
2444 new = DTRACESPEC_INACTIVE;
2446 new = DTRACESPEC_DISCARDING;
2453 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2454 current, new) != current);
2456 buf->dtb_offset = 0;
2461 * Note: not called from probe context. This function is called
2462 * asynchronously from cross call context to clean any speculations that are
2463 * in the COMMITTINGMANY or DISCARDING states. These speculations may not be
2464 * transitioned back to the INACTIVE state until all CPUs have cleaned the
2468 dtrace_speculation_clean_here(dtrace_state_t *state)
2470 dtrace_icookie_t cookie;
2471 processorid_t cpu = curcpu;
2472 dtrace_buffer_t *dest = &state->dts_buffer[cpu];
2475 cookie = dtrace_interrupt_disable();
2477 if (dest->dtb_tomax == NULL) {
2478 dtrace_interrupt_enable(cookie);
2482 for (i = 0; i < state->dts_nspeculations; i++) {
2483 dtrace_speculation_t *spec = &state->dts_speculations[i];
2484 dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];
2486 if (src->dtb_tomax == NULL)
2489 if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
2490 src->dtb_offset = 0;
2494 if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2497 if (src->dtb_offset == 0)
2500 dtrace_speculation_commit(state, cpu, i + 1);
2503 dtrace_interrupt_enable(cookie);
2507 * Note: not called from probe context. This function is called
2508 * asynchronously (and at a regular interval) to clean any speculations that
2509 * are in the COMMITTINGMANY or DISCARDING states. If it discovers that there
2510 * is work to be done, it cross calls all CPUs to perform that work;
2511 * COMMITMANY and DISCARDING speculations may not be transitioned back to the
2512 * INACTIVE state until they have been cleaned by all CPUs.
2515 dtrace_speculation_clean(dtrace_state_t *state)
2520 for (i = 0; i < state->dts_nspeculations; i++) {
2521 dtrace_speculation_t *spec = &state->dts_speculations[i];
2523 ASSERT(!spec->dtsp_cleaning);
2525 if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
2526 spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2530 spec->dtsp_cleaning = 1;
2536 dtrace_xcall(DTRACE_CPUALL,
2537 (dtrace_xcall_t)dtrace_speculation_clean_here, state);
2540 * We now know that all CPUs have committed or discarded their
2541 * speculation buffers, as appropriate. We can now set the state
2544 for (i = 0; i < state->dts_nspeculations; i++) {
2545 dtrace_speculation_t *spec = &state->dts_speculations[i];
2546 dtrace_speculation_state_t current, new;
2548 if (!spec->dtsp_cleaning)
2551 current = spec->dtsp_state;
2552 ASSERT(current == DTRACESPEC_DISCARDING ||
2553 current == DTRACESPEC_COMMITTINGMANY);
2555 new = DTRACESPEC_INACTIVE;
2557 rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, current, new);
2558 ASSERT(rv == current);
2559 spec->dtsp_cleaning = 0;
2564 * Called as part of a speculate() to get the speculative buffer associated
2565 * with a given speculation. Returns NULL if the specified speculation is not
2566 * in an ACTIVE state. If the speculation is in the ACTIVEONE state -- and
2567 * the active CPU is not the specified CPU -- the speculation will be
2568 * atomically transitioned into the ACTIVEMANY state.
2570 static dtrace_buffer_t *
2571 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
2572 dtrace_specid_t which)
2574 dtrace_speculation_t *spec;
2575 dtrace_speculation_state_t current, new = 0;
2576 dtrace_buffer_t *buf;
2581 if (which > state->dts_nspeculations) {
2582 cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2586 spec = &state->dts_speculations[which - 1];
2587 buf = &spec->dtsp_buffer[cpuid];
2590 current = spec->dtsp_state;
2593 case DTRACESPEC_INACTIVE:
2594 case DTRACESPEC_COMMITTINGMANY:
2595 case DTRACESPEC_DISCARDING:
2598 case DTRACESPEC_COMMITTING:
2599 ASSERT(buf->dtb_offset == 0);
2602 case DTRACESPEC_ACTIVEONE:
2604 * This speculation is currently active on one CPU.
2605 * Check the offset in the buffer; if it's non-zero,
2606 * that CPU must be us (and we leave the state alone).
2607 * If it's zero, assume that we're starting on a new
2608 * CPU -- and change the state to indicate that the
2609 * speculation is active on more than one CPU.
2611 if (buf->dtb_offset != 0)
2614 new = DTRACESPEC_ACTIVEMANY;
2617 case DTRACESPEC_ACTIVEMANY:
2620 case DTRACESPEC_ACTIVE:
2621 new = DTRACESPEC_ACTIVEONE;
2627 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2628 current, new) != current);
2630 ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
2635 * Return a string. In the event that the user lacks the privilege to access
2636 * arbitrary kernel memory, we copy the string out to scratch memory so that we
2637 * don't fail access checking.
2639 * dtrace_dif_variable() uses this routine as a helper for various
2640 * builtin values such as 'execname' and 'probefunc.'
2643 dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state,
2644 dtrace_mstate_t *mstate)
2646 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
2651 * The easy case: this probe is allowed to read all of memory, so
2652 * we can just return this as a vanilla pointer.
2654 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
2658 * This is the tougher case: we copy the string in question from
2659 * kernel memory into scratch memory and return it that way: this
2660 * ensures that we won't trip up when access checking tests the
2661 * BYREF return value.
2663 strsz = dtrace_strlen((char *)addr, size) + 1;
2665 if (mstate->dtms_scratch_ptr + strsz >
2666 mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2667 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2671 dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2673 ret = mstate->dtms_scratch_ptr;
2674 mstate->dtms_scratch_ptr += strsz;
2679 * Return a string from a memoy address which is known to have one or
2680 * more concatenated, individually zero terminated, sub-strings.
2681 * In the event that the user lacks the privilege to access
2682 * arbitrary kernel memory, we copy the string out to scratch memory so that we
2683 * don't fail access checking.
2685 * dtrace_dif_variable() uses this routine as a helper for various
2686 * builtin values such as 'execargs'.
2689 dtrace_dif_varstrz(uintptr_t addr, size_t strsz, dtrace_state_t *state,
2690 dtrace_mstate_t *mstate)
2696 if (mstate->dtms_scratch_ptr + strsz >
2697 mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2698 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2702 dtrace_bcopy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2705 /* Replace sub-string termination characters with a space. */
2706 for (p = (char *) mstate->dtms_scratch_ptr, i = 0; i < strsz - 1;
2711 ret = mstate->dtms_scratch_ptr;
2712 mstate->dtms_scratch_ptr += strsz;
2717 * This function implements the DIF emulator's variable lookups. The emulator
2718 * passes a reserved variable identifier and optional built-in array index.
2721 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
2725 * If we're accessing one of the uncached arguments, we'll turn this
2726 * into a reference in the args array.
2728 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
2729 ndx = v - DIF_VAR_ARG0;
2735 ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
2736 if (ndx >= sizeof (mstate->dtms_arg) /
2737 sizeof (mstate->dtms_arg[0])) {
2738 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2739 dtrace_provider_t *pv;
2742 pv = mstate->dtms_probe->dtpr_provider;
2743 if (pv->dtpv_pops.dtps_getargval != NULL)
2744 val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
2745 mstate->dtms_probe->dtpr_id,
2746 mstate->dtms_probe->dtpr_arg, ndx, aframes);
2748 val = dtrace_getarg(ndx, aframes);
2751 * This is regrettably required to keep the compiler
2752 * from tail-optimizing the call to dtrace_getarg().
2753 * The condition always evaluates to true, but the
2754 * compiler has no way of figuring that out a priori.
2755 * (None of this would be necessary if the compiler
2756 * could be relied upon to _always_ tail-optimize
2757 * the call to dtrace_getarg() -- but it can't.)
2759 if (mstate->dtms_probe != NULL)
2765 return (mstate->dtms_arg[ndx]);
2768 case DIF_VAR_UREGS: {
2771 if (!dtrace_priv_proc(state))
2774 if ((lwp = curthread->t_lwp) == NULL) {
2775 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
2776 cpu_core[curcpu].cpuc_dtrace_illval = NULL;
2780 return (dtrace_getreg(lwp->lwp_regs, ndx));
2784 case DIF_VAR_UREGS: {
2785 struct trapframe *tframe;
2787 if (!dtrace_priv_proc(state))
2790 if ((tframe = curthread->td_frame) == NULL) {
2791 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
2792 cpu_core[curcpu].cpuc_dtrace_illval = 0;
2796 return (dtrace_getreg(tframe, ndx));
2800 case DIF_VAR_CURTHREAD:
2801 if (!dtrace_priv_kernel(state))
2803 return ((uint64_t)(uintptr_t)curthread);
2805 case DIF_VAR_TIMESTAMP:
2806 if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
2807 mstate->dtms_timestamp = dtrace_gethrtime();
2808 mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
2810 return (mstate->dtms_timestamp);
2812 case DIF_VAR_VTIMESTAMP:
2813 ASSERT(dtrace_vtime_references != 0);
2814 return (curthread->t_dtrace_vtime);
2816 case DIF_VAR_WALLTIMESTAMP:
2817 if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
2818 mstate->dtms_walltimestamp = dtrace_gethrestime();
2819 mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
2821 return (mstate->dtms_walltimestamp);
2825 if (!dtrace_priv_kernel(state))
2827 if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
2828 mstate->dtms_ipl = dtrace_getipl();
2829 mstate->dtms_present |= DTRACE_MSTATE_IPL;
2831 return (mstate->dtms_ipl);
2835 ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
2836 return (mstate->dtms_epid);
2839 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2840 return (mstate->dtms_probe->dtpr_id);
2842 case DIF_VAR_STACKDEPTH:
2843 if (!dtrace_priv_kernel(state))
2845 if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
2846 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2848 mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
2849 mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
2851 return (mstate->dtms_stackdepth);
2853 case DIF_VAR_USTACKDEPTH:
2854 if (!dtrace_priv_proc(state))
2856 if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
2858 * See comment in DIF_VAR_PID.
2860 if (DTRACE_ANCHORED(mstate->dtms_probe) &&
2862 mstate->dtms_ustackdepth = 0;
2864 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2865 mstate->dtms_ustackdepth =
2866 dtrace_getustackdepth();
2867 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2869 mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
2871 return (mstate->dtms_ustackdepth);
2873 case DIF_VAR_CALLER:
2874 if (!dtrace_priv_kernel(state))
2876 if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
2877 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2879 if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
2881 * If this is an unanchored probe, we are
2882 * required to go through the slow path:
2883 * dtrace_caller() only guarantees correct
2884 * results for anchored probes.
2886 pc_t caller[2] = {0, 0};
2888 dtrace_getpcstack(caller, 2, aframes,
2889 (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
2890 mstate->dtms_caller = caller[1];
2891 } else if ((mstate->dtms_caller =
2892 dtrace_caller(aframes)) == -1) {
2894 * We have failed to do this the quick way;
2895 * we must resort to the slower approach of
2896 * calling dtrace_getpcstack().
2900 dtrace_getpcstack(&caller, 1, aframes, NULL);
2901 mstate->dtms_caller = caller;
2904 mstate->dtms_present |= DTRACE_MSTATE_CALLER;
2906 return (mstate->dtms_caller);
2908 case DIF_VAR_UCALLER:
2909 if (!dtrace_priv_proc(state))
2912 if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
2916 * dtrace_getupcstack() fills in the first uint64_t
2917 * with the current PID. The second uint64_t will
2918 * be the program counter at user-level. The third
2919 * uint64_t will contain the caller, which is what
2923 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2924 dtrace_getupcstack(ustack, 3);
2925 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2926 mstate->dtms_ucaller = ustack[2];
2927 mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
2930 return (mstate->dtms_ucaller);
2932 case DIF_VAR_PROBEPROV:
2933 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2934 return (dtrace_dif_varstr(
2935 (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name,
2938 case DIF_VAR_PROBEMOD:
2939 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2940 return (dtrace_dif_varstr(
2941 (uintptr_t)mstate->dtms_probe->dtpr_mod,
2944 case DIF_VAR_PROBEFUNC:
2945 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2946 return (dtrace_dif_varstr(
2947 (uintptr_t)mstate->dtms_probe->dtpr_func,
2950 case DIF_VAR_PROBENAME:
2951 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2952 return (dtrace_dif_varstr(
2953 (uintptr_t)mstate->dtms_probe->dtpr_name,
2957 if (!dtrace_priv_proc(state))
2962 * Note that we are assuming that an unanchored probe is
2963 * always due to a high-level interrupt. (And we're assuming
2964 * that there is only a single high level interrupt.)
2966 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
2967 return (pid0.pid_id);
2970 * It is always safe to dereference one's own t_procp pointer:
2971 * it always points to a valid, allocated proc structure.
2972 * Further, it is always safe to dereference the p_pidp member
2973 * of one's own proc structure. (These are truisms becuase
2974 * threads and processes don't clean up their own state --
2975 * they leave that task to whomever reaps them.)
2977 return ((uint64_t)curthread->t_procp->p_pidp->pid_id);
2979 return ((uint64_t)curproc->p_pid);
2983 if (!dtrace_priv_proc(state))
2988 * See comment in DIF_VAR_PID.
2990 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
2991 return (pid0.pid_id);
2994 * It is always safe to dereference one's own t_procp pointer:
2995 * it always points to a valid, allocated proc structure.
2996 * (This is true because threads don't clean up their own
2997 * state -- they leave that task to whomever reaps them.)
2999 return ((uint64_t)curthread->t_procp->p_ppid);
3001 return ((uint64_t)curproc->p_pptr->p_pid);
3007 * See comment in DIF_VAR_PID.
3009 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3013 return ((uint64_t)curthread->t_tid);
3015 case DIF_VAR_EXECARGS: {
3016 struct pargs *p_args = curthread->td_proc->p_args;
3021 return (dtrace_dif_varstrz(
3022 (uintptr_t) p_args->ar_args, p_args->ar_length, state, mstate));
3025 case DIF_VAR_EXECNAME:
3027 if (!dtrace_priv_proc(state))
3031 * See comment in DIF_VAR_PID.
3033 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3034 return ((uint64_t)(uintptr_t)p0.p_user.u_comm);
3037 * It is always safe to dereference one's own t_procp pointer:
3038 * it always points to a valid, allocated proc structure.
3039 * (This is true because threads don't clean up their own
3040 * state -- they leave that task to whomever reaps them.)
3042 return (dtrace_dif_varstr(
3043 (uintptr_t)curthread->t_procp->p_user.u_comm,
3046 return (dtrace_dif_varstr(
3047 (uintptr_t) curthread->td_proc->p_comm, state, mstate));
3050 case DIF_VAR_ZONENAME:
3052 if (!dtrace_priv_proc(state))
3056 * See comment in DIF_VAR_PID.
3058 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3059 return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);
3062 * It is always safe to dereference one's own t_procp pointer:
3063 * it always points to a valid, allocated proc structure.
3064 * (This is true because threads don't clean up their own
3065 * state -- they leave that task to whomever reaps them.)
3067 return (dtrace_dif_varstr(
3068 (uintptr_t)curthread->t_procp->p_zone->zone_name,
3075 if (!dtrace_priv_proc(state))
3080 * See comment in DIF_VAR_PID.
3082 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3083 return ((uint64_t)p0.p_cred->cr_uid);
3087 * It is always safe to dereference one's own t_procp pointer:
3088 * it always points to a valid, allocated proc structure.
3089 * (This is true because threads don't clean up their own
3090 * state -- they leave that task to whomever reaps them.)
3092 * Additionally, it is safe to dereference one's own process
3093 * credential, since this is never NULL after process birth.
3095 return ((uint64_t)curthread->t_procp->p_cred->cr_uid);
3098 if (!dtrace_priv_proc(state))
3103 * See comment in DIF_VAR_PID.
3105 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3106 return ((uint64_t)p0.p_cred->cr_gid);
3110 * It is always safe to dereference one's own t_procp pointer:
3111 * it always points to a valid, allocated proc structure.
3112 * (This is true because threads don't clean up their own
3113 * state -- they leave that task to whomever reaps them.)
3115 * Additionally, it is safe to dereference one's own process
3116 * credential, since this is never NULL after process birth.
3118 return ((uint64_t)curthread->t_procp->p_cred->cr_gid);
3120 case DIF_VAR_ERRNO: {
3123 if (!dtrace_priv_proc(state))
3127 * See comment in DIF_VAR_PID.
3129 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3133 * It is always safe to dereference one's own t_lwp pointer in
3134 * the event that this pointer is non-NULL. (This is true
3135 * because threads and lwps don't clean up their own state --
3136 * they leave that task to whomever reaps them.)
3138 if ((lwp = curthread->t_lwp) == NULL)
3141 return ((uint64_t)lwp->lwp_errno);
3143 return (curthread->td_errno);
3147 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
3153 * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
3154 * Notice that we don't bother validating the proper number of arguments or
3155 * their types in the tuple stack. This isn't needed because all argument
3156 * interpretation is safe because of our load safety -- the worst that can
3157 * happen is that a bogus program can obtain bogus results.
3160 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
3161 dtrace_key_t *tupregs, int nargs,
3162 dtrace_mstate_t *mstate, dtrace_state_t *state)
3164 volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
3165 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
3166 dtrace_vstate_t *vstate = &state->dts_vstate;
3179 struct thread *lowner;
3181 struct lock_object *li;
3188 regs[rd] = (dtrace_gethrtime() * 2416 + 374441) % 1771875;
3192 case DIF_SUBR_MUTEX_OWNED:
3193 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3199 m.mx = dtrace_load64(tupregs[0].dttk_value);
3200 if (MUTEX_TYPE_ADAPTIVE(&m.mi))
3201 regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
3203 regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
3206 case DIF_SUBR_MUTEX_OWNER:
3207 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3213 m.mx = dtrace_load64(tupregs[0].dttk_value);
3214 if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
3215 MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
3216 regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
3221 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3222 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3228 m.mx = dtrace_load64(tupregs[0].dttk_value);
3229 regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
3232 case DIF_SUBR_MUTEX_TYPE_SPIN:
3233 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3239 m.mx = dtrace_load64(tupregs[0].dttk_value);
3240 regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
3243 case DIF_SUBR_RW_READ_HELD: {
3246 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3252 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3253 regs[rd] = _RW_READ_HELD(&r.ri, tmp);
3257 case DIF_SUBR_RW_WRITE_HELD:
3258 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3264 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3265 regs[rd] = _RW_WRITE_HELD(&r.ri);
3268 case DIF_SUBR_RW_ISWRITER:
3269 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3275 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3276 regs[rd] = _RW_ISWRITER(&r.ri);
3280 case DIF_SUBR_MUTEX_OWNED:
3281 if (!dtrace_canload(tupregs[0].dttk_value,
3282 sizeof (struct lock_object), mstate, vstate)) {
3286 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3287 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3290 case DIF_SUBR_MUTEX_OWNER:
3291 if (!dtrace_canload(tupregs[0].dttk_value,
3292 sizeof (struct lock_object), mstate, vstate)) {
3296 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3297 LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3298 regs[rd] = (uintptr_t)lowner;
3301 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3302 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
3307 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3308 /* XXX - should be only LC_SLEEPABLE? */
3309 regs[rd] = (LOCK_CLASS(l.li)->lc_flags &
3310 (LC_SLEEPLOCK | LC_SLEEPABLE)) != 0;
3313 case DIF_SUBR_MUTEX_TYPE_SPIN:
3314 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
3319 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3320 regs[rd] = (LOCK_CLASS(l.li)->lc_flags & LC_SPINLOCK) != 0;
3323 case DIF_SUBR_RW_READ_HELD:
3324 case DIF_SUBR_SX_SHARED_HELD:
3325 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3330 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3331 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
3335 case DIF_SUBR_RW_WRITE_HELD:
3336 case DIF_SUBR_SX_EXCLUSIVE_HELD:
3337 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3342 l.lx = dtrace_loadptr(tupregs[0].dttk_value);
3343 LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3344 regs[rd] = (lowner == curthread);
3347 case DIF_SUBR_RW_ISWRITER:
3348 case DIF_SUBR_SX_ISEXCLUSIVE:
3349 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3354 l.lx = dtrace_loadptr(tupregs[0].dttk_value);
3355 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
3358 #endif /* ! defined(sun) */
3360 case DIF_SUBR_BCOPY: {
3362 * We need to be sure that the destination is in the scratch
3363 * region -- no other region is allowed.
3365 uintptr_t src = tupregs[0].dttk_value;
3366 uintptr_t dest = tupregs[1].dttk_value;
3367 size_t size = tupregs[2].dttk_value;
3369 if (!dtrace_inscratch(dest, size, mstate)) {
3370 *flags |= CPU_DTRACE_BADADDR;
3375 if (!dtrace_canload(src, size, mstate, vstate)) {
3380 dtrace_bcopy((void *)src, (void *)dest, size);
3384 case DIF_SUBR_ALLOCA:
3385 case DIF_SUBR_COPYIN: {
3386 uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
3388 tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
3389 size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;
3392 * This action doesn't require any credential checks since
3393 * probes will not activate in user contexts to which the
3394 * enabling user does not have permissions.
3398 * Rounding up the user allocation size could have overflowed
3399 * a large, bogus allocation (like -1ULL) to 0.
3401 if (scratch_size < size ||
3402 !DTRACE_INSCRATCH(mstate, scratch_size)) {
3403 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3408 if (subr == DIF_SUBR_COPYIN) {
3409 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3410 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3411 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3414 mstate->dtms_scratch_ptr += scratch_size;
3419 case DIF_SUBR_COPYINTO: {
3420 uint64_t size = tupregs[1].dttk_value;
3421 uintptr_t dest = tupregs[2].dttk_value;
3424 * This action doesn't require any credential checks since
3425 * probes will not activate in user contexts to which the
3426 * enabling user does not have permissions.
3428 if (!dtrace_inscratch(dest, size, mstate)) {
3429 *flags |= CPU_DTRACE_BADADDR;
3434 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3435 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3436 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3440 case DIF_SUBR_COPYINSTR: {
3441 uintptr_t dest = mstate->dtms_scratch_ptr;
3442 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3444 if (nargs > 1 && tupregs[1].dttk_value < size)
3445 size = tupregs[1].dttk_value + 1;
3448 * This action doesn't require any credential checks since
3449 * probes will not activate in user contexts to which the
3450 * enabling user does not have permissions.
3452 if (!DTRACE_INSCRATCH(mstate, size)) {
3453 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3458 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3459 dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags);
3460 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3462 ((char *)dest)[size - 1] = '\0';
3463 mstate->dtms_scratch_ptr += size;
3469 case DIF_SUBR_MSGSIZE:
3470 case DIF_SUBR_MSGDSIZE: {
3471 uintptr_t baddr = tupregs[0].dttk_value, daddr;
3472 uintptr_t wptr, rptr;
3476 while (baddr != 0 && !(*flags & CPU_DTRACE_FAULT)) {
3478 if (!dtrace_canload(baddr, sizeof (mblk_t), mstate,
3484 wptr = dtrace_loadptr(baddr +
3485 offsetof(mblk_t, b_wptr));
3487 rptr = dtrace_loadptr(baddr +
3488 offsetof(mblk_t, b_rptr));
3491 *flags |= CPU_DTRACE_BADADDR;
3492 *illval = tupregs[0].dttk_value;
3496 daddr = dtrace_loadptr(baddr +
3497 offsetof(mblk_t, b_datap));
3499 baddr = dtrace_loadptr(baddr +
3500 offsetof(mblk_t, b_cont));
3503 * We want to prevent against denial-of-service here,
3504 * so we're only going to search the list for
3505 * dtrace_msgdsize_max mblks.
3507 if (cont++ > dtrace_msgdsize_max) {
3508 *flags |= CPU_DTRACE_ILLOP;
3512 if (subr == DIF_SUBR_MSGDSIZE) {
3513 if (dtrace_load8(daddr +
3514 offsetof(dblk_t, db_type)) != M_DATA)
3518 count += wptr - rptr;
3521 if (!(*flags & CPU_DTRACE_FAULT))
3528 case DIF_SUBR_PROGENYOF: {
3529 pid_t pid = tupregs[0].dttk_value;
3533 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3535 for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
3537 if (p->p_pidp->pid_id == pid) {
3539 if (p->p_pid == pid) {
3546 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3552 case DIF_SUBR_SPECULATION:
3553 regs[rd] = dtrace_speculation(state);
3556 case DIF_SUBR_COPYOUT: {
3557 uintptr_t kaddr = tupregs[0].dttk_value;
3558 uintptr_t uaddr = tupregs[1].dttk_value;
3559 uint64_t size = tupregs[2].dttk_value;
3561 if (!dtrace_destructive_disallow &&
3562 dtrace_priv_proc_control(state) &&
3563 !dtrace_istoxic(kaddr, size)) {
3564 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3565 dtrace_copyout(kaddr, uaddr, size, flags);
3566 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3571 case DIF_SUBR_COPYOUTSTR: {
3572 uintptr_t kaddr = tupregs[0].dttk_value;
3573 uintptr_t uaddr = tupregs[1].dttk_value;
3574 uint64_t size = tupregs[2].dttk_value;
3576 if (!dtrace_destructive_disallow &&
3577 dtrace_priv_proc_control(state) &&
3578 !dtrace_istoxic(kaddr, size)) {
3579 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3580 dtrace_copyoutstr(kaddr, uaddr, size, flags);
3581 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3586 case DIF_SUBR_STRLEN: {
3588 uintptr_t addr = (uintptr_t)tupregs[0].dttk_value;
3589 sz = dtrace_strlen((char *)addr,
3590 state->dts_options[DTRACEOPT_STRSIZE]);
3592 if (!dtrace_canload(addr, sz + 1, mstate, vstate)) {
3602 case DIF_SUBR_STRCHR:
3603 case DIF_SUBR_STRRCHR: {
3605 * We're going to iterate over the string looking for the
3606 * specified character. We will iterate until we have reached
3607 * the string length or we have found the character. If this
3608 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
3609 * of the specified character instead of the first.
3611 uintptr_t saddr = tupregs[0].dttk_value;
3612 uintptr_t addr = tupregs[0].dttk_value;
3613 uintptr_t limit = addr + state->dts_options[DTRACEOPT_STRSIZE];
3614 char c, target = (char)tupregs[1].dttk_value;
3616 for (regs[rd] = 0; addr < limit; addr++) {
3617 if ((c = dtrace_load8(addr)) == target) {
3620 if (subr == DIF_SUBR_STRCHR)
3628 if (!dtrace_canload(saddr, addr - saddr, mstate, vstate)) {
3636 case DIF_SUBR_STRSTR:
3637 case DIF_SUBR_INDEX:
3638 case DIF_SUBR_RINDEX: {
3640 * We're going to iterate over the string looking for the
3641 * specified string. We will iterate until we have reached
3642 * the string length or we have found the string. (Yes, this
3643 * is done in the most naive way possible -- but considering
3644 * that the string we're searching for is likely to be
3645 * relatively short, the complexity of Rabin-Karp or similar
3646 * hardly seems merited.)
3648 char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
3649 char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
3650 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3651 size_t len = dtrace_strlen(addr, size);
3652 size_t sublen = dtrace_strlen(substr, size);
3653 char *limit = addr + len, *orig = addr;
3654 int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
3657 regs[rd] = notfound;
3659 if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) {
3664 if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate,
3671 * strstr() and index()/rindex() have similar semantics if
3672 * both strings are the empty string: strstr() returns a
3673 * pointer to the (empty) string, and index() and rindex()
3674 * both return index 0 (regardless of any position argument).
3676 if (sublen == 0 && len == 0) {
3677 if (subr == DIF_SUBR_STRSTR)
3678 regs[rd] = (uintptr_t)addr;
3684 if (subr != DIF_SUBR_STRSTR) {
3685 if (subr == DIF_SUBR_RINDEX) {
3692 * Both index() and rindex() take an optional position
3693 * argument that denotes the starting position.
3696 int64_t pos = (int64_t)tupregs[2].dttk_value;
3699 * If the position argument to index() is
3700 * negative, Perl implicitly clamps it at
3701 * zero. This semantic is a little surprising
3702 * given the special meaning of negative
3703 * positions to similar Perl functions like
3704 * substr(), but it appears to reflect a
3705 * notion that index() can start from a
3706 * negative index and increment its way up to
3707 * the string. Given this notion, Perl's
3708 * rindex() is at least self-consistent in
3709 * that it implicitly clamps positions greater
3710 * than the string length to be the string
3711 * length. Where Perl completely loses
3712 * coherence, however, is when the specified
3713 * substring is the empty string (""). In
3714 * this case, even if the position is
3715 * negative, rindex() returns 0 -- and even if
3716 * the position is greater than the length,
3717 * index() returns the string length. These
3718 * semantics violate the notion that index()
3719 * should never return a value less than the
3720 * specified position and that rindex() should
3721 * never return a value greater than the
3722 * specified position. (One assumes that
3723 * these semantics are artifacts of Perl's
3724 * implementation and not the results of
3725 * deliberate design -- it beggars belief that
3726 * even Larry Wall could desire such oddness.)
3727 * While in the abstract one would wish for
3728 * consistent position semantics across
3729 * substr(), index() and rindex() -- or at the
3730 * very least self-consistent position
3731 * semantics for index() and rindex() -- we
3732 * instead opt to keep with the extant Perl
3733 * semantics, in all their broken glory. (Do
3734 * we have more desire to maintain Perl's
3735 * semantics than Perl does? Probably.)
3737 if (subr == DIF_SUBR_RINDEX) {
3761 for (regs[rd] = notfound; addr != limit; addr += inc) {
3762 if (dtrace_strncmp(addr, substr, sublen) == 0) {
3763 if (subr != DIF_SUBR_STRSTR) {
3765 * As D index() and rindex() are
3766 * modeled on Perl (and not on awk),
3767 * we return a zero-based (and not a
3768 * one-based) index. (For you Perl
3769 * weenies: no, we're not going to add
3770 * $[ -- and shouldn't you be at a con
3773 regs[rd] = (uintptr_t)(addr - orig);
3777 ASSERT(subr == DIF_SUBR_STRSTR);
3778 regs[rd] = (uintptr_t)addr;
3786 case DIF_SUBR_STRTOK: {
3787 uintptr_t addr = tupregs[0].dttk_value;
3788 uintptr_t tokaddr = tupregs[1].dttk_value;
3789 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3790 uintptr_t limit, toklimit = tokaddr + size;
3791 uint8_t c = 0, tokmap[32]; /* 256 / 8 */
3792 char *dest = (char *)mstate->dtms_scratch_ptr;
3796 * Check both the token buffer and (later) the input buffer,
3797 * since both could be non-scratch addresses.
3799 if (!dtrace_strcanload(tokaddr, size, mstate, vstate)) {
3804 if (!DTRACE_INSCRATCH(mstate, size)) {
3805 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3812 * If the address specified is NULL, we use our saved
3813 * strtok pointer from the mstate. Note that this
3814 * means that the saved strtok pointer is _only_
3815 * valid within multiple enablings of the same probe --
3816 * it behaves like an implicit clause-local variable.
3818 addr = mstate->dtms_strtok;
3821 * If the user-specified address is non-NULL we must
3822 * access check it. This is the only time we have
3823 * a chance to do so, since this address may reside
3824 * in the string table of this clause-- future calls
3825 * (when we fetch addr from mstate->dtms_strtok)
3826 * would fail this access check.
3828 if (!dtrace_strcanload(addr, size, mstate, vstate)) {
3835 * First, zero the token map, and then process the token
3836 * string -- setting a bit in the map for every character
3837 * found in the token string.
3839 for (i = 0; i < sizeof (tokmap); i++)
3842 for (; tokaddr < toklimit; tokaddr++) {
3843 if ((c = dtrace_load8(tokaddr)) == '\0')
3846 ASSERT((c >> 3) < sizeof (tokmap));
3847 tokmap[c >> 3] |= (1 << (c & 0x7));
3850 for (limit = addr + size; addr < limit; addr++) {
3852 * We're looking for a character that is _not_ contained
3853 * in the token string.
3855 if ((c = dtrace_load8(addr)) == '\0')
3858 if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
3864 * We reached the end of the string without finding
3865 * any character that was not in the token string.
3866 * We return NULL in this case, and we set the saved
3867 * address to NULL as well.
3870 mstate->dtms_strtok = 0;
3875 * From here on, we're copying into the destination string.
3877 for (i = 0; addr < limit && i < size - 1; addr++) {
3878 if ((c = dtrace_load8(addr)) == '\0')
3881 if (tokmap[c >> 3] & (1 << (c & 0x7)))
3890 regs[rd] = (uintptr_t)dest;
3891 mstate->dtms_scratch_ptr += size;
3892 mstate->dtms_strtok = addr;
3896 case DIF_SUBR_SUBSTR: {
3897 uintptr_t s = tupregs[0].dttk_value;
3898 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3899 char *d = (char *)mstate->dtms_scratch_ptr;
3900 int64_t index = (int64_t)tupregs[1].dttk_value;
3901 int64_t remaining = (int64_t)tupregs[2].dttk_value;
3902 size_t len = dtrace_strlen((char *)s, size);
3905 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
3910 if (!DTRACE_INSCRATCH(mstate, size)) {
3911 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3917 remaining = (int64_t)size;
3922 if (index < 0 && index + remaining > 0) {
3928 if (index >= len || index < 0) {
3930 } else if (remaining < 0) {
3931 remaining += len - index;
3932 } else if (index + remaining > size) {
3933 remaining = size - index;
3936 for (i = 0; i < remaining; i++) {
3937 if ((d[i] = dtrace_load8(s + index + i)) == '\0')
3943 mstate->dtms_scratch_ptr += size;
3944 regs[rd] = (uintptr_t)d;
3949 case DIF_SUBR_GETMAJOR:
3951 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
3953 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
3957 case DIF_SUBR_GETMINOR:
3959 regs[rd] = tupregs[0].dttk_value & MAXMIN64;
3961 regs[rd] = tupregs[0].dttk_value & MAXMIN;
3965 case DIF_SUBR_DDI_PATHNAME: {
3967 * This one is a galactic mess. We are going to roughly
3968 * emulate ddi_pathname(), but it's made more complicated
3969 * by the fact that we (a) want to include the minor name and
3970 * (b) must proceed iteratively instead of recursively.
3972 uintptr_t dest = mstate->dtms_scratch_ptr;
3973 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3974 char *start = (char *)dest, *end = start + size - 1;
3975 uintptr_t daddr = tupregs[0].dttk_value;
3976 int64_t minor = (int64_t)tupregs[1].dttk_value;
3978 int i, len, depth = 0;
3981 * Due to all the pointer jumping we do and context we must
3982 * rely upon, we just mandate that the user must have kernel
3983 * read privileges to use this routine.
3985 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
3986 *flags |= CPU_DTRACE_KPRIV;
3991 if (!DTRACE_INSCRATCH(mstate, size)) {
3992 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4000 * We want to have a name for the minor. In order to do this,
4001 * we need to walk the minor list from the devinfo. We want
4002 * to be sure that we don't infinitely walk a circular list,
4003 * so we check for circularity by sending a scout pointer
4004 * ahead two elements for every element that we iterate over;
4005 * if the list is circular, these will ultimately point to the
4006 * same element. You may recognize this little trick as the
4007 * answer to a stupid interview question -- one that always
4008 * seems to be asked by those who had to have it laboriously
4009 * explained to them, and who can't even concisely describe
4010 * the conditions under which one would be forced to resort to
4011 * this technique. Needless to say, those conditions are
4012 * found here -- and probably only here. Is this the only use
4013 * of this infamous trick in shipping, production code? If it
4014 * isn't, it probably should be...
4017 uintptr_t maddr = dtrace_loadptr(daddr +
4018 offsetof(struct dev_info, devi_minor));
4020 uintptr_t next = offsetof(struct ddi_minor_data, next);
4021 uintptr_t name = offsetof(struct ddi_minor_data,
4022 d_minor) + offsetof(struct ddi_minor, name);
4023 uintptr_t dev = offsetof(struct ddi_minor_data,
4024 d_minor) + offsetof(struct ddi_minor, dev);
4028 scout = dtrace_loadptr(maddr + next);
4030 while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4033 m = dtrace_load64(maddr + dev) & MAXMIN64;
4035 m = dtrace_load32(maddr + dev) & MAXMIN;
4038 maddr = dtrace_loadptr(maddr + next);
4043 scout = dtrace_loadptr(scout + next);
4048 scout = dtrace_loadptr(scout + next);
4053 if (scout == maddr) {
4054 *flags |= CPU_DTRACE_ILLOP;
4062 * We have the minor data. Now we need to
4063 * copy the minor's name into the end of the
4066 s = (char *)dtrace_loadptr(maddr + name);
4067 len = dtrace_strlen(s, size);
4069 if (*flags & CPU_DTRACE_FAULT)
4073 if ((end -= (len + 1)) < start)
4079 for (i = 1; i <= len; i++)
4080 end[i] = dtrace_load8((uintptr_t)s++);
4085 while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4086 ddi_node_state_t devi_state;
4088 devi_state = dtrace_load32(daddr +
4089 offsetof(struct dev_info, devi_node_state));
4091 if (*flags & CPU_DTRACE_FAULT)
4094 if (devi_state >= DS_INITIALIZED) {
4095 s = (char *)dtrace_loadptr(daddr +
4096 offsetof(struct dev_info, devi_addr));
4097 len = dtrace_strlen(s, size);
4099 if (*flags & CPU_DTRACE_FAULT)
4103 if ((end -= (len + 1)) < start)
4109 for (i = 1; i <= len; i++)
4110 end[i] = dtrace_load8((uintptr_t)s++);
4114 * Now for the node name...
4116 s = (char *)dtrace_loadptr(daddr +
4117 offsetof(struct dev_info, devi_node_name));
4119 daddr = dtrace_loadptr(daddr +
4120 offsetof(struct dev_info, devi_parent));
4123 * If our parent is NULL (that is, if we're the root
4124 * node), we're going to use the special path
4130 len = dtrace_strlen(s, size);
4131 if (*flags & CPU_DTRACE_FAULT)
4134 if ((end -= (len + 1)) < start)
4137 for (i = 1; i <= len; i++)
4138 end[i] = dtrace_load8((uintptr_t)s++);
4141 if (depth++ > dtrace_devdepth_max) {
4142 *flags |= CPU_DTRACE_ILLOP;
4148 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4151 regs[rd] = (uintptr_t)end;
4152 mstate->dtms_scratch_ptr += size;
4159 case DIF_SUBR_STRJOIN: {
4160 char *d = (char *)mstate->dtms_scratch_ptr;
4161 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4162 uintptr_t s1 = tupregs[0].dttk_value;
4163 uintptr_t s2 = tupregs[1].dttk_value;
4166 if (!dtrace_strcanload(s1, size, mstate, vstate) ||
4167 !dtrace_strcanload(s2, size, mstate, vstate)) {
4172 if (!DTRACE_INSCRATCH(mstate, size)) {
4173 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4180 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4185 if ((d[i++] = dtrace_load8(s1++)) == '\0') {
4193 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4198 if ((d[i++] = dtrace_load8(s2++)) == '\0')
4203 mstate->dtms_scratch_ptr += i;
4204 regs[rd] = (uintptr_t)d;
4210 case DIF_SUBR_LLTOSTR: {
4211 int64_t i = (int64_t)tupregs[0].dttk_value;
4212 int64_t val = i < 0 ? i * -1 : i;
4213 uint64_t size = 22; /* enough room for 2^64 in decimal */
4214 char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
4216 if (!DTRACE_INSCRATCH(mstate, size)) {
4217 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4222 for (*end-- = '\0'; val; val /= 10)
4223 *end-- = '0' + (val % 10);
4231 regs[rd] = (uintptr_t)end + 1;
4232 mstate->dtms_scratch_ptr += size;
4236 case DIF_SUBR_HTONS:
4237 case DIF_SUBR_NTOHS:
4238 #if BYTE_ORDER == BIG_ENDIAN
4239 regs[rd] = (uint16_t)tupregs[0].dttk_value;
4241 regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
4246 case DIF_SUBR_HTONL:
4247 case DIF_SUBR_NTOHL:
4248 #if BYTE_ORDER == BIG_ENDIAN
4249 regs[rd] = (uint32_t)tupregs[0].dttk_value;
4251 regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
4256 case DIF_SUBR_HTONLL:
4257 case DIF_SUBR_NTOHLL:
4258 #if BYTE_ORDER == BIG_ENDIAN
4259 regs[rd] = (uint64_t)tupregs[0].dttk_value;
4261 regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
4266 case DIF_SUBR_DIRNAME:
4267 case DIF_SUBR_BASENAME: {
4268 char *dest = (char *)mstate->dtms_scratch_ptr;
4269 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4270 uintptr_t src = tupregs[0].dttk_value;
4271 int i, j, len = dtrace_strlen((char *)src, size);
4272 int lastbase = -1, firstbase = -1, lastdir = -1;
4275 if (!dtrace_canload(src, len + 1, mstate, vstate)) {
4280 if (!DTRACE_INSCRATCH(mstate, size)) {
4281 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4287 * The basename and dirname for a zero-length string is
4292 src = (uintptr_t)".";
4296 * Start from the back of the string, moving back toward the
4297 * front until we see a character that isn't a slash. That
4298 * character is the last character in the basename.
4300 for (i = len - 1; i >= 0; i--) {
4301 if (dtrace_load8(src + i) != '/')
4309 * Starting from the last character in the basename, move
4310 * towards the front until we find a slash. The character
4311 * that we processed immediately before that is the first
4312 * character in the basename.
4314 for (; i >= 0; i--) {
4315 if (dtrace_load8(src + i) == '/')
4323 * Now keep going until we find a non-slash character. That
4324 * character is the last character in the dirname.
4326 for (; i >= 0; i--) {
4327 if (dtrace_load8(src + i) != '/')
4334 ASSERT(!(lastbase == -1 && firstbase != -1));
4335 ASSERT(!(firstbase == -1 && lastdir != -1));
4337 if (lastbase == -1) {
4339 * We didn't find a non-slash character. We know that
4340 * the length is non-zero, so the whole string must be
4341 * slashes. In either the dirname or the basename
4342 * case, we return '/'.
4344 ASSERT(firstbase == -1);
4345 firstbase = lastbase = lastdir = 0;
4348 if (firstbase == -1) {
4350 * The entire string consists only of a basename
4351 * component. If we're looking for dirname, we need
4352 * to change our string to be just "."; if we're
4353 * looking for a basename, we'll just set the first
4354 * character of the basename to be 0.
4356 if (subr == DIF_SUBR_DIRNAME) {
4357 ASSERT(lastdir == -1);
4358 src = (uintptr_t)".";
4365 if (subr == DIF_SUBR_DIRNAME) {
4366 if (lastdir == -1) {
4368 * We know that we have a slash in the name --
4369 * or lastdir would be set to 0, above. And
4370 * because lastdir is -1, we know that this
4371 * slash must be the first character. (That
4372 * is, the full string must be of the form
4373 * "/basename".) In this case, the last
4374 * character of the directory name is 0.
4382 ASSERT(subr == DIF_SUBR_BASENAME);
4383 ASSERT(firstbase != -1 && lastbase != -1);
4388 for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
4389 dest[j] = dtrace_load8(src + i);
4392 regs[rd] = (uintptr_t)dest;
4393 mstate->dtms_scratch_ptr += size;
4397 case DIF_SUBR_CLEANPATH: {
4398 char *dest = (char *)mstate->dtms_scratch_ptr, c;
4399 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4400 uintptr_t src = tupregs[0].dttk_value;
4403 if (!dtrace_strcanload(src, size, mstate, vstate)) {
4408 if (!DTRACE_INSCRATCH(mstate, size)) {
4409 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4415 * Move forward, loading each character.
4418 c = dtrace_load8(src + i++);
4420 if (j + 5 >= size) /* 5 = strlen("/..c\0") */
4428 c = dtrace_load8(src + i++);
4432 * We have two slashes -- we can just advance
4433 * to the next character.
4440 * This is not "." and it's not ".." -- we can
4441 * just store the "/" and this character and
4449 c = dtrace_load8(src + i++);
4453 * This is a "/./" component. We're not going
4454 * to store anything in the destination buffer;
4455 * we're just going to go to the next component.
4462 * This is not ".." -- we can just store the
4463 * "/." and this character and continue
4472 c = dtrace_load8(src + i++);
4474 if (c != '/' && c != '\0') {
4476 * This is not ".." -- it's "..[mumble]".
4477 * We'll store the "/.." and this character
4478 * and continue processing.
4488 * This is "/../" or "/..\0". We need to back up
4489 * our destination pointer until we find a "/".
4492 while (j != 0 && dest[--j] != '/')
4497 } while (c != '\0');
4500 regs[rd] = (uintptr_t)dest;
4501 mstate->dtms_scratch_ptr += size;
4505 case DIF_SUBR_INET_NTOA:
4506 case DIF_SUBR_INET_NTOA6:
4507 case DIF_SUBR_INET_NTOP: {
4512 if (subr == DIF_SUBR_INET_NTOP) {
4513 af = (int)tupregs[0].dttk_value;
4516 af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
4520 if (af == AF_INET) {
4525 * Safely load the IPv4 address.
4527 ip4 = dtrace_load32(tupregs[argi].dttk_value);
4530 * Check an IPv4 string will fit in scratch.
4532 size = INET_ADDRSTRLEN;
4533 if (!DTRACE_INSCRATCH(mstate, size)) {
4534 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4538 base = (char *)mstate->dtms_scratch_ptr;
4539 end = (char *)mstate->dtms_scratch_ptr + size - 1;
4542 * Stringify as a dotted decimal quad.
4545 ptr8 = (uint8_t *)&ip4;
4546 for (i = 3; i >= 0; i--) {
4552 for (; val; val /= 10) {
4553 *end-- = '0' + (val % 10);
4560 ASSERT(end + 1 >= base);
4562 } else if (af == AF_INET6) {
4563 struct in6_addr ip6;
4564 int firstzero, tryzero, numzero, v6end;
4566 const char digits[] = "0123456789abcdef";
4569 * Stringify using RFC 1884 convention 2 - 16 bit
4570 * hexadecimal values with a zero-run compression.
4571 * Lower case hexadecimal digits are used.
4572 * eg, fe80::214:4fff:fe0b:76c8.
4573 * The IPv4 embedded form is returned for inet_ntop,
4574 * just the IPv4 string is returned for inet_ntoa6.
4578 * Safely load the IPv6 address.
4581 (void *)(uintptr_t)tupregs[argi].dttk_value,
4582 (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));
4585 * Check an IPv6 string will fit in scratch.
4587 size = INET6_ADDRSTRLEN;
4588 if (!DTRACE_INSCRATCH(mstate, size)) {
4589 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4593 base = (char *)mstate->dtms_scratch_ptr;
4594 end = (char *)mstate->dtms_scratch_ptr + size - 1;
4598 * Find the longest run of 16 bit zero values
4599 * for the single allowed zero compression - "::".
4604 for (i = 0; i < sizeof (struct in6_addr); i++) {
4606 if (ip6._S6_un._S6_u8[i] == 0 &&
4608 if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
4610 tryzero == -1 && i % 2 == 0) {
4615 if (tryzero != -1 &&
4617 (ip6._S6_un._S6_u8[i] != 0 ||
4619 (ip6.__u6_addr.__u6_addr8[i] != 0 ||
4621 i == sizeof (struct in6_addr) - 1)) {
4623 if (i - tryzero <= numzero) {
4628 firstzero = tryzero;
4629 numzero = i - i % 2 - tryzero;
4633 if (ip6._S6_un._S6_u8[i] == 0 &&
4635 if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
4637 i == sizeof (struct in6_addr) - 1)
4641 ASSERT(firstzero + numzero <= sizeof (struct in6_addr));
4644 * Check for an IPv4 embedded address.
4646 v6end = sizeof (struct in6_addr) - 2;
4647 if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
4648 IN6_IS_ADDR_V4COMPAT(&ip6)) {
4649 for (i = sizeof (struct in6_addr) - 1;
4650 i >= DTRACE_V4MAPPED_OFFSET; i--) {
4651 ASSERT(end >= base);
4654 val = ip6._S6_un._S6_u8[i];
4656 val = ip6.__u6_addr.__u6_addr8[i];
4662 for (; val; val /= 10) {
4663 *end-- = '0' + val % 10;
4667 if (i > DTRACE_V4MAPPED_OFFSET)
4671 if (subr == DIF_SUBR_INET_NTOA6)
4675 * Set v6end to skip the IPv4 address that
4676 * we have already stringified.
4682 * Build the IPv6 string by working through the
4683 * address in reverse.
4685 for (i = v6end; i >= 0; i -= 2) {
4686 ASSERT(end >= base);
4688 if (i == firstzero + numzero - 2) {
4695 if (i < 14 && i != firstzero - 2)
4699 val = (ip6._S6_un._S6_u8[i] << 8) +
4700 ip6._S6_un._S6_u8[i + 1];
4702 val = (ip6.__u6_addr.__u6_addr8[i] << 8) +
4703 ip6.__u6_addr.__u6_addr8[i + 1];
4709 for (; val; val /= 16) {
4710 *end-- = digits[val % 16];
4714 ASSERT(end + 1 >= base);
4718 * The user didn't use AH_INET or AH_INET6.
4720 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
4725 inetout: regs[rd] = (uintptr_t)end + 1;
4726 mstate->dtms_scratch_ptr += size;
4730 case DIF_SUBR_MEMREF: {
4731 uintptr_t size = 2 * sizeof(uintptr_t);
4732 uintptr_t *memref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
4733 size_t scratch_size = ((uintptr_t) memref - mstate->dtms_scratch_ptr) + size;
4735 /* address and length */
4736 memref[0] = tupregs[0].dttk_value;
4737 memref[1] = tupregs[1].dttk_value;
4739 regs[rd] = (uintptr_t) memref;
4740 mstate->dtms_scratch_ptr += scratch_size;
4744 case DIF_SUBR_TYPEREF: {
4745 uintptr_t size = 4 * sizeof(uintptr_t);
4746 uintptr_t *typeref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
4747 size_t scratch_size = ((uintptr_t) typeref - mstate->dtms_scratch_ptr) + size;
4749 /* address, num_elements, type_str, type_len */
4750 typeref[0] = tupregs[0].dttk_value;
4751 typeref[1] = tupregs[1].dttk_value;
4752 typeref[2] = tupregs[2].dttk_value;
4753 typeref[3] = tupregs[3].dttk_value;
4755 regs[rd] = (uintptr_t) typeref;
4756 mstate->dtms_scratch_ptr += scratch_size;
4763 * Emulate the execution of DTrace IR instructions specified by the given
4764 * DIF object. This function is deliberately void of assertions as all of
4765 * the necessary checks are handled by a call to dtrace_difo_validate().
4768 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
4769 dtrace_vstate_t *vstate, dtrace_state_t *state)
4771 const dif_instr_t *text = difo->dtdo_buf;
4772 const uint_t textlen = difo->dtdo_len;
4773 const char *strtab = difo->dtdo_strtab;
4774 const uint64_t *inttab = difo->dtdo_inttab;
4777 dtrace_statvar_t *svar;
4778 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
4780 volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
4781 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
4783 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
4784 uint64_t regs[DIF_DIR_NREGS];
4787 uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
4789 uint_t pc = 0, id, opc = 0;
4795 * We stash the current DIF object into the machine state: we need it
4796 * for subsequent access checking.
4798 mstate->dtms_difo = difo;
4800 regs[DIF_REG_R0] = 0; /* %r0 is fixed at zero */
4802 while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
4806 r1 = DIF_INSTR_R1(instr);
4807 r2 = DIF_INSTR_R2(instr);
4808 rd = DIF_INSTR_RD(instr);
4810 switch (DIF_INSTR_OP(instr)) {
4812 regs[rd] = regs[r1] | regs[r2];
4815 regs[rd] = regs[r1] ^ regs[r2];
4818 regs[rd] = regs[r1] & regs[r2];
4821 regs[rd] = regs[r1] << regs[r2];
4824 regs[rd] = regs[r1] >> regs[r2];
4827 regs[rd] = regs[r1] - regs[r2];
4830 regs[rd] = regs[r1] + regs[r2];
4833 regs[rd] = regs[r1] * regs[r2];
4836 if (regs[r2] == 0) {
4838 *flags |= CPU_DTRACE_DIVZERO;
4840 regs[rd] = (int64_t)regs[r1] /
4846 if (regs[r2] == 0) {
4848 *flags |= CPU_DTRACE_DIVZERO;
4850 regs[rd] = regs[r1] / regs[r2];
4855 if (regs[r2] == 0) {
4857 *flags |= CPU_DTRACE_DIVZERO;
4859 regs[rd] = (int64_t)regs[r1] %
4865 if (regs[r2] == 0) {
4867 *flags |= CPU_DTRACE_DIVZERO;
4869 regs[rd] = regs[r1] % regs[r2];
4874 regs[rd] = ~regs[r1];
4877 regs[rd] = regs[r1];
4880 cc_r = regs[r1] - regs[r2];
4884 cc_c = regs[r1] < regs[r2];
4887 cc_n = cc_v = cc_c = 0;
4888 cc_z = regs[r1] == 0;
4891 pc = DIF_INSTR_LABEL(instr);
4895 pc = DIF_INSTR_LABEL(instr);
4899 pc = DIF_INSTR_LABEL(instr);
4902 if ((cc_z | (cc_n ^ cc_v)) == 0)
4903 pc = DIF_INSTR_LABEL(instr);
4906 if ((cc_c | cc_z) == 0)
4907 pc = DIF_INSTR_LABEL(instr);
4910 if ((cc_n ^ cc_v) == 0)
4911 pc = DIF_INSTR_LABEL(instr);
4915 pc = DIF_INSTR_LABEL(instr);
4919 pc = DIF_INSTR_LABEL(instr);
4923 pc = DIF_INSTR_LABEL(instr);
4926 if (cc_z | (cc_n ^ cc_v))
4927 pc = DIF_INSTR_LABEL(instr);
4931 pc = DIF_INSTR_LABEL(instr);
4934 if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
4935 *flags |= CPU_DTRACE_KPRIV;
4941 regs[rd] = (int8_t)dtrace_load8(regs[r1]);
4944 if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
4945 *flags |= CPU_DTRACE_KPRIV;
4951 regs[rd] = (int16_t)dtrace_load16(regs[r1]);
4954 if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
4955 *flags |= CPU_DTRACE_KPRIV;
4961 regs[rd] = (int32_t)dtrace_load32(regs[r1]);
4964 if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
4965 *flags |= CPU_DTRACE_KPRIV;
4971 regs[rd] = dtrace_load8(regs[r1]);
4974 if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
4975 *flags |= CPU_DTRACE_KPRIV;
4981 regs[rd] = dtrace_load16(regs[r1]);
4984 if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
4985 *flags |= CPU_DTRACE_KPRIV;
4991 regs[rd] = dtrace_load32(regs[r1]);
4994 if (!dtrace_canstore(regs[r1], 8, mstate, vstate)) {
4995 *flags |= CPU_DTRACE_KPRIV;
5001 regs[rd] = dtrace_load64(regs[r1]);
5005 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5008 regs[rd] = (int16_t)
5009 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5012 regs[rd] = (int32_t)
5013 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5017 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5021 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5025 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5029 dtrace_fuword64((void *)(uintptr_t)regs[r1]);
5038 regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
5041 regs[rd] = (uint64_t)(uintptr_t)
5042 (strtab + DIF_INSTR_STRING(instr));
5045 size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
5046 uintptr_t s1 = regs[r1];
5047 uintptr_t s2 = regs[r2];
5050 !dtrace_strcanload(s1, sz, mstate, vstate))
5053 !dtrace_strcanload(s2, sz, mstate, vstate))
5056 cc_r = dtrace_strncmp((char *)s1, (char *)s2, sz);
5064 regs[rd] = dtrace_dif_variable(mstate, state,
5068 id = DIF_INSTR_VAR(instr);
5070 if (id >= DIF_VAR_OTHER_UBASE) {
5073 id -= DIF_VAR_OTHER_UBASE;
5074 svar = vstate->dtvs_globals[id];
5075 ASSERT(svar != NULL);
5076 v = &svar->dtsv_var;
5078 if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
5079 regs[rd] = svar->dtsv_data;
5083 a = (uintptr_t)svar->dtsv_data;
5085 if (*(uint8_t *)a == UINT8_MAX) {
5087 * If the 0th byte is set to UINT8_MAX
5088 * then this is to be treated as a
5089 * reference to a NULL variable.
5093 regs[rd] = a + sizeof (uint64_t);
5099 regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
5103 id = DIF_INSTR_VAR(instr);
5105 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5106 id -= DIF_VAR_OTHER_UBASE;
5108 svar = vstate->dtvs_globals[id];
5109 ASSERT(svar != NULL);
5110 v = &svar->dtsv_var;
5112 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5113 uintptr_t a = (uintptr_t)svar->dtsv_data;
5116 ASSERT(svar->dtsv_size != 0);
5118 if (regs[rd] == 0) {
5119 *(uint8_t *)a = UINT8_MAX;
5123 a += sizeof (uint64_t);
5125 if (!dtrace_vcanload(
5126 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5130 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5131 (void *)a, &v->dtdv_type);
5135 svar->dtsv_data = regs[rd];
5140 * There are no DTrace built-in thread-local arrays at
5141 * present. This opcode is saved for future work.
5143 *flags |= CPU_DTRACE_ILLOP;
5148 id = DIF_INSTR_VAR(instr);
5150 if (id < DIF_VAR_OTHER_UBASE) {
5152 * For now, this has no meaning.
5158 id -= DIF_VAR_OTHER_UBASE;
5160 ASSERT(id < vstate->dtvs_nlocals);
5161 ASSERT(vstate->dtvs_locals != NULL);
5163 svar = vstate->dtvs_locals[id];
5164 ASSERT(svar != NULL);
5165 v = &svar->dtsv_var;
5167 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5168 uintptr_t a = (uintptr_t)svar->dtsv_data;
5169 size_t sz = v->dtdv_type.dtdt_size;
5171 sz += sizeof (uint64_t);
5172 ASSERT(svar->dtsv_size == NCPU * sz);
5175 if (*(uint8_t *)a == UINT8_MAX) {
5177 * If the 0th byte is set to UINT8_MAX
5178 * then this is to be treated as a
5179 * reference to a NULL variable.
5183 regs[rd] = a + sizeof (uint64_t);
5189 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5190 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5191 regs[rd] = tmp[curcpu];
5195 id = DIF_INSTR_VAR(instr);
5197 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5198 id -= DIF_VAR_OTHER_UBASE;
5199 ASSERT(id < vstate->dtvs_nlocals);
5201 ASSERT(vstate->dtvs_locals != NULL);
5202 svar = vstate->dtvs_locals[id];
5203 ASSERT(svar != NULL);
5204 v = &svar->dtsv_var;
5206 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5207 uintptr_t a = (uintptr_t)svar->dtsv_data;
5208 size_t sz = v->dtdv_type.dtdt_size;
5210 sz += sizeof (uint64_t);
5211 ASSERT(svar->dtsv_size == NCPU * sz);
5214 if (regs[rd] == 0) {
5215 *(uint8_t *)a = UINT8_MAX;
5219 a += sizeof (uint64_t);
5222 if (!dtrace_vcanload(
5223 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5227 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5228 (void *)a, &v->dtdv_type);
5232 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5233 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5234 tmp[curcpu] = regs[rd];
5238 dtrace_dynvar_t *dvar;
5241 id = DIF_INSTR_VAR(instr);
5242 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5243 id -= DIF_VAR_OTHER_UBASE;
5244 v = &vstate->dtvs_tlocals[id];
5246 key = &tupregs[DIF_DTR_NREGS];
5247 key[0].dttk_value = (uint64_t)id;
5248 key[0].dttk_size = 0;
5249 DTRACE_TLS_THRKEY(key[1].dttk_value);
5250 key[1].dttk_size = 0;
5252 dvar = dtrace_dynvar(dstate, 2, key,
5253 sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
5261 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5262 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5264 regs[rd] = *((uint64_t *)dvar->dtdv_data);
5271 dtrace_dynvar_t *dvar;
5274 id = DIF_INSTR_VAR(instr);
5275 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5276 id -= DIF_VAR_OTHER_UBASE;
5278 key = &tupregs[DIF_DTR_NREGS];
5279 key[0].dttk_value = (uint64_t)id;
5280 key[0].dttk_size = 0;
5281 DTRACE_TLS_THRKEY(key[1].dttk_value);
5282 key[1].dttk_size = 0;
5283 v = &vstate->dtvs_tlocals[id];
5285 dvar = dtrace_dynvar(dstate, 2, key,
5286 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5287 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5288 regs[rd] ? DTRACE_DYNVAR_ALLOC :
5289 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5292 * Given that we're storing to thread-local data,
5293 * we need to flush our predicate cache.
5295 curthread->t_predcache = 0;
5300 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5301 if (!dtrace_vcanload(
5302 (void *)(uintptr_t)regs[rd],
5303 &v->dtdv_type, mstate, vstate))
5306 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5307 dvar->dtdv_data, &v->dtdv_type);
5309 *((uint64_t *)dvar->dtdv_data) = regs[rd];
5316 regs[rd] = (int64_t)regs[r1] >> regs[r2];
5320 dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
5321 regs, tupregs, ttop, mstate, state);
5325 if (ttop == DIF_DTR_NREGS) {
5326 *flags |= CPU_DTRACE_TUPOFLOW;
5330 if (r1 == DIF_TYPE_STRING) {
5332 * If this is a string type and the size is 0,
5333 * we'll use the system-wide default string
5334 * size. Note that we are _not_ looking at
5335 * the value of the DTRACEOPT_STRSIZE option;
5336 * had this been set, we would expect to have
5337 * a non-zero size value in the "pushtr".
5339 tupregs[ttop].dttk_size =
5340 dtrace_strlen((char *)(uintptr_t)regs[rd],
5341 regs[r2] ? regs[r2] :
5342 dtrace_strsize_default) + 1;
5344 tupregs[ttop].dttk_size = regs[r2];
5347 tupregs[ttop++].dttk_value = regs[rd];
5351 if (ttop == DIF_DTR_NREGS) {
5352 *flags |= CPU_DTRACE_TUPOFLOW;
5356 tupregs[ttop].dttk_value = regs[rd];
5357 tupregs[ttop++].dttk_size = 0;
5365 case DIF_OP_FLUSHTS:
5370 case DIF_OP_LDTAA: {
5371 dtrace_dynvar_t *dvar;
5372 dtrace_key_t *key = tupregs;
5373 uint_t nkeys = ttop;
5375 id = DIF_INSTR_VAR(instr);
5376 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5377 id -= DIF_VAR_OTHER_UBASE;
5379 key[nkeys].dttk_value = (uint64_t)id;
5380 key[nkeys++].dttk_size = 0;
5382 if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
5383 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5384 key[nkeys++].dttk_size = 0;
5385 v = &vstate->dtvs_tlocals[id];
5387 v = &vstate->dtvs_globals[id]->dtsv_var;
5390 dvar = dtrace_dynvar(dstate, nkeys, key,
5391 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5392 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5393 DTRACE_DYNVAR_NOALLOC, mstate, vstate);
5400 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5401 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5403 regs[rd] = *((uint64_t *)dvar->dtdv_data);
5410 case DIF_OP_STTAA: {
5411 dtrace_dynvar_t *dvar;
5412 dtrace_key_t *key = tupregs;
5413 uint_t nkeys = ttop;
5415 id = DIF_INSTR_VAR(instr);
5416 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5417 id -= DIF_VAR_OTHER_UBASE;
5419 key[nkeys].dttk_value = (uint64_t)id;
5420 key[nkeys++].dttk_size = 0;
5422 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
5423 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5424 key[nkeys++].dttk_size = 0;
5425 v = &vstate->dtvs_tlocals[id];
5427 v = &vstate->dtvs_globals[id]->dtsv_var;
5430 dvar = dtrace_dynvar(dstate, nkeys, key,
5431 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5432 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5433 regs[rd] ? DTRACE_DYNVAR_ALLOC :
5434 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5439 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5440 if (!dtrace_vcanload(
5441 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5445 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5446 dvar->dtdv_data, &v->dtdv_type);
5448 *((uint64_t *)dvar->dtdv_data) = regs[rd];
5454 case DIF_OP_ALLOCS: {
5455 uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5456 size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
5459 * Rounding up the user allocation size could have
5460 * overflowed large, bogus allocations (like -1ULL) to
5463 if (size < regs[r1] ||
5464 !DTRACE_INSCRATCH(mstate, size)) {
5465 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5470 dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
5471 mstate->dtms_scratch_ptr += size;
5477 if (!dtrace_canstore(regs[rd], regs[r2],
5479 *flags |= CPU_DTRACE_BADADDR;
5484 if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
5487 dtrace_bcopy((void *)(uintptr_t)regs[r1],
5488 (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
5492 if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
5493 *flags |= CPU_DTRACE_BADADDR;
5497 *((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
5501 if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
5502 *flags |= CPU_DTRACE_BADADDR;
5507 *flags |= CPU_DTRACE_BADALIGN;
5511 *((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
5515 if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
5516 *flags |= CPU_DTRACE_BADADDR;
5521 *flags |= CPU_DTRACE_BADALIGN;
5525 *((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
5529 if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
5530 *flags |= CPU_DTRACE_BADADDR;
5535 *flags |= CPU_DTRACE_BADALIGN;
5539 *((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
5544 if (!(*flags & CPU_DTRACE_FAULT))
5547 mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
5548 mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
5554 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
5556 dtrace_probe_t *probe = ecb->dte_probe;
5557 dtrace_provider_t *prov = probe->dtpr_provider;
5558 char c[DTRACE_FULLNAMELEN + 80], *str;
5559 char *msg = "dtrace: breakpoint action at probe ";
5560 char *ecbmsg = " (ecb ";
5561 uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
5562 uintptr_t val = (uintptr_t)ecb;
5563 int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
5565 if (dtrace_destructive_disallow)
5569 * It's impossible to be taking action on the NULL probe.
5571 ASSERT(probe != NULL);
5574 * This is a poor man's (destitute man's?) sprintf(): we want to
5575 * print the provider name, module name, function name and name of
5576 * the probe, along with the hex address of the ECB with the breakpoint
5577 * action -- all of which we must place in the character buffer by
5580 while (*msg != '\0')
5583 for (str = prov->dtpv_name; *str != '\0'; str++)
5587 for (str = probe->dtpr_mod; *str != '\0'; str++)
5591 for (str = probe->dtpr_func; *str != '\0'; str++)
5595 for (str = probe->dtpr_name; *str != '\0'; str++)
5598 while (*ecbmsg != '\0')
5601 while (shift >= 0) {
5602 mask = (uintptr_t)0xf << shift;
5604 if (val >= ((uintptr_t)1 << shift))
5605 c[i++] = "0123456789abcdef"[(val & mask) >> shift];
5615 kdb_enter(KDB_WHY_DTRACE, "breakpoint action");
5620 dtrace_action_panic(dtrace_ecb_t *ecb)
5622 dtrace_probe_t *probe = ecb->dte_probe;
5625 * It's impossible to be taking action on the NULL probe.
5627 ASSERT(probe != NULL);
5629 if (dtrace_destructive_disallow)
5632 if (dtrace_panicked != NULL)
5635 if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
5639 * We won the right to panic. (We want to be sure that only one
5640 * thread calls panic() from dtrace_probe(), and that panic() is
5641 * called exactly once.)
5643 dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
5644 probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
5645 probe->dtpr_func, probe->dtpr_name, (void *)ecb);
5649 dtrace_action_raise(uint64_t sig)
5651 if (dtrace_destructive_disallow)
5655 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
5661 * raise() has a queue depth of 1 -- we ignore all subsequent
5662 * invocations of the raise() action.
5664 if (curthread->t_dtrace_sig == 0)
5665 curthread->t_dtrace_sig = (uint8_t)sig;
5667 curthread->t_sig_check = 1;
5670 struct proc *p = curproc;
5672 kern_psignal(p, sig);
5678 dtrace_action_stop(void)
5680 if (dtrace_destructive_disallow)
5684 if (!curthread->t_dtrace_stop) {
5685 curthread->t_dtrace_stop = 1;
5686 curthread->t_sig_check = 1;
5690 struct proc *p = curproc;
5692 kern_psignal(p, SIGSTOP);
5698 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
5701 volatile uint16_t *flags;
5705 cpu_t *cpu = &solaris_cpu[curcpu];
5708 if (dtrace_destructive_disallow)
5711 flags = (volatile uint16_t *)&cpu_core[cpu->cpu_id].cpuc_dtrace_flags;
5713 now = dtrace_gethrtime();
5715 if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
5717 * We need to advance the mark to the current time.
5719 cpu->cpu_dtrace_chillmark = now;
5720 cpu->cpu_dtrace_chilled = 0;
5724 * Now check to see if the requested chill time would take us over
5725 * the maximum amount of time allowed in the chill interval. (Or
5726 * worse, if the calculation itself induces overflow.)
5728 if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
5729 cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
5730 *flags |= CPU_DTRACE_ILLOP;
5734 while (dtrace_gethrtime() - now < val)
5738 * Normally, we assure that the value of the variable "timestamp" does
5739 * not change within an ECB. The presence of chill() represents an
5740 * exception to this rule, however.
5742 mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
5743 cpu->cpu_dtrace_chilled += val;
5747 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
5748 uint64_t *buf, uint64_t arg)
5750 int nframes = DTRACE_USTACK_NFRAMES(arg);
5751 int strsize = DTRACE_USTACK_STRSIZE(arg);
5752 uint64_t *pcs = &buf[1], *fps;
5753 char *str = (char *)&pcs[nframes];
5754 int size, offs = 0, i, j;
5755 uintptr_t old = mstate->dtms_scratch_ptr, saved;
5756 uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
5760 * Should be taking a faster path if string space has not been
5763 ASSERT(strsize != 0);
5766 * We will first allocate some temporary space for the frame pointers.
5768 fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5769 size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
5770 (nframes * sizeof (uint64_t));
5772 if (!DTRACE_INSCRATCH(mstate, size)) {
5774 * Not enough room for our frame pointers -- need to indicate
5775 * that we ran out of scratch space.
5777 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5781 mstate->dtms_scratch_ptr += size;
5782 saved = mstate->dtms_scratch_ptr;
5785 * Now get a stack with both program counters and frame pointers.
5787 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5788 dtrace_getufpstack(buf, fps, nframes + 1);
5789 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5792 * If that faulted, we're cooked.
5794 if (*flags & CPU_DTRACE_FAULT)
5798 * Now we want to walk up the stack, calling the USTACK helper. For
5799 * each iteration, we restore the scratch pointer.
5801 for (i = 0; i < nframes; i++) {
5802 mstate->dtms_scratch_ptr = saved;
5804 if (offs >= strsize)
5807 sym = (char *)(uintptr_t)dtrace_helper(
5808 DTRACE_HELPER_ACTION_USTACK,
5809 mstate, state, pcs[i], fps[i]);
5812 * If we faulted while running the helper, we're going to
5813 * clear the fault and null out the corresponding string.
5815 if (*flags & CPU_DTRACE_FAULT) {
5816 *flags &= ~CPU_DTRACE_FAULT;
5826 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5829 * Now copy in the string that the helper returned to us.
5831 for (j = 0; offs + j < strsize; j++) {
5832 if ((str[offs + j] = sym[j]) == '\0')
5836 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5841 if (offs >= strsize) {
5843 * If we didn't have room for all of the strings, we don't
5844 * abort processing -- this needn't be a fatal error -- but we
5845 * still want to increment a counter (dts_stkstroverflows) to
5846 * allow this condition to be warned about. (If this is from
5847 * a jstack() action, it is easily tuned via jstackstrsize.)
5849 dtrace_error(&state->dts_stkstroverflows);
5852 while (offs < strsize)
5856 mstate->dtms_scratch_ptr = old;
5860 * If you're looking for the epicenter of DTrace, you just found it. This
5861 * is the function called by the provider to fire a probe -- from which all
5862 * subsequent probe-context DTrace activity emanates.
5865 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
5866 uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
5868 processorid_t cpuid;
5869 dtrace_icookie_t cookie;
5870 dtrace_probe_t *probe;
5871 dtrace_mstate_t mstate;
5873 dtrace_action_t *act;
5877 volatile uint16_t *flags;
5882 * Kick out immediately if this CPU is still being born (in which case
5883 * curthread will be set to -1) or the current thread can't allow
5884 * probes in its current context.
5886 if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
5890 cookie = dtrace_interrupt_disable();
5891 probe = dtrace_probes[id - 1];
5893 onintr = CPU_ON_INTR(CPU);
5895 if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
5896 probe->dtpr_predcache == curthread->t_predcache) {
5898 * We have hit in the predicate cache; we know that
5899 * this predicate would evaluate to be false.
5901 dtrace_interrupt_enable(cookie);
5906 if (panic_quiesce) {
5908 if (panicstr != NULL) {
5911 * We don't trace anything if we're panicking.
5913 dtrace_interrupt_enable(cookie);
5917 now = dtrace_gethrtime();
5918 vtime = dtrace_vtime_references != 0;
5920 if (vtime && curthread->t_dtrace_start)
5921 curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
5923 mstate.dtms_difo = NULL;
5924 mstate.dtms_probe = probe;
5925 mstate.dtms_strtok = 0;
5926 mstate.dtms_arg[0] = arg0;
5927 mstate.dtms_arg[1] = arg1;
5928 mstate.dtms_arg[2] = arg2;
5929 mstate.dtms_arg[3] = arg3;
5930 mstate.dtms_arg[4] = arg4;
5932 flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
5934 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
5935 dtrace_predicate_t *pred = ecb->dte_predicate;
5936 dtrace_state_t *state = ecb->dte_state;
5937 dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
5938 dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
5939 dtrace_vstate_t *vstate = &state->dts_vstate;
5940 dtrace_provider_t *prov = probe->dtpr_provider;
5945 * A little subtlety with the following (seemingly innocuous)
5946 * declaration of the automatic 'val': by looking at the
5947 * code, you might think that it could be declared in the
5948 * action processing loop, below. (That is, it's only used in
5949 * the action processing loop.) However, it must be declared
5950 * out of that scope because in the case of DIF expression
5951 * arguments to aggregating actions, one iteration of the
5952 * action loop will use the last iteration's value.
5956 mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
5957 *flags &= ~CPU_DTRACE_ERROR;
5959 if (prov == dtrace_provider) {
5961 * If dtrace itself is the provider of this probe,
5962 * we're only going to continue processing the ECB if
5963 * arg0 (the dtrace_state_t) is equal to the ECB's
5964 * creating state. (This prevents disjoint consumers
5965 * from seeing one another's metaprobes.)
5967 if (arg0 != (uint64_t)(uintptr_t)state)
5971 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
5973 * We're not currently active. If our provider isn't
5974 * the dtrace pseudo provider, we're not interested.
5976 if (prov != dtrace_provider)
5980 * Now we must further check if we are in the BEGIN
5981 * probe. If we are, we will only continue processing
5982 * if we're still in WARMUP -- if one BEGIN enabling
5983 * has invoked the exit() action, we don't want to
5984 * evaluate subsequent BEGIN enablings.
5986 if (probe->dtpr_id == dtrace_probeid_begin &&
5987 state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
5988 ASSERT(state->dts_activity ==
5989 DTRACE_ACTIVITY_DRAINING);
5994 if (ecb->dte_cond) {
5996 * If the dte_cond bits indicate that this
5997 * consumer is only allowed to see user-mode firings
5998 * of this probe, call the provider's dtps_usermode()
5999 * entry point to check that the probe was fired
6000 * while in a user context. Skip this ECB if that's
6003 if ((ecb->dte_cond & DTRACE_COND_USERMODE) &&
6004 prov->dtpv_pops.dtps_usermode(prov->dtpv_arg,
6005 probe->dtpr_id, probe->dtpr_arg) == 0)
6010 * This is more subtle than it looks. We have to be
6011 * absolutely certain that CRED() isn't going to
6012 * change out from under us so it's only legit to
6013 * examine that structure if we're in constrained
6014 * situations. Currently, the only times we'll this
6015 * check is if a non-super-user has enabled the
6016 * profile or syscall providers -- providers that
6017 * allow visibility of all processes. For the
6018 * profile case, the check above will ensure that
6019 * we're examining a user context.
6021 if (ecb->dte_cond & DTRACE_COND_OWNER) {
6024 ecb->dte_state->dts_cred.dcr_cred;
6027 ASSERT(s_cr != NULL);
6029 if ((cr = CRED()) == NULL ||
6030 s_cr->cr_uid != cr->cr_uid ||
6031 s_cr->cr_uid != cr->cr_ruid ||
6032 s_cr->cr_uid != cr->cr_suid ||
6033 s_cr->cr_gid != cr->cr_gid ||
6034 s_cr->cr_gid != cr->cr_rgid ||
6035 s_cr->cr_gid != cr->cr_sgid ||
6036 (proc = ttoproc(curthread)) == NULL ||
6037 (proc->p_flag & SNOCD))
6041 if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
6044 ecb->dte_state->dts_cred.dcr_cred;
6046 ASSERT(s_cr != NULL);
6048 if ((cr = CRED()) == NULL ||
6049 s_cr->cr_zone->zone_id !=
6050 cr->cr_zone->zone_id)
6056 if (now - state->dts_alive > dtrace_deadman_timeout) {
6058 * We seem to be dead. Unless we (a) have kernel
6059 * destructive permissions (b) have expicitly enabled
6060 * destructive actions and (c) destructive actions have
6061 * not been disabled, we're going to transition into
6062 * the KILLED state, from which no further processing
6063 * on this state will be performed.
6065 if (!dtrace_priv_kernel_destructive(state) ||
6066 !state->dts_cred.dcr_destructive ||
6067 dtrace_destructive_disallow) {
6068 void *activity = &state->dts_activity;
6069 dtrace_activity_t current;
6072 current = state->dts_activity;
6073 } while (dtrace_cas32(activity, current,
6074 DTRACE_ACTIVITY_KILLED) != current);
6080 if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
6081 ecb->dte_alignment, state, &mstate)) < 0)
6084 tomax = buf->dtb_tomax;
6085 ASSERT(tomax != NULL);
6087 if (ecb->dte_size != 0)
6088 DTRACE_STORE(uint32_t, tomax, offs, ecb->dte_epid);
6090 mstate.dtms_epid = ecb->dte_epid;
6091 mstate.dtms_present |= DTRACE_MSTATE_EPID;
6093 if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
6094 mstate.dtms_access = DTRACE_ACCESS_KERNEL;
6096 mstate.dtms_access = 0;
6099 dtrace_difo_t *dp = pred->dtp_difo;
6102 rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
6104 if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
6105 dtrace_cacheid_t cid = probe->dtpr_predcache;
6107 if (cid != DTRACE_CACHEIDNONE && !onintr) {
6109 * Update the predicate cache...
6111 ASSERT(cid == pred->dtp_cacheid);
6112 curthread->t_predcache = cid;
6119 for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
6120 act != NULL; act = act->dta_next) {
6123 dtrace_recdesc_t *rec = &act->dta_rec;
6125 size = rec->dtrd_size;
6126 valoffs = offs + rec->dtrd_offset;
6128 if (DTRACEACT_ISAGG(act->dta_kind)) {
6130 dtrace_aggregation_t *agg;
6132 agg = (dtrace_aggregation_t *)act;
6134 if ((dp = act->dta_difo) != NULL)
6135 v = dtrace_dif_emulate(dp,
6136 &mstate, vstate, state);
6138 if (*flags & CPU_DTRACE_ERROR)
6142 * Note that we always pass the expression
6143 * value from the previous iteration of the
6144 * action loop. This value will only be used
6145 * if there is an expression argument to the
6146 * aggregating action, denoted by the
6147 * dtag_hasarg field.
6149 dtrace_aggregate(agg, buf,
6150 offs, aggbuf, v, val);
6154 switch (act->dta_kind) {
6155 case DTRACEACT_STOP:
6156 if (dtrace_priv_proc_destructive(state))
6157 dtrace_action_stop();
6160 case DTRACEACT_BREAKPOINT:
6161 if (dtrace_priv_kernel_destructive(state))
6162 dtrace_action_breakpoint(ecb);
6165 case DTRACEACT_PANIC:
6166 if (dtrace_priv_kernel_destructive(state))
6167 dtrace_action_panic(ecb);
6170 case DTRACEACT_STACK:
6171 if (!dtrace_priv_kernel(state))
6174 dtrace_getpcstack((pc_t *)(tomax + valoffs),
6175 size / sizeof (pc_t), probe->dtpr_aframes,
6176 DTRACE_ANCHORED(probe) ? NULL :
6180 case DTRACEACT_JSTACK:
6181 case DTRACEACT_USTACK:
6182 if (!dtrace_priv_proc(state))
6186 * See comment in DIF_VAR_PID.
6188 if (DTRACE_ANCHORED(mstate.dtms_probe) &&
6190 int depth = DTRACE_USTACK_NFRAMES(
6193 dtrace_bzero((void *)(tomax + valoffs),
6194 DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
6195 + depth * sizeof (uint64_t));
6200 if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
6201 curproc->p_dtrace_helpers != NULL) {
6203 * This is the slow path -- we have
6204 * allocated string space, and we're
6205 * getting the stack of a process that
6206 * has helpers. Call into a separate
6207 * routine to perform this processing.
6209 dtrace_action_ustack(&mstate, state,
6210 (uint64_t *)(tomax + valoffs),
6215 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6216 dtrace_getupcstack((uint64_t *)
6218 DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
6219 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6229 val = dtrace_dif_emulate(dp, &mstate, vstate, state);
6231 if (*flags & CPU_DTRACE_ERROR)
6234 switch (act->dta_kind) {
6235 case DTRACEACT_SPECULATE:
6236 ASSERT(buf == &state->dts_buffer[cpuid]);
6237 buf = dtrace_speculation_buffer(state,
6241 *flags |= CPU_DTRACE_DROP;
6245 offs = dtrace_buffer_reserve(buf,
6246 ecb->dte_needed, ecb->dte_alignment,
6250 *flags |= CPU_DTRACE_DROP;
6254 tomax = buf->dtb_tomax;
6255 ASSERT(tomax != NULL);
6257 if (ecb->dte_size != 0)
6258 DTRACE_STORE(uint32_t, tomax, offs,
6262 case DTRACEACT_PRINTM: {
6263 /* The DIF returns a 'memref'. */
6264 uintptr_t *memref = (uintptr_t *)(uintptr_t) val;
6266 /* Get the size from the memref. */
6270 * Check if the size exceeds the allocated
6273 if (size + sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
6275 *flags |= CPU_DTRACE_DROP;
6279 /* Store the size in the buffer first. */
6280 DTRACE_STORE(uintptr_t, tomax,
6284 * Offset the buffer address to the start
6287 valoffs += sizeof(uintptr_t);
6290 * Reset to the memory address rather than
6291 * the memref array, then let the BYREF
6292 * code below do the work to store the
6293 * memory data in the buffer.
6299 case DTRACEACT_PRINTT: {
6300 /* The DIF returns a 'typeref'. */
6301 uintptr_t *typeref = (uintptr_t *)(uintptr_t) val;
6306 * Get the type string length and round it
6307 * up so that the data that follows is
6308 * aligned for easy access.
6310 size_t typs = strlen((char *) typeref[2]) + 1;
6311 typs = roundup(typs, sizeof(uintptr_t));
6314 *Get the size from the typeref using the
6315 * number of elements and the type size.
6317 size = typeref[1] * typeref[3];
6320 * Check if the size exceeds the allocated
6323 if (size + typs + 2 * sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
6325 *flags |= CPU_DTRACE_DROP;
6329 /* Store the size in the buffer first. */
6330 DTRACE_STORE(uintptr_t, tomax,
6332 valoffs += sizeof(uintptr_t);
6334 /* Store the type size in the buffer. */
6335 DTRACE_STORE(uintptr_t, tomax,
6336 valoffs, typeref[3]);
6337 valoffs += sizeof(uintptr_t);
6341 for (s = 0; s < typs; s++) {
6343 c = dtrace_load8(val++);
6345 DTRACE_STORE(uint8_t, tomax,
6350 * Reset to the memory address rather than
6351 * the typeref array, then let the BYREF
6352 * code below do the work to store the
6353 * memory data in the buffer.
6359 case DTRACEACT_CHILL:
6360 if (dtrace_priv_kernel_destructive(state))
6361 dtrace_action_chill(&mstate, val);
6364 case DTRACEACT_RAISE:
6365 if (dtrace_priv_proc_destructive(state))
6366 dtrace_action_raise(val);
6369 case DTRACEACT_COMMIT:
6373 * We need to commit our buffer state.
6376 buf->dtb_offset = offs + ecb->dte_size;
6377 buf = &state->dts_buffer[cpuid];
6378 dtrace_speculation_commit(state, cpuid, val);
6382 case DTRACEACT_DISCARD:
6383 dtrace_speculation_discard(state, cpuid, val);
6386 case DTRACEACT_DIFEXPR:
6387 case DTRACEACT_LIBACT:
6388 case DTRACEACT_PRINTF:
6389 case DTRACEACT_PRINTA:
6390 case DTRACEACT_SYSTEM:
6391 case DTRACEACT_FREOPEN:
6396 if (!dtrace_priv_kernel(state))
6400 case DTRACEACT_USYM:
6401 case DTRACEACT_UMOD:
6402 case DTRACEACT_UADDR: {
6404 struct pid *pid = curthread->t_procp->p_pidp;
6407 if (!dtrace_priv_proc(state))
6410 DTRACE_STORE(uint64_t, tomax,
6412 valoffs, (uint64_t)pid->pid_id);
6414 valoffs, (uint64_t) curproc->p_pid);
6416 DTRACE_STORE(uint64_t, tomax,
6417 valoffs + sizeof (uint64_t), val);
6422 case DTRACEACT_EXIT: {
6424 * For the exit action, we are going to attempt
6425 * to atomically set our activity to be
6426 * draining. If this fails (either because
6427 * another CPU has beat us to the exit action,
6428 * or because our current activity is something
6429 * other than ACTIVE or WARMUP), we will
6430 * continue. This assures that the exit action
6431 * can be successfully recorded at most once
6432 * when we're in the ACTIVE state. If we're
6433 * encountering the exit() action while in
6434 * COOLDOWN, however, we want to honor the new
6435 * status code. (We know that we're the only
6436 * thread in COOLDOWN, so there is no race.)
6438 void *activity = &state->dts_activity;
6439 dtrace_activity_t current = state->dts_activity;
6441 if (current == DTRACE_ACTIVITY_COOLDOWN)
6444 if (current != DTRACE_ACTIVITY_WARMUP)
6445 current = DTRACE_ACTIVITY_ACTIVE;
6447 if (dtrace_cas32(activity, current,
6448 DTRACE_ACTIVITY_DRAINING) != current) {
6449 *flags |= CPU_DTRACE_DROP;
6460 if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF) {
6461 uintptr_t end = valoffs + size;
6463 if (!dtrace_vcanload((void *)(uintptr_t)val,
6464 &dp->dtdo_rtype, &mstate, vstate))
6468 * If this is a string, we're going to only
6469 * load until we find the zero byte -- after
6470 * which we'll store zero bytes.
6472 if (dp->dtdo_rtype.dtdt_kind ==
6475 int intuple = act->dta_intuple;
6478 for (s = 0; s < size; s++) {
6480 c = dtrace_load8(val++);
6482 DTRACE_STORE(uint8_t, tomax,
6485 if (c == '\0' && intuple)
6492 while (valoffs < end) {
6493 DTRACE_STORE(uint8_t, tomax, valoffs++,
6494 dtrace_load8(val++));
6504 case sizeof (uint8_t):
6505 DTRACE_STORE(uint8_t, tomax, valoffs, val);
6507 case sizeof (uint16_t):
6508 DTRACE_STORE(uint16_t, tomax, valoffs, val);
6510 case sizeof (uint32_t):
6511 DTRACE_STORE(uint32_t, tomax, valoffs, val);
6513 case sizeof (uint64_t):
6514 DTRACE_STORE(uint64_t, tomax, valoffs, val);
6518 * Any other size should have been returned by
6519 * reference, not by value.
6526 if (*flags & CPU_DTRACE_DROP)
6529 if (*flags & CPU_DTRACE_FAULT) {
6531 dtrace_action_t *err;
6535 if (probe->dtpr_id == dtrace_probeid_error) {
6537 * There's nothing we can do -- we had an
6538 * error on the error probe. We bump an
6539 * error counter to at least indicate that
6540 * this condition happened.
6542 dtrace_error(&state->dts_dblerrors);
6548 * Before recursing on dtrace_probe(), we
6549 * need to explicitly clear out our start
6550 * time to prevent it from being accumulated
6551 * into t_dtrace_vtime.
6553 curthread->t_dtrace_start = 0;
6557 * Iterate over the actions to figure out which action
6558 * we were processing when we experienced the error.
6559 * Note that act points _past_ the faulting action; if
6560 * act is ecb->dte_action, the fault was in the
6561 * predicate, if it's ecb->dte_action->dta_next it's
6562 * in action #1, and so on.
6564 for (err = ecb->dte_action, ndx = 0;
6565 err != act; err = err->dta_next, ndx++)
6568 dtrace_probe_error(state, ecb->dte_epid, ndx,
6569 (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
6570 mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
6571 cpu_core[cpuid].cpuc_dtrace_illval);
6577 buf->dtb_offset = offs + ecb->dte_size;
6581 curthread->t_dtrace_start = dtrace_gethrtime();
6583 dtrace_interrupt_enable(cookie);
6587 * DTrace Probe Hashing Functions
6589 * The functions in this section (and indeed, the functions in remaining
6590 * sections) are not _called_ from probe context. (Any exceptions to this are
6591 * marked with a "Note:".) Rather, they are called from elsewhere in the
6592 * DTrace framework to look-up probes in, add probes to and remove probes from
6593 * the DTrace probe hashes. (Each probe is hashed by each element of the
6594 * probe tuple -- allowing for fast lookups, regardless of what was
6598 dtrace_hash_str(const char *p)
6604 hval = (hval << 4) + *p++;
6605 if ((g = (hval & 0xf0000000)) != 0)
6612 static dtrace_hash_t *
6613 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
6615 dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
6617 hash->dth_stroffs = stroffs;
6618 hash->dth_nextoffs = nextoffs;
6619 hash->dth_prevoffs = prevoffs;
6622 hash->dth_mask = hash->dth_size - 1;
6624 hash->dth_tab = kmem_zalloc(hash->dth_size *
6625 sizeof (dtrace_hashbucket_t *), KM_SLEEP);
6631 dtrace_hash_destroy(dtrace_hash_t *hash)
6636 for (i = 0; i < hash->dth_size; i++)
6637 ASSERT(hash->dth_tab[i] == NULL);
6640 kmem_free(hash->dth_tab,
6641 hash->dth_size * sizeof (dtrace_hashbucket_t *));
6642 kmem_free(hash, sizeof (dtrace_hash_t));
6646 dtrace_hash_resize(dtrace_hash_t *hash)
6648 int size = hash->dth_size, i, ndx;
6649 int new_size = hash->dth_size << 1;
6650 int new_mask = new_size - 1;
6651 dtrace_hashbucket_t **new_tab, *bucket, *next;
6653 ASSERT((new_size & new_mask) == 0);
6655 new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
6657 for (i = 0; i < size; i++) {
6658 for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
6659 dtrace_probe_t *probe = bucket->dthb_chain;
6661 ASSERT(probe != NULL);
6662 ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
6664 next = bucket->dthb_next;
6665 bucket->dthb_next = new_tab[ndx];
6666 new_tab[ndx] = bucket;
6670 kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
6671 hash->dth_tab = new_tab;
6672 hash->dth_size = new_size;
6673 hash->dth_mask = new_mask;
6677 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
6679 int hashval = DTRACE_HASHSTR(hash, new);
6680 int ndx = hashval & hash->dth_mask;
6681 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6682 dtrace_probe_t **nextp, **prevp;
6684 for (; bucket != NULL; bucket = bucket->dthb_next) {
6685 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
6689 if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
6690 dtrace_hash_resize(hash);
6691 dtrace_hash_add(hash, new);
6695 bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
6696 bucket->dthb_next = hash->dth_tab[ndx];
6697 hash->dth_tab[ndx] = bucket;
6698 hash->dth_nbuckets++;
6701 nextp = DTRACE_HASHNEXT(hash, new);
6702 ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
6703 *nextp = bucket->dthb_chain;
6705 if (bucket->dthb_chain != NULL) {
6706 prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
6707 ASSERT(*prevp == NULL);
6711 bucket->dthb_chain = new;
6715 static dtrace_probe_t *
6716 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
6718 int hashval = DTRACE_HASHSTR(hash, template);
6719 int ndx = hashval & hash->dth_mask;
6720 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6722 for (; bucket != NULL; bucket = bucket->dthb_next) {
6723 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6724 return (bucket->dthb_chain);
6731 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
6733 int hashval = DTRACE_HASHSTR(hash, template);
6734 int ndx = hashval & hash->dth_mask;
6735 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6737 for (; bucket != NULL; bucket = bucket->dthb_next) {
6738 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6739 return (bucket->dthb_len);
6746 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
6748 int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
6749 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6751 dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
6752 dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
6755 * Find the bucket that we're removing this probe from.
6757 for (; bucket != NULL; bucket = bucket->dthb_next) {
6758 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
6762 ASSERT(bucket != NULL);
6764 if (*prevp == NULL) {
6765 if (*nextp == NULL) {
6767 * The removed probe was the only probe on this
6768 * bucket; we need to remove the bucket.
6770 dtrace_hashbucket_t *b = hash->dth_tab[ndx];
6772 ASSERT(bucket->dthb_chain == probe);
6776 hash->dth_tab[ndx] = bucket->dthb_next;
6778 while (b->dthb_next != bucket)
6780 b->dthb_next = bucket->dthb_next;
6783 ASSERT(hash->dth_nbuckets > 0);
6784 hash->dth_nbuckets--;
6785 kmem_free(bucket, sizeof (dtrace_hashbucket_t));
6789 bucket->dthb_chain = *nextp;
6791 *(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
6795 *(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
6799 * DTrace Utility Functions
6801 * These are random utility functions that are _not_ called from probe context.
6804 dtrace_badattr(const dtrace_attribute_t *a)
6806 return (a->dtat_name > DTRACE_STABILITY_MAX ||
6807 a->dtat_data > DTRACE_STABILITY_MAX ||
6808 a->dtat_class > DTRACE_CLASS_MAX);
6812 * Return a duplicate copy of a string. If the specified string is NULL,
6813 * this function returns a zero-length string.
6816 dtrace_strdup(const char *str)
6818 char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
6821 (void) strcpy(new, str);
6826 #define DTRACE_ISALPHA(c) \
6827 (((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
6830 dtrace_badname(const char *s)
6834 if (s == NULL || (c = *s++) == '\0')
6837 if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
6840 while ((c = *s++) != '\0') {
6841 if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
6842 c != '-' && c != '_' && c != '.' && c != '`')
6850 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
6855 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
6857 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
6859 priv = DTRACE_PRIV_ALL;
6861 *uidp = crgetuid(cr);
6862 *zoneidp = crgetzoneid(cr);
6865 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
6866 priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
6867 else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
6868 priv |= DTRACE_PRIV_USER;
6869 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
6870 priv |= DTRACE_PRIV_PROC;
6871 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
6872 priv |= DTRACE_PRIV_OWNER;
6873 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
6874 priv |= DTRACE_PRIV_ZONEOWNER;
6877 priv = DTRACE_PRIV_ALL;
6883 #ifdef DTRACE_ERRDEBUG
6885 dtrace_errdebug(const char *str)
6887 int hval = dtrace_hash_str(str) % DTRACE_ERRHASHSZ;
6890 mutex_enter(&dtrace_errlock);
6891 dtrace_errlast = str;
6892 dtrace_errthread = curthread;
6894 while (occupied++ < DTRACE_ERRHASHSZ) {
6895 if (dtrace_errhash[hval].dter_msg == str) {
6896 dtrace_errhash[hval].dter_count++;
6900 if (dtrace_errhash[hval].dter_msg != NULL) {
6901 hval = (hval + 1) % DTRACE_ERRHASHSZ;
6905 dtrace_errhash[hval].dter_msg = str;
6906 dtrace_errhash[hval].dter_count = 1;
6910 panic("dtrace: undersized error hash");
6912 mutex_exit(&dtrace_errlock);
6917 * DTrace Matching Functions
6919 * These functions are used to match groups of probes, given some elements of
6920 * a probe tuple, or some globbed expressions for elements of a probe tuple.
6923 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
6926 if (priv != DTRACE_PRIV_ALL) {
6927 uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
6928 uint32_t match = priv & ppriv;
6931 * No PRIV_DTRACE_* privileges...
6933 if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
6934 DTRACE_PRIV_KERNEL)) == 0)
6938 * No matching bits, but there were bits to match...
6940 if (match == 0 && ppriv != 0)
6944 * Need to have permissions to the process, but don't...
6946 if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
6947 uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
6952 * Need to be in the same zone unless we possess the
6953 * privilege to examine all zones.
6955 if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
6956 zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
6965 * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
6966 * consists of input pattern strings and an ops-vector to evaluate them.
6967 * This function returns >0 for match, 0 for no match, and <0 for error.
6970 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
6971 uint32_t priv, uid_t uid, zoneid_t zoneid)
6973 dtrace_provider_t *pvp = prp->dtpr_provider;
6976 if (pvp->dtpv_defunct)
6979 if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
6982 if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
6985 if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
6988 if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
6991 if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
6998 * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
6999 * interface for matching a glob pattern 'p' to an input string 's'. Unlike
7000 * libc's version, the kernel version only applies to 8-bit ASCII strings.
7001 * In addition, all of the recursion cases except for '*' matching have been
7002 * unwound. For '*', we still implement recursive evaluation, but a depth
7003 * counter is maintained and matching is aborted if we recurse too deep.
7004 * The function returns 0 if no match, >0 if match, and <0 if recursion error.
7007 dtrace_match_glob(const char *s, const char *p, int depth)
7013 if (depth > DTRACE_PROBEKEY_MAXDEPTH)
7017 s = ""; /* treat NULL as empty string */
7026 if ((c = *p++) == '\0')
7027 return (s1 == '\0');
7031 int ok = 0, notflag = 0;
7042 if ((c = *p++) == '\0')
7046 if (c == '-' && lc != '\0' && *p != ']') {
7047 if ((c = *p++) == '\0')
7049 if (c == '\\' && (c = *p++) == '\0')
7053 if (s1 < lc || s1 > c)
7057 } else if (lc <= s1 && s1 <= c)
7060 } else if (c == '\\' && (c = *p++) == '\0')
7063 lc = c; /* save left-hand 'c' for next iteration */
7073 if ((c = *p++) == '\0')
7085 if ((c = *p++) == '\0')
7101 p++; /* consecutive *'s are identical to a single one */
7106 for (s = olds; *s != '\0'; s++) {
7107 if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
7117 dtrace_match_string(const char *s, const char *p, int depth)
7119 return (s != NULL && strcmp(s, p) == 0);
7124 dtrace_match_nul(const char *s, const char *p, int depth)
7126 return (1); /* always match the empty pattern */
7131 dtrace_match_nonzero(const char *s, const char *p, int depth)
7133 return (s != NULL && s[0] != '\0');
7137 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
7138 zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
7140 dtrace_probe_t template, *probe;
7141 dtrace_hash_t *hash = NULL;
7142 int len, best = INT_MAX, nmatched = 0;
7145 ASSERT(MUTEX_HELD(&dtrace_lock));
7148 * If the probe ID is specified in the key, just lookup by ID and
7149 * invoke the match callback once if a matching probe is found.
7151 if (pkp->dtpk_id != DTRACE_IDNONE) {
7152 if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
7153 dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
7154 (void) (*matched)(probe, arg);
7160 template.dtpr_mod = (char *)pkp->dtpk_mod;
7161 template.dtpr_func = (char *)pkp->dtpk_func;
7162 template.dtpr_name = (char *)pkp->dtpk_name;
7165 * We want to find the most distinct of the module name, function
7166 * name, and name. So for each one that is not a glob pattern or
7167 * empty string, we perform a lookup in the corresponding hash and
7168 * use the hash table with the fewest collisions to do our search.
7170 if (pkp->dtpk_mmatch == &dtrace_match_string &&
7171 (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
7173 hash = dtrace_bymod;
7176 if (pkp->dtpk_fmatch == &dtrace_match_string &&
7177 (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
7179 hash = dtrace_byfunc;
7182 if (pkp->dtpk_nmatch == &dtrace_match_string &&
7183 (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
7185 hash = dtrace_byname;
7189 * If we did not select a hash table, iterate over every probe and
7190 * invoke our callback for each one that matches our input probe key.
7193 for (i = 0; i < dtrace_nprobes; i++) {
7194 if ((probe = dtrace_probes[i]) == NULL ||
7195 dtrace_match_probe(probe, pkp, priv, uid,
7201 if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
7209 * If we selected a hash table, iterate over each probe of the same key
7210 * name and invoke the callback for every probe that matches the other
7211 * attributes of our input probe key.
7213 for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
7214 probe = *(DTRACE_HASHNEXT(hash, probe))) {
7216 if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
7221 if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
7229 * Return the function pointer dtrace_probecmp() should use to compare the
7230 * specified pattern with a string. For NULL or empty patterns, we select
7231 * dtrace_match_nul(). For glob pattern strings, we use dtrace_match_glob().
7232 * For non-empty non-glob strings, we use dtrace_match_string().
7234 static dtrace_probekey_f *
7235 dtrace_probekey_func(const char *p)
7239 if (p == NULL || *p == '\0')
7240 return (&dtrace_match_nul);
7242 while ((c = *p++) != '\0') {
7243 if (c == '[' || c == '?' || c == '*' || c == '\\')
7244 return (&dtrace_match_glob);
7247 return (&dtrace_match_string);
7251 * Build a probe comparison key for use with dtrace_match_probe() from the
7252 * given probe description. By convention, a null key only matches anchored
7253 * probes: if each field is the empty string, reset dtpk_fmatch to
7254 * dtrace_match_nonzero().
7257 dtrace_probekey(dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
7259 pkp->dtpk_prov = pdp->dtpd_provider;
7260 pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
7262 pkp->dtpk_mod = pdp->dtpd_mod;
7263 pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
7265 pkp->dtpk_func = pdp->dtpd_func;
7266 pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
7268 pkp->dtpk_name = pdp->dtpd_name;
7269 pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
7271 pkp->dtpk_id = pdp->dtpd_id;
7273 if (pkp->dtpk_id == DTRACE_IDNONE &&
7274 pkp->dtpk_pmatch == &dtrace_match_nul &&
7275 pkp->dtpk_mmatch == &dtrace_match_nul &&
7276 pkp->dtpk_fmatch == &dtrace_match_nul &&
7277 pkp->dtpk_nmatch == &dtrace_match_nul)
7278 pkp->dtpk_fmatch = &dtrace_match_nonzero;
7282 * DTrace Provider-to-Framework API Functions
7284 * These functions implement much of the Provider-to-Framework API, as
7285 * described in <sys/dtrace.h>. The parts of the API not in this section are
7286 * the functions in the API for probe management (found below), and
7287 * dtrace_probe() itself (found above).
7291 * Register the calling provider with the DTrace framework. This should
7292 * generally be called by DTrace providers in their attach(9E) entry point.
7295 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
7296 cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
7298 dtrace_provider_t *provider;
7300 if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
7301 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7302 "arguments", name ? name : "<NULL>");
7306 if (name[0] == '\0' || dtrace_badname(name)) {
7307 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7308 "provider name", name);
7312 if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
7313 pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
7314 pops->dtps_destroy == NULL ||
7315 ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
7316 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7317 "provider ops", name);
7321 if (dtrace_badattr(&pap->dtpa_provider) ||
7322 dtrace_badattr(&pap->dtpa_mod) ||
7323 dtrace_badattr(&pap->dtpa_func) ||
7324 dtrace_badattr(&pap->dtpa_name) ||
7325 dtrace_badattr(&pap->dtpa_args)) {
7326 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7327 "provider attributes", name);
7331 if (priv & ~DTRACE_PRIV_ALL) {
7332 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7333 "privilege attributes", name);
7337 if ((priv & DTRACE_PRIV_KERNEL) &&
7338 (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
7339 pops->dtps_usermode == NULL) {
7340 cmn_err(CE_WARN, "failed to register provider '%s': need "
7341 "dtps_usermode() op for given privilege attributes", name);
7345 provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
7346 provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
7347 (void) strcpy(provider->dtpv_name, name);
7349 provider->dtpv_attr = *pap;
7350 provider->dtpv_priv.dtpp_flags = priv;
7352 provider->dtpv_priv.dtpp_uid = crgetuid(cr);
7353 provider->dtpv_priv.dtpp_zoneid = crgetzoneid(cr);
7355 provider->dtpv_pops = *pops;
7357 if (pops->dtps_provide == NULL) {
7358 ASSERT(pops->dtps_provide_module != NULL);
7359 provider->dtpv_pops.dtps_provide =
7360 (void (*)(void *, dtrace_probedesc_t *))dtrace_nullop;
7363 if (pops->dtps_provide_module == NULL) {
7364 ASSERT(pops->dtps_provide != NULL);
7365 provider->dtpv_pops.dtps_provide_module =
7366 (void (*)(void *, modctl_t *))dtrace_nullop;
7369 if (pops->dtps_suspend == NULL) {
7370 ASSERT(pops->dtps_resume == NULL);
7371 provider->dtpv_pops.dtps_suspend =
7372 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7373 provider->dtpv_pops.dtps_resume =
7374 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7377 provider->dtpv_arg = arg;
7378 *idp = (dtrace_provider_id_t)provider;
7380 if (pops == &dtrace_provider_ops) {
7381 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7382 ASSERT(MUTEX_HELD(&dtrace_lock));
7383 ASSERT(dtrace_anon.dta_enabling == NULL);
7386 * We make sure that the DTrace provider is at the head of
7387 * the provider chain.
7389 provider->dtpv_next = dtrace_provider;
7390 dtrace_provider = provider;
7394 mutex_enter(&dtrace_provider_lock);
7395 mutex_enter(&dtrace_lock);
7398 * If there is at least one provider registered, we'll add this
7399 * provider after the first provider.
7401 if (dtrace_provider != NULL) {
7402 provider->dtpv_next = dtrace_provider->dtpv_next;
7403 dtrace_provider->dtpv_next = provider;
7405 dtrace_provider = provider;
7408 if (dtrace_retained != NULL) {
7409 dtrace_enabling_provide(provider);
7412 * Now we need to call dtrace_enabling_matchall() -- which
7413 * will acquire cpu_lock and dtrace_lock. We therefore need
7414 * to drop all of our locks before calling into it...
7416 mutex_exit(&dtrace_lock);
7417 mutex_exit(&dtrace_provider_lock);
7418 dtrace_enabling_matchall();
7423 mutex_exit(&dtrace_lock);
7424 mutex_exit(&dtrace_provider_lock);
7430 * Unregister the specified provider from the DTrace framework. This should
7431 * generally be called by DTrace providers in their detach(9E) entry point.
7434 dtrace_unregister(dtrace_provider_id_t id)
7436 dtrace_provider_t *old = (dtrace_provider_t *)id;
7437 dtrace_provider_t *prev = NULL;
7439 dtrace_probe_t *probe, *first = NULL;
7441 if (old->dtpv_pops.dtps_enable ==
7442 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop) {
7444 * If DTrace itself is the provider, we're called with locks
7447 ASSERT(old == dtrace_provider);
7449 ASSERT(dtrace_devi != NULL);
7451 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7452 ASSERT(MUTEX_HELD(&dtrace_lock));
7455 if (dtrace_provider->dtpv_next != NULL) {
7457 * There's another provider here; return failure.
7462 mutex_enter(&dtrace_provider_lock);
7463 mutex_enter(&mod_lock);
7464 mutex_enter(&dtrace_lock);
7468 * If anyone has /dev/dtrace open, or if there are anonymous enabled
7469 * probes, we refuse to let providers slither away, unless this
7470 * provider has already been explicitly invalidated.
7472 if (!old->dtpv_defunct &&
7473 (dtrace_opens || (dtrace_anon.dta_state != NULL &&
7474 dtrace_anon.dta_state->dts_necbs > 0))) {
7476 mutex_exit(&dtrace_lock);
7477 mutex_exit(&mod_lock);
7478 mutex_exit(&dtrace_provider_lock);
7484 * Attempt to destroy the probes associated with this provider.
7486 for (i = 0; i < dtrace_nprobes; i++) {
7487 if ((probe = dtrace_probes[i]) == NULL)
7490 if (probe->dtpr_provider != old)
7493 if (probe->dtpr_ecb == NULL)
7497 * We have at least one ECB; we can't remove this provider.
7500 mutex_exit(&dtrace_lock);
7501 mutex_exit(&mod_lock);
7502 mutex_exit(&dtrace_provider_lock);
7508 * All of the probes for this provider are disabled; we can safely
7509 * remove all of them from their hash chains and from the probe array.
7511 for (i = 0; i < dtrace_nprobes; i++) {
7512 if ((probe = dtrace_probes[i]) == NULL)
7515 if (probe->dtpr_provider != old)
7518 dtrace_probes[i] = NULL;
7520 dtrace_hash_remove(dtrace_bymod, probe);
7521 dtrace_hash_remove(dtrace_byfunc, probe);
7522 dtrace_hash_remove(dtrace_byname, probe);
7524 if (first == NULL) {
7526 probe->dtpr_nextmod = NULL;
7528 probe->dtpr_nextmod = first;
7534 * The provider's probes have been removed from the hash chains and
7535 * from the probe array. Now issue a dtrace_sync() to be sure that
7536 * everyone has cleared out from any probe array processing.
7540 for (probe = first; probe != NULL; probe = first) {
7541 first = probe->dtpr_nextmod;
7543 old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
7545 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7546 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7547 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7549 vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
7551 free_unr(dtrace_arena, probe->dtpr_id);
7553 kmem_free(probe, sizeof (dtrace_probe_t));
7556 if ((prev = dtrace_provider) == old) {
7558 ASSERT(self || dtrace_devi == NULL);
7559 ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
7561 dtrace_provider = old->dtpv_next;
7563 while (prev != NULL && prev->dtpv_next != old)
7564 prev = prev->dtpv_next;
7567 panic("attempt to unregister non-existent "
7568 "dtrace provider %p\n", (void *)id);
7571 prev->dtpv_next = old->dtpv_next;
7575 mutex_exit(&dtrace_lock);
7576 mutex_exit(&mod_lock);
7577 mutex_exit(&dtrace_provider_lock);
7580 kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
7581 kmem_free(old, sizeof (dtrace_provider_t));
7587 * Invalidate the specified provider. All subsequent probe lookups for the
7588 * specified provider will fail, but its probes will not be removed.
7591 dtrace_invalidate(dtrace_provider_id_t id)
7593 dtrace_provider_t *pvp = (dtrace_provider_t *)id;
7595 ASSERT(pvp->dtpv_pops.dtps_enable !=
7596 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
7598 mutex_enter(&dtrace_provider_lock);
7599 mutex_enter(&dtrace_lock);
7601 pvp->dtpv_defunct = 1;
7603 mutex_exit(&dtrace_lock);
7604 mutex_exit(&dtrace_provider_lock);
7608 * Indicate whether or not DTrace has attached.
7611 dtrace_attached(void)
7614 * dtrace_provider will be non-NULL iff the DTrace driver has
7615 * attached. (It's non-NULL because DTrace is always itself a
7618 return (dtrace_provider != NULL);
7622 * Remove all the unenabled probes for the given provider. This function is
7623 * not unlike dtrace_unregister(), except that it doesn't remove the provider
7624 * -- just as many of its associated probes as it can.
7627 dtrace_condense(dtrace_provider_id_t id)
7629 dtrace_provider_t *prov = (dtrace_provider_t *)id;
7631 dtrace_probe_t *probe;
7634 * Make sure this isn't the dtrace provider itself.
7636 ASSERT(prov->dtpv_pops.dtps_enable !=
7637 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
7639 mutex_enter(&dtrace_provider_lock);
7640 mutex_enter(&dtrace_lock);
7643 * Attempt to destroy the probes associated with this provider.
7645 for (i = 0; i < dtrace_nprobes; i++) {
7646 if ((probe = dtrace_probes[i]) == NULL)
7649 if (probe->dtpr_provider != prov)
7652 if (probe->dtpr_ecb != NULL)
7655 dtrace_probes[i] = NULL;
7657 dtrace_hash_remove(dtrace_bymod, probe);
7658 dtrace_hash_remove(dtrace_byfunc, probe);
7659 dtrace_hash_remove(dtrace_byname, probe);
7661 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
7663 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7664 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7665 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7666 kmem_free(probe, sizeof (dtrace_probe_t));
7668 vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
7670 free_unr(dtrace_arena, i + 1);
7674 mutex_exit(&dtrace_lock);
7675 mutex_exit(&dtrace_provider_lock);
7681 * DTrace Probe Management Functions
7683 * The functions in this section perform the DTrace probe management,
7684 * including functions to create probes, look-up probes, and call into the
7685 * providers to request that probes be provided. Some of these functions are
7686 * in the Provider-to-Framework API; these functions can be identified by the
7687 * fact that they are not declared "static".
7691 * Create a probe with the specified module name, function name, and name.
7694 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
7695 const char *func, const char *name, int aframes, void *arg)
7697 dtrace_probe_t *probe, **probes;
7698 dtrace_provider_t *provider = (dtrace_provider_t *)prov;
7701 if (provider == dtrace_provider) {
7702 ASSERT(MUTEX_HELD(&dtrace_lock));
7704 mutex_enter(&dtrace_lock);
7708 id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
7709 VM_BESTFIT | VM_SLEEP);
7711 id = alloc_unr(dtrace_arena);
7713 probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
7715 probe->dtpr_id = id;
7716 probe->dtpr_gen = dtrace_probegen++;
7717 probe->dtpr_mod = dtrace_strdup(mod);
7718 probe->dtpr_func = dtrace_strdup(func);
7719 probe->dtpr_name = dtrace_strdup(name);
7720 probe->dtpr_arg = arg;
7721 probe->dtpr_aframes = aframes;
7722 probe->dtpr_provider = provider;
7724 dtrace_hash_add(dtrace_bymod, probe);
7725 dtrace_hash_add(dtrace_byfunc, probe);
7726 dtrace_hash_add(dtrace_byname, probe);
7728 if (id - 1 >= dtrace_nprobes) {
7729 size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
7730 size_t nsize = osize << 1;
7734 ASSERT(dtrace_probes == NULL);
7735 nsize = sizeof (dtrace_probe_t *);
7738 probes = kmem_zalloc(nsize, KM_SLEEP);
7740 if (dtrace_probes == NULL) {
7742 dtrace_probes = probes;
7745 dtrace_probe_t **oprobes = dtrace_probes;
7747 bcopy(oprobes, probes, osize);
7748 dtrace_membar_producer();
7749 dtrace_probes = probes;
7754 * All CPUs are now seeing the new probes array; we can
7755 * safely free the old array.
7757 kmem_free(oprobes, osize);
7758 dtrace_nprobes <<= 1;
7761 ASSERT(id - 1 < dtrace_nprobes);
7764 ASSERT(dtrace_probes[id - 1] == NULL);
7765 dtrace_probes[id - 1] = probe;
7767 if (provider != dtrace_provider)
7768 mutex_exit(&dtrace_lock);
7773 static dtrace_probe_t *
7774 dtrace_probe_lookup_id(dtrace_id_t id)
7776 ASSERT(MUTEX_HELD(&dtrace_lock));
7778 if (id == 0 || id > dtrace_nprobes)
7781 return (dtrace_probes[id - 1]);
7785 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
7787 *((dtrace_id_t *)arg) = probe->dtpr_id;
7789 return (DTRACE_MATCH_DONE);
7793 * Look up a probe based on provider and one or more of module name, function
7794 * name and probe name.
7797 dtrace_probe_lookup(dtrace_provider_id_t prid, char *mod,
7798 char *func, char *name)
7800 dtrace_probekey_t pkey;
7804 pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
7805 pkey.dtpk_pmatch = &dtrace_match_string;
7806 pkey.dtpk_mod = mod;
7807 pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
7808 pkey.dtpk_func = func;
7809 pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
7810 pkey.dtpk_name = name;
7811 pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
7812 pkey.dtpk_id = DTRACE_IDNONE;
7814 mutex_enter(&dtrace_lock);
7815 match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
7816 dtrace_probe_lookup_match, &id);
7817 mutex_exit(&dtrace_lock);
7819 ASSERT(match == 1 || match == 0);
7820 return (match ? id : 0);
7824 * Returns the probe argument associated with the specified probe.
7827 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
7829 dtrace_probe_t *probe;
7832 mutex_enter(&dtrace_lock);
7834 if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
7835 probe->dtpr_provider == (dtrace_provider_t *)id)
7836 rval = probe->dtpr_arg;
7838 mutex_exit(&dtrace_lock);
7844 * Copy a probe into a probe description.
7847 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
7849 bzero(pdp, sizeof (dtrace_probedesc_t));
7850 pdp->dtpd_id = prp->dtpr_id;
7852 (void) strncpy(pdp->dtpd_provider,
7853 prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
7855 (void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
7856 (void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
7857 (void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
7862 dtrace_probe_provide_cb(linker_file_t lf, void *arg)
7864 dtrace_provider_t *prv = (dtrace_provider_t *) arg;
7866 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, lf);
7874 * Called to indicate that a probe -- or probes -- should be provided by a
7875 * specfied provider. If the specified description is NULL, the provider will
7876 * be told to provide all of its probes. (This is done whenever a new
7877 * consumer comes along, or whenever a retained enabling is to be matched.) If
7878 * the specified description is non-NULL, the provider is given the
7879 * opportunity to dynamically provide the specified probe, allowing providers
7880 * to support the creation of probes on-the-fly. (So-called _autocreated_
7881 * probes.) If the provider is NULL, the operations will be applied to all
7882 * providers; if the provider is non-NULL the operations will only be applied
7883 * to the specified provider. The dtrace_provider_lock must be held, and the
7884 * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
7885 * will need to grab the dtrace_lock when it reenters the framework through
7886 * dtrace_probe_lookup(), dtrace_probe_create(), etc.
7889 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
7896 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7900 prv = dtrace_provider;
7905 * First, call the blanket provide operation.
7907 prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
7910 * Now call the per-module provide operation. We will grab
7911 * mod_lock to prevent the list from being modified. Note
7912 * that this also prevents the mod_busy bits from changing.
7913 * (mod_busy can only be changed with mod_lock held.)
7915 mutex_enter(&mod_lock);
7920 if (ctl->mod_busy || ctl->mod_mp == NULL)
7923 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
7925 } while ((ctl = ctl->mod_next) != &modules);
7927 (void) linker_file_foreach(dtrace_probe_provide_cb, prv);
7930 mutex_exit(&mod_lock);
7931 } while (all && (prv = prv->dtpv_next) != NULL);
7936 * Iterate over each probe, and call the Framework-to-Provider API function
7940 dtrace_probe_foreach(uintptr_t offs)
7942 dtrace_provider_t *prov;
7943 void (*func)(void *, dtrace_id_t, void *);
7944 dtrace_probe_t *probe;
7945 dtrace_icookie_t cookie;
7949 * We disable interrupts to walk through the probe array. This is
7950 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
7951 * won't see stale data.
7953 cookie = dtrace_interrupt_disable();
7955 for (i = 0; i < dtrace_nprobes; i++) {
7956 if ((probe = dtrace_probes[i]) == NULL)
7959 if (probe->dtpr_ecb == NULL) {
7961 * This probe isn't enabled -- don't call the function.
7966 prov = probe->dtpr_provider;
7967 func = *((void(**)(void *, dtrace_id_t, void *))
7968 ((uintptr_t)&prov->dtpv_pops + offs));
7970 func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
7973 dtrace_interrupt_enable(cookie);
7978 dtrace_probe_enable(dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
7980 dtrace_probekey_t pkey;
7985 ASSERT(MUTEX_HELD(&dtrace_lock));
7986 dtrace_ecb_create_cache = NULL;
7990 * If we're passed a NULL description, we're being asked to
7991 * create an ECB with a NULL probe.
7993 (void) dtrace_ecb_create_enable(NULL, enab);
7997 dtrace_probekey(desc, &pkey);
7998 dtrace_cred2priv(enab->dten_vstate->dtvs_state->dts_cred.dcr_cred,
7999 &priv, &uid, &zoneid);
8001 return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
8006 * DTrace Helper Provider Functions
8009 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
8011 attr->dtat_name = DOF_ATTR_NAME(dofattr);
8012 attr->dtat_data = DOF_ATTR_DATA(dofattr);
8013 attr->dtat_class = DOF_ATTR_CLASS(dofattr);
8017 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
8018 const dof_provider_t *dofprov, char *strtab)
8020 hprov->dthpv_provname = strtab + dofprov->dofpv_name;
8021 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
8022 dofprov->dofpv_provattr);
8023 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
8024 dofprov->dofpv_modattr);
8025 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
8026 dofprov->dofpv_funcattr);
8027 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
8028 dofprov->dofpv_nameattr);
8029 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
8030 dofprov->dofpv_argsattr);
8034 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8036 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8037 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8038 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
8039 dof_provider_t *provider;
8041 uint32_t *off, *enoff;
8045 dtrace_helper_provdesc_t dhpv;
8046 dtrace_helper_probedesc_t dhpb;
8047 dtrace_meta_t *meta = dtrace_meta_pid;
8048 dtrace_mops_t *mops = &meta->dtm_mops;
8051 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8052 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8053 provider->dofpv_strtab * dof->dofh_secsize);
8054 prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8055 provider->dofpv_probes * dof->dofh_secsize);
8056 arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8057 provider->dofpv_prargs * dof->dofh_secsize);
8058 off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8059 provider->dofpv_proffs * dof->dofh_secsize);
8061 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8062 off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
8063 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
8067 * See dtrace_helper_provider_validate().
8069 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
8070 provider->dofpv_prenoffs != DOF_SECT_NONE) {
8071 enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8072 provider->dofpv_prenoffs * dof->dofh_secsize);
8073 enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
8076 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
8079 * Create the provider.
8081 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8083 if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
8089 * Create the probes.
8091 for (i = 0; i < nprobes; i++) {
8092 probe = (dof_probe_t *)(uintptr_t)(daddr +
8093 prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
8095 dhpb.dthpb_mod = dhp->dofhp_mod;
8096 dhpb.dthpb_func = strtab + probe->dofpr_func;
8097 dhpb.dthpb_name = strtab + probe->dofpr_name;
8098 dhpb.dthpb_base = probe->dofpr_addr;
8099 dhpb.dthpb_offs = off + probe->dofpr_offidx;
8100 dhpb.dthpb_noffs = probe->dofpr_noffs;
8101 if (enoff != NULL) {
8102 dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
8103 dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
8105 dhpb.dthpb_enoffs = NULL;
8106 dhpb.dthpb_nenoffs = 0;
8108 dhpb.dthpb_args = arg + probe->dofpr_argidx;
8109 dhpb.dthpb_nargc = probe->dofpr_nargc;
8110 dhpb.dthpb_xargc = probe->dofpr_xargc;
8111 dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
8112 dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
8114 mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
8119 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
8121 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8122 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8125 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8127 for (i = 0; i < dof->dofh_secnum; i++) {
8128 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8129 dof->dofh_secoff + i * dof->dofh_secsize);
8131 if (sec->dofs_type != DOF_SECT_PROVIDER)
8134 dtrace_helper_provide_one(dhp, sec, pid);
8138 * We may have just created probes, so we must now rematch against
8139 * any retained enablings. Note that this call will acquire both
8140 * cpu_lock and dtrace_lock; the fact that we are holding
8141 * dtrace_meta_lock now is what defines the ordering with respect to
8142 * these three locks.
8144 dtrace_enabling_matchall();
8148 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8150 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8151 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8153 dof_provider_t *provider;
8155 dtrace_helper_provdesc_t dhpv;
8156 dtrace_meta_t *meta = dtrace_meta_pid;
8157 dtrace_mops_t *mops = &meta->dtm_mops;
8159 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8160 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8161 provider->dofpv_strtab * dof->dofh_secsize);
8163 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8166 * Create the provider.
8168 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8170 mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
8176 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
8178 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8179 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8182 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8184 for (i = 0; i < dof->dofh_secnum; i++) {
8185 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8186 dof->dofh_secoff + i * dof->dofh_secsize);
8188 if (sec->dofs_type != DOF_SECT_PROVIDER)
8191 dtrace_helper_provider_remove_one(dhp, sec, pid);
8196 * DTrace Meta Provider-to-Framework API Functions
8198 * These functions implement the Meta Provider-to-Framework API, as described
8199 * in <sys/dtrace.h>.
8202 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
8203 dtrace_meta_provider_id_t *idp)
8205 dtrace_meta_t *meta;
8206 dtrace_helpers_t *help, *next;
8209 *idp = DTRACE_METAPROVNONE;
8212 * We strictly don't need the name, but we hold onto it for
8213 * debuggability. All hail error queues!
8216 cmn_err(CE_WARN, "failed to register meta-provider: "
8222 mops->dtms_create_probe == NULL ||
8223 mops->dtms_provide_pid == NULL ||
8224 mops->dtms_remove_pid == NULL) {
8225 cmn_err(CE_WARN, "failed to register meta-register %s: "
8226 "invalid ops", name);
8230 meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
8231 meta->dtm_mops = *mops;
8232 meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
8233 (void) strcpy(meta->dtm_name, name);
8234 meta->dtm_arg = arg;
8236 mutex_enter(&dtrace_meta_lock);
8237 mutex_enter(&dtrace_lock);
8239 if (dtrace_meta_pid != NULL) {
8240 mutex_exit(&dtrace_lock);
8241 mutex_exit(&dtrace_meta_lock);
8242 cmn_err(CE_WARN, "failed to register meta-register %s: "
8243 "user-land meta-provider exists", name);
8244 kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
8245 kmem_free(meta, sizeof (dtrace_meta_t));
8249 dtrace_meta_pid = meta;
8250 *idp = (dtrace_meta_provider_id_t)meta;
8253 * If there are providers and probes ready to go, pass them
8254 * off to the new meta provider now.
8257 help = dtrace_deferred_pid;
8258 dtrace_deferred_pid = NULL;
8260 mutex_exit(&dtrace_lock);
8262 while (help != NULL) {
8263 for (i = 0; i < help->dthps_nprovs; i++) {
8264 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
8268 next = help->dthps_next;
8269 help->dthps_next = NULL;
8270 help->dthps_prev = NULL;
8271 help->dthps_deferred = 0;
8275 mutex_exit(&dtrace_meta_lock);
8281 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
8283 dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
8285 mutex_enter(&dtrace_meta_lock);
8286 mutex_enter(&dtrace_lock);
8288 if (old == dtrace_meta_pid) {
8289 pp = &dtrace_meta_pid;
8291 panic("attempt to unregister non-existent "
8292 "dtrace meta-provider %p\n", (void *)old);
8295 if (old->dtm_count != 0) {
8296 mutex_exit(&dtrace_lock);
8297 mutex_exit(&dtrace_meta_lock);
8303 mutex_exit(&dtrace_lock);
8304 mutex_exit(&dtrace_meta_lock);
8306 kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
8307 kmem_free(old, sizeof (dtrace_meta_t));
8314 * DTrace DIF Object Functions
8317 dtrace_difo_err(uint_t pc, const char *format, ...)
8319 if (dtrace_err_verbose) {
8322 (void) uprintf("dtrace DIF object error: [%u]: ", pc);
8323 va_start(alist, format);
8324 (void) vuprintf(format, alist);
8328 #ifdef DTRACE_ERRDEBUG
8329 dtrace_errdebug(format);
8335 * Validate a DTrace DIF object by checking the IR instructions. The following
8336 * rules are currently enforced by dtrace_difo_validate():
8338 * 1. Each instruction must have a valid opcode
8339 * 2. Each register, string, variable, or subroutine reference must be valid
8340 * 3. No instruction can modify register %r0 (must be zero)
8341 * 4. All instruction reserved bits must be set to zero
8342 * 5. The last instruction must be a "ret" instruction
8343 * 6. All branch targets must reference a valid instruction _after_ the branch
8346 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
8350 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8354 kcheckload = cr == NULL ||
8355 (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
8357 dp->dtdo_destructive = 0;
8359 for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
8360 dif_instr_t instr = dp->dtdo_buf[pc];
8362 uint_t r1 = DIF_INSTR_R1(instr);
8363 uint_t r2 = DIF_INSTR_R2(instr);
8364 uint_t rd = DIF_INSTR_RD(instr);
8365 uint_t rs = DIF_INSTR_RS(instr);
8366 uint_t label = DIF_INSTR_LABEL(instr);
8367 uint_t v = DIF_INSTR_VAR(instr);
8368 uint_t subr = DIF_INSTR_SUBR(instr);
8369 uint_t type = DIF_INSTR_TYPE(instr);
8370 uint_t op = DIF_INSTR_OP(instr);
8388 err += efunc(pc, "invalid register %u\n", r1);
8390 err += efunc(pc, "invalid register %u\n", r2);
8392 err += efunc(pc, "invalid register %u\n", rd);
8394 err += efunc(pc, "cannot write to %r0\n");
8400 err += efunc(pc, "invalid register %u\n", r1);
8402 err += efunc(pc, "non-zero reserved bits\n");
8404 err += efunc(pc, "invalid register %u\n", rd);
8406 err += efunc(pc, "cannot write to %r0\n");
8416 err += efunc(pc, "invalid register %u\n", r1);
8418 err += efunc(pc, "non-zero reserved bits\n");
8420 err += efunc(pc, "invalid register %u\n", rd);
8422 err += efunc(pc, "cannot write to %r0\n");
8424 dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
8425 DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
8435 err += efunc(pc, "invalid register %u\n", r1);
8437 err += efunc(pc, "non-zero reserved bits\n");
8439 err += efunc(pc, "invalid register %u\n", rd);
8441 err += efunc(pc, "cannot write to %r0\n");
8451 err += efunc(pc, "invalid register %u\n", r1);
8453 err += efunc(pc, "non-zero reserved bits\n");
8455 err += efunc(pc, "invalid register %u\n", rd);
8457 err += efunc(pc, "cannot write to %r0\n");
8464 err += efunc(pc, "invalid register %u\n", r1);
8466 err += efunc(pc, "non-zero reserved bits\n");
8468 err += efunc(pc, "invalid register %u\n", rd);
8470 err += efunc(pc, "cannot write to 0 address\n");
8475 err += efunc(pc, "invalid register %u\n", r1);
8477 err += efunc(pc, "invalid register %u\n", r2);
8479 err += efunc(pc, "non-zero reserved bits\n");
8483 err += efunc(pc, "invalid register %u\n", r1);
8484 if (r2 != 0 || rd != 0)
8485 err += efunc(pc, "non-zero reserved bits\n");
8498 if (label >= dp->dtdo_len) {
8499 err += efunc(pc, "invalid branch target %u\n",
8503 err += efunc(pc, "backward branch to %u\n",
8508 if (r1 != 0 || r2 != 0)
8509 err += efunc(pc, "non-zero reserved bits\n");
8511 err += efunc(pc, "invalid register %u\n", rd);
8515 case DIF_OP_FLUSHTS:
8516 if (r1 != 0 || r2 != 0 || rd != 0)
8517 err += efunc(pc, "non-zero reserved bits\n");
8520 if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
8521 err += efunc(pc, "invalid integer ref %u\n",
8522 DIF_INSTR_INTEGER(instr));
8525 err += efunc(pc, "invalid register %u\n", rd);
8527 err += efunc(pc, "cannot write to %r0\n");
8530 if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
8531 err += efunc(pc, "invalid string ref %u\n",
8532 DIF_INSTR_STRING(instr));
8535 err += efunc(pc, "invalid register %u\n", rd);
8537 err += efunc(pc, "cannot write to %r0\n");
8541 if (r1 > DIF_VAR_ARRAY_MAX)
8542 err += efunc(pc, "invalid array %u\n", r1);
8544 err += efunc(pc, "invalid register %u\n", r2);
8546 err += efunc(pc, "invalid register %u\n", rd);
8548 err += efunc(pc, "cannot write to %r0\n");
8555 if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
8556 err += efunc(pc, "invalid variable %u\n", v);
8558 err += efunc(pc, "invalid register %u\n", rd);
8560 err += efunc(pc, "cannot write to %r0\n");
8567 if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
8568 err += efunc(pc, "invalid variable %u\n", v);
8570 err += efunc(pc, "invalid register %u\n", rd);
8573 if (subr > DIF_SUBR_MAX)
8574 err += efunc(pc, "invalid subr %u\n", subr);
8576 err += efunc(pc, "invalid register %u\n", rd);
8578 err += efunc(pc, "cannot write to %r0\n");
8580 if (subr == DIF_SUBR_COPYOUT ||
8581 subr == DIF_SUBR_COPYOUTSTR) {
8582 dp->dtdo_destructive = 1;
8586 if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
8587 err += efunc(pc, "invalid ref type %u\n", type);
8589 err += efunc(pc, "invalid register %u\n", r2);
8591 err += efunc(pc, "invalid register %u\n", rs);
8594 if (type != DIF_TYPE_CTF)
8595 err += efunc(pc, "invalid val type %u\n", type);
8597 err += efunc(pc, "invalid register %u\n", r2);
8599 err += efunc(pc, "invalid register %u\n", rs);
8602 err += efunc(pc, "invalid opcode %u\n",
8603 DIF_INSTR_OP(instr));
8607 if (dp->dtdo_len != 0 &&
8608 DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
8609 err += efunc(dp->dtdo_len - 1,
8610 "expected 'ret' as last DIF instruction\n");
8613 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) {
8615 * If we're not returning by reference, the size must be either
8616 * 0 or the size of one of the base types.
8618 switch (dp->dtdo_rtype.dtdt_size) {
8620 case sizeof (uint8_t):
8621 case sizeof (uint16_t):
8622 case sizeof (uint32_t):
8623 case sizeof (uint64_t):
8627 err += efunc(dp->dtdo_len - 1, "bad return size");
8631 for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
8632 dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
8633 dtrace_diftype_t *vt, *et;
8636 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
8637 v->dtdv_scope != DIFV_SCOPE_THREAD &&
8638 v->dtdv_scope != DIFV_SCOPE_LOCAL) {
8639 err += efunc(i, "unrecognized variable scope %d\n",
8644 if (v->dtdv_kind != DIFV_KIND_ARRAY &&
8645 v->dtdv_kind != DIFV_KIND_SCALAR) {
8646 err += efunc(i, "unrecognized variable type %d\n",
8651 if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
8652 err += efunc(i, "%d exceeds variable id limit\n", id);
8656 if (id < DIF_VAR_OTHER_UBASE)
8660 * For user-defined variables, we need to check that this
8661 * definition is identical to any previous definition that we
8664 ndx = id - DIF_VAR_OTHER_UBASE;
8666 switch (v->dtdv_scope) {
8667 case DIFV_SCOPE_GLOBAL:
8668 if (ndx < vstate->dtvs_nglobals) {
8669 dtrace_statvar_t *svar;
8671 if ((svar = vstate->dtvs_globals[ndx]) != NULL)
8672 existing = &svar->dtsv_var;
8677 case DIFV_SCOPE_THREAD:
8678 if (ndx < vstate->dtvs_ntlocals)
8679 existing = &vstate->dtvs_tlocals[ndx];
8682 case DIFV_SCOPE_LOCAL:
8683 if (ndx < vstate->dtvs_nlocals) {
8684 dtrace_statvar_t *svar;
8686 if ((svar = vstate->dtvs_locals[ndx]) != NULL)
8687 existing = &svar->dtsv_var;
8695 if (vt->dtdt_flags & DIF_TF_BYREF) {
8696 if (vt->dtdt_size == 0) {
8697 err += efunc(i, "zero-sized variable\n");
8701 if (v->dtdv_scope == DIFV_SCOPE_GLOBAL &&
8702 vt->dtdt_size > dtrace_global_maxsize) {
8703 err += efunc(i, "oversized by-ref global\n");
8708 if (existing == NULL || existing->dtdv_id == 0)
8711 ASSERT(existing->dtdv_id == v->dtdv_id);
8712 ASSERT(existing->dtdv_scope == v->dtdv_scope);
8714 if (existing->dtdv_kind != v->dtdv_kind)
8715 err += efunc(i, "%d changed variable kind\n", id);
8717 et = &existing->dtdv_type;
8719 if (vt->dtdt_flags != et->dtdt_flags) {
8720 err += efunc(i, "%d changed variable type flags\n", id);
8724 if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
8725 err += efunc(i, "%d changed variable type size\n", id);
8734 * Validate a DTrace DIF object that it is to be used as a helper. Helpers
8735 * are much more constrained than normal DIFOs. Specifically, they may
8738 * 1. Make calls to subroutines other than copyin(), copyinstr() or
8739 * miscellaneous string routines
8740 * 2. Access DTrace variables other than the args[] array, and the
8741 * curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
8742 * 3. Have thread-local variables.
8743 * 4. Have dynamic variables.
8746 dtrace_difo_validate_helper(dtrace_difo_t *dp)
8748 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8752 for (pc = 0; pc < dp->dtdo_len; pc++) {
8753 dif_instr_t instr = dp->dtdo_buf[pc];
8755 uint_t v = DIF_INSTR_VAR(instr);
8756 uint_t subr = DIF_INSTR_SUBR(instr);
8757 uint_t op = DIF_INSTR_OP(instr);
8812 case DIF_OP_FLUSHTS:
8824 if (v >= DIF_VAR_OTHER_UBASE)
8827 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
8830 if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
8831 v == DIF_VAR_PPID || v == DIF_VAR_TID ||
8832 v == DIF_VAR_EXECARGS ||
8833 v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
8834 v == DIF_VAR_UID || v == DIF_VAR_GID)
8837 err += efunc(pc, "illegal variable %u\n", v);
8844 err += efunc(pc, "illegal dynamic variable load\n");
8850 err += efunc(pc, "illegal dynamic variable store\n");
8854 if (subr == DIF_SUBR_ALLOCA ||
8855 subr == DIF_SUBR_BCOPY ||
8856 subr == DIF_SUBR_COPYIN ||
8857 subr == DIF_SUBR_COPYINTO ||
8858 subr == DIF_SUBR_COPYINSTR ||
8859 subr == DIF_SUBR_INDEX ||
8860 subr == DIF_SUBR_INET_NTOA ||
8861 subr == DIF_SUBR_INET_NTOA6 ||
8862 subr == DIF_SUBR_INET_NTOP ||
8863 subr == DIF_SUBR_LLTOSTR ||
8864 subr == DIF_SUBR_RINDEX ||
8865 subr == DIF_SUBR_STRCHR ||
8866 subr == DIF_SUBR_STRJOIN ||
8867 subr == DIF_SUBR_STRRCHR ||
8868 subr == DIF_SUBR_STRSTR ||
8869 subr == DIF_SUBR_HTONS ||
8870 subr == DIF_SUBR_HTONL ||
8871 subr == DIF_SUBR_HTONLL ||
8872 subr == DIF_SUBR_NTOHS ||
8873 subr == DIF_SUBR_NTOHL ||
8874 subr == DIF_SUBR_NTOHLL ||
8875 subr == DIF_SUBR_MEMREF ||
8876 subr == DIF_SUBR_TYPEREF)
8879 err += efunc(pc, "invalid subr %u\n", subr);
8883 err += efunc(pc, "invalid opcode %u\n",
8884 DIF_INSTR_OP(instr));
8892 * Returns 1 if the expression in the DIF object can be cached on a per-thread
8896 dtrace_difo_cacheable(dtrace_difo_t *dp)
8903 for (i = 0; i < dp->dtdo_varlen; i++) {
8904 dtrace_difv_t *v = &dp->dtdo_vartab[i];
8906 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
8909 switch (v->dtdv_id) {
8910 case DIF_VAR_CURTHREAD:
8913 case DIF_VAR_EXECARGS:
8914 case DIF_VAR_EXECNAME:
8915 case DIF_VAR_ZONENAME:
8924 * This DIF object may be cacheable. Now we need to look for any
8925 * array loading instructions, any memory loading instructions, or
8926 * any stores to thread-local variables.
8928 for (i = 0; i < dp->dtdo_len; i++) {
8929 uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
8931 if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
8932 (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
8933 (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
8934 op == DIF_OP_LDGA || op == DIF_OP_STTS)
8942 dtrace_difo_hold(dtrace_difo_t *dp)
8946 ASSERT(MUTEX_HELD(&dtrace_lock));
8949 ASSERT(dp->dtdo_refcnt != 0);
8952 * We need to check this DIF object for references to the variable
8953 * DIF_VAR_VTIMESTAMP.
8955 for (i = 0; i < dp->dtdo_varlen; i++) {
8956 dtrace_difv_t *v = &dp->dtdo_vartab[i];
8958 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
8961 if (dtrace_vtime_references++ == 0)
8962 dtrace_vtime_enable();
8967 * This routine calculates the dynamic variable chunksize for a given DIF
8968 * object. The calculation is not fool-proof, and can probably be tricked by
8969 * malicious DIF -- but it works for all compiler-generated DIF. Because this
8970 * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
8971 * if a dynamic variable size exceeds the chunksize.
8974 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
8977 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
8978 const dif_instr_t *text = dp->dtdo_buf;
8984 for (pc = 0; pc < dp->dtdo_len; pc++) {
8985 dif_instr_t instr = text[pc];
8986 uint_t op = DIF_INSTR_OP(instr);
8987 uint_t rd = DIF_INSTR_RD(instr);
8988 uint_t r1 = DIF_INSTR_R1(instr);
8992 dtrace_key_t *key = tupregs;
8996 sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
9001 key = &tupregs[DIF_DTR_NREGS];
9002 key[0].dttk_size = 0;
9003 key[1].dttk_size = 0;
9005 scope = DIFV_SCOPE_THREAD;
9012 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
9013 key[nkeys++].dttk_size = 0;
9015 key[nkeys++].dttk_size = 0;
9017 if (op == DIF_OP_STTAA) {
9018 scope = DIFV_SCOPE_THREAD;
9020 scope = DIFV_SCOPE_GLOBAL;
9026 if (ttop == DIF_DTR_NREGS)
9029 if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
9031 * If the register for the size of the "pushtr"
9032 * is %r0 (or the value is 0) and the type is
9033 * a string, we'll use the system-wide default
9036 tupregs[ttop++].dttk_size =
9037 dtrace_strsize_default;
9042 tupregs[ttop++].dttk_size = sval;
9048 if (ttop == DIF_DTR_NREGS)
9051 tupregs[ttop++].dttk_size = 0;
9054 case DIF_OP_FLUSHTS:
9071 * We have a dynamic variable allocation; calculate its size.
9073 for (ksize = 0, i = 0; i < nkeys; i++)
9074 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
9076 size = sizeof (dtrace_dynvar_t);
9077 size += sizeof (dtrace_key_t) * (nkeys - 1);
9081 * Now we need to determine the size of the stored data.
9083 id = DIF_INSTR_VAR(instr);
9085 for (i = 0; i < dp->dtdo_varlen; i++) {
9086 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9088 if (v->dtdv_id == id && v->dtdv_scope == scope) {
9089 size += v->dtdv_type.dtdt_size;
9094 if (i == dp->dtdo_varlen)
9098 * We have the size. If this is larger than the chunk size
9099 * for our dynamic variable state, reset the chunk size.
9101 size = P2ROUNDUP(size, sizeof (uint64_t));
9103 if (size > vstate->dtvs_dynvars.dtds_chunksize)
9104 vstate->dtvs_dynvars.dtds_chunksize = size;
9109 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9111 int i, oldsvars, osz, nsz, otlocals, ntlocals;
9114 ASSERT(MUTEX_HELD(&dtrace_lock));
9115 ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
9117 for (i = 0; i < dp->dtdo_varlen; i++) {
9118 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9119 dtrace_statvar_t *svar, ***svarp = NULL;
9121 uint8_t scope = v->dtdv_scope;
9124 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9127 id -= DIF_VAR_OTHER_UBASE;
9130 case DIFV_SCOPE_THREAD:
9131 while (id >= (otlocals = vstate->dtvs_ntlocals)) {
9132 dtrace_difv_t *tlocals;
9134 if ((ntlocals = (otlocals << 1)) == 0)
9137 osz = otlocals * sizeof (dtrace_difv_t);
9138 nsz = ntlocals * sizeof (dtrace_difv_t);
9140 tlocals = kmem_zalloc(nsz, KM_SLEEP);
9143 bcopy(vstate->dtvs_tlocals,
9145 kmem_free(vstate->dtvs_tlocals, osz);
9148 vstate->dtvs_tlocals = tlocals;
9149 vstate->dtvs_ntlocals = ntlocals;
9152 vstate->dtvs_tlocals[id] = *v;
9155 case DIFV_SCOPE_LOCAL:
9156 np = &vstate->dtvs_nlocals;
9157 svarp = &vstate->dtvs_locals;
9159 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9160 dsize = NCPU * (v->dtdv_type.dtdt_size +
9163 dsize = NCPU * sizeof (uint64_t);
9167 case DIFV_SCOPE_GLOBAL:
9168 np = &vstate->dtvs_nglobals;
9169 svarp = &vstate->dtvs_globals;
9171 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9172 dsize = v->dtdv_type.dtdt_size +
9181 while (id >= (oldsvars = *np)) {
9182 dtrace_statvar_t **statics;
9183 int newsvars, oldsize, newsize;
9185 if ((newsvars = (oldsvars << 1)) == 0)
9188 oldsize = oldsvars * sizeof (dtrace_statvar_t *);
9189 newsize = newsvars * sizeof (dtrace_statvar_t *);
9191 statics = kmem_zalloc(newsize, KM_SLEEP);
9194 bcopy(*svarp, statics, oldsize);
9195 kmem_free(*svarp, oldsize);
9202 if ((svar = (*svarp)[id]) == NULL) {
9203 svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
9204 svar->dtsv_var = *v;
9206 if ((svar->dtsv_size = dsize) != 0) {
9207 svar->dtsv_data = (uint64_t)(uintptr_t)
9208 kmem_zalloc(dsize, KM_SLEEP);
9211 (*svarp)[id] = svar;
9214 svar->dtsv_refcnt++;
9217 dtrace_difo_chunksize(dp, vstate);
9218 dtrace_difo_hold(dp);
9221 static dtrace_difo_t *
9222 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9227 ASSERT(dp->dtdo_buf != NULL);
9228 ASSERT(dp->dtdo_refcnt != 0);
9230 new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
9232 ASSERT(dp->dtdo_buf != NULL);
9233 sz = dp->dtdo_len * sizeof (dif_instr_t);
9234 new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
9235 bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
9236 new->dtdo_len = dp->dtdo_len;
9238 if (dp->dtdo_strtab != NULL) {
9239 ASSERT(dp->dtdo_strlen != 0);
9240 new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
9241 bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
9242 new->dtdo_strlen = dp->dtdo_strlen;
9245 if (dp->dtdo_inttab != NULL) {
9246 ASSERT(dp->dtdo_intlen != 0);
9247 sz = dp->dtdo_intlen * sizeof (uint64_t);
9248 new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
9249 bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
9250 new->dtdo_intlen = dp->dtdo_intlen;
9253 if (dp->dtdo_vartab != NULL) {
9254 ASSERT(dp->dtdo_varlen != 0);
9255 sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
9256 new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
9257 bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
9258 new->dtdo_varlen = dp->dtdo_varlen;
9261 dtrace_difo_init(new, vstate);
9266 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9270 ASSERT(dp->dtdo_refcnt == 0);
9272 for (i = 0; i < dp->dtdo_varlen; i++) {
9273 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9274 dtrace_statvar_t *svar, **svarp = NULL;
9276 uint8_t scope = v->dtdv_scope;
9280 case DIFV_SCOPE_THREAD:
9283 case DIFV_SCOPE_LOCAL:
9284 np = &vstate->dtvs_nlocals;
9285 svarp = vstate->dtvs_locals;
9288 case DIFV_SCOPE_GLOBAL:
9289 np = &vstate->dtvs_nglobals;
9290 svarp = vstate->dtvs_globals;
9297 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9300 id -= DIF_VAR_OTHER_UBASE;
9304 ASSERT(svar != NULL);
9305 ASSERT(svar->dtsv_refcnt > 0);
9307 if (--svar->dtsv_refcnt > 0)
9310 if (svar->dtsv_size != 0) {
9311 ASSERT(svar->dtsv_data != 0);
9312 kmem_free((void *)(uintptr_t)svar->dtsv_data,
9316 kmem_free(svar, sizeof (dtrace_statvar_t));
9320 if (dp->dtdo_buf != NULL)
9321 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
9322 if (dp->dtdo_inttab != NULL)
9323 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
9324 if (dp->dtdo_strtab != NULL)
9325 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
9326 if (dp->dtdo_vartab != NULL)
9327 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
9329 kmem_free(dp, sizeof (dtrace_difo_t));
9333 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9337 ASSERT(MUTEX_HELD(&dtrace_lock));
9338 ASSERT(dp->dtdo_refcnt != 0);
9340 for (i = 0; i < dp->dtdo_varlen; i++) {
9341 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9343 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9346 ASSERT(dtrace_vtime_references > 0);
9347 if (--dtrace_vtime_references == 0)
9348 dtrace_vtime_disable();
9351 if (--dp->dtdo_refcnt == 0)
9352 dtrace_difo_destroy(dp, vstate);
9356 * DTrace Format Functions
9359 dtrace_format_add(dtrace_state_t *state, char *str)
9362 uint16_t ndx, len = strlen(str) + 1;
9364 fmt = kmem_zalloc(len, KM_SLEEP);
9365 bcopy(str, fmt, len);
9367 for (ndx = 0; ndx < state->dts_nformats; ndx++) {
9368 if (state->dts_formats[ndx] == NULL) {
9369 state->dts_formats[ndx] = fmt;
9374 if (state->dts_nformats == USHRT_MAX) {
9376 * This is only likely if a denial-of-service attack is being
9377 * attempted. As such, it's okay to fail silently here.
9379 kmem_free(fmt, len);
9384 * For simplicity, we always resize the formats array to be exactly the
9385 * number of formats.
9387 ndx = state->dts_nformats++;
9388 new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
9390 if (state->dts_formats != NULL) {
9392 bcopy(state->dts_formats, new, ndx * sizeof (char *));
9393 kmem_free(state->dts_formats, ndx * sizeof (char *));
9396 state->dts_formats = new;
9397 state->dts_formats[ndx] = fmt;
9403 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
9407 ASSERT(state->dts_formats != NULL);
9408 ASSERT(format <= state->dts_nformats);
9409 ASSERT(state->dts_formats[format - 1] != NULL);
9411 fmt = state->dts_formats[format - 1];
9412 kmem_free(fmt, strlen(fmt) + 1);
9413 state->dts_formats[format - 1] = NULL;
9417 dtrace_format_destroy(dtrace_state_t *state)
9421 if (state->dts_nformats == 0) {
9422 ASSERT(state->dts_formats == NULL);
9426 ASSERT(state->dts_formats != NULL);
9428 for (i = 0; i < state->dts_nformats; i++) {
9429 char *fmt = state->dts_formats[i];
9434 kmem_free(fmt, strlen(fmt) + 1);
9437 kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
9438 state->dts_nformats = 0;
9439 state->dts_formats = NULL;
9443 * DTrace Predicate Functions
9445 static dtrace_predicate_t *
9446 dtrace_predicate_create(dtrace_difo_t *dp)
9448 dtrace_predicate_t *pred;
9450 ASSERT(MUTEX_HELD(&dtrace_lock));
9451 ASSERT(dp->dtdo_refcnt != 0);
9453 pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
9454 pred->dtp_difo = dp;
9455 pred->dtp_refcnt = 1;
9457 if (!dtrace_difo_cacheable(dp))
9460 if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
9462 * This is only theoretically possible -- we have had 2^32
9463 * cacheable predicates on this machine. We cannot allow any
9464 * more predicates to become cacheable: as unlikely as it is,
9465 * there may be a thread caching a (now stale) predicate cache
9466 * ID. (N.B.: the temptation is being successfully resisted to
9467 * have this cmn_err() "Holy shit -- we executed this code!")
9472 pred->dtp_cacheid = dtrace_predcache_id++;
9478 dtrace_predicate_hold(dtrace_predicate_t *pred)
9480 ASSERT(MUTEX_HELD(&dtrace_lock));
9481 ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
9482 ASSERT(pred->dtp_refcnt > 0);
9488 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
9490 dtrace_difo_t *dp = pred->dtp_difo;
9492 ASSERT(MUTEX_HELD(&dtrace_lock));
9493 ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
9494 ASSERT(pred->dtp_refcnt > 0);
9496 if (--pred->dtp_refcnt == 0) {
9497 dtrace_difo_release(pred->dtp_difo, vstate);
9498 kmem_free(pred, sizeof (dtrace_predicate_t));
9503 * DTrace Action Description Functions
9505 static dtrace_actdesc_t *
9506 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
9507 uint64_t uarg, uint64_t arg)
9509 dtrace_actdesc_t *act;
9512 ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
9513 arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
9516 act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
9517 act->dtad_kind = kind;
9518 act->dtad_ntuple = ntuple;
9519 act->dtad_uarg = uarg;
9520 act->dtad_arg = arg;
9521 act->dtad_refcnt = 1;
9527 dtrace_actdesc_hold(dtrace_actdesc_t *act)
9529 ASSERT(act->dtad_refcnt >= 1);
9534 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
9536 dtrace_actkind_t kind = act->dtad_kind;
9539 ASSERT(act->dtad_refcnt >= 1);
9541 if (--act->dtad_refcnt != 0)
9544 if ((dp = act->dtad_difo) != NULL)
9545 dtrace_difo_release(dp, vstate);
9547 if (DTRACEACT_ISPRINTFLIKE(kind)) {
9548 char *str = (char *)(uintptr_t)act->dtad_arg;
9551 ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
9552 (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
9556 kmem_free(str, strlen(str) + 1);
9559 kmem_free(act, sizeof (dtrace_actdesc_t));
9563 * DTrace ECB Functions
9565 static dtrace_ecb_t *
9566 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
9571 ASSERT(MUTEX_HELD(&dtrace_lock));
9573 ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
9574 ecb->dte_predicate = NULL;
9575 ecb->dte_probe = probe;
9578 * The default size is the size of the default action: recording
9581 ecb->dte_size = ecb->dte_needed = sizeof (dtrace_epid_t);
9582 ecb->dte_alignment = sizeof (dtrace_epid_t);
9584 epid = state->dts_epid++;
9586 if (epid - 1 >= state->dts_necbs) {
9587 dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
9588 int necbs = state->dts_necbs << 1;
9590 ASSERT(epid == state->dts_necbs + 1);
9593 ASSERT(oecbs == NULL);
9597 ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
9600 bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
9602 dtrace_membar_producer();
9603 state->dts_ecbs = ecbs;
9605 if (oecbs != NULL) {
9607 * If this state is active, we must dtrace_sync()
9608 * before we can free the old dts_ecbs array: we're
9609 * coming in hot, and there may be active ring
9610 * buffer processing (which indexes into the dts_ecbs
9611 * array) on another CPU.
9613 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
9616 kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
9619 dtrace_membar_producer();
9620 state->dts_necbs = necbs;
9623 ecb->dte_state = state;
9625 ASSERT(state->dts_ecbs[epid - 1] == NULL);
9626 dtrace_membar_producer();
9627 state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
9633 dtrace_ecb_enable(dtrace_ecb_t *ecb)
9635 dtrace_probe_t *probe = ecb->dte_probe;
9637 ASSERT(MUTEX_HELD(&cpu_lock));
9638 ASSERT(MUTEX_HELD(&dtrace_lock));
9639 ASSERT(ecb->dte_next == NULL);
9641 if (probe == NULL) {
9643 * This is the NULL probe -- there's nothing to do.
9648 if (probe->dtpr_ecb == NULL) {
9649 dtrace_provider_t *prov = probe->dtpr_provider;
9652 * We're the first ECB on this probe.
9654 probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
9656 if (ecb->dte_predicate != NULL)
9657 probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
9659 prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
9660 probe->dtpr_id, probe->dtpr_arg);
9663 * This probe is already active. Swing the last pointer to
9664 * point to the new ECB, and issue a dtrace_sync() to assure
9665 * that all CPUs have seen the change.
9667 ASSERT(probe->dtpr_ecb_last != NULL);
9668 probe->dtpr_ecb_last->dte_next = ecb;
9669 probe->dtpr_ecb_last = ecb;
9670 probe->dtpr_predcache = 0;
9677 dtrace_ecb_resize(dtrace_ecb_t *ecb)
9679 uint32_t maxalign = sizeof (dtrace_epid_t);
9680 uint32_t align = sizeof (uint8_t), offs, diff;
9681 dtrace_action_t *act;
9683 uint32_t aggbase = UINT32_MAX;
9684 dtrace_state_t *state = ecb->dte_state;
9687 * If we record anything, we always record the epid. (And we always
9690 offs = sizeof (dtrace_epid_t);
9691 ecb->dte_size = ecb->dte_needed = sizeof (dtrace_epid_t);
9693 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
9694 dtrace_recdesc_t *rec = &act->dta_rec;
9696 if ((align = rec->dtrd_alignment) > maxalign)
9699 if (!wastuple && act->dta_intuple) {
9701 * This is the first record in a tuple. Align the
9702 * offset to be at offset 4 in an 8-byte aligned
9705 diff = offs + sizeof (dtrace_aggid_t);
9707 if ((diff = (diff & (sizeof (uint64_t) - 1))))
9708 offs += sizeof (uint64_t) - diff;
9710 aggbase = offs - sizeof (dtrace_aggid_t);
9711 ASSERT(!(aggbase & (sizeof (uint64_t) - 1)));
9715 if (rec->dtrd_size != 0 && (diff = (offs & (align - 1)))) {
9717 * The current offset is not properly aligned; align it.
9719 offs += align - diff;
9722 rec->dtrd_offset = offs;
9724 if (offs + rec->dtrd_size > ecb->dte_needed) {
9725 ecb->dte_needed = offs + rec->dtrd_size;
9727 if (ecb->dte_needed > state->dts_needed)
9728 state->dts_needed = ecb->dte_needed;
9731 if (DTRACEACT_ISAGG(act->dta_kind)) {
9732 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
9733 dtrace_action_t *first = agg->dtag_first, *prev;
9735 ASSERT(rec->dtrd_size != 0 && first != NULL);
9737 ASSERT(aggbase != UINT32_MAX);
9739 agg->dtag_base = aggbase;
9741 while ((prev = first->dta_prev) != NULL &&
9742 DTRACEACT_ISAGG(prev->dta_kind)) {
9743 agg = (dtrace_aggregation_t *)prev;
9744 first = agg->dtag_first;
9748 offs = prev->dta_rec.dtrd_offset +
9749 prev->dta_rec.dtrd_size;
9751 offs = sizeof (dtrace_epid_t);
9755 if (!act->dta_intuple)
9756 ecb->dte_size = offs + rec->dtrd_size;
9758 offs += rec->dtrd_size;
9761 wastuple = act->dta_intuple;
9764 if ((act = ecb->dte_action) != NULL &&
9765 !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
9766 ecb->dte_size == sizeof (dtrace_epid_t)) {
9768 * If the size is still sizeof (dtrace_epid_t), then all
9769 * actions store no data; set the size to 0.
9771 ecb->dte_alignment = maxalign;
9775 * If the needed space is still sizeof (dtrace_epid_t), then
9776 * all actions need no additional space; set the needed
9779 if (ecb->dte_needed == sizeof (dtrace_epid_t))
9780 ecb->dte_needed = 0;
9786 * Set our alignment, and make sure that the dte_size and dte_needed
9787 * are aligned to the size of an EPID.
9789 ecb->dte_alignment = maxalign;
9790 ecb->dte_size = (ecb->dte_size + (sizeof (dtrace_epid_t) - 1)) &
9791 ~(sizeof (dtrace_epid_t) - 1);
9792 ecb->dte_needed = (ecb->dte_needed + (sizeof (dtrace_epid_t) - 1)) &
9793 ~(sizeof (dtrace_epid_t) - 1);
9794 ASSERT(ecb->dte_size <= ecb->dte_needed);
9797 static dtrace_action_t *
9798 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
9800 dtrace_aggregation_t *agg;
9801 size_t size = sizeof (uint64_t);
9802 int ntuple = desc->dtad_ntuple;
9803 dtrace_action_t *act;
9804 dtrace_recdesc_t *frec;
9805 dtrace_aggid_t aggid;
9806 dtrace_state_t *state = ecb->dte_state;
9808 agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
9809 agg->dtag_ecb = ecb;
9811 ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
9813 switch (desc->dtad_kind) {
9815 agg->dtag_initial = INT64_MAX;
9816 agg->dtag_aggregate = dtrace_aggregate_min;
9820 agg->dtag_initial = INT64_MIN;
9821 agg->dtag_aggregate = dtrace_aggregate_max;
9824 case DTRACEAGG_COUNT:
9825 agg->dtag_aggregate = dtrace_aggregate_count;
9828 case DTRACEAGG_QUANTIZE:
9829 agg->dtag_aggregate = dtrace_aggregate_quantize;
9830 size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
9834 case DTRACEAGG_LQUANTIZE: {
9835 uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
9836 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
9838 agg->dtag_initial = desc->dtad_arg;
9839 agg->dtag_aggregate = dtrace_aggregate_lquantize;
9841 if (step == 0 || levels == 0)
9844 size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
9849 agg->dtag_aggregate = dtrace_aggregate_avg;
9850 size = sizeof (uint64_t) * 2;
9853 case DTRACEAGG_STDDEV:
9854 agg->dtag_aggregate = dtrace_aggregate_stddev;
9855 size = sizeof (uint64_t) * 4;
9859 agg->dtag_aggregate = dtrace_aggregate_sum;
9866 agg->dtag_action.dta_rec.dtrd_size = size;
9872 * We must make sure that we have enough actions for the n-tuple.
9874 for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
9875 if (DTRACEACT_ISAGG(act->dta_kind))
9878 if (--ntuple == 0) {
9880 * This is the action with which our n-tuple begins.
9882 agg->dtag_first = act;
9888 * This n-tuple is short by ntuple elements. Return failure.
9890 ASSERT(ntuple != 0);
9892 kmem_free(agg, sizeof (dtrace_aggregation_t));
9897 * If the last action in the tuple has a size of zero, it's actually
9898 * an expression argument for the aggregating action.
9900 ASSERT(ecb->dte_action_last != NULL);
9901 act = ecb->dte_action_last;
9903 if (act->dta_kind == DTRACEACT_DIFEXPR) {
9904 ASSERT(act->dta_difo != NULL);
9906 if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
9907 agg->dtag_hasarg = 1;
9911 * We need to allocate an id for this aggregation.
9914 aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
9915 VM_BESTFIT | VM_SLEEP);
9917 aggid = alloc_unr(state->dts_aggid_arena);
9920 if (aggid - 1 >= state->dts_naggregations) {
9921 dtrace_aggregation_t **oaggs = state->dts_aggregations;
9922 dtrace_aggregation_t **aggs;
9923 int naggs = state->dts_naggregations << 1;
9924 int onaggs = state->dts_naggregations;
9926 ASSERT(aggid == state->dts_naggregations + 1);
9929 ASSERT(oaggs == NULL);
9933 aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
9935 if (oaggs != NULL) {
9936 bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
9937 kmem_free(oaggs, onaggs * sizeof (*aggs));
9940 state->dts_aggregations = aggs;
9941 state->dts_naggregations = naggs;
9944 ASSERT(state->dts_aggregations[aggid - 1] == NULL);
9945 state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
9947 frec = &agg->dtag_first->dta_rec;
9948 if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
9949 frec->dtrd_alignment = sizeof (dtrace_aggid_t);
9951 for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
9952 ASSERT(!act->dta_intuple);
9953 act->dta_intuple = 1;
9956 return (&agg->dtag_action);
9960 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
9962 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
9963 dtrace_state_t *state = ecb->dte_state;
9964 dtrace_aggid_t aggid = agg->dtag_id;
9966 ASSERT(DTRACEACT_ISAGG(act->dta_kind));
9968 vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
9970 free_unr(state->dts_aggid_arena, aggid);
9973 ASSERT(state->dts_aggregations[aggid - 1] == agg);
9974 state->dts_aggregations[aggid - 1] = NULL;
9976 kmem_free(agg, sizeof (dtrace_aggregation_t));
9980 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
9982 dtrace_action_t *action, *last;
9983 dtrace_difo_t *dp = desc->dtad_difo;
9984 uint32_t size = 0, align = sizeof (uint8_t), mask;
9985 uint16_t format = 0;
9986 dtrace_recdesc_t *rec;
9987 dtrace_state_t *state = ecb->dte_state;
9988 dtrace_optval_t *opt = state->dts_options, nframes = 0, strsize;
9989 uint64_t arg = desc->dtad_arg;
9991 ASSERT(MUTEX_HELD(&dtrace_lock));
9992 ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
9994 if (DTRACEACT_ISAGG(desc->dtad_kind)) {
9996 * If this is an aggregating action, there must be neither
9997 * a speculate nor a commit on the action chain.
9999 dtrace_action_t *act;
10001 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
10002 if (act->dta_kind == DTRACEACT_COMMIT)
10005 if (act->dta_kind == DTRACEACT_SPECULATE)
10009 action = dtrace_ecb_aggregation_create(ecb, desc);
10011 if (action == NULL)
10014 if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
10015 (desc->dtad_kind == DTRACEACT_DIFEXPR &&
10016 dp != NULL && dp->dtdo_destructive)) {
10017 state->dts_destructive = 1;
10020 switch (desc->dtad_kind) {
10021 case DTRACEACT_PRINTF:
10022 case DTRACEACT_PRINTA:
10023 case DTRACEACT_SYSTEM:
10024 case DTRACEACT_FREOPEN:
10026 * We know that our arg is a string -- turn it into a
10030 ASSERT(desc->dtad_kind == DTRACEACT_PRINTA);
10035 ASSERT(arg > KERNELBASE);
10037 format = dtrace_format_add(state,
10038 (char *)(uintptr_t)arg);
10042 case DTRACEACT_LIBACT:
10043 case DTRACEACT_DIFEXPR:
10047 if ((size = dp->dtdo_rtype.dtdt_size) != 0)
10050 if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
10051 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10054 size = opt[DTRACEOPT_STRSIZE];
10059 case DTRACEACT_STACK:
10060 if ((nframes = arg) == 0) {
10061 nframes = opt[DTRACEOPT_STACKFRAMES];
10062 ASSERT(nframes > 0);
10066 size = nframes * sizeof (pc_t);
10069 case DTRACEACT_JSTACK:
10070 if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
10071 strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
10073 if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
10074 nframes = opt[DTRACEOPT_JSTACKFRAMES];
10076 arg = DTRACE_USTACK_ARG(nframes, strsize);
10079 case DTRACEACT_USTACK:
10080 if (desc->dtad_kind != DTRACEACT_JSTACK &&
10081 (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
10082 strsize = DTRACE_USTACK_STRSIZE(arg);
10083 nframes = opt[DTRACEOPT_USTACKFRAMES];
10084 ASSERT(nframes > 0);
10085 arg = DTRACE_USTACK_ARG(nframes, strsize);
10089 * Save a slot for the pid.
10091 size = (nframes + 1) * sizeof (uint64_t);
10092 size += DTRACE_USTACK_STRSIZE(arg);
10093 size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
10097 case DTRACEACT_SYM:
10098 case DTRACEACT_MOD:
10099 if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
10100 sizeof (uint64_t)) ||
10101 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10105 case DTRACEACT_USYM:
10106 case DTRACEACT_UMOD:
10107 case DTRACEACT_UADDR:
10109 (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
10110 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10114 * We have a slot for the pid, plus a slot for the
10115 * argument. To keep things simple (aligned with
10116 * bitness-neutral sizing), we store each as a 64-bit
10119 size = 2 * sizeof (uint64_t);
10122 case DTRACEACT_STOP:
10123 case DTRACEACT_BREAKPOINT:
10124 case DTRACEACT_PANIC:
10127 case DTRACEACT_CHILL:
10128 case DTRACEACT_DISCARD:
10129 case DTRACEACT_RAISE:
10134 case DTRACEACT_EXIT:
10136 (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
10137 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10141 case DTRACEACT_SPECULATE:
10142 if (ecb->dte_size > sizeof (dtrace_epid_t))
10148 state->dts_speculates = 1;
10151 case DTRACEACT_PRINTM:
10152 size = dp->dtdo_rtype.dtdt_size;
10155 case DTRACEACT_PRINTT:
10156 size = dp->dtdo_rtype.dtdt_size;
10159 case DTRACEACT_COMMIT: {
10160 dtrace_action_t *act = ecb->dte_action;
10162 for (; act != NULL; act = act->dta_next) {
10163 if (act->dta_kind == DTRACEACT_COMMIT)
10176 if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
10178 * If this is a data-storing action or a speculate,
10179 * we must be sure that there isn't a commit on the
10182 dtrace_action_t *act = ecb->dte_action;
10184 for (; act != NULL; act = act->dta_next) {
10185 if (act->dta_kind == DTRACEACT_COMMIT)
10190 action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
10191 action->dta_rec.dtrd_size = size;
10194 action->dta_refcnt = 1;
10195 rec = &action->dta_rec;
10196 size = rec->dtrd_size;
10198 for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
10199 if (!(size & mask)) {
10205 action->dta_kind = desc->dtad_kind;
10207 if ((action->dta_difo = dp) != NULL)
10208 dtrace_difo_hold(dp);
10210 rec->dtrd_action = action->dta_kind;
10211 rec->dtrd_arg = arg;
10212 rec->dtrd_uarg = desc->dtad_uarg;
10213 rec->dtrd_alignment = (uint16_t)align;
10214 rec->dtrd_format = format;
10216 if ((last = ecb->dte_action_last) != NULL) {
10217 ASSERT(ecb->dte_action != NULL);
10218 action->dta_prev = last;
10219 last->dta_next = action;
10221 ASSERT(ecb->dte_action == NULL);
10222 ecb->dte_action = action;
10225 ecb->dte_action_last = action;
10231 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
10233 dtrace_action_t *act = ecb->dte_action, *next;
10234 dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
10238 if (act != NULL && act->dta_refcnt > 1) {
10239 ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
10242 for (; act != NULL; act = next) {
10243 next = act->dta_next;
10244 ASSERT(next != NULL || act == ecb->dte_action_last);
10245 ASSERT(act->dta_refcnt == 1);
10247 if ((format = act->dta_rec.dtrd_format) != 0)
10248 dtrace_format_remove(ecb->dte_state, format);
10250 if ((dp = act->dta_difo) != NULL)
10251 dtrace_difo_release(dp, vstate);
10253 if (DTRACEACT_ISAGG(act->dta_kind)) {
10254 dtrace_ecb_aggregation_destroy(ecb, act);
10256 kmem_free(act, sizeof (dtrace_action_t));
10261 ecb->dte_action = NULL;
10262 ecb->dte_action_last = NULL;
10263 ecb->dte_size = sizeof (dtrace_epid_t);
10267 dtrace_ecb_disable(dtrace_ecb_t *ecb)
10270 * We disable the ECB by removing it from its probe.
10272 dtrace_ecb_t *pecb, *prev = NULL;
10273 dtrace_probe_t *probe = ecb->dte_probe;
10275 ASSERT(MUTEX_HELD(&dtrace_lock));
10277 if (probe == NULL) {
10279 * This is the NULL probe; there is nothing to disable.
10284 for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
10290 ASSERT(pecb != NULL);
10292 if (prev == NULL) {
10293 probe->dtpr_ecb = ecb->dte_next;
10295 prev->dte_next = ecb->dte_next;
10298 if (ecb == probe->dtpr_ecb_last) {
10299 ASSERT(ecb->dte_next == NULL);
10300 probe->dtpr_ecb_last = prev;
10304 * The ECB has been disconnected from the probe; now sync to assure
10305 * that all CPUs have seen the change before returning.
10309 if (probe->dtpr_ecb == NULL) {
10311 * That was the last ECB on the probe; clear the predicate
10312 * cache ID for the probe, disable it and sync one more time
10313 * to assure that we'll never hit it again.
10315 dtrace_provider_t *prov = probe->dtpr_provider;
10317 ASSERT(ecb->dte_next == NULL);
10318 ASSERT(probe->dtpr_ecb_last == NULL);
10319 probe->dtpr_predcache = DTRACE_CACHEIDNONE;
10320 prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
10321 probe->dtpr_id, probe->dtpr_arg);
10325 * There is at least one ECB remaining on the probe. If there
10326 * is _exactly_ one, set the probe's predicate cache ID to be
10327 * the predicate cache ID of the remaining ECB.
10329 ASSERT(probe->dtpr_ecb_last != NULL);
10330 ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
10332 if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
10333 dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
10335 ASSERT(probe->dtpr_ecb->dte_next == NULL);
10338 probe->dtpr_predcache = p->dtp_cacheid;
10341 ecb->dte_next = NULL;
10346 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
10348 dtrace_state_t *state = ecb->dte_state;
10349 dtrace_vstate_t *vstate = &state->dts_vstate;
10350 dtrace_predicate_t *pred;
10351 dtrace_epid_t epid = ecb->dte_epid;
10353 ASSERT(MUTEX_HELD(&dtrace_lock));
10354 ASSERT(ecb->dte_next == NULL);
10355 ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
10357 if ((pred = ecb->dte_predicate) != NULL)
10358 dtrace_predicate_release(pred, vstate);
10360 dtrace_ecb_action_remove(ecb);
10362 ASSERT(state->dts_ecbs[epid - 1] == ecb);
10363 state->dts_ecbs[epid - 1] = NULL;
10365 kmem_free(ecb, sizeof (dtrace_ecb_t));
10368 static dtrace_ecb_t *
10369 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
10370 dtrace_enabling_t *enab)
10373 dtrace_predicate_t *pred;
10374 dtrace_actdesc_t *act;
10375 dtrace_provider_t *prov;
10376 dtrace_ecbdesc_t *desc = enab->dten_current;
10378 ASSERT(MUTEX_HELD(&dtrace_lock));
10379 ASSERT(state != NULL);
10381 ecb = dtrace_ecb_add(state, probe);
10382 ecb->dte_uarg = desc->dted_uarg;
10384 if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
10385 dtrace_predicate_hold(pred);
10386 ecb->dte_predicate = pred;
10389 if (probe != NULL) {
10391 * If the provider shows more leg than the consumer is old
10392 * enough to see, we need to enable the appropriate implicit
10393 * predicate bits to prevent the ecb from activating at
10396 * Providers specifying DTRACE_PRIV_USER at register time
10397 * are stating that they need the /proc-style privilege
10398 * model to be enforced, and this is what DTRACE_COND_OWNER
10399 * and DTRACE_COND_ZONEOWNER will then do at probe time.
10401 prov = probe->dtpr_provider;
10402 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
10403 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10404 ecb->dte_cond |= DTRACE_COND_OWNER;
10406 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
10407 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10408 ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
10411 * If the provider shows us kernel innards and the user
10412 * is lacking sufficient privilege, enable the
10413 * DTRACE_COND_USERMODE implicit predicate.
10415 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
10416 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
10417 ecb->dte_cond |= DTRACE_COND_USERMODE;
10420 if (dtrace_ecb_create_cache != NULL) {
10422 * If we have a cached ecb, we'll use its action list instead
10423 * of creating our own (saving both time and space).
10425 dtrace_ecb_t *cached = dtrace_ecb_create_cache;
10426 dtrace_action_t *act = cached->dte_action;
10429 ASSERT(act->dta_refcnt > 0);
10431 ecb->dte_action = act;
10432 ecb->dte_action_last = cached->dte_action_last;
10433 ecb->dte_needed = cached->dte_needed;
10434 ecb->dte_size = cached->dte_size;
10435 ecb->dte_alignment = cached->dte_alignment;
10441 for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
10442 if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
10443 dtrace_ecb_destroy(ecb);
10448 dtrace_ecb_resize(ecb);
10450 return (dtrace_ecb_create_cache = ecb);
10454 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
10457 dtrace_enabling_t *enab = arg;
10458 dtrace_state_t *state = enab->dten_vstate->dtvs_state;
10460 ASSERT(state != NULL);
10462 if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
10464 * This probe was created in a generation for which this
10465 * enabling has previously created ECBs; we don't want to
10466 * enable it again, so just kick out.
10468 return (DTRACE_MATCH_NEXT);
10471 if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
10472 return (DTRACE_MATCH_DONE);
10474 dtrace_ecb_enable(ecb);
10475 return (DTRACE_MATCH_NEXT);
10478 static dtrace_ecb_t *
10479 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
10483 ASSERT(MUTEX_HELD(&dtrace_lock));
10485 if (id == 0 || id > state->dts_necbs)
10488 ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
10489 ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
10491 return (state->dts_ecbs[id - 1]);
10494 static dtrace_aggregation_t *
10495 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
10497 dtrace_aggregation_t *agg;
10499 ASSERT(MUTEX_HELD(&dtrace_lock));
10501 if (id == 0 || id > state->dts_naggregations)
10504 ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
10505 ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
10506 agg->dtag_id == id);
10508 return (state->dts_aggregations[id - 1]);
10512 * DTrace Buffer Functions
10514 * The following functions manipulate DTrace buffers. Most of these functions
10515 * are called in the context of establishing or processing consumer state;
10516 * exceptions are explicitly noted.
10520 * Note: called from cross call context. This function switches the two
10521 * buffers on a given CPU. The atomicity of this operation is assured by
10522 * disabling interrupts while the actual switch takes place; the disabling of
10523 * interrupts serializes the execution with any execution of dtrace_probe() on
10527 dtrace_buffer_switch(dtrace_buffer_t *buf)
10529 caddr_t tomax = buf->dtb_tomax;
10530 caddr_t xamot = buf->dtb_xamot;
10531 dtrace_icookie_t cookie;
10533 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
10534 ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
10536 cookie = dtrace_interrupt_disable();
10537 buf->dtb_tomax = xamot;
10538 buf->dtb_xamot = tomax;
10539 buf->dtb_xamot_drops = buf->dtb_drops;
10540 buf->dtb_xamot_offset = buf->dtb_offset;
10541 buf->dtb_xamot_errors = buf->dtb_errors;
10542 buf->dtb_xamot_flags = buf->dtb_flags;
10543 buf->dtb_offset = 0;
10544 buf->dtb_drops = 0;
10545 buf->dtb_errors = 0;
10546 buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
10547 dtrace_interrupt_enable(cookie);
10551 * Note: called from cross call context. This function activates a buffer
10552 * on a CPU. As with dtrace_buffer_switch(), the atomicity of the operation
10553 * is guaranteed by the disabling of interrupts.
10556 dtrace_buffer_activate(dtrace_state_t *state)
10558 dtrace_buffer_t *buf;
10559 dtrace_icookie_t cookie = dtrace_interrupt_disable();
10561 buf = &state->dts_buffer[curcpu];
10563 if (buf->dtb_tomax != NULL) {
10565 * We might like to assert that the buffer is marked inactive,
10566 * but this isn't necessarily true: the buffer for the CPU
10567 * that processes the BEGIN probe has its buffer activated
10568 * manually. In this case, we take the (harmless) action
10569 * re-clearing the bit INACTIVE bit.
10571 buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
10574 dtrace_interrupt_enable(cookie);
10578 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
10584 dtrace_buffer_t *buf;
10587 ASSERT(MUTEX_HELD(&cpu_lock));
10588 ASSERT(MUTEX_HELD(&dtrace_lock));
10590 if (size > dtrace_nonroot_maxsize &&
10591 !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
10597 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10600 buf = &bufs[cp->cpu_id];
10603 * If there is already a buffer allocated for this CPU, it
10604 * is only possible that this is a DR event. In this case,
10606 if (buf->dtb_tomax != NULL) {
10607 ASSERT(buf->dtb_size == size);
10611 ASSERT(buf->dtb_xamot == NULL);
10613 if ((buf->dtb_tomax = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10616 buf->dtb_size = size;
10617 buf->dtb_flags = flags;
10618 buf->dtb_offset = 0;
10619 buf->dtb_drops = 0;
10621 if (flags & DTRACEBUF_NOSWITCH)
10624 if ((buf->dtb_xamot = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10626 } while ((cp = cp->cpu_next) != cpu_list);
10634 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10637 buf = &bufs[cp->cpu_id];
10639 if (buf->dtb_xamot != NULL) {
10640 ASSERT(buf->dtb_tomax != NULL);
10641 ASSERT(buf->dtb_size == size);
10642 kmem_free(buf->dtb_xamot, size);
10645 if (buf->dtb_tomax != NULL) {
10646 ASSERT(buf->dtb_size == size);
10647 kmem_free(buf->dtb_tomax, size);
10650 buf->dtb_tomax = NULL;
10651 buf->dtb_xamot = NULL;
10653 } while ((cp = cp->cpu_next) != cpu_list);
10659 #if defined(__amd64__)
10661 * FreeBSD isn't good at limiting the amount of memory we
10662 * ask to malloc, so let's place a limit here before trying
10663 * to do something that might well end in tears at bedtime.
10665 if (size > physmem * PAGE_SIZE / (128 * (mp_maxid + 1)))
10669 ASSERT(MUTEX_HELD(&dtrace_lock));
10671 if (cpu != DTRACE_CPUALL && cpu != i)
10677 * If there is already a buffer allocated for this CPU, it
10678 * is only possible that this is a DR event. In this case,
10679 * the buffer size must match our specified size.
10681 if (buf->dtb_tomax != NULL) {
10682 ASSERT(buf->dtb_size == size);
10686 ASSERT(buf->dtb_xamot == NULL);
10688 if ((buf->dtb_tomax = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10691 buf->dtb_size = size;
10692 buf->dtb_flags = flags;
10693 buf->dtb_offset = 0;
10694 buf->dtb_drops = 0;
10696 if (flags & DTRACEBUF_NOSWITCH)
10699 if ((buf->dtb_xamot = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10707 * Error allocating memory, so free the buffers that were
10708 * allocated before the failed allocation.
10711 if (cpu != DTRACE_CPUALL && cpu != i)
10716 if (buf->dtb_xamot != NULL) {
10717 ASSERT(buf->dtb_tomax != NULL);
10718 ASSERT(buf->dtb_size == size);
10719 kmem_free(buf->dtb_xamot, size);
10722 if (buf->dtb_tomax != NULL) {
10723 ASSERT(buf->dtb_size == size);
10724 kmem_free(buf->dtb_tomax, size);
10727 buf->dtb_tomax = NULL;
10728 buf->dtb_xamot = NULL;
10738 * Note: called from probe context. This function just increments the drop
10739 * count on a buffer. It has been made a function to allow for the
10740 * possibility of understanding the source of mysterious drop counts. (A
10741 * problem for which one may be particularly disappointed that DTrace cannot
10742 * be used to understand DTrace.)
10745 dtrace_buffer_drop(dtrace_buffer_t *buf)
10751 * Note: called from probe context. This function is called to reserve space
10752 * in a buffer. If mstate is non-NULL, sets the scratch base and size in the
10753 * mstate. Returns the new offset in the buffer, or a negative value if an
10754 * error has occurred.
10757 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
10758 dtrace_state_t *state, dtrace_mstate_t *mstate)
10760 intptr_t offs = buf->dtb_offset, soffs;
10765 if (buf->dtb_flags & DTRACEBUF_INACTIVE)
10768 if ((tomax = buf->dtb_tomax) == NULL) {
10769 dtrace_buffer_drop(buf);
10773 if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
10774 while (offs & (align - 1)) {
10776 * Assert that our alignment is off by a number which
10777 * is itself sizeof (uint32_t) aligned.
10779 ASSERT(!((align - (offs & (align - 1))) &
10780 (sizeof (uint32_t) - 1)));
10781 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
10782 offs += sizeof (uint32_t);
10785 if ((soffs = offs + needed) > buf->dtb_size) {
10786 dtrace_buffer_drop(buf);
10790 if (mstate == NULL)
10793 mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
10794 mstate->dtms_scratch_size = buf->dtb_size - soffs;
10795 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
10800 if (buf->dtb_flags & DTRACEBUF_FILL) {
10801 if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
10802 (buf->dtb_flags & DTRACEBUF_FULL))
10807 total = needed + (offs & (align - 1));
10810 * For a ring buffer, life is quite a bit more complicated. Before
10811 * we can store any padding, we need to adjust our wrapping offset.
10812 * (If we've never before wrapped or we're not about to, no adjustment
10815 if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
10816 offs + total > buf->dtb_size) {
10817 woffs = buf->dtb_xamot_offset;
10819 if (offs + total > buf->dtb_size) {
10821 * We can't fit in the end of the buffer. First, a
10822 * sanity check that we can fit in the buffer at all.
10824 if (total > buf->dtb_size) {
10825 dtrace_buffer_drop(buf);
10830 * We're going to be storing at the top of the buffer,
10831 * so now we need to deal with the wrapped offset. We
10832 * only reset our wrapped offset to 0 if it is
10833 * currently greater than the current offset. If it
10834 * is less than the current offset, it is because a
10835 * previous allocation induced a wrap -- but the
10836 * allocation didn't subsequently take the space due
10837 * to an error or false predicate evaluation. In this
10838 * case, we'll just leave the wrapped offset alone: if
10839 * the wrapped offset hasn't been advanced far enough
10840 * for this allocation, it will be adjusted in the
10843 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
10851 * Now we know that we're going to be storing to the
10852 * top of the buffer and that there is room for us
10853 * there. We need to clear the buffer from the current
10854 * offset to the end (there may be old gunk there).
10856 while (offs < buf->dtb_size)
10860 * We need to set our offset to zero. And because we
10861 * are wrapping, we need to set the bit indicating as
10862 * much. We can also adjust our needed space back
10863 * down to the space required by the ECB -- we know
10864 * that the top of the buffer is aligned.
10868 buf->dtb_flags |= DTRACEBUF_WRAPPED;
10871 * There is room for us in the buffer, so we simply
10872 * need to check the wrapped offset.
10874 if (woffs < offs) {
10876 * The wrapped offset is less than the offset.
10877 * This can happen if we allocated buffer space
10878 * that induced a wrap, but then we didn't
10879 * subsequently take the space due to an error
10880 * or false predicate evaluation. This is
10881 * okay; we know that _this_ allocation isn't
10882 * going to induce a wrap. We still can't
10883 * reset the wrapped offset to be zero,
10884 * however: the space may have been trashed in
10885 * the previous failed probe attempt. But at
10886 * least the wrapped offset doesn't need to
10887 * be adjusted at all...
10893 while (offs + total > woffs) {
10894 dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
10897 if (epid == DTRACE_EPIDNONE) {
10898 size = sizeof (uint32_t);
10900 ASSERT(epid <= state->dts_necbs);
10901 ASSERT(state->dts_ecbs[epid - 1] != NULL);
10903 size = state->dts_ecbs[epid - 1]->dte_size;
10906 ASSERT(woffs + size <= buf->dtb_size);
10909 if (woffs + size == buf->dtb_size) {
10911 * We've reached the end of the buffer; we want
10912 * to set the wrapped offset to 0 and break
10913 * out. However, if the offs is 0, then we're
10914 * in a strange edge-condition: the amount of
10915 * space that we want to reserve plus the size
10916 * of the record that we're overwriting is
10917 * greater than the size of the buffer. This
10918 * is problematic because if we reserve the
10919 * space but subsequently don't consume it (due
10920 * to a failed predicate or error) the wrapped
10921 * offset will be 0 -- yet the EPID at offset 0
10922 * will not be committed. This situation is
10923 * relatively easy to deal with: if we're in
10924 * this case, the buffer is indistinguishable
10925 * from one that hasn't wrapped; we need only
10926 * finish the job by clearing the wrapped bit,
10927 * explicitly setting the offset to be 0, and
10928 * zero'ing out the old data in the buffer.
10931 buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
10932 buf->dtb_offset = 0;
10935 while (woffs < buf->dtb_size)
10936 tomax[woffs++] = 0;
10947 * We have a wrapped offset. It may be that the wrapped offset
10948 * has become zero -- that's okay.
10950 buf->dtb_xamot_offset = woffs;
10955 * Now we can plow the buffer with any necessary padding.
10957 while (offs & (align - 1)) {
10959 * Assert that our alignment is off by a number which
10960 * is itself sizeof (uint32_t) aligned.
10962 ASSERT(!((align - (offs & (align - 1))) &
10963 (sizeof (uint32_t) - 1)));
10964 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
10965 offs += sizeof (uint32_t);
10968 if (buf->dtb_flags & DTRACEBUF_FILL) {
10969 if (offs + needed > buf->dtb_size - state->dts_reserve) {
10970 buf->dtb_flags |= DTRACEBUF_FULL;
10975 if (mstate == NULL)
10979 * For ring buffers and fill buffers, the scratch space is always
10980 * the inactive buffer.
10982 mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
10983 mstate->dtms_scratch_size = buf->dtb_size;
10984 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
10990 dtrace_buffer_polish(dtrace_buffer_t *buf)
10992 ASSERT(buf->dtb_flags & DTRACEBUF_RING);
10993 ASSERT(MUTEX_HELD(&dtrace_lock));
10995 if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
10999 * We need to polish the ring buffer. There are three cases:
11001 * - The first (and presumably most common) is that there is no gap
11002 * between the buffer offset and the wrapped offset. In this case,
11003 * there is nothing in the buffer that isn't valid data; we can
11004 * mark the buffer as polished and return.
11006 * - The second (less common than the first but still more common
11007 * than the third) is that there is a gap between the buffer offset
11008 * and the wrapped offset, and the wrapped offset is larger than the
11009 * buffer offset. This can happen because of an alignment issue, or
11010 * can happen because of a call to dtrace_buffer_reserve() that
11011 * didn't subsequently consume the buffer space. In this case,
11012 * we need to zero the data from the buffer offset to the wrapped
11015 * - The third (and least common) is that there is a gap between the
11016 * buffer offset and the wrapped offset, but the wrapped offset is
11017 * _less_ than the buffer offset. This can only happen because a
11018 * call to dtrace_buffer_reserve() induced a wrap, but the space
11019 * was not subsequently consumed. In this case, we need to zero the
11020 * space from the offset to the end of the buffer _and_ from the
11021 * top of the buffer to the wrapped offset.
11023 if (buf->dtb_offset < buf->dtb_xamot_offset) {
11024 bzero(buf->dtb_tomax + buf->dtb_offset,
11025 buf->dtb_xamot_offset - buf->dtb_offset);
11028 if (buf->dtb_offset > buf->dtb_xamot_offset) {
11029 bzero(buf->dtb_tomax + buf->dtb_offset,
11030 buf->dtb_size - buf->dtb_offset);
11031 bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
11036 dtrace_buffer_free(dtrace_buffer_t *bufs)
11040 for (i = 0; i < NCPU; i++) {
11041 dtrace_buffer_t *buf = &bufs[i];
11043 if (buf->dtb_tomax == NULL) {
11044 ASSERT(buf->dtb_xamot == NULL);
11045 ASSERT(buf->dtb_size == 0);
11049 if (buf->dtb_xamot != NULL) {
11050 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
11051 kmem_free(buf->dtb_xamot, buf->dtb_size);
11054 kmem_free(buf->dtb_tomax, buf->dtb_size);
11056 buf->dtb_tomax = NULL;
11057 buf->dtb_xamot = NULL;
11062 * DTrace Enabling Functions
11064 static dtrace_enabling_t *
11065 dtrace_enabling_create(dtrace_vstate_t *vstate)
11067 dtrace_enabling_t *enab;
11069 enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
11070 enab->dten_vstate = vstate;
11076 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
11078 dtrace_ecbdesc_t **ndesc;
11079 size_t osize, nsize;
11082 * We can't add to enablings after we've enabled them, or after we've
11085 ASSERT(enab->dten_probegen == 0);
11086 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11088 if (enab->dten_ndesc < enab->dten_maxdesc) {
11089 enab->dten_desc[enab->dten_ndesc++] = ecb;
11093 osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11095 if (enab->dten_maxdesc == 0) {
11096 enab->dten_maxdesc = 1;
11098 enab->dten_maxdesc <<= 1;
11101 ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
11103 nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11104 ndesc = kmem_zalloc(nsize, KM_SLEEP);
11105 bcopy(enab->dten_desc, ndesc, osize);
11106 if (enab->dten_desc != NULL)
11107 kmem_free(enab->dten_desc, osize);
11109 enab->dten_desc = ndesc;
11110 enab->dten_desc[enab->dten_ndesc++] = ecb;
11114 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
11115 dtrace_probedesc_t *pd)
11117 dtrace_ecbdesc_t *new;
11118 dtrace_predicate_t *pred;
11119 dtrace_actdesc_t *act;
11122 * We're going to create a new ECB description that matches the
11123 * specified ECB in every way, but has the specified probe description.
11125 new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
11127 if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
11128 dtrace_predicate_hold(pred);
11130 for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
11131 dtrace_actdesc_hold(act);
11133 new->dted_action = ecb->dted_action;
11134 new->dted_pred = ecb->dted_pred;
11135 new->dted_probe = *pd;
11136 new->dted_uarg = ecb->dted_uarg;
11138 dtrace_enabling_add(enab, new);
11142 dtrace_enabling_dump(dtrace_enabling_t *enab)
11146 for (i = 0; i < enab->dten_ndesc; i++) {
11147 dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
11149 cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
11150 desc->dtpd_provider, desc->dtpd_mod,
11151 desc->dtpd_func, desc->dtpd_name);
11156 dtrace_enabling_destroy(dtrace_enabling_t *enab)
11159 dtrace_ecbdesc_t *ep;
11160 dtrace_vstate_t *vstate = enab->dten_vstate;
11162 ASSERT(MUTEX_HELD(&dtrace_lock));
11164 for (i = 0; i < enab->dten_ndesc; i++) {
11165 dtrace_actdesc_t *act, *next;
11166 dtrace_predicate_t *pred;
11168 ep = enab->dten_desc[i];
11170 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
11171 dtrace_predicate_release(pred, vstate);
11173 for (act = ep->dted_action; act != NULL; act = next) {
11174 next = act->dtad_next;
11175 dtrace_actdesc_release(act, vstate);
11178 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
11181 if (enab->dten_desc != NULL)
11182 kmem_free(enab->dten_desc,
11183 enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
11186 * If this was a retained enabling, decrement the dts_nretained count
11187 * and take it off of the dtrace_retained list.
11189 if (enab->dten_prev != NULL || enab->dten_next != NULL ||
11190 dtrace_retained == enab) {
11191 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11192 ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
11193 enab->dten_vstate->dtvs_state->dts_nretained--;
11196 if (enab->dten_prev == NULL) {
11197 if (dtrace_retained == enab) {
11198 dtrace_retained = enab->dten_next;
11200 if (dtrace_retained != NULL)
11201 dtrace_retained->dten_prev = NULL;
11204 ASSERT(enab != dtrace_retained);
11205 ASSERT(dtrace_retained != NULL);
11206 enab->dten_prev->dten_next = enab->dten_next;
11209 if (enab->dten_next != NULL) {
11210 ASSERT(dtrace_retained != NULL);
11211 enab->dten_next->dten_prev = enab->dten_prev;
11214 kmem_free(enab, sizeof (dtrace_enabling_t));
11218 dtrace_enabling_retain(dtrace_enabling_t *enab)
11220 dtrace_state_t *state;
11222 ASSERT(MUTEX_HELD(&dtrace_lock));
11223 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11224 ASSERT(enab->dten_vstate != NULL);
11226 state = enab->dten_vstate->dtvs_state;
11227 ASSERT(state != NULL);
11230 * We only allow each state to retain dtrace_retain_max enablings.
11232 if (state->dts_nretained >= dtrace_retain_max)
11235 state->dts_nretained++;
11237 if (dtrace_retained == NULL) {
11238 dtrace_retained = enab;
11242 enab->dten_next = dtrace_retained;
11243 dtrace_retained->dten_prev = enab;
11244 dtrace_retained = enab;
11250 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
11251 dtrace_probedesc_t *create)
11253 dtrace_enabling_t *new, *enab;
11254 int found = 0, err = ENOENT;
11256 ASSERT(MUTEX_HELD(&dtrace_lock));
11257 ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
11258 ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
11259 ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
11260 ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
11262 new = dtrace_enabling_create(&state->dts_vstate);
11265 * Iterate over all retained enablings, looking for enablings that
11266 * match the specified state.
11268 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11272 * dtvs_state can only be NULL for helper enablings -- and
11273 * helper enablings can't be retained.
11275 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11277 if (enab->dten_vstate->dtvs_state != state)
11281 * Now iterate over each probe description; we're looking for
11282 * an exact match to the specified probe description.
11284 for (i = 0; i < enab->dten_ndesc; i++) {
11285 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11286 dtrace_probedesc_t *pd = &ep->dted_probe;
11288 if (strcmp(pd->dtpd_provider, match->dtpd_provider))
11291 if (strcmp(pd->dtpd_mod, match->dtpd_mod))
11294 if (strcmp(pd->dtpd_func, match->dtpd_func))
11297 if (strcmp(pd->dtpd_name, match->dtpd_name))
11301 * We have a winning probe! Add it to our growing
11305 dtrace_enabling_addlike(new, ep, create);
11309 if (!found || (err = dtrace_enabling_retain(new)) != 0) {
11310 dtrace_enabling_destroy(new);
11318 dtrace_enabling_retract(dtrace_state_t *state)
11320 dtrace_enabling_t *enab, *next;
11322 ASSERT(MUTEX_HELD(&dtrace_lock));
11325 * Iterate over all retained enablings, destroy the enablings retained
11326 * for the specified state.
11328 for (enab = dtrace_retained; enab != NULL; enab = next) {
11329 next = enab->dten_next;
11332 * dtvs_state can only be NULL for helper enablings -- and
11333 * helper enablings can't be retained.
11335 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11337 if (enab->dten_vstate->dtvs_state == state) {
11338 ASSERT(state->dts_nretained > 0);
11339 dtrace_enabling_destroy(enab);
11343 ASSERT(state->dts_nretained == 0);
11347 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
11352 ASSERT(MUTEX_HELD(&cpu_lock));
11353 ASSERT(MUTEX_HELD(&dtrace_lock));
11355 for (i = 0; i < enab->dten_ndesc; i++) {
11356 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11358 enab->dten_current = ep;
11359 enab->dten_error = 0;
11361 matched += dtrace_probe_enable(&ep->dted_probe, enab);
11363 if (enab->dten_error != 0) {
11365 * If we get an error half-way through enabling the
11366 * probes, we kick out -- perhaps with some number of
11367 * them enabled. Leaving enabled probes enabled may
11368 * be slightly confusing for user-level, but we expect
11369 * that no one will attempt to actually drive on in
11370 * the face of such errors. If this is an anonymous
11371 * enabling (indicated with a NULL nmatched pointer),
11372 * we cmn_err() a message. We aren't expecting to
11373 * get such an error -- such as it can exist at all,
11374 * it would be a result of corrupted DOF in the driver
11377 if (nmatched == NULL) {
11378 cmn_err(CE_WARN, "dtrace_enabling_match() "
11379 "error on %p: %d", (void *)ep,
11383 return (enab->dten_error);
11387 enab->dten_probegen = dtrace_probegen;
11388 if (nmatched != NULL)
11389 *nmatched = matched;
11395 dtrace_enabling_matchall(void)
11397 dtrace_enabling_t *enab;
11399 mutex_enter(&cpu_lock);
11400 mutex_enter(&dtrace_lock);
11403 * Iterate over all retained enablings to see if any probes match
11404 * against them. We only perform this operation on enablings for which
11405 * we have sufficient permissions by virtue of being in the global zone
11406 * or in the same zone as the DTrace client. Because we can be called
11407 * after dtrace_detach() has been called, we cannot assert that there
11408 * are retained enablings. We can safely load from dtrace_retained,
11409 * however: the taskq_destroy() at the end of dtrace_detach() will
11410 * block pending our completion.
11412 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11414 cred_t *cr = enab->dten_vstate->dtvs_state->dts_cred.dcr_cred;
11416 if (INGLOBALZONE(curproc) || getzoneid() == crgetzoneid(cr))
11418 (void) dtrace_enabling_match(enab, NULL);
11421 mutex_exit(&dtrace_lock);
11422 mutex_exit(&cpu_lock);
11426 * If an enabling is to be enabled without having matched probes (that is, if
11427 * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
11428 * enabling must be _primed_ by creating an ECB for every ECB description.
11429 * This must be done to assure that we know the number of speculations, the
11430 * number of aggregations, the minimum buffer size needed, etc. before we
11431 * transition out of DTRACE_ACTIVITY_INACTIVE. To do this without actually
11432 * enabling any probes, we create ECBs for every ECB decription, but with a
11433 * NULL probe -- which is exactly what this function does.
11436 dtrace_enabling_prime(dtrace_state_t *state)
11438 dtrace_enabling_t *enab;
11441 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11442 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11444 if (enab->dten_vstate->dtvs_state != state)
11448 * We don't want to prime an enabling more than once, lest
11449 * we allow a malicious user to induce resource exhaustion.
11450 * (The ECBs that result from priming an enabling aren't
11451 * leaked -- but they also aren't deallocated until the
11452 * consumer state is destroyed.)
11454 if (enab->dten_primed)
11457 for (i = 0; i < enab->dten_ndesc; i++) {
11458 enab->dten_current = enab->dten_desc[i];
11459 (void) dtrace_probe_enable(NULL, enab);
11462 enab->dten_primed = 1;
11467 * Called to indicate that probes should be provided due to retained
11468 * enablings. This is implemented in terms of dtrace_probe_provide(), but it
11469 * must take an initial lap through the enabling calling the dtps_provide()
11470 * entry point explicitly to allow for autocreated probes.
11473 dtrace_enabling_provide(dtrace_provider_t *prv)
11476 dtrace_probedesc_t desc;
11478 ASSERT(MUTEX_HELD(&dtrace_lock));
11479 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
11483 prv = dtrace_provider;
11487 dtrace_enabling_t *enab = dtrace_retained;
11488 void *parg = prv->dtpv_arg;
11490 for (; enab != NULL; enab = enab->dten_next) {
11491 for (i = 0; i < enab->dten_ndesc; i++) {
11492 desc = enab->dten_desc[i]->dted_probe;
11493 mutex_exit(&dtrace_lock);
11494 prv->dtpv_pops.dtps_provide(parg, &desc);
11495 mutex_enter(&dtrace_lock);
11498 } while (all && (prv = prv->dtpv_next) != NULL);
11500 mutex_exit(&dtrace_lock);
11501 dtrace_probe_provide(NULL, all ? NULL : prv);
11502 mutex_enter(&dtrace_lock);
11506 * DTrace DOF Functions
11510 dtrace_dof_error(dof_hdr_t *dof, const char *str)
11512 if (dtrace_err_verbose)
11513 cmn_err(CE_WARN, "failed to process DOF: %s", str);
11515 #ifdef DTRACE_ERRDEBUG
11516 dtrace_errdebug(str);
11521 * Create DOF out of a currently enabled state. Right now, we only create
11522 * DOF containing the run-time options -- but this could be expanded to create
11523 * complete DOF representing the enabled state.
11526 dtrace_dof_create(dtrace_state_t *state)
11530 dof_optdesc_t *opt;
11531 int i, len = sizeof (dof_hdr_t) +
11532 roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
11533 sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11535 ASSERT(MUTEX_HELD(&dtrace_lock));
11537 dof = kmem_zalloc(len, KM_SLEEP);
11538 dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
11539 dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
11540 dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
11541 dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
11543 dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
11544 dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
11545 dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
11546 dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
11547 dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
11548 dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
11550 dof->dofh_flags = 0;
11551 dof->dofh_hdrsize = sizeof (dof_hdr_t);
11552 dof->dofh_secsize = sizeof (dof_sec_t);
11553 dof->dofh_secnum = 1; /* only DOF_SECT_OPTDESC */
11554 dof->dofh_secoff = sizeof (dof_hdr_t);
11555 dof->dofh_loadsz = len;
11556 dof->dofh_filesz = len;
11560 * Fill in the option section header...
11562 sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
11563 sec->dofs_type = DOF_SECT_OPTDESC;
11564 sec->dofs_align = sizeof (uint64_t);
11565 sec->dofs_flags = DOF_SECF_LOAD;
11566 sec->dofs_entsize = sizeof (dof_optdesc_t);
11568 opt = (dof_optdesc_t *)((uintptr_t)sec +
11569 roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
11571 sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
11572 sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11574 for (i = 0; i < DTRACEOPT_MAX; i++) {
11575 opt[i].dofo_option = i;
11576 opt[i].dofo_strtab = DOF_SECIDX_NONE;
11577 opt[i].dofo_value = state->dts_options[i];
11584 dtrace_dof_copyin(uintptr_t uarg, int *errp)
11586 dof_hdr_t hdr, *dof;
11588 ASSERT(!MUTEX_HELD(&dtrace_lock));
11591 * First, we're going to copyin() the sizeof (dof_hdr_t).
11593 if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
11594 dtrace_dof_error(NULL, "failed to copyin DOF header");
11600 * Now we'll allocate the entire DOF and copy it in -- provided
11601 * that the length isn't outrageous.
11603 if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
11604 dtrace_dof_error(&hdr, "load size exceeds maximum");
11609 if (hdr.dofh_loadsz < sizeof (hdr)) {
11610 dtrace_dof_error(&hdr, "invalid load size");
11615 dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
11617 if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0) {
11618 kmem_free(dof, hdr.dofh_loadsz);
11627 static __inline uchar_t
11628 dtrace_dof_char(char c) {
11647 return (c - 'A' + 10);
11654 return (c - 'a' + 10);
11656 /* Should not reach here. */
11662 dtrace_dof_property(const char *name)
11666 unsigned int len, i;
11671 * Unfortunately, array of values in .conf files are always (and
11672 * only) interpreted to be integer arrays. We must read our DOF
11673 * as an integer array, and then squeeze it into a byte array.
11675 if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
11676 (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
11679 for (i = 0; i < len; i++)
11680 buf[i] = (uchar_t)(((int *)buf)[i]);
11682 if (len < sizeof (dof_hdr_t)) {
11683 ddi_prop_free(buf);
11684 dtrace_dof_error(NULL, "truncated header");
11688 if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
11689 ddi_prop_free(buf);
11690 dtrace_dof_error(NULL, "truncated DOF");
11694 if (loadsz >= dtrace_dof_maxsize) {
11695 ddi_prop_free(buf);
11696 dtrace_dof_error(NULL, "oversized DOF");
11700 dof = kmem_alloc(loadsz, KM_SLEEP);
11701 bcopy(buf, dof, loadsz);
11702 ddi_prop_free(buf);
11707 if ((p_env = getenv(name)) == NULL)
11710 len = strlen(p_env) / 2;
11712 buf = kmem_alloc(len, KM_SLEEP);
11714 dof = (dof_hdr_t *) buf;
11718 for (i = 0; i < len; i++) {
11719 buf[i] = (dtrace_dof_char(p[0]) << 4) |
11720 dtrace_dof_char(p[1]);
11726 if (len < sizeof (dof_hdr_t)) {
11728 dtrace_dof_error(NULL, "truncated header");
11732 if (len < (loadsz = dof->dofh_loadsz)) {
11734 dtrace_dof_error(NULL, "truncated DOF");
11738 if (loadsz >= dtrace_dof_maxsize) {
11740 dtrace_dof_error(NULL, "oversized DOF");
11749 dtrace_dof_destroy(dof_hdr_t *dof)
11751 kmem_free(dof, dof->dofh_loadsz);
11755 * Return the dof_sec_t pointer corresponding to a given section index. If the
11756 * index is not valid, dtrace_dof_error() is called and NULL is returned. If
11757 * a type other than DOF_SECT_NONE is specified, the header is checked against
11758 * this type and NULL is returned if the types do not match.
11761 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
11763 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
11764 ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
11766 if (i >= dof->dofh_secnum) {
11767 dtrace_dof_error(dof, "referenced section index is invalid");
11771 if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
11772 dtrace_dof_error(dof, "referenced section is not loadable");
11776 if (type != DOF_SECT_NONE && type != sec->dofs_type) {
11777 dtrace_dof_error(dof, "referenced section is the wrong type");
11784 static dtrace_probedesc_t *
11785 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
11787 dof_probedesc_t *probe;
11789 uintptr_t daddr = (uintptr_t)dof;
11793 if (sec->dofs_type != DOF_SECT_PROBEDESC) {
11794 dtrace_dof_error(dof, "invalid probe section");
11798 if (sec->dofs_align != sizeof (dof_secidx_t)) {
11799 dtrace_dof_error(dof, "bad alignment in probe description");
11803 if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
11804 dtrace_dof_error(dof, "truncated probe description");
11808 probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
11809 strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
11811 if (strtab == NULL)
11814 str = daddr + strtab->dofs_offset;
11815 size = strtab->dofs_size;
11817 if (probe->dofp_provider >= strtab->dofs_size) {
11818 dtrace_dof_error(dof, "corrupt probe provider");
11822 (void) strncpy(desc->dtpd_provider,
11823 (char *)(str + probe->dofp_provider),
11824 MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
11826 if (probe->dofp_mod >= strtab->dofs_size) {
11827 dtrace_dof_error(dof, "corrupt probe module");
11831 (void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
11832 MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
11834 if (probe->dofp_func >= strtab->dofs_size) {
11835 dtrace_dof_error(dof, "corrupt probe function");
11839 (void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
11840 MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
11842 if (probe->dofp_name >= strtab->dofs_size) {
11843 dtrace_dof_error(dof, "corrupt probe name");
11847 (void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
11848 MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
11853 static dtrace_difo_t *
11854 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
11859 dof_difohdr_t *dofd;
11860 uintptr_t daddr = (uintptr_t)dof;
11861 size_t max = dtrace_difo_maxsize;
11864 static const struct {
11872 { DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
11873 offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
11874 sizeof (dif_instr_t), "multiple DIF sections" },
11876 { DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
11877 offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
11878 sizeof (uint64_t), "multiple integer tables" },
11880 { DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
11881 offsetof(dtrace_difo_t, dtdo_strlen), 0,
11882 sizeof (char), "multiple string tables" },
11884 { DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
11885 offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
11886 sizeof (uint_t), "multiple variable tables" },
11888 { DOF_SECT_NONE, 0, 0, 0, 0, NULL }
11891 if (sec->dofs_type != DOF_SECT_DIFOHDR) {
11892 dtrace_dof_error(dof, "invalid DIFO header section");
11896 if (sec->dofs_align != sizeof (dof_secidx_t)) {
11897 dtrace_dof_error(dof, "bad alignment in DIFO header");
11901 if (sec->dofs_size < sizeof (dof_difohdr_t) ||
11902 sec->dofs_size % sizeof (dof_secidx_t)) {
11903 dtrace_dof_error(dof, "bad size in DIFO header");
11907 dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
11908 n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
11910 dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
11911 dp->dtdo_rtype = dofd->dofd_rtype;
11913 for (l = 0; l < n; l++) {
11918 if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
11919 dofd->dofd_links[l])) == NULL)
11920 goto err; /* invalid section link */
11922 if (ttl + subsec->dofs_size > max) {
11923 dtrace_dof_error(dof, "exceeds maximum size");
11927 ttl += subsec->dofs_size;
11929 for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
11930 if (subsec->dofs_type != difo[i].section)
11933 if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
11934 dtrace_dof_error(dof, "section not loaded");
11938 if (subsec->dofs_align != difo[i].align) {
11939 dtrace_dof_error(dof, "bad alignment");
11943 bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
11944 lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
11946 if (*bufp != NULL) {
11947 dtrace_dof_error(dof, difo[i].msg);
11951 if (difo[i].entsize != subsec->dofs_entsize) {
11952 dtrace_dof_error(dof, "entry size mismatch");
11956 if (subsec->dofs_entsize != 0 &&
11957 (subsec->dofs_size % subsec->dofs_entsize) != 0) {
11958 dtrace_dof_error(dof, "corrupt entry size");
11962 *lenp = subsec->dofs_size;
11963 *bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
11964 bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
11965 *bufp, subsec->dofs_size);
11967 if (subsec->dofs_entsize != 0)
11968 *lenp /= subsec->dofs_entsize;
11974 * If we encounter a loadable DIFO sub-section that is not
11975 * known to us, assume this is a broken program and fail.
11977 if (difo[i].section == DOF_SECT_NONE &&
11978 (subsec->dofs_flags & DOF_SECF_LOAD)) {
11979 dtrace_dof_error(dof, "unrecognized DIFO subsection");
11984 if (dp->dtdo_buf == NULL) {
11986 * We can't have a DIF object without DIF text.
11988 dtrace_dof_error(dof, "missing DIF text");
11993 * Before we validate the DIF object, run through the variable table
11994 * looking for the strings -- if any of their size are under, we'll set
11995 * their size to be the system-wide default string size. Note that
11996 * this should _not_ happen if the "strsize" option has been set --
11997 * in this case, the compiler should have set the size to reflect the
11998 * setting of the option.
12000 for (i = 0; i < dp->dtdo_varlen; i++) {
12001 dtrace_difv_t *v = &dp->dtdo_vartab[i];
12002 dtrace_diftype_t *t = &v->dtdv_type;
12004 if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
12007 if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
12008 t->dtdt_size = dtrace_strsize_default;
12011 if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
12014 dtrace_difo_init(dp, vstate);
12018 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
12019 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
12020 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
12021 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
12023 kmem_free(dp, sizeof (dtrace_difo_t));
12027 static dtrace_predicate_t *
12028 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12033 if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
12036 return (dtrace_predicate_create(dp));
12039 static dtrace_actdesc_t *
12040 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12043 dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
12044 dof_actdesc_t *desc;
12045 dof_sec_t *difosec;
12047 uintptr_t daddr = (uintptr_t)dof;
12049 dtrace_actkind_t kind;
12051 if (sec->dofs_type != DOF_SECT_ACTDESC) {
12052 dtrace_dof_error(dof, "invalid action section");
12056 if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
12057 dtrace_dof_error(dof, "truncated action description");
12061 if (sec->dofs_align != sizeof (uint64_t)) {
12062 dtrace_dof_error(dof, "bad alignment in action description");
12066 if (sec->dofs_size < sec->dofs_entsize) {
12067 dtrace_dof_error(dof, "section entry size exceeds total size");
12071 if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
12072 dtrace_dof_error(dof, "bad entry size in action description");
12076 if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
12077 dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
12081 for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
12082 desc = (dof_actdesc_t *)(daddr +
12083 (uintptr_t)sec->dofs_offset + offs);
12084 kind = (dtrace_actkind_t)desc->dofa_kind;
12086 if (DTRACEACT_ISPRINTFLIKE(kind) &&
12087 (kind != DTRACEACT_PRINTA ||
12088 desc->dofa_strtab != DOF_SECIDX_NONE)) {
12094 * printf()-like actions must have a format string.
12096 if ((strtab = dtrace_dof_sect(dof,
12097 DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
12100 str = (char *)((uintptr_t)dof +
12101 (uintptr_t)strtab->dofs_offset);
12103 for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
12104 if (str[i] == '\0')
12108 if (i >= strtab->dofs_size) {
12109 dtrace_dof_error(dof, "bogus format string");
12113 if (i == desc->dofa_arg) {
12114 dtrace_dof_error(dof, "empty format string");
12118 i -= desc->dofa_arg;
12119 fmt = kmem_alloc(i + 1, KM_SLEEP);
12120 bcopy(&str[desc->dofa_arg], fmt, i + 1);
12121 arg = (uint64_t)(uintptr_t)fmt;
12123 if (kind == DTRACEACT_PRINTA) {
12124 ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
12127 arg = desc->dofa_arg;
12131 act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
12132 desc->dofa_uarg, arg);
12134 if (last != NULL) {
12135 last->dtad_next = act;
12142 if (desc->dofa_difo == DOF_SECIDX_NONE)
12145 if ((difosec = dtrace_dof_sect(dof,
12146 DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
12149 act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
12151 if (act->dtad_difo == NULL)
12155 ASSERT(first != NULL);
12159 for (act = first; act != NULL; act = next) {
12160 next = act->dtad_next;
12161 dtrace_actdesc_release(act, vstate);
12167 static dtrace_ecbdesc_t *
12168 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12171 dtrace_ecbdesc_t *ep;
12172 dof_ecbdesc_t *ecb;
12173 dtrace_probedesc_t *desc;
12174 dtrace_predicate_t *pred = NULL;
12176 if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
12177 dtrace_dof_error(dof, "truncated ECB description");
12181 if (sec->dofs_align != sizeof (uint64_t)) {
12182 dtrace_dof_error(dof, "bad alignment in ECB description");
12186 ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
12187 sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
12192 ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
12193 ep->dted_uarg = ecb->dofe_uarg;
12194 desc = &ep->dted_probe;
12196 if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
12199 if (ecb->dofe_pred != DOF_SECIDX_NONE) {
12200 if ((sec = dtrace_dof_sect(dof,
12201 DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
12204 if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
12207 ep->dted_pred.dtpdd_predicate = pred;
12210 if (ecb->dofe_actions != DOF_SECIDX_NONE) {
12211 if ((sec = dtrace_dof_sect(dof,
12212 DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
12215 ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
12217 if (ep->dted_action == NULL)
12225 dtrace_predicate_release(pred, vstate);
12226 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
12231 * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
12232 * specified DOF. At present, this amounts to simply adding 'ubase' to the
12233 * site of any user SETX relocations to account for load object base address.
12234 * In the future, if we need other relocations, this function can be extended.
12237 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase)
12239 uintptr_t daddr = (uintptr_t)dof;
12240 dof_relohdr_t *dofr =
12241 (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12242 dof_sec_t *ss, *rs, *ts;
12246 if (sec->dofs_size < sizeof (dof_relohdr_t) ||
12247 sec->dofs_align != sizeof (dof_secidx_t)) {
12248 dtrace_dof_error(dof, "invalid relocation header");
12252 ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
12253 rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
12254 ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
12256 if (ss == NULL || rs == NULL || ts == NULL)
12257 return (-1); /* dtrace_dof_error() has been called already */
12259 if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
12260 rs->dofs_align != sizeof (uint64_t)) {
12261 dtrace_dof_error(dof, "invalid relocation section");
12265 r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
12266 n = rs->dofs_size / rs->dofs_entsize;
12268 for (i = 0; i < n; i++) {
12269 uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
12271 switch (r->dofr_type) {
12272 case DOF_RELO_NONE:
12274 case DOF_RELO_SETX:
12275 if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
12276 sizeof (uint64_t) > ts->dofs_size) {
12277 dtrace_dof_error(dof, "bad relocation offset");
12281 if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
12282 dtrace_dof_error(dof, "misaligned setx relo");
12286 *(uint64_t *)taddr += ubase;
12289 dtrace_dof_error(dof, "invalid relocation type");
12293 r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
12300 * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
12301 * header: it should be at the front of a memory region that is at least
12302 * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
12303 * size. It need not be validated in any other way.
12306 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
12307 dtrace_enabling_t **enabp, uint64_t ubase, int noprobes)
12309 uint64_t len = dof->dofh_loadsz, seclen;
12310 uintptr_t daddr = (uintptr_t)dof;
12311 dtrace_ecbdesc_t *ep;
12312 dtrace_enabling_t *enab;
12315 ASSERT(MUTEX_HELD(&dtrace_lock));
12316 ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
12319 * Check the DOF header identification bytes. In addition to checking
12320 * valid settings, we also verify that unused bits/bytes are zeroed so
12321 * we can use them later without fear of regressing existing binaries.
12323 if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
12324 DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
12325 dtrace_dof_error(dof, "DOF magic string mismatch");
12329 if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
12330 dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
12331 dtrace_dof_error(dof, "DOF has invalid data model");
12335 if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
12336 dtrace_dof_error(dof, "DOF encoding mismatch");
12340 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
12341 dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
12342 dtrace_dof_error(dof, "DOF version mismatch");
12346 if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
12347 dtrace_dof_error(dof, "DOF uses unsupported instruction set");
12351 if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
12352 dtrace_dof_error(dof, "DOF uses too many integer registers");
12356 if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
12357 dtrace_dof_error(dof, "DOF uses too many tuple registers");
12361 for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
12362 if (dof->dofh_ident[i] != 0) {
12363 dtrace_dof_error(dof, "DOF has invalid ident byte set");
12368 if (dof->dofh_flags & ~DOF_FL_VALID) {
12369 dtrace_dof_error(dof, "DOF has invalid flag bits set");
12373 if (dof->dofh_secsize == 0) {
12374 dtrace_dof_error(dof, "zero section header size");
12379 * Check that the section headers don't exceed the amount of DOF
12380 * data. Note that we cast the section size and number of sections
12381 * to uint64_t's to prevent possible overflow in the multiplication.
12383 seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
12385 if (dof->dofh_secoff > len || seclen > len ||
12386 dof->dofh_secoff + seclen > len) {
12387 dtrace_dof_error(dof, "truncated section headers");
12391 if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
12392 dtrace_dof_error(dof, "misaligned section headers");
12396 if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
12397 dtrace_dof_error(dof, "misaligned section size");
12402 * Take an initial pass through the section headers to be sure that
12403 * the headers don't have stray offsets. If the 'noprobes' flag is
12404 * set, do not permit sections relating to providers, probes, or args.
12406 for (i = 0; i < dof->dofh_secnum; i++) {
12407 dof_sec_t *sec = (dof_sec_t *)(daddr +
12408 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12411 switch (sec->dofs_type) {
12412 case DOF_SECT_PROVIDER:
12413 case DOF_SECT_PROBES:
12414 case DOF_SECT_PRARGS:
12415 case DOF_SECT_PROFFS:
12416 dtrace_dof_error(dof, "illegal sections "
12422 if (!(sec->dofs_flags & DOF_SECF_LOAD))
12423 continue; /* just ignore non-loadable sections */
12425 if (sec->dofs_align & (sec->dofs_align - 1)) {
12426 dtrace_dof_error(dof, "bad section alignment");
12430 if (sec->dofs_offset & (sec->dofs_align - 1)) {
12431 dtrace_dof_error(dof, "misaligned section");
12435 if (sec->dofs_offset > len || sec->dofs_size > len ||
12436 sec->dofs_offset + sec->dofs_size > len) {
12437 dtrace_dof_error(dof, "corrupt section header");
12441 if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
12442 sec->dofs_offset + sec->dofs_size - 1) != '\0') {
12443 dtrace_dof_error(dof, "non-terminating string table");
12449 * Take a second pass through the sections and locate and perform any
12450 * relocations that are present. We do this after the first pass to
12451 * be sure that all sections have had their headers validated.
12453 for (i = 0; i < dof->dofh_secnum; i++) {
12454 dof_sec_t *sec = (dof_sec_t *)(daddr +
12455 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12457 if (!(sec->dofs_flags & DOF_SECF_LOAD))
12458 continue; /* skip sections that are not loadable */
12460 switch (sec->dofs_type) {
12461 case DOF_SECT_URELHDR:
12462 if (dtrace_dof_relocate(dof, sec, ubase) != 0)
12468 if ((enab = *enabp) == NULL)
12469 enab = *enabp = dtrace_enabling_create(vstate);
12471 for (i = 0; i < dof->dofh_secnum; i++) {
12472 dof_sec_t *sec = (dof_sec_t *)(daddr +
12473 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12475 if (sec->dofs_type != DOF_SECT_ECBDESC)
12478 if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
12479 dtrace_enabling_destroy(enab);
12484 dtrace_enabling_add(enab, ep);
12491 * Process DOF for any options. This routine assumes that the DOF has been
12492 * at least processed by dtrace_dof_slurp().
12495 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
12500 dof_optdesc_t *desc;
12502 for (i = 0; i < dof->dofh_secnum; i++) {
12503 dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
12504 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12506 if (sec->dofs_type != DOF_SECT_OPTDESC)
12509 if (sec->dofs_align != sizeof (uint64_t)) {
12510 dtrace_dof_error(dof, "bad alignment in "
12511 "option description");
12515 if ((entsize = sec->dofs_entsize) == 0) {
12516 dtrace_dof_error(dof, "zeroed option entry size");
12520 if (entsize < sizeof (dof_optdesc_t)) {
12521 dtrace_dof_error(dof, "bad option entry size");
12525 for (offs = 0; offs < sec->dofs_size; offs += entsize) {
12526 desc = (dof_optdesc_t *)((uintptr_t)dof +
12527 (uintptr_t)sec->dofs_offset + offs);
12529 if (desc->dofo_strtab != DOF_SECIDX_NONE) {
12530 dtrace_dof_error(dof, "non-zero option string");
12534 if (desc->dofo_value == DTRACEOPT_UNSET) {
12535 dtrace_dof_error(dof, "unset option");
12539 if ((rval = dtrace_state_option(state,
12540 desc->dofo_option, desc->dofo_value)) != 0) {
12541 dtrace_dof_error(dof, "rejected option");
12551 * DTrace Consumer State Functions
12554 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
12556 size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
12559 dtrace_dynvar_t *dvar, *next, *start;
12562 ASSERT(MUTEX_HELD(&dtrace_lock));
12563 ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
12565 bzero(dstate, sizeof (dtrace_dstate_t));
12567 if ((dstate->dtds_chunksize = chunksize) == 0)
12568 dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
12570 if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
12573 if ((base = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
12576 dstate->dtds_size = size;
12577 dstate->dtds_base = base;
12578 dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
12579 bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
12581 hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
12583 if (hashsize != 1 && (hashsize & 1))
12586 dstate->dtds_hashsize = hashsize;
12587 dstate->dtds_hash = dstate->dtds_base;
12590 * Set all of our hash buckets to point to the single sink, and (if
12591 * it hasn't already been set), set the sink's hash value to be the
12592 * sink sentinel value. The sink is needed for dynamic variable
12593 * lookups to know that they have iterated over an entire, valid hash
12596 for (i = 0; i < hashsize; i++)
12597 dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
12599 if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
12600 dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
12603 * Determine number of active CPUs. Divide free list evenly among
12606 start = (dtrace_dynvar_t *)
12607 ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
12608 limit = (uintptr_t)base + size;
12610 maxper = (limit - (uintptr_t)start) / NCPU;
12611 maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
12616 for (i = 0; i < NCPU; i++) {
12618 dstate->dtds_percpu[i].dtdsc_free = dvar = start;
12621 * If we don't even have enough chunks to make it once through
12622 * NCPUs, we're just going to allocate everything to the first
12623 * CPU. And if we're on the last CPU, we're going to allocate
12624 * whatever is left over. In either case, we set the limit to
12625 * be the limit of the dynamic variable space.
12627 if (maxper == 0 || i == NCPU - 1) {
12628 limit = (uintptr_t)base + size;
12631 limit = (uintptr_t)start + maxper;
12632 start = (dtrace_dynvar_t *)limit;
12635 ASSERT(limit <= (uintptr_t)base + size);
12638 next = (dtrace_dynvar_t *)((uintptr_t)dvar +
12639 dstate->dtds_chunksize);
12641 if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
12644 dvar->dtdv_next = next;
12656 dtrace_dstate_fini(dtrace_dstate_t *dstate)
12658 ASSERT(MUTEX_HELD(&cpu_lock));
12660 if (dstate->dtds_base == NULL)
12663 kmem_free(dstate->dtds_base, dstate->dtds_size);
12664 kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
12668 dtrace_vstate_fini(dtrace_vstate_t *vstate)
12671 * Logical XOR, where are you?
12673 ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
12675 if (vstate->dtvs_nglobals > 0) {
12676 kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
12677 sizeof (dtrace_statvar_t *));
12680 if (vstate->dtvs_ntlocals > 0) {
12681 kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
12682 sizeof (dtrace_difv_t));
12685 ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
12687 if (vstate->dtvs_nlocals > 0) {
12688 kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
12689 sizeof (dtrace_statvar_t *));
12695 dtrace_state_clean(dtrace_state_t *state)
12697 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
12700 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
12701 dtrace_speculation_clean(state);
12705 dtrace_state_deadman(dtrace_state_t *state)
12711 now = dtrace_gethrtime();
12713 if (state != dtrace_anon.dta_state &&
12714 now - state->dts_laststatus >= dtrace_deadman_user)
12718 * We must be sure that dts_alive never appears to be less than the
12719 * value upon entry to dtrace_state_deadman(), and because we lack a
12720 * dtrace_cas64(), we cannot store to it atomically. We thus instead
12721 * store INT64_MAX to it, followed by a memory barrier, followed by
12722 * the new value. This assures that dts_alive never appears to be
12723 * less than its true value, regardless of the order in which the
12724 * stores to the underlying storage are issued.
12726 state->dts_alive = INT64_MAX;
12727 dtrace_membar_producer();
12728 state->dts_alive = now;
12732 dtrace_state_clean(void *arg)
12734 dtrace_state_t *state = arg;
12735 dtrace_optval_t *opt = state->dts_options;
12737 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
12740 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
12741 dtrace_speculation_clean(state);
12743 callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
12744 dtrace_state_clean, state);
12748 dtrace_state_deadman(void *arg)
12750 dtrace_state_t *state = arg;
12755 dtrace_debug_output();
12757 now = dtrace_gethrtime();
12759 if (state != dtrace_anon.dta_state &&
12760 now - state->dts_laststatus >= dtrace_deadman_user)
12764 * We must be sure that dts_alive never appears to be less than the
12765 * value upon entry to dtrace_state_deadman(), and because we lack a
12766 * dtrace_cas64(), we cannot store to it atomically. We thus instead
12767 * store INT64_MAX to it, followed by a memory barrier, followed by
12768 * the new value. This assures that dts_alive never appears to be
12769 * less than its true value, regardless of the order in which the
12770 * stores to the underlying storage are issued.
12772 state->dts_alive = INT64_MAX;
12773 dtrace_membar_producer();
12774 state->dts_alive = now;
12776 callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
12777 dtrace_state_deadman, state);
12781 static dtrace_state_t *
12783 dtrace_state_create(dev_t *devp, cred_t *cr)
12785 dtrace_state_create(struct cdev *dev)
12796 dtrace_state_t *state;
12797 dtrace_optval_t *opt;
12798 int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
12800 ASSERT(MUTEX_HELD(&dtrace_lock));
12801 ASSERT(MUTEX_HELD(&cpu_lock));
12804 minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
12805 VM_BESTFIT | VM_SLEEP);
12807 if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
12808 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
12812 state = ddi_get_soft_state(dtrace_softstate, minor);
12819 /* Allocate memory for the state. */
12820 state = kmem_zalloc(sizeof(dtrace_state_t), KM_SLEEP);
12823 state->dts_epid = DTRACE_EPIDNONE + 1;
12825 (void) snprintf(c, sizeof (c), "dtrace_aggid_%d", m);
12827 state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
12828 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
12830 if (devp != NULL) {
12831 major = getemajor(*devp);
12833 major = ddi_driver_major(dtrace_devi);
12836 state->dts_dev = makedevice(major, minor);
12839 *devp = state->dts_dev;
12841 state->dts_aggid_arena = new_unrhdr(1, INT_MAX, &dtrace_unr_mtx);
12842 state->dts_dev = dev;
12846 * We allocate NCPU buffers. On the one hand, this can be quite
12847 * a bit of memory per instance (nearly 36K on a Starcat). On the
12848 * other hand, it saves an additional memory reference in the probe
12851 state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
12852 state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
12855 state->dts_cleaner = CYCLIC_NONE;
12856 state->dts_deadman = CYCLIC_NONE;
12858 callout_init(&state->dts_cleaner, CALLOUT_MPSAFE);
12859 callout_init(&state->dts_deadman, CALLOUT_MPSAFE);
12861 state->dts_vstate.dtvs_state = state;
12863 for (i = 0; i < DTRACEOPT_MAX; i++)
12864 state->dts_options[i] = DTRACEOPT_UNSET;
12867 * Set the default options.
12869 opt = state->dts_options;
12870 opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
12871 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
12872 opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
12873 opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
12874 opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
12875 opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
12876 opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
12877 opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
12878 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
12879 opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
12880 opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
12881 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
12882 opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
12883 opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
12885 state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
12888 * Depending on the user credentials, we set flag bits which alter probe
12889 * visibility or the amount of destructiveness allowed. In the case of
12890 * actual anonymous tracing, or the possession of all privileges, all of
12891 * the normal checks are bypassed.
12893 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
12894 state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
12895 state->dts_cred.dcr_action = DTRACE_CRA_ALL;
12898 * Set up the credentials for this instantiation. We take a
12899 * hold on the credential to prevent it from disappearing on
12900 * us; this in turn prevents the zone_t referenced by this
12901 * credential from disappearing. This means that we can
12902 * examine the credential and the zone from probe context.
12905 state->dts_cred.dcr_cred = cr;
12908 * CRA_PROC means "we have *some* privilege for dtrace" and
12909 * unlocks the use of variables like pid, zonename, etc.
12911 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
12912 PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
12913 state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
12917 * dtrace_user allows use of syscall and profile providers.
12918 * If the user also has proc_owner and/or proc_zone, we
12919 * extend the scope to include additional visibility and
12920 * destructive power.
12922 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
12923 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
12924 state->dts_cred.dcr_visible |=
12925 DTRACE_CRV_ALLPROC;
12927 state->dts_cred.dcr_action |=
12928 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
12931 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
12932 state->dts_cred.dcr_visible |=
12933 DTRACE_CRV_ALLZONE;
12935 state->dts_cred.dcr_action |=
12936 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
12940 * If we have all privs in whatever zone this is,
12941 * we can do destructive things to processes which
12942 * have altered credentials.
12945 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
12946 cr->cr_zone->zone_privset)) {
12947 state->dts_cred.dcr_action |=
12948 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
12954 * Holding the dtrace_kernel privilege also implies that
12955 * the user has the dtrace_user privilege from a visibility
12956 * perspective. But without further privileges, some
12957 * destructive actions are not available.
12959 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
12961 * Make all probes in all zones visible. However,
12962 * this doesn't mean that all actions become available
12965 state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
12966 DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
12968 state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
12971 * Holding proc_owner means that destructive actions
12972 * for *this* zone are allowed.
12974 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
12975 state->dts_cred.dcr_action |=
12976 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
12979 * Holding proc_zone means that destructive actions
12980 * for this user/group ID in all zones is allowed.
12982 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
12983 state->dts_cred.dcr_action |=
12984 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
12988 * If we have all privs in whatever zone this is,
12989 * we can do destructive things to processes which
12990 * have altered credentials.
12992 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
12993 cr->cr_zone->zone_privset)) {
12994 state->dts_cred.dcr_action |=
12995 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13001 * Holding the dtrace_proc privilege gives control over fasttrap
13002 * and pid providers. We need to grant wider destructive
13003 * privileges in the event that the user has proc_owner and/or
13006 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13007 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13008 state->dts_cred.dcr_action |=
13009 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13011 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13012 state->dts_cred.dcr_action |=
13013 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13021 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
13023 dtrace_optval_t *opt = state->dts_options, size;
13024 processorid_t cpu = 0;;
13025 int flags = 0, rval;
13027 ASSERT(MUTEX_HELD(&dtrace_lock));
13028 ASSERT(MUTEX_HELD(&cpu_lock));
13029 ASSERT(which < DTRACEOPT_MAX);
13030 ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
13031 (state == dtrace_anon.dta_state &&
13032 state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
13034 if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
13037 if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
13038 cpu = opt[DTRACEOPT_CPU];
13040 if (which == DTRACEOPT_SPECSIZE)
13041 flags |= DTRACEBUF_NOSWITCH;
13043 if (which == DTRACEOPT_BUFSIZE) {
13044 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
13045 flags |= DTRACEBUF_RING;
13047 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
13048 flags |= DTRACEBUF_FILL;
13050 if (state != dtrace_anon.dta_state ||
13051 state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
13052 flags |= DTRACEBUF_INACTIVE;
13055 for (size = opt[which]; size >= sizeof (uint64_t); size >>= 1) {
13057 * The size must be 8-byte aligned. If the size is not 8-byte
13058 * aligned, drop it down by the difference.
13060 if (size & (sizeof (uint64_t) - 1))
13061 size -= size & (sizeof (uint64_t) - 1);
13063 if (size < state->dts_reserve) {
13065 * Buffers always must be large enough to accommodate
13066 * their prereserved space. We return E2BIG instead
13067 * of ENOMEM in this case to allow for user-level
13068 * software to differentiate the cases.
13073 rval = dtrace_buffer_alloc(buf, size, flags, cpu);
13075 if (rval != ENOMEM) {
13080 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13088 dtrace_state_buffers(dtrace_state_t *state)
13090 dtrace_speculation_t *spec = state->dts_speculations;
13093 if ((rval = dtrace_state_buffer(state, state->dts_buffer,
13094 DTRACEOPT_BUFSIZE)) != 0)
13097 if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
13098 DTRACEOPT_AGGSIZE)) != 0)
13101 for (i = 0; i < state->dts_nspeculations; i++) {
13102 if ((rval = dtrace_state_buffer(state,
13103 spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
13111 dtrace_state_prereserve(dtrace_state_t *state)
13114 dtrace_probe_t *probe;
13116 state->dts_reserve = 0;
13118 if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
13122 * If our buffer policy is a "fill" buffer policy, we need to set the
13123 * prereserved space to be the space required by the END probes.
13125 probe = dtrace_probes[dtrace_probeid_end - 1];
13126 ASSERT(probe != NULL);
13128 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
13129 if (ecb->dte_state != state)
13132 state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
13137 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
13139 dtrace_optval_t *opt = state->dts_options, sz, nspec;
13140 dtrace_speculation_t *spec;
13141 dtrace_buffer_t *buf;
13143 cyc_handler_t hdlr;
13146 int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13147 dtrace_icookie_t cookie;
13149 mutex_enter(&cpu_lock);
13150 mutex_enter(&dtrace_lock);
13152 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
13158 * Before we can perform any checks, we must prime all of the
13159 * retained enablings that correspond to this state.
13161 dtrace_enabling_prime(state);
13163 if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
13168 dtrace_state_prereserve(state);
13171 * Now we want to do is try to allocate our speculations.
13172 * We do not automatically resize the number of speculations; if
13173 * this fails, we will fail the operation.
13175 nspec = opt[DTRACEOPT_NSPEC];
13176 ASSERT(nspec != DTRACEOPT_UNSET);
13178 if (nspec > INT_MAX) {
13183 spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t), KM_NOSLEEP);
13185 if (spec == NULL) {
13190 state->dts_speculations = spec;
13191 state->dts_nspeculations = (int)nspec;
13193 for (i = 0; i < nspec; i++) {
13194 if ((buf = kmem_zalloc(bufsize, KM_NOSLEEP)) == NULL) {
13199 spec[i].dtsp_buffer = buf;
13202 if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
13203 if (dtrace_anon.dta_state == NULL) {
13208 if (state->dts_necbs != 0) {
13213 state->dts_anon = dtrace_anon_grab();
13214 ASSERT(state->dts_anon != NULL);
13215 state = state->dts_anon;
13218 * We want "grabanon" to be set in the grabbed state, so we'll
13219 * copy that option value from the grabbing state into the
13222 state->dts_options[DTRACEOPT_GRABANON] =
13223 opt[DTRACEOPT_GRABANON];
13225 *cpu = dtrace_anon.dta_beganon;
13228 * If the anonymous state is active (as it almost certainly
13229 * is if the anonymous enabling ultimately matched anything),
13230 * we don't allow any further option processing -- but we
13231 * don't return failure.
13233 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13237 if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
13238 opt[DTRACEOPT_AGGSIZE] != 0) {
13239 if (state->dts_aggregations == NULL) {
13241 * We're not going to create an aggregation buffer
13242 * because we don't have any ECBs that contain
13243 * aggregations -- set this option to 0.
13245 opt[DTRACEOPT_AGGSIZE] = 0;
13248 * If we have an aggregation buffer, we must also have
13249 * a buffer to use as scratch.
13251 if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
13252 opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
13253 opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
13258 if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
13259 opt[DTRACEOPT_SPECSIZE] != 0) {
13260 if (!state->dts_speculates) {
13262 * We're not going to create speculation buffers
13263 * because we don't have any ECBs that actually
13264 * speculate -- set the speculation size to 0.
13266 opt[DTRACEOPT_SPECSIZE] = 0;
13271 * The bare minimum size for any buffer that we're actually going to
13272 * do anything to is sizeof (uint64_t).
13274 sz = sizeof (uint64_t);
13276 if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
13277 (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
13278 (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
13280 * A buffer size has been explicitly set to 0 (or to a size
13281 * that will be adjusted to 0) and we need the space -- we
13282 * need to return failure. We return ENOSPC to differentiate
13283 * it from failing to allocate a buffer due to failure to meet
13284 * the reserve (for which we return E2BIG).
13290 if ((rval = dtrace_state_buffers(state)) != 0)
13293 if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
13294 sz = dtrace_dstate_defsize;
13297 rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
13302 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13304 } while (sz >>= 1);
13306 opt[DTRACEOPT_DYNVARSIZE] = sz;
13311 if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
13312 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
13314 if (opt[DTRACEOPT_CLEANRATE] == 0)
13315 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13317 if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
13318 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
13320 if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
13321 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13323 state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
13325 hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
13326 hdlr.cyh_arg = state;
13327 hdlr.cyh_level = CY_LOW_LEVEL;
13330 when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
13332 state->dts_cleaner = cyclic_add(&hdlr, &when);
13334 hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
13335 hdlr.cyh_arg = state;
13336 hdlr.cyh_level = CY_LOW_LEVEL;
13339 when.cyt_interval = dtrace_deadman_interval;
13341 state->dts_deadman = cyclic_add(&hdlr, &when);
13343 callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
13344 dtrace_state_clean, state);
13345 callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
13346 dtrace_state_deadman, state);
13349 state->dts_activity = DTRACE_ACTIVITY_WARMUP;
13352 * Now it's time to actually fire the BEGIN probe. We need to disable
13353 * interrupts here both to record the CPU on which we fired the BEGIN
13354 * probe (the data from this CPU will be processed first at user
13355 * level) and to manually activate the buffer for this CPU.
13357 cookie = dtrace_interrupt_disable();
13359 ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
13360 state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
13362 dtrace_probe(dtrace_probeid_begin,
13363 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13364 dtrace_interrupt_enable(cookie);
13366 * We may have had an exit action from a BEGIN probe; only change our
13367 * state to ACTIVE if we're still in WARMUP.
13369 ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
13370 state->dts_activity == DTRACE_ACTIVITY_DRAINING);
13372 if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
13373 state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
13376 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
13377 * want each CPU to transition its principal buffer out of the
13378 * INACTIVE state. Doing this assures that no CPU will suddenly begin
13379 * processing an ECB halfway down a probe's ECB chain; all CPUs will
13380 * atomically transition from processing none of a state's ECBs to
13381 * processing all of them.
13383 dtrace_xcall(DTRACE_CPUALL,
13384 (dtrace_xcall_t)dtrace_buffer_activate, state);
13388 dtrace_buffer_free(state->dts_buffer);
13389 dtrace_buffer_free(state->dts_aggbuffer);
13391 if ((nspec = state->dts_nspeculations) == 0) {
13392 ASSERT(state->dts_speculations == NULL);
13396 spec = state->dts_speculations;
13397 ASSERT(spec != NULL);
13399 for (i = 0; i < state->dts_nspeculations; i++) {
13400 if ((buf = spec[i].dtsp_buffer) == NULL)
13403 dtrace_buffer_free(buf);
13404 kmem_free(buf, bufsize);
13407 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13408 state->dts_nspeculations = 0;
13409 state->dts_speculations = NULL;
13412 mutex_exit(&dtrace_lock);
13413 mutex_exit(&cpu_lock);
13419 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
13421 dtrace_icookie_t cookie;
13423 ASSERT(MUTEX_HELD(&dtrace_lock));
13425 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
13426 state->dts_activity != DTRACE_ACTIVITY_DRAINING)
13430 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
13431 * to be sure that every CPU has seen it. See below for the details
13432 * on why this is done.
13434 state->dts_activity = DTRACE_ACTIVITY_DRAINING;
13438 * By this point, it is impossible for any CPU to be still processing
13439 * with DTRACE_ACTIVITY_ACTIVE. We can thus set our activity to
13440 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
13441 * other CPU in dtrace_buffer_reserve(). This allows dtrace_probe()
13442 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
13443 * iff we're in the END probe.
13445 state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
13447 ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
13450 * Finally, we can release the reserve and call the END probe. We
13451 * disable interrupts across calling the END probe to allow us to
13452 * return the CPU on which we actually called the END probe. This
13453 * allows user-land to be sure that this CPU's principal buffer is
13456 state->dts_reserve = 0;
13458 cookie = dtrace_interrupt_disable();
13460 dtrace_probe(dtrace_probeid_end,
13461 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13462 dtrace_interrupt_enable(cookie);
13464 state->dts_activity = DTRACE_ACTIVITY_STOPPED;
13471 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
13472 dtrace_optval_t val)
13474 ASSERT(MUTEX_HELD(&dtrace_lock));
13476 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13479 if (option >= DTRACEOPT_MAX)
13482 if (option != DTRACEOPT_CPU && val < 0)
13486 case DTRACEOPT_DESTRUCTIVE:
13487 if (dtrace_destructive_disallow)
13490 state->dts_cred.dcr_destructive = 1;
13493 case DTRACEOPT_BUFSIZE:
13494 case DTRACEOPT_DYNVARSIZE:
13495 case DTRACEOPT_AGGSIZE:
13496 case DTRACEOPT_SPECSIZE:
13497 case DTRACEOPT_STRSIZE:
13501 if (val >= LONG_MAX) {
13503 * If this is an otherwise negative value, set it to
13504 * the highest multiple of 128m less than LONG_MAX.
13505 * Technically, we're adjusting the size without
13506 * regard to the buffer resizing policy, but in fact,
13507 * this has no effect -- if we set the buffer size to
13508 * ~LONG_MAX and the buffer policy is ultimately set to
13509 * be "manual", the buffer allocation is guaranteed to
13510 * fail, if only because the allocation requires two
13511 * buffers. (We set the the size to the highest
13512 * multiple of 128m because it ensures that the size
13513 * will remain a multiple of a megabyte when
13514 * repeatedly halved -- all the way down to 15m.)
13516 val = LONG_MAX - (1 << 27) + 1;
13520 state->dts_options[option] = val;
13526 dtrace_state_destroy(dtrace_state_t *state)
13529 dtrace_vstate_t *vstate = &state->dts_vstate;
13531 minor_t minor = getminor(state->dts_dev);
13533 int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13534 dtrace_speculation_t *spec = state->dts_speculations;
13535 int nspec = state->dts_nspeculations;
13538 ASSERT(MUTEX_HELD(&dtrace_lock));
13539 ASSERT(MUTEX_HELD(&cpu_lock));
13542 * First, retract any retained enablings for this state.
13544 dtrace_enabling_retract(state);
13545 ASSERT(state->dts_nretained == 0);
13547 if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
13548 state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
13550 * We have managed to come into dtrace_state_destroy() on a
13551 * hot enabling -- almost certainly because of a disorderly
13552 * shutdown of a consumer. (That is, a consumer that is
13553 * exiting without having called dtrace_stop().) In this case,
13554 * we're going to set our activity to be KILLED, and then
13555 * issue a sync to be sure that everyone is out of probe
13556 * context before we start blowing away ECBs.
13558 state->dts_activity = DTRACE_ACTIVITY_KILLED;
13563 * Release the credential hold we took in dtrace_state_create().
13565 if (state->dts_cred.dcr_cred != NULL)
13566 crfree(state->dts_cred.dcr_cred);
13569 * Now we can safely disable and destroy any enabled probes. Because
13570 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
13571 * (especially if they're all enabled), we take two passes through the
13572 * ECBs: in the first, we disable just DTRACE_PRIV_KERNEL probes, and
13573 * in the second we disable whatever is left over.
13575 for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
13576 for (i = 0; i < state->dts_necbs; i++) {
13577 if ((ecb = state->dts_ecbs[i]) == NULL)
13580 if (match && ecb->dte_probe != NULL) {
13581 dtrace_probe_t *probe = ecb->dte_probe;
13582 dtrace_provider_t *prov = probe->dtpr_provider;
13584 if (!(prov->dtpv_priv.dtpp_flags & match))
13588 dtrace_ecb_disable(ecb);
13589 dtrace_ecb_destroy(ecb);
13597 * Before we free the buffers, perform one more sync to assure that
13598 * every CPU is out of probe context.
13602 dtrace_buffer_free(state->dts_buffer);
13603 dtrace_buffer_free(state->dts_aggbuffer);
13605 for (i = 0; i < nspec; i++)
13606 dtrace_buffer_free(spec[i].dtsp_buffer);
13609 if (state->dts_cleaner != CYCLIC_NONE)
13610 cyclic_remove(state->dts_cleaner);
13612 if (state->dts_deadman != CYCLIC_NONE)
13613 cyclic_remove(state->dts_deadman);
13615 callout_stop(&state->dts_cleaner);
13616 callout_drain(&state->dts_cleaner);
13617 callout_stop(&state->dts_deadman);
13618 callout_drain(&state->dts_deadman);
13621 dtrace_dstate_fini(&vstate->dtvs_dynvars);
13622 dtrace_vstate_fini(vstate);
13623 if (state->dts_ecbs != NULL)
13624 kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
13626 if (state->dts_aggregations != NULL) {
13628 for (i = 0; i < state->dts_naggregations; i++)
13629 ASSERT(state->dts_aggregations[i] == NULL);
13631 ASSERT(state->dts_naggregations > 0);
13632 kmem_free(state->dts_aggregations,
13633 state->dts_naggregations * sizeof (dtrace_aggregation_t *));
13636 kmem_free(state->dts_buffer, bufsize);
13637 kmem_free(state->dts_aggbuffer, bufsize);
13639 for (i = 0; i < nspec; i++)
13640 kmem_free(spec[i].dtsp_buffer, bufsize);
13643 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13645 dtrace_format_destroy(state);
13647 if (state->dts_aggid_arena != NULL) {
13649 vmem_destroy(state->dts_aggid_arena);
13651 delete_unrhdr(state->dts_aggid_arena);
13653 state->dts_aggid_arena = NULL;
13656 ddi_soft_state_free(dtrace_softstate, minor);
13657 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
13662 * DTrace Anonymous Enabling Functions
13664 static dtrace_state_t *
13665 dtrace_anon_grab(void)
13667 dtrace_state_t *state;
13669 ASSERT(MUTEX_HELD(&dtrace_lock));
13671 if ((state = dtrace_anon.dta_state) == NULL) {
13672 ASSERT(dtrace_anon.dta_enabling == NULL);
13676 ASSERT(dtrace_anon.dta_enabling != NULL);
13677 ASSERT(dtrace_retained != NULL);
13679 dtrace_enabling_destroy(dtrace_anon.dta_enabling);
13680 dtrace_anon.dta_enabling = NULL;
13681 dtrace_anon.dta_state = NULL;
13687 dtrace_anon_property(void)
13690 dtrace_state_t *state;
13692 char c[32]; /* enough for "dof-data-" + digits */
13694 ASSERT(MUTEX_HELD(&dtrace_lock));
13695 ASSERT(MUTEX_HELD(&cpu_lock));
13697 for (i = 0; ; i++) {
13698 (void) snprintf(c, sizeof (c), "dof-data-%d", i);
13700 dtrace_err_verbose = 1;
13702 if ((dof = dtrace_dof_property(c)) == NULL) {
13703 dtrace_err_verbose = 0;
13709 * We want to create anonymous state, so we need to transition
13710 * the kernel debugger to indicate that DTrace is active. If
13711 * this fails (e.g. because the debugger has modified text in
13712 * some way), we won't continue with the processing.
13714 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
13715 cmn_err(CE_NOTE, "kernel debugger active; anonymous "
13716 "enabling ignored.");
13717 dtrace_dof_destroy(dof);
13723 * If we haven't allocated an anonymous state, we'll do so now.
13725 if ((state = dtrace_anon.dta_state) == NULL) {
13727 state = dtrace_state_create(NULL, NULL);
13729 state = dtrace_state_create(NULL);
13731 dtrace_anon.dta_state = state;
13733 if (state == NULL) {
13735 * This basically shouldn't happen: the only
13736 * failure mode from dtrace_state_create() is a
13737 * failure of ddi_soft_state_zalloc() that
13738 * itself should never happen. Still, the
13739 * interface allows for a failure mode, and
13740 * we want to fail as gracefully as possible:
13741 * we'll emit an error message and cease
13742 * processing anonymous state in this case.
13744 cmn_err(CE_WARN, "failed to create "
13745 "anonymous state");
13746 dtrace_dof_destroy(dof);
13751 rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
13752 &dtrace_anon.dta_enabling, 0, B_TRUE);
13755 rv = dtrace_dof_options(dof, state);
13757 dtrace_err_verbose = 0;
13758 dtrace_dof_destroy(dof);
13762 * This is malformed DOF; chuck any anonymous state
13765 ASSERT(dtrace_anon.dta_enabling == NULL);
13766 dtrace_state_destroy(state);
13767 dtrace_anon.dta_state = NULL;
13771 ASSERT(dtrace_anon.dta_enabling != NULL);
13774 if (dtrace_anon.dta_enabling != NULL) {
13778 * dtrace_enabling_retain() can only fail because we are
13779 * trying to retain more enablings than are allowed -- but
13780 * we only have one anonymous enabling, and we are guaranteed
13781 * to be allowed at least one retained enabling; we assert
13782 * that dtrace_enabling_retain() returns success.
13784 rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
13787 dtrace_enabling_dump(dtrace_anon.dta_enabling);
13792 * DTrace Helper Functions
13795 dtrace_helper_trace(dtrace_helper_action_t *helper,
13796 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
13798 uint32_t size, next, nnext, i;
13799 dtrace_helptrace_t *ent;
13800 uint16_t flags = cpu_core[curcpu].cpuc_dtrace_flags;
13802 if (!dtrace_helptrace_enabled)
13805 ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
13808 * What would a tracing framework be without its own tracing
13809 * framework? (Well, a hell of a lot simpler, for starters...)
13811 size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
13812 sizeof (uint64_t) - sizeof (uint64_t);
13815 * Iterate until we can allocate a slot in the trace buffer.
13818 next = dtrace_helptrace_next;
13820 if (next + size < dtrace_helptrace_bufsize) {
13821 nnext = next + size;
13825 } while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
13828 * We have our slot; fill it in.
13833 ent = (dtrace_helptrace_t *)&dtrace_helptrace_buffer[next];
13834 ent->dtht_helper = helper;
13835 ent->dtht_where = where;
13836 ent->dtht_nlocals = vstate->dtvs_nlocals;
13838 ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
13839 mstate->dtms_fltoffs : -1;
13840 ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
13841 ent->dtht_illval = cpu_core[curcpu].cpuc_dtrace_illval;
13843 for (i = 0; i < vstate->dtvs_nlocals; i++) {
13844 dtrace_statvar_t *svar;
13846 if ((svar = vstate->dtvs_locals[i]) == NULL)
13849 ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
13850 ent->dtht_locals[i] =
13851 ((uint64_t *)(uintptr_t)svar->dtsv_data)[curcpu];
13856 dtrace_helper(int which, dtrace_mstate_t *mstate,
13857 dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
13859 uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
13860 uint64_t sarg0 = mstate->dtms_arg[0];
13861 uint64_t sarg1 = mstate->dtms_arg[1];
13863 dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
13864 dtrace_helper_action_t *helper;
13865 dtrace_vstate_t *vstate;
13866 dtrace_difo_t *pred;
13867 int i, trace = dtrace_helptrace_enabled;
13869 ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
13871 if (helpers == NULL)
13874 if ((helper = helpers->dthps_actions[which]) == NULL)
13877 vstate = &helpers->dthps_vstate;
13878 mstate->dtms_arg[0] = arg0;
13879 mstate->dtms_arg[1] = arg1;
13882 * Now iterate over each helper. If its predicate evaluates to 'true',
13883 * we'll call the corresponding actions. Note that the below calls
13884 * to dtrace_dif_emulate() may set faults in machine state. This is
13885 * okay: our caller (the outer dtrace_dif_emulate()) will simply plow
13886 * the stored DIF offset with its own (which is the desired behavior).
13887 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
13888 * from machine state; this is okay, too.
13890 for (; helper != NULL; helper = helper->dtha_next) {
13891 if ((pred = helper->dtha_predicate) != NULL) {
13893 dtrace_helper_trace(helper, mstate, vstate, 0);
13895 if (!dtrace_dif_emulate(pred, mstate, vstate, state))
13898 if (*flags & CPU_DTRACE_FAULT)
13902 for (i = 0; i < helper->dtha_nactions; i++) {
13904 dtrace_helper_trace(helper,
13905 mstate, vstate, i + 1);
13907 rval = dtrace_dif_emulate(helper->dtha_actions[i],
13908 mstate, vstate, state);
13910 if (*flags & CPU_DTRACE_FAULT)
13916 dtrace_helper_trace(helper, mstate, vstate,
13917 DTRACE_HELPTRACE_NEXT);
13921 dtrace_helper_trace(helper, mstate, vstate,
13922 DTRACE_HELPTRACE_DONE);
13925 * Restore the arg0 that we saved upon entry.
13927 mstate->dtms_arg[0] = sarg0;
13928 mstate->dtms_arg[1] = sarg1;
13934 dtrace_helper_trace(helper, mstate, vstate,
13935 DTRACE_HELPTRACE_ERR);
13938 * Restore the arg0 that we saved upon entry.
13940 mstate->dtms_arg[0] = sarg0;
13941 mstate->dtms_arg[1] = sarg1;
13947 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
13948 dtrace_vstate_t *vstate)
13952 if (helper->dtha_predicate != NULL)
13953 dtrace_difo_release(helper->dtha_predicate, vstate);
13955 for (i = 0; i < helper->dtha_nactions; i++) {
13956 ASSERT(helper->dtha_actions[i] != NULL);
13957 dtrace_difo_release(helper->dtha_actions[i], vstate);
13960 kmem_free(helper->dtha_actions,
13961 helper->dtha_nactions * sizeof (dtrace_difo_t *));
13962 kmem_free(helper, sizeof (dtrace_helper_action_t));
13966 dtrace_helper_destroygen(int gen)
13968 proc_t *p = curproc;
13969 dtrace_helpers_t *help = p->p_dtrace_helpers;
13970 dtrace_vstate_t *vstate;
13973 ASSERT(MUTEX_HELD(&dtrace_lock));
13975 if (help == NULL || gen > help->dthps_generation)
13978 vstate = &help->dthps_vstate;
13980 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
13981 dtrace_helper_action_t *last = NULL, *h, *next;
13983 for (h = help->dthps_actions[i]; h != NULL; h = next) {
13984 next = h->dtha_next;
13986 if (h->dtha_generation == gen) {
13987 if (last != NULL) {
13988 last->dtha_next = next;
13990 help->dthps_actions[i] = next;
13993 dtrace_helper_action_destroy(h, vstate);
14001 * Interate until we've cleared out all helper providers with the
14002 * given generation number.
14005 dtrace_helper_provider_t *prov;
14008 * Look for a helper provider with the right generation. We
14009 * have to start back at the beginning of the list each time
14010 * because we drop dtrace_lock. It's unlikely that we'll make
14011 * more than two passes.
14013 for (i = 0; i < help->dthps_nprovs; i++) {
14014 prov = help->dthps_provs[i];
14016 if (prov->dthp_generation == gen)
14021 * If there were no matches, we're done.
14023 if (i == help->dthps_nprovs)
14027 * Move the last helper provider into this slot.
14029 help->dthps_nprovs--;
14030 help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
14031 help->dthps_provs[help->dthps_nprovs] = NULL;
14033 mutex_exit(&dtrace_lock);
14036 * If we have a meta provider, remove this helper provider.
14038 mutex_enter(&dtrace_meta_lock);
14039 if (dtrace_meta_pid != NULL) {
14040 ASSERT(dtrace_deferred_pid == NULL);
14041 dtrace_helper_provider_remove(&prov->dthp_prov,
14044 mutex_exit(&dtrace_meta_lock);
14046 dtrace_helper_provider_destroy(prov);
14048 mutex_enter(&dtrace_lock);
14055 dtrace_helper_validate(dtrace_helper_action_t *helper)
14060 if ((dp = helper->dtha_predicate) != NULL)
14061 err += dtrace_difo_validate_helper(dp);
14063 for (i = 0; i < helper->dtha_nactions; i++)
14064 err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
14070 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep)
14072 dtrace_helpers_t *help;
14073 dtrace_helper_action_t *helper, *last;
14074 dtrace_actdesc_t *act;
14075 dtrace_vstate_t *vstate;
14076 dtrace_predicate_t *pred;
14077 int count = 0, nactions = 0, i;
14079 if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
14082 help = curproc->p_dtrace_helpers;
14083 last = help->dthps_actions[which];
14084 vstate = &help->dthps_vstate;
14086 for (count = 0; last != NULL; last = last->dtha_next) {
14088 if (last->dtha_next == NULL)
14093 * If we already have dtrace_helper_actions_max helper actions for this
14094 * helper action type, we'll refuse to add a new one.
14096 if (count >= dtrace_helper_actions_max)
14099 helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
14100 helper->dtha_generation = help->dthps_generation;
14102 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
14103 ASSERT(pred->dtp_difo != NULL);
14104 dtrace_difo_hold(pred->dtp_difo);
14105 helper->dtha_predicate = pred->dtp_difo;
14108 for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
14109 if (act->dtad_kind != DTRACEACT_DIFEXPR)
14112 if (act->dtad_difo == NULL)
14118 helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
14119 (helper->dtha_nactions = nactions), KM_SLEEP);
14121 for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
14122 dtrace_difo_hold(act->dtad_difo);
14123 helper->dtha_actions[i++] = act->dtad_difo;
14126 if (!dtrace_helper_validate(helper))
14129 if (last == NULL) {
14130 help->dthps_actions[which] = helper;
14132 last->dtha_next = helper;
14135 if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
14136 dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
14137 dtrace_helptrace_next = 0;
14142 dtrace_helper_action_destroy(helper, vstate);
14147 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
14148 dof_helper_t *dofhp)
14150 ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
14152 mutex_enter(&dtrace_meta_lock);
14153 mutex_enter(&dtrace_lock);
14155 if (!dtrace_attached() || dtrace_meta_pid == NULL) {
14157 * If the dtrace module is loaded but not attached, or if
14158 * there aren't isn't a meta provider registered to deal with
14159 * these provider descriptions, we need to postpone creating
14160 * the actual providers until later.
14163 if (help->dthps_next == NULL && help->dthps_prev == NULL &&
14164 dtrace_deferred_pid != help) {
14165 help->dthps_deferred = 1;
14166 help->dthps_pid = p->p_pid;
14167 help->dthps_next = dtrace_deferred_pid;
14168 help->dthps_prev = NULL;
14169 if (dtrace_deferred_pid != NULL)
14170 dtrace_deferred_pid->dthps_prev = help;
14171 dtrace_deferred_pid = help;
14174 mutex_exit(&dtrace_lock);
14176 } else if (dofhp != NULL) {
14178 * If the dtrace module is loaded and we have a particular
14179 * helper provider description, pass that off to the
14183 mutex_exit(&dtrace_lock);
14185 dtrace_helper_provide(dofhp, p->p_pid);
14189 * Otherwise, just pass all the helper provider descriptions
14190 * off to the meta provider.
14194 mutex_exit(&dtrace_lock);
14196 for (i = 0; i < help->dthps_nprovs; i++) {
14197 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
14202 mutex_exit(&dtrace_meta_lock);
14206 dtrace_helper_provider_add(dof_helper_t *dofhp, int gen)
14208 dtrace_helpers_t *help;
14209 dtrace_helper_provider_t *hprov, **tmp_provs;
14210 uint_t tmp_maxprovs, i;
14212 ASSERT(MUTEX_HELD(&dtrace_lock));
14214 help = curproc->p_dtrace_helpers;
14215 ASSERT(help != NULL);
14218 * If we already have dtrace_helper_providers_max helper providers,
14219 * we're refuse to add a new one.
14221 if (help->dthps_nprovs >= dtrace_helper_providers_max)
14225 * Check to make sure this isn't a duplicate.
14227 for (i = 0; i < help->dthps_nprovs; i++) {
14228 if (dofhp->dofhp_addr ==
14229 help->dthps_provs[i]->dthp_prov.dofhp_addr)
14233 hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
14234 hprov->dthp_prov = *dofhp;
14235 hprov->dthp_ref = 1;
14236 hprov->dthp_generation = gen;
14239 * Allocate a bigger table for helper providers if it's already full.
14241 if (help->dthps_maxprovs == help->dthps_nprovs) {
14242 tmp_maxprovs = help->dthps_maxprovs;
14243 tmp_provs = help->dthps_provs;
14245 if (help->dthps_maxprovs == 0)
14246 help->dthps_maxprovs = 2;
14248 help->dthps_maxprovs *= 2;
14249 if (help->dthps_maxprovs > dtrace_helper_providers_max)
14250 help->dthps_maxprovs = dtrace_helper_providers_max;
14252 ASSERT(tmp_maxprovs < help->dthps_maxprovs);
14254 help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
14255 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
14257 if (tmp_provs != NULL) {
14258 bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
14259 sizeof (dtrace_helper_provider_t *));
14260 kmem_free(tmp_provs, tmp_maxprovs *
14261 sizeof (dtrace_helper_provider_t *));
14265 help->dthps_provs[help->dthps_nprovs] = hprov;
14266 help->dthps_nprovs++;
14272 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
14274 mutex_enter(&dtrace_lock);
14276 if (--hprov->dthp_ref == 0) {
14278 mutex_exit(&dtrace_lock);
14279 dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
14280 dtrace_dof_destroy(dof);
14281 kmem_free(hprov, sizeof (dtrace_helper_provider_t));
14283 mutex_exit(&dtrace_lock);
14288 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
14290 uintptr_t daddr = (uintptr_t)dof;
14291 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
14292 dof_provider_t *provider;
14293 dof_probe_t *probe;
14295 char *strtab, *typestr;
14296 dof_stridx_t typeidx;
14298 uint_t nprobes, j, k;
14300 ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
14302 if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
14303 dtrace_dof_error(dof, "misaligned section offset");
14308 * The section needs to be large enough to contain the DOF provider
14309 * structure appropriate for the given version.
14311 if (sec->dofs_size <
14312 ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
14313 offsetof(dof_provider_t, dofpv_prenoffs) :
14314 sizeof (dof_provider_t))) {
14315 dtrace_dof_error(dof, "provider section too small");
14319 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
14320 str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
14321 prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
14322 arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
14323 off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
14325 if (str_sec == NULL || prb_sec == NULL ||
14326 arg_sec == NULL || off_sec == NULL)
14331 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
14332 provider->dofpv_prenoffs != DOF_SECT_NONE &&
14333 (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
14334 provider->dofpv_prenoffs)) == NULL)
14337 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
14339 if (provider->dofpv_name >= str_sec->dofs_size ||
14340 strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
14341 dtrace_dof_error(dof, "invalid provider name");
14345 if (prb_sec->dofs_entsize == 0 ||
14346 prb_sec->dofs_entsize > prb_sec->dofs_size) {
14347 dtrace_dof_error(dof, "invalid entry size");
14351 if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
14352 dtrace_dof_error(dof, "misaligned entry size");
14356 if (off_sec->dofs_entsize != sizeof (uint32_t)) {
14357 dtrace_dof_error(dof, "invalid entry size");
14361 if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
14362 dtrace_dof_error(dof, "misaligned section offset");
14366 if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
14367 dtrace_dof_error(dof, "invalid entry size");
14371 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
14373 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
14376 * Take a pass through the probes to check for errors.
14378 for (j = 0; j < nprobes; j++) {
14379 probe = (dof_probe_t *)(uintptr_t)(daddr +
14380 prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
14382 if (probe->dofpr_func >= str_sec->dofs_size) {
14383 dtrace_dof_error(dof, "invalid function name");
14387 if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
14388 dtrace_dof_error(dof, "function name too long");
14392 if (probe->dofpr_name >= str_sec->dofs_size ||
14393 strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
14394 dtrace_dof_error(dof, "invalid probe name");
14399 * The offset count must not wrap the index, and the offsets
14400 * must also not overflow the section's data.
14402 if (probe->dofpr_offidx + probe->dofpr_noffs <
14403 probe->dofpr_offidx ||
14404 (probe->dofpr_offidx + probe->dofpr_noffs) *
14405 off_sec->dofs_entsize > off_sec->dofs_size) {
14406 dtrace_dof_error(dof, "invalid probe offset");
14410 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
14412 * If there's no is-enabled offset section, make sure
14413 * there aren't any is-enabled offsets. Otherwise
14414 * perform the same checks as for probe offsets
14415 * (immediately above).
14417 if (enoff_sec == NULL) {
14418 if (probe->dofpr_enoffidx != 0 ||
14419 probe->dofpr_nenoffs != 0) {
14420 dtrace_dof_error(dof, "is-enabled "
14421 "offsets with null section");
14424 } else if (probe->dofpr_enoffidx +
14425 probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
14426 (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
14427 enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
14428 dtrace_dof_error(dof, "invalid is-enabled "
14433 if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
14434 dtrace_dof_error(dof, "zero probe and "
14435 "is-enabled offsets");
14438 } else if (probe->dofpr_noffs == 0) {
14439 dtrace_dof_error(dof, "zero probe offsets");
14443 if (probe->dofpr_argidx + probe->dofpr_xargc <
14444 probe->dofpr_argidx ||
14445 (probe->dofpr_argidx + probe->dofpr_xargc) *
14446 arg_sec->dofs_entsize > arg_sec->dofs_size) {
14447 dtrace_dof_error(dof, "invalid args");
14451 typeidx = probe->dofpr_nargv;
14452 typestr = strtab + probe->dofpr_nargv;
14453 for (k = 0; k < probe->dofpr_nargc; k++) {
14454 if (typeidx >= str_sec->dofs_size) {
14455 dtrace_dof_error(dof, "bad "
14456 "native argument type");
14460 typesz = strlen(typestr) + 1;
14461 if (typesz > DTRACE_ARGTYPELEN) {
14462 dtrace_dof_error(dof, "native "
14463 "argument type too long");
14470 typeidx = probe->dofpr_xargv;
14471 typestr = strtab + probe->dofpr_xargv;
14472 for (k = 0; k < probe->dofpr_xargc; k++) {
14473 if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
14474 dtrace_dof_error(dof, "bad "
14475 "native argument index");
14479 if (typeidx >= str_sec->dofs_size) {
14480 dtrace_dof_error(dof, "bad "
14481 "translated argument type");
14485 typesz = strlen(typestr) + 1;
14486 if (typesz > DTRACE_ARGTYPELEN) {
14487 dtrace_dof_error(dof, "translated argument "
14501 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp)
14503 dtrace_helpers_t *help;
14504 dtrace_vstate_t *vstate;
14505 dtrace_enabling_t *enab = NULL;
14506 int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
14507 uintptr_t daddr = (uintptr_t)dof;
14509 ASSERT(MUTEX_HELD(&dtrace_lock));
14511 if ((help = curproc->p_dtrace_helpers) == NULL)
14512 help = dtrace_helpers_create(curproc);
14514 vstate = &help->dthps_vstate;
14516 if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab,
14517 dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) {
14518 dtrace_dof_destroy(dof);
14523 * Look for helper providers and validate their descriptions.
14526 for (i = 0; i < dof->dofh_secnum; i++) {
14527 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
14528 dof->dofh_secoff + i * dof->dofh_secsize);
14530 if (sec->dofs_type != DOF_SECT_PROVIDER)
14533 if (dtrace_helper_provider_validate(dof, sec) != 0) {
14534 dtrace_enabling_destroy(enab);
14535 dtrace_dof_destroy(dof);
14544 * Now we need to walk through the ECB descriptions in the enabling.
14546 for (i = 0; i < enab->dten_ndesc; i++) {
14547 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
14548 dtrace_probedesc_t *desc = &ep->dted_probe;
14550 if (strcmp(desc->dtpd_provider, "dtrace") != 0)
14553 if (strcmp(desc->dtpd_mod, "helper") != 0)
14556 if (strcmp(desc->dtpd_func, "ustack") != 0)
14559 if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
14562 * Adding this helper action failed -- we are now going
14563 * to rip out the entire generation and return failure.
14565 (void) dtrace_helper_destroygen(help->dthps_generation);
14566 dtrace_enabling_destroy(enab);
14567 dtrace_dof_destroy(dof);
14574 if (nhelpers < enab->dten_ndesc)
14575 dtrace_dof_error(dof, "unmatched helpers");
14577 gen = help->dthps_generation++;
14578 dtrace_enabling_destroy(enab);
14580 if (dhp != NULL && nprovs > 0) {
14581 dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
14582 if (dtrace_helper_provider_add(dhp, gen) == 0) {
14583 mutex_exit(&dtrace_lock);
14584 dtrace_helper_provider_register(curproc, help, dhp);
14585 mutex_enter(&dtrace_lock);
14592 dtrace_dof_destroy(dof);
14597 static dtrace_helpers_t *
14598 dtrace_helpers_create(proc_t *p)
14600 dtrace_helpers_t *help;
14602 ASSERT(MUTEX_HELD(&dtrace_lock));
14603 ASSERT(p->p_dtrace_helpers == NULL);
14605 help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
14606 help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
14607 DTRACE_NHELPER_ACTIONS, KM_SLEEP);
14609 p->p_dtrace_helpers = help;
14619 dtrace_helpers_destroy(proc_t *p)
14621 dtrace_helpers_t *help;
14622 dtrace_vstate_t *vstate;
14624 proc_t *p = curproc;
14628 mutex_enter(&dtrace_lock);
14630 ASSERT(p->p_dtrace_helpers != NULL);
14631 ASSERT(dtrace_helpers > 0);
14633 help = p->p_dtrace_helpers;
14634 vstate = &help->dthps_vstate;
14637 * We're now going to lose the help from this process.
14639 p->p_dtrace_helpers = NULL;
14643 * Destory the helper actions.
14645 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14646 dtrace_helper_action_t *h, *next;
14648 for (h = help->dthps_actions[i]; h != NULL; h = next) {
14649 next = h->dtha_next;
14650 dtrace_helper_action_destroy(h, vstate);
14655 mutex_exit(&dtrace_lock);
14658 * Destroy the helper providers.
14660 if (help->dthps_maxprovs > 0) {
14661 mutex_enter(&dtrace_meta_lock);
14662 if (dtrace_meta_pid != NULL) {
14663 ASSERT(dtrace_deferred_pid == NULL);
14665 for (i = 0; i < help->dthps_nprovs; i++) {
14666 dtrace_helper_provider_remove(
14667 &help->dthps_provs[i]->dthp_prov, p->p_pid);
14670 mutex_enter(&dtrace_lock);
14671 ASSERT(help->dthps_deferred == 0 ||
14672 help->dthps_next != NULL ||
14673 help->dthps_prev != NULL ||
14674 help == dtrace_deferred_pid);
14677 * Remove the helper from the deferred list.
14679 if (help->dthps_next != NULL)
14680 help->dthps_next->dthps_prev = help->dthps_prev;
14681 if (help->dthps_prev != NULL)
14682 help->dthps_prev->dthps_next = help->dthps_next;
14683 if (dtrace_deferred_pid == help) {
14684 dtrace_deferred_pid = help->dthps_next;
14685 ASSERT(help->dthps_prev == NULL);
14688 mutex_exit(&dtrace_lock);
14691 mutex_exit(&dtrace_meta_lock);
14693 for (i = 0; i < help->dthps_nprovs; i++) {
14694 dtrace_helper_provider_destroy(help->dthps_provs[i]);
14697 kmem_free(help->dthps_provs, help->dthps_maxprovs *
14698 sizeof (dtrace_helper_provider_t *));
14701 mutex_enter(&dtrace_lock);
14703 dtrace_vstate_fini(&help->dthps_vstate);
14704 kmem_free(help->dthps_actions,
14705 sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
14706 kmem_free(help, sizeof (dtrace_helpers_t));
14709 mutex_exit(&dtrace_lock);
14716 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
14718 dtrace_helpers_t *help, *newhelp;
14719 dtrace_helper_action_t *helper, *new, *last;
14721 dtrace_vstate_t *vstate;
14722 int i, j, sz, hasprovs = 0;
14724 mutex_enter(&dtrace_lock);
14725 ASSERT(from->p_dtrace_helpers != NULL);
14726 ASSERT(dtrace_helpers > 0);
14728 help = from->p_dtrace_helpers;
14729 newhelp = dtrace_helpers_create(to);
14730 ASSERT(to->p_dtrace_helpers != NULL);
14732 newhelp->dthps_generation = help->dthps_generation;
14733 vstate = &newhelp->dthps_vstate;
14736 * Duplicate the helper actions.
14738 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14739 if ((helper = help->dthps_actions[i]) == NULL)
14742 for (last = NULL; helper != NULL; helper = helper->dtha_next) {
14743 new = kmem_zalloc(sizeof (dtrace_helper_action_t),
14745 new->dtha_generation = helper->dtha_generation;
14747 if ((dp = helper->dtha_predicate) != NULL) {
14748 dp = dtrace_difo_duplicate(dp, vstate);
14749 new->dtha_predicate = dp;
14752 new->dtha_nactions = helper->dtha_nactions;
14753 sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
14754 new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
14756 for (j = 0; j < new->dtha_nactions; j++) {
14757 dtrace_difo_t *dp = helper->dtha_actions[j];
14759 ASSERT(dp != NULL);
14760 dp = dtrace_difo_duplicate(dp, vstate);
14761 new->dtha_actions[j] = dp;
14764 if (last != NULL) {
14765 last->dtha_next = new;
14767 newhelp->dthps_actions[i] = new;
14775 * Duplicate the helper providers and register them with the
14776 * DTrace framework.
14778 if (help->dthps_nprovs > 0) {
14779 newhelp->dthps_nprovs = help->dthps_nprovs;
14780 newhelp->dthps_maxprovs = help->dthps_nprovs;
14781 newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
14782 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
14783 for (i = 0; i < newhelp->dthps_nprovs; i++) {
14784 newhelp->dthps_provs[i] = help->dthps_provs[i];
14785 newhelp->dthps_provs[i]->dthp_ref++;
14791 mutex_exit(&dtrace_lock);
14794 dtrace_helper_provider_register(to, newhelp, NULL);
14799 * DTrace Hook Functions
14802 dtrace_module_loaded(modctl_t *ctl)
14804 dtrace_provider_t *prv;
14806 mutex_enter(&dtrace_provider_lock);
14807 mutex_enter(&mod_lock);
14809 ASSERT(ctl->mod_busy);
14812 * We're going to call each providers per-module provide operation
14813 * specifying only this module.
14815 for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
14816 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
14818 mutex_exit(&mod_lock);
14819 mutex_exit(&dtrace_provider_lock);
14822 * If we have any retained enablings, we need to match against them.
14823 * Enabling probes requires that cpu_lock be held, and we cannot hold
14824 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
14825 * module. (In particular, this happens when loading scheduling
14826 * classes.) So if we have any retained enablings, we need to dispatch
14827 * our task queue to do the match for us.
14829 mutex_enter(&dtrace_lock);
14831 if (dtrace_retained == NULL) {
14832 mutex_exit(&dtrace_lock);
14836 (void) taskq_dispatch(dtrace_taskq,
14837 (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
14839 mutex_exit(&dtrace_lock);
14842 * And now, for a little heuristic sleaze: in general, we want to
14843 * match modules as soon as they load. However, we cannot guarantee
14844 * this, because it would lead us to the lock ordering violation
14845 * outlined above. The common case, of course, is that cpu_lock is
14846 * _not_ held -- so we delay here for a clock tick, hoping that that's
14847 * long enough for the task queue to do its work. If it's not, it's
14848 * not a serious problem -- it just means that the module that we
14849 * just loaded may not be immediately instrumentable.
14855 dtrace_module_unloaded(modctl_t *ctl)
14857 dtrace_probe_t template, *probe, *first, *next;
14858 dtrace_provider_t *prov;
14860 template.dtpr_mod = ctl->mod_modname;
14862 mutex_enter(&dtrace_provider_lock);
14863 mutex_enter(&mod_lock);
14864 mutex_enter(&dtrace_lock);
14866 if (dtrace_bymod == NULL) {
14868 * The DTrace module is loaded (obviously) but not attached;
14869 * we don't have any work to do.
14871 mutex_exit(&dtrace_provider_lock);
14872 mutex_exit(&mod_lock);
14873 mutex_exit(&dtrace_lock);
14877 for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
14878 probe != NULL; probe = probe->dtpr_nextmod) {
14879 if (probe->dtpr_ecb != NULL) {
14880 mutex_exit(&dtrace_provider_lock);
14881 mutex_exit(&mod_lock);
14882 mutex_exit(&dtrace_lock);
14885 * This shouldn't _actually_ be possible -- we're
14886 * unloading a module that has an enabled probe in it.
14887 * (It's normally up to the provider to make sure that
14888 * this can't happen.) However, because dtps_enable()
14889 * doesn't have a failure mode, there can be an
14890 * enable/unload race. Upshot: we don't want to
14891 * assert, but we're not going to disable the
14894 if (dtrace_err_verbose) {
14895 cmn_err(CE_WARN, "unloaded module '%s' had "
14896 "enabled probes", ctl->mod_modname);
14905 for (first = NULL; probe != NULL; probe = next) {
14906 ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
14908 dtrace_probes[probe->dtpr_id - 1] = NULL;
14910 next = probe->dtpr_nextmod;
14911 dtrace_hash_remove(dtrace_bymod, probe);
14912 dtrace_hash_remove(dtrace_byfunc, probe);
14913 dtrace_hash_remove(dtrace_byname, probe);
14915 if (first == NULL) {
14917 probe->dtpr_nextmod = NULL;
14919 probe->dtpr_nextmod = first;
14925 * We've removed all of the module's probes from the hash chains and
14926 * from the probe array. Now issue a dtrace_sync() to be sure that
14927 * everyone has cleared out from any probe array processing.
14931 for (probe = first; probe != NULL; probe = first) {
14932 first = probe->dtpr_nextmod;
14933 prov = probe->dtpr_provider;
14934 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
14936 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
14937 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
14938 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
14939 vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
14940 kmem_free(probe, sizeof (dtrace_probe_t));
14943 mutex_exit(&dtrace_lock);
14944 mutex_exit(&mod_lock);
14945 mutex_exit(&dtrace_provider_lock);
14949 dtrace_suspend(void)
14951 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
14955 dtrace_resume(void)
14957 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
14962 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
14964 ASSERT(MUTEX_HELD(&cpu_lock));
14965 mutex_enter(&dtrace_lock);
14969 dtrace_state_t *state;
14970 dtrace_optval_t *opt, rs, c;
14973 * For now, we only allocate a new buffer for anonymous state.
14975 if ((state = dtrace_anon.dta_state) == NULL)
14978 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
14981 opt = state->dts_options;
14982 c = opt[DTRACEOPT_CPU];
14984 if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
14988 * Regardless of what the actual policy is, we're going to
14989 * temporarily set our resize policy to be manual. We're
14990 * also going to temporarily set our CPU option to denote
14991 * the newly configured CPU.
14993 rs = opt[DTRACEOPT_BUFRESIZE];
14994 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
14995 opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
14997 (void) dtrace_state_buffers(state);
14999 opt[DTRACEOPT_BUFRESIZE] = rs;
15000 opt[DTRACEOPT_CPU] = c;
15007 * We don't free the buffer in the CPU_UNCONFIG case. (The
15008 * buffer will be freed when the consumer exits.)
15016 mutex_exit(&dtrace_lock);
15022 dtrace_cpu_setup_initial(processorid_t cpu)
15024 (void) dtrace_cpu_setup(CPU_CONFIG, cpu);
15029 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
15031 if (dtrace_toxranges >= dtrace_toxranges_max) {
15033 dtrace_toxrange_t *range;
15035 osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15038 ASSERT(dtrace_toxrange == NULL);
15039 ASSERT(dtrace_toxranges_max == 0);
15040 dtrace_toxranges_max = 1;
15042 dtrace_toxranges_max <<= 1;
15045 nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15046 range = kmem_zalloc(nsize, KM_SLEEP);
15048 if (dtrace_toxrange != NULL) {
15049 ASSERT(osize != 0);
15050 bcopy(dtrace_toxrange, range, osize);
15051 kmem_free(dtrace_toxrange, osize);
15054 dtrace_toxrange = range;
15057 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == 0);
15058 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == 0);
15060 dtrace_toxrange[dtrace_toxranges].dtt_base = base;
15061 dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
15062 dtrace_toxranges++;
15066 * DTrace Driver Cookbook Functions
15071 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
15073 dtrace_provider_id_t id;
15074 dtrace_state_t *state = NULL;
15075 dtrace_enabling_t *enab;
15077 mutex_enter(&cpu_lock);
15078 mutex_enter(&dtrace_provider_lock);
15079 mutex_enter(&dtrace_lock);
15081 if (ddi_soft_state_init(&dtrace_softstate,
15082 sizeof (dtrace_state_t), 0) != 0) {
15083 cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
15084 mutex_exit(&cpu_lock);
15085 mutex_exit(&dtrace_provider_lock);
15086 mutex_exit(&dtrace_lock);
15087 return (DDI_FAILURE);
15090 if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
15091 DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
15092 ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
15093 DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
15094 cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
15095 ddi_remove_minor_node(devi, NULL);
15096 ddi_soft_state_fini(&dtrace_softstate);
15097 mutex_exit(&cpu_lock);
15098 mutex_exit(&dtrace_provider_lock);
15099 mutex_exit(&dtrace_lock);
15100 return (DDI_FAILURE);
15103 ddi_report_dev(devi);
15104 dtrace_devi = devi;
15106 dtrace_modload = dtrace_module_loaded;
15107 dtrace_modunload = dtrace_module_unloaded;
15108 dtrace_cpu_init = dtrace_cpu_setup_initial;
15109 dtrace_helpers_cleanup = dtrace_helpers_destroy;
15110 dtrace_helpers_fork = dtrace_helpers_duplicate;
15111 dtrace_cpustart_init = dtrace_suspend;
15112 dtrace_cpustart_fini = dtrace_resume;
15113 dtrace_debugger_init = dtrace_suspend;
15114 dtrace_debugger_fini = dtrace_resume;
15116 register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
15118 ASSERT(MUTEX_HELD(&cpu_lock));
15120 dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
15121 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
15122 dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
15123 UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
15124 VM_SLEEP | VMC_IDENTIFIER);
15125 dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
15128 dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
15129 sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
15130 NULL, NULL, NULL, NULL, NULL, 0);
15132 ASSERT(MUTEX_HELD(&cpu_lock));
15133 dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
15134 offsetof(dtrace_probe_t, dtpr_nextmod),
15135 offsetof(dtrace_probe_t, dtpr_prevmod));
15137 dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
15138 offsetof(dtrace_probe_t, dtpr_nextfunc),
15139 offsetof(dtrace_probe_t, dtpr_prevfunc));
15141 dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
15142 offsetof(dtrace_probe_t, dtpr_nextname),
15143 offsetof(dtrace_probe_t, dtpr_prevname));
15145 if (dtrace_retain_max < 1) {
15146 cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; "
15147 "setting to 1", dtrace_retain_max);
15148 dtrace_retain_max = 1;
15152 * Now discover our toxic ranges.
15154 dtrace_toxic_ranges(dtrace_toxrange_add);
15157 * Before we register ourselves as a provider to our own framework,
15158 * we would like to assert that dtrace_provider is NULL -- but that's
15159 * not true if we were loaded as a dependency of a DTrace provider.
15160 * Once we've registered, we can assert that dtrace_provider is our
15163 (void) dtrace_register("dtrace", &dtrace_provider_attr,
15164 DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
15166 ASSERT(dtrace_provider != NULL);
15167 ASSERT((dtrace_provider_id_t)dtrace_provider == id);
15169 dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
15170 dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
15171 dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
15172 dtrace_provider, NULL, NULL, "END", 0, NULL);
15173 dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
15174 dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
15176 dtrace_anon_property();
15177 mutex_exit(&cpu_lock);
15180 * If DTrace helper tracing is enabled, we need to allocate the
15181 * trace buffer and initialize the values.
15183 if (dtrace_helptrace_enabled) {
15184 ASSERT(dtrace_helptrace_buffer == NULL);
15185 dtrace_helptrace_buffer =
15186 kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
15187 dtrace_helptrace_next = 0;
15191 * If there are already providers, we must ask them to provide their
15192 * probes, and then match any anonymous enabling against them. Note
15193 * that there should be no other retained enablings at this time:
15194 * the only retained enablings at this time should be the anonymous
15197 if (dtrace_anon.dta_enabling != NULL) {
15198 ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
15200 dtrace_enabling_provide(NULL);
15201 state = dtrace_anon.dta_state;
15204 * We couldn't hold cpu_lock across the above call to
15205 * dtrace_enabling_provide(), but we must hold it to actually
15206 * enable the probes. We have to drop all of our locks, pick
15207 * up cpu_lock, and regain our locks before matching the
15208 * retained anonymous enabling.
15210 mutex_exit(&dtrace_lock);
15211 mutex_exit(&dtrace_provider_lock);
15213 mutex_enter(&cpu_lock);
15214 mutex_enter(&dtrace_provider_lock);
15215 mutex_enter(&dtrace_lock);
15217 if ((enab = dtrace_anon.dta_enabling) != NULL)
15218 (void) dtrace_enabling_match(enab, NULL);
15220 mutex_exit(&cpu_lock);
15223 mutex_exit(&dtrace_lock);
15224 mutex_exit(&dtrace_provider_lock);
15226 if (state != NULL) {
15228 * If we created any anonymous state, set it going now.
15230 (void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
15233 return (DDI_SUCCESS);
15238 #if __FreeBSD_version >= 800039
15240 dtrace_dtr(void *data __unused)
15249 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
15251 dtrace_open(struct cdev *dev, int oflags, int devtype, struct thread *td)
15254 dtrace_state_t *state;
15260 if (getminor(*devp) == DTRACEMNRN_HELPER)
15264 * If this wasn't an open with the "helper" minor, then it must be
15265 * the "dtrace" minor.
15267 ASSERT(getminor(*devp) == DTRACEMNRN_DTRACE);
15269 cred_t *cred_p = NULL;
15271 #if __FreeBSD_version < 800039
15273 * The first minor device is the one that is cloned so there is
15274 * nothing more to do here.
15276 if (dev2unit(dev) == 0)
15280 * Devices are cloned, so if the DTrace state has already
15281 * been allocated, that means this device belongs to a
15282 * different client. Each client should open '/dev/dtrace'
15283 * to get a cloned device.
15285 if (dev->si_drv1 != NULL)
15289 cred_p = dev->si_cred;
15293 * If no DTRACE_PRIV_* bits are set in the credential, then the
15294 * caller lacks sufficient permission to do anything with DTrace.
15296 dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
15297 if (priv == DTRACE_PRIV_NONE) {
15299 #if __FreeBSD_version < 800039
15300 /* Destroy the cloned device. */
15309 * Ask all providers to provide all their probes.
15311 mutex_enter(&dtrace_provider_lock);
15312 dtrace_probe_provide(NULL, NULL);
15313 mutex_exit(&dtrace_provider_lock);
15315 mutex_enter(&cpu_lock);
15316 mutex_enter(&dtrace_lock);
15318 dtrace_membar_producer();
15322 * If the kernel debugger is active (that is, if the kernel debugger
15323 * modified text in some way), we won't allow the open.
15325 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
15327 mutex_exit(&cpu_lock);
15328 mutex_exit(&dtrace_lock);
15332 state = dtrace_state_create(devp, cred_p);
15334 state = dtrace_state_create(dev);
15335 #if __FreeBSD_version < 800039
15336 dev->si_drv1 = state;
15338 devfs_set_cdevpriv(state, dtrace_dtr);
15342 mutex_exit(&cpu_lock);
15344 if (state == NULL) {
15346 if (--dtrace_opens == 0)
15347 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15351 mutex_exit(&dtrace_lock);
15353 #if __FreeBSD_version < 800039
15354 /* Destroy the cloned device. */
15361 mutex_exit(&dtrace_lock);
15369 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
15371 dtrace_close(struct cdev *dev, int flags, int fmt __unused, struct thread *td)
15375 minor_t minor = getminor(dev);
15376 dtrace_state_t *state;
15378 if (minor == DTRACEMNRN_HELPER)
15381 state = ddi_get_soft_state(dtrace_softstate, minor);
15383 #if __FreeBSD_version < 800039
15384 dtrace_state_t *state = dev->si_drv1;
15386 /* Check if this is not a cloned device. */
15387 if (dev2unit(dev) == 0)
15390 dtrace_state_t *state;
15391 devfs_get_cdevpriv((void **) &state);
15396 mutex_enter(&cpu_lock);
15397 mutex_enter(&dtrace_lock);
15399 if (state != NULL) {
15400 if (state->dts_anon) {
15402 * There is anonymous state. Destroy that first.
15404 ASSERT(dtrace_anon.dta_state == NULL);
15405 dtrace_state_destroy(state->dts_anon);
15408 dtrace_state_destroy(state);
15411 kmem_free(state, 0);
15412 #if __FreeBSD_version < 800039
15413 dev->si_drv1 = NULL;
15415 devfs_clear_cdevpriv();
15420 ASSERT(dtrace_opens > 0);
15422 if (--dtrace_opens == 0)
15423 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15428 mutex_exit(&dtrace_lock);
15429 mutex_exit(&cpu_lock);
15431 #if __FreeBSD_version < 800039
15432 /* Schedule this cloned device to be destroyed. */
15433 destroy_dev_sched(dev);
15442 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
15445 dof_helper_t help, *dhp = NULL;
15448 case DTRACEHIOC_ADDDOF:
15449 if (copyin((void *)arg, &help, sizeof (help)) != 0) {
15450 dtrace_dof_error(NULL, "failed to copyin DOF helper");
15455 arg = (intptr_t)help.dofhp_dof;
15458 case DTRACEHIOC_ADD: {
15459 dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
15464 mutex_enter(&dtrace_lock);
15467 * dtrace_helper_slurp() takes responsibility for the dof --
15468 * it may free it now or it may save it and free it later.
15470 if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
15477 mutex_exit(&dtrace_lock);
15481 case DTRACEHIOC_REMOVE: {
15482 mutex_enter(&dtrace_lock);
15483 rval = dtrace_helper_destroygen(arg);
15484 mutex_exit(&dtrace_lock);
15498 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
15500 minor_t minor = getminor(dev);
15501 dtrace_state_t *state;
15504 if (minor == DTRACEMNRN_HELPER)
15505 return (dtrace_ioctl_helper(cmd, arg, rv));
15507 state = ddi_get_soft_state(dtrace_softstate, minor);
15509 if (state->dts_anon) {
15510 ASSERT(dtrace_anon.dta_state == NULL);
15511 state = state->dts_anon;
15515 case DTRACEIOC_PROVIDER: {
15516 dtrace_providerdesc_t pvd;
15517 dtrace_provider_t *pvp;
15519 if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
15522 pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
15523 mutex_enter(&dtrace_provider_lock);
15525 for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
15526 if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
15530 mutex_exit(&dtrace_provider_lock);
15535 bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
15536 bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
15538 if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
15544 case DTRACEIOC_EPROBE: {
15545 dtrace_eprobedesc_t epdesc;
15547 dtrace_action_t *act;
15553 if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
15556 mutex_enter(&dtrace_lock);
15558 if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
15559 mutex_exit(&dtrace_lock);
15563 if (ecb->dte_probe == NULL) {
15564 mutex_exit(&dtrace_lock);
15568 epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
15569 epdesc.dtepd_uarg = ecb->dte_uarg;
15570 epdesc.dtepd_size = ecb->dte_size;
15572 nrecs = epdesc.dtepd_nrecs;
15573 epdesc.dtepd_nrecs = 0;
15574 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15575 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15578 epdesc.dtepd_nrecs++;
15582 * Now that we have the size, we need to allocate a temporary
15583 * buffer in which to store the complete description. We need
15584 * the temporary buffer to be able to drop dtrace_lock()
15585 * across the copyout(), below.
15587 size = sizeof (dtrace_eprobedesc_t) +
15588 (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
15590 buf = kmem_alloc(size, KM_SLEEP);
15591 dest = (uintptr_t)buf;
15593 bcopy(&epdesc, (void *)dest, sizeof (epdesc));
15594 dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
15596 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15597 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15603 bcopy(&act->dta_rec, (void *)dest,
15604 sizeof (dtrace_recdesc_t));
15605 dest += sizeof (dtrace_recdesc_t);
15608 mutex_exit(&dtrace_lock);
15610 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
15611 kmem_free(buf, size);
15615 kmem_free(buf, size);
15619 case DTRACEIOC_AGGDESC: {
15620 dtrace_aggdesc_t aggdesc;
15621 dtrace_action_t *act;
15622 dtrace_aggregation_t *agg;
15625 dtrace_recdesc_t *lrec;
15630 if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
15633 mutex_enter(&dtrace_lock);
15635 if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
15636 mutex_exit(&dtrace_lock);
15640 aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
15642 nrecs = aggdesc.dtagd_nrecs;
15643 aggdesc.dtagd_nrecs = 0;
15645 offs = agg->dtag_base;
15646 lrec = &agg->dtag_action.dta_rec;
15647 aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
15649 for (act = agg->dtag_first; ; act = act->dta_next) {
15650 ASSERT(act->dta_intuple ||
15651 DTRACEACT_ISAGG(act->dta_kind));
15654 * If this action has a record size of zero, it
15655 * denotes an argument to the aggregating action.
15656 * Because the presence of this record doesn't (or
15657 * shouldn't) affect the way the data is interpreted,
15658 * we don't copy it out to save user-level the
15659 * confusion of dealing with a zero-length record.
15661 if (act->dta_rec.dtrd_size == 0) {
15662 ASSERT(agg->dtag_hasarg);
15666 aggdesc.dtagd_nrecs++;
15668 if (act == &agg->dtag_action)
15673 * Now that we have the size, we need to allocate a temporary
15674 * buffer in which to store the complete description. We need
15675 * the temporary buffer to be able to drop dtrace_lock()
15676 * across the copyout(), below.
15678 size = sizeof (dtrace_aggdesc_t) +
15679 (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
15681 buf = kmem_alloc(size, KM_SLEEP);
15682 dest = (uintptr_t)buf;
15684 bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
15685 dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
15687 for (act = agg->dtag_first; ; act = act->dta_next) {
15688 dtrace_recdesc_t rec = act->dta_rec;
15691 * See the comment in the above loop for why we pass
15692 * over zero-length records.
15694 if (rec.dtrd_size == 0) {
15695 ASSERT(agg->dtag_hasarg);
15702 rec.dtrd_offset -= offs;
15703 bcopy(&rec, (void *)dest, sizeof (rec));
15704 dest += sizeof (dtrace_recdesc_t);
15706 if (act == &agg->dtag_action)
15710 mutex_exit(&dtrace_lock);
15712 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
15713 kmem_free(buf, size);
15717 kmem_free(buf, size);
15721 case DTRACEIOC_ENABLE: {
15723 dtrace_enabling_t *enab = NULL;
15724 dtrace_vstate_t *vstate;
15730 * If a NULL argument has been passed, we take this as our
15731 * cue to reevaluate our enablings.
15734 dtrace_enabling_matchall();
15739 if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
15742 mutex_enter(&cpu_lock);
15743 mutex_enter(&dtrace_lock);
15744 vstate = &state->dts_vstate;
15746 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
15747 mutex_exit(&dtrace_lock);
15748 mutex_exit(&cpu_lock);
15749 dtrace_dof_destroy(dof);
15753 if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
15754 mutex_exit(&dtrace_lock);
15755 mutex_exit(&cpu_lock);
15756 dtrace_dof_destroy(dof);
15760 if ((rval = dtrace_dof_options(dof, state)) != 0) {
15761 dtrace_enabling_destroy(enab);
15762 mutex_exit(&dtrace_lock);
15763 mutex_exit(&cpu_lock);
15764 dtrace_dof_destroy(dof);
15768 if ((err = dtrace_enabling_match(enab, rv)) == 0) {
15769 err = dtrace_enabling_retain(enab);
15771 dtrace_enabling_destroy(enab);
15774 mutex_exit(&cpu_lock);
15775 mutex_exit(&dtrace_lock);
15776 dtrace_dof_destroy(dof);
15781 case DTRACEIOC_REPLICATE: {
15782 dtrace_repldesc_t desc;
15783 dtrace_probedesc_t *match = &desc.dtrpd_match;
15784 dtrace_probedesc_t *create = &desc.dtrpd_create;
15787 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15790 match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
15791 match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
15792 match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
15793 match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
15795 create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
15796 create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
15797 create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
15798 create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
15800 mutex_enter(&dtrace_lock);
15801 err = dtrace_enabling_replicate(state, match, create);
15802 mutex_exit(&dtrace_lock);
15807 case DTRACEIOC_PROBEMATCH:
15808 case DTRACEIOC_PROBES: {
15809 dtrace_probe_t *probe = NULL;
15810 dtrace_probedesc_t desc;
15811 dtrace_probekey_t pkey;
15818 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15821 desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
15822 desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
15823 desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
15824 desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
15827 * Before we attempt to match this probe, we want to give
15828 * all providers the opportunity to provide it.
15830 if (desc.dtpd_id == DTRACE_IDNONE) {
15831 mutex_enter(&dtrace_provider_lock);
15832 dtrace_probe_provide(&desc, NULL);
15833 mutex_exit(&dtrace_provider_lock);
15837 if (cmd == DTRACEIOC_PROBEMATCH) {
15838 dtrace_probekey(&desc, &pkey);
15839 pkey.dtpk_id = DTRACE_IDNONE;
15842 dtrace_cred2priv(cr, &priv, &uid, &zoneid);
15844 mutex_enter(&dtrace_lock);
15846 if (cmd == DTRACEIOC_PROBEMATCH) {
15847 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
15848 if ((probe = dtrace_probes[i - 1]) != NULL &&
15849 (m = dtrace_match_probe(probe, &pkey,
15850 priv, uid, zoneid)) != 0)
15855 mutex_exit(&dtrace_lock);
15860 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
15861 if ((probe = dtrace_probes[i - 1]) != NULL &&
15862 dtrace_match_priv(probe, priv, uid, zoneid))
15867 if (probe == NULL) {
15868 mutex_exit(&dtrace_lock);
15872 dtrace_probe_description(probe, &desc);
15873 mutex_exit(&dtrace_lock);
15875 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15881 case DTRACEIOC_PROBEARG: {
15882 dtrace_argdesc_t desc;
15883 dtrace_probe_t *probe;
15884 dtrace_provider_t *prov;
15886 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15889 if (desc.dtargd_id == DTRACE_IDNONE)
15892 if (desc.dtargd_ndx == DTRACE_ARGNONE)
15895 mutex_enter(&dtrace_provider_lock);
15896 mutex_enter(&mod_lock);
15897 mutex_enter(&dtrace_lock);
15899 if (desc.dtargd_id > dtrace_nprobes) {
15900 mutex_exit(&dtrace_lock);
15901 mutex_exit(&mod_lock);
15902 mutex_exit(&dtrace_provider_lock);
15906 if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
15907 mutex_exit(&dtrace_lock);
15908 mutex_exit(&mod_lock);
15909 mutex_exit(&dtrace_provider_lock);
15913 mutex_exit(&dtrace_lock);
15915 prov = probe->dtpr_provider;
15917 if (prov->dtpv_pops.dtps_getargdesc == NULL) {
15919 * There isn't any typed information for this probe.
15920 * Set the argument number to DTRACE_ARGNONE.
15922 desc.dtargd_ndx = DTRACE_ARGNONE;
15924 desc.dtargd_native[0] = '\0';
15925 desc.dtargd_xlate[0] = '\0';
15926 desc.dtargd_mapping = desc.dtargd_ndx;
15928 prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
15929 probe->dtpr_id, probe->dtpr_arg, &desc);
15932 mutex_exit(&mod_lock);
15933 mutex_exit(&dtrace_provider_lock);
15935 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15941 case DTRACEIOC_GO: {
15942 processorid_t cpuid;
15943 rval = dtrace_state_go(state, &cpuid);
15948 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
15954 case DTRACEIOC_STOP: {
15955 processorid_t cpuid;
15957 mutex_enter(&dtrace_lock);
15958 rval = dtrace_state_stop(state, &cpuid);
15959 mutex_exit(&dtrace_lock);
15964 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
15970 case DTRACEIOC_DOFGET: {
15971 dof_hdr_t hdr, *dof;
15974 if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
15977 mutex_enter(&dtrace_lock);
15978 dof = dtrace_dof_create(state);
15979 mutex_exit(&dtrace_lock);
15981 len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
15982 rval = copyout(dof, (void *)arg, len);
15983 dtrace_dof_destroy(dof);
15985 return (rval == 0 ? 0 : EFAULT);
15988 case DTRACEIOC_AGGSNAP:
15989 case DTRACEIOC_BUFSNAP: {
15990 dtrace_bufdesc_t desc;
15992 dtrace_buffer_t *buf;
15994 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15997 if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
16000 mutex_enter(&dtrace_lock);
16002 if (cmd == DTRACEIOC_BUFSNAP) {
16003 buf = &state->dts_buffer[desc.dtbd_cpu];
16005 buf = &state->dts_aggbuffer[desc.dtbd_cpu];
16008 if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
16009 size_t sz = buf->dtb_offset;
16011 if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
16012 mutex_exit(&dtrace_lock);
16017 * If this buffer has already been consumed, we're
16018 * going to indicate that there's nothing left here
16021 if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
16022 mutex_exit(&dtrace_lock);
16024 desc.dtbd_size = 0;
16025 desc.dtbd_drops = 0;
16026 desc.dtbd_errors = 0;
16027 desc.dtbd_oldest = 0;
16028 sz = sizeof (desc);
16030 if (copyout(&desc, (void *)arg, sz) != 0)
16037 * If this is a ring buffer that has wrapped, we want
16038 * to copy the whole thing out.
16040 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
16041 dtrace_buffer_polish(buf);
16042 sz = buf->dtb_size;
16045 if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
16046 mutex_exit(&dtrace_lock);
16050 desc.dtbd_size = sz;
16051 desc.dtbd_drops = buf->dtb_drops;
16052 desc.dtbd_errors = buf->dtb_errors;
16053 desc.dtbd_oldest = buf->dtb_xamot_offset;
16055 mutex_exit(&dtrace_lock);
16057 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16060 buf->dtb_flags |= DTRACEBUF_CONSUMED;
16065 if (buf->dtb_tomax == NULL) {
16066 ASSERT(buf->dtb_xamot == NULL);
16067 mutex_exit(&dtrace_lock);
16071 cached = buf->dtb_tomax;
16072 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
16074 dtrace_xcall(desc.dtbd_cpu,
16075 (dtrace_xcall_t)dtrace_buffer_switch, buf);
16077 state->dts_errors += buf->dtb_xamot_errors;
16080 * If the buffers did not actually switch, then the cross call
16081 * did not take place -- presumably because the given CPU is
16082 * not in the ready set. If this is the case, we'll return
16085 if (buf->dtb_tomax == cached) {
16086 ASSERT(buf->dtb_xamot != cached);
16087 mutex_exit(&dtrace_lock);
16091 ASSERT(cached == buf->dtb_xamot);
16094 * We have our snapshot; now copy it out.
16096 if (copyout(buf->dtb_xamot, desc.dtbd_data,
16097 buf->dtb_xamot_offset) != 0) {
16098 mutex_exit(&dtrace_lock);
16102 desc.dtbd_size = buf->dtb_xamot_offset;
16103 desc.dtbd_drops = buf->dtb_xamot_drops;
16104 desc.dtbd_errors = buf->dtb_xamot_errors;
16105 desc.dtbd_oldest = 0;
16107 mutex_exit(&dtrace_lock);
16110 * Finally, copy out the buffer description.
16112 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16118 case DTRACEIOC_CONF: {
16119 dtrace_conf_t conf;
16121 bzero(&conf, sizeof (conf));
16122 conf.dtc_difversion = DIF_VERSION;
16123 conf.dtc_difintregs = DIF_DIR_NREGS;
16124 conf.dtc_diftupregs = DIF_DTR_NREGS;
16125 conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
16127 if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
16133 case DTRACEIOC_STATUS: {
16134 dtrace_status_t stat;
16135 dtrace_dstate_t *dstate;
16140 * See the comment in dtrace_state_deadman() for the reason
16141 * for setting dts_laststatus to INT64_MAX before setting
16142 * it to the correct value.
16144 state->dts_laststatus = INT64_MAX;
16145 dtrace_membar_producer();
16146 state->dts_laststatus = dtrace_gethrtime();
16148 bzero(&stat, sizeof (stat));
16150 mutex_enter(&dtrace_lock);
16152 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
16153 mutex_exit(&dtrace_lock);
16157 if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
16158 stat.dtst_exiting = 1;
16160 nerrs = state->dts_errors;
16161 dstate = &state->dts_vstate.dtvs_dynvars;
16163 for (i = 0; i < NCPU; i++) {
16164 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
16166 stat.dtst_dyndrops += dcpu->dtdsc_drops;
16167 stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
16168 stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
16170 if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
16171 stat.dtst_filled++;
16173 nerrs += state->dts_buffer[i].dtb_errors;
16175 for (j = 0; j < state->dts_nspeculations; j++) {
16176 dtrace_speculation_t *spec;
16177 dtrace_buffer_t *buf;
16179 spec = &state->dts_speculations[j];
16180 buf = &spec->dtsp_buffer[i];
16181 stat.dtst_specdrops += buf->dtb_xamot_drops;
16185 stat.dtst_specdrops_busy = state->dts_speculations_busy;
16186 stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
16187 stat.dtst_stkstroverflows = state->dts_stkstroverflows;
16188 stat.dtst_dblerrors = state->dts_dblerrors;
16190 (state->dts_activity == DTRACE_ACTIVITY_KILLED);
16191 stat.dtst_errors = nerrs;
16193 mutex_exit(&dtrace_lock);
16195 if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
16201 case DTRACEIOC_FORMAT: {
16202 dtrace_fmtdesc_t fmt;
16206 if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
16209 mutex_enter(&dtrace_lock);
16211 if (fmt.dtfd_format == 0 ||
16212 fmt.dtfd_format > state->dts_nformats) {
16213 mutex_exit(&dtrace_lock);
16218 * Format strings are allocated contiguously and they are
16219 * never freed; if a format index is less than the number
16220 * of formats, we can assert that the format map is non-NULL
16221 * and that the format for the specified index is non-NULL.
16223 ASSERT(state->dts_formats != NULL);
16224 str = state->dts_formats[fmt.dtfd_format - 1];
16225 ASSERT(str != NULL);
16227 len = strlen(str) + 1;
16229 if (len > fmt.dtfd_length) {
16230 fmt.dtfd_length = len;
16232 if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
16233 mutex_exit(&dtrace_lock);
16237 if (copyout(str, fmt.dtfd_string, len) != 0) {
16238 mutex_exit(&dtrace_lock);
16243 mutex_exit(&dtrace_lock);
16256 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
16258 dtrace_state_t *state;
16265 return (DDI_SUCCESS);
16268 return (DDI_FAILURE);
16271 mutex_enter(&cpu_lock);
16272 mutex_enter(&dtrace_provider_lock);
16273 mutex_enter(&dtrace_lock);
16275 ASSERT(dtrace_opens == 0);
16277 if (dtrace_helpers > 0) {
16278 mutex_exit(&dtrace_provider_lock);
16279 mutex_exit(&dtrace_lock);
16280 mutex_exit(&cpu_lock);
16281 return (DDI_FAILURE);
16284 if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
16285 mutex_exit(&dtrace_provider_lock);
16286 mutex_exit(&dtrace_lock);
16287 mutex_exit(&cpu_lock);
16288 return (DDI_FAILURE);
16291 dtrace_provider = NULL;
16293 if ((state = dtrace_anon_grab()) != NULL) {
16295 * If there were ECBs on this state, the provider should
16296 * have not been allowed to detach; assert that there is
16299 ASSERT(state->dts_necbs == 0);
16300 dtrace_state_destroy(state);
16303 * If we're being detached with anonymous state, we need to
16304 * indicate to the kernel debugger that DTrace is now inactive.
16306 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
16309 bzero(&dtrace_anon, sizeof (dtrace_anon_t));
16310 unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
16311 dtrace_cpu_init = NULL;
16312 dtrace_helpers_cleanup = NULL;
16313 dtrace_helpers_fork = NULL;
16314 dtrace_cpustart_init = NULL;
16315 dtrace_cpustart_fini = NULL;
16316 dtrace_debugger_init = NULL;
16317 dtrace_debugger_fini = NULL;
16318 dtrace_modload = NULL;
16319 dtrace_modunload = NULL;
16321 mutex_exit(&cpu_lock);
16323 if (dtrace_helptrace_enabled) {
16324 kmem_free(dtrace_helptrace_buffer, dtrace_helptrace_bufsize);
16325 dtrace_helptrace_buffer = NULL;
16328 kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
16329 dtrace_probes = NULL;
16330 dtrace_nprobes = 0;
16332 dtrace_hash_destroy(dtrace_bymod);
16333 dtrace_hash_destroy(dtrace_byfunc);
16334 dtrace_hash_destroy(dtrace_byname);
16335 dtrace_bymod = NULL;
16336 dtrace_byfunc = NULL;
16337 dtrace_byname = NULL;
16339 kmem_cache_destroy(dtrace_state_cache);
16340 vmem_destroy(dtrace_minor);
16341 vmem_destroy(dtrace_arena);
16343 if (dtrace_toxrange != NULL) {
16344 kmem_free(dtrace_toxrange,
16345 dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
16346 dtrace_toxrange = NULL;
16347 dtrace_toxranges = 0;
16348 dtrace_toxranges_max = 0;
16351 ddi_remove_minor_node(dtrace_devi, NULL);
16352 dtrace_devi = NULL;
16354 ddi_soft_state_fini(&dtrace_softstate);
16356 ASSERT(dtrace_vtime_references == 0);
16357 ASSERT(dtrace_opens == 0);
16358 ASSERT(dtrace_retained == NULL);
16360 mutex_exit(&dtrace_lock);
16361 mutex_exit(&dtrace_provider_lock);
16364 * We don't destroy the task queue until after we have dropped our
16365 * locks (taskq_destroy() may block on running tasks). To prevent
16366 * attempting to do work after we have effectively detached but before
16367 * the task queue has been destroyed, all tasks dispatched via the
16368 * task queue must check that DTrace is still attached before
16369 * performing any operation.
16371 taskq_destroy(dtrace_taskq);
16372 dtrace_taskq = NULL;
16374 return (DDI_SUCCESS);
16381 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
16386 case DDI_INFO_DEVT2DEVINFO:
16387 *result = (void *)dtrace_devi;
16388 error = DDI_SUCCESS;
16390 case DDI_INFO_DEVT2INSTANCE:
16391 *result = (void *)0;
16392 error = DDI_SUCCESS;
16395 error = DDI_FAILURE;
16402 static struct cb_ops dtrace_cb_ops = {
16403 dtrace_open, /* open */
16404 dtrace_close, /* close */
16405 nulldev, /* strategy */
16406 nulldev, /* print */
16410 dtrace_ioctl, /* ioctl */
16411 nodev, /* devmap */
16413 nodev, /* segmap */
16414 nochpoll, /* poll */
16415 ddi_prop_op, /* cb_prop_op */
16417 D_NEW | D_MP /* Driver compatibility flag */
16420 static struct dev_ops dtrace_ops = {
16421 DEVO_REV, /* devo_rev */
16423 dtrace_info, /* get_dev_info */
16424 nulldev, /* identify */
16425 nulldev, /* probe */
16426 dtrace_attach, /* attach */
16427 dtrace_detach, /* detach */
16429 &dtrace_cb_ops, /* driver operations */
16430 NULL, /* bus operations */
16431 nodev /* dev power */
16434 static struct modldrv modldrv = {
16435 &mod_driverops, /* module type (this is a pseudo driver) */
16436 "Dynamic Tracing", /* name of module */
16437 &dtrace_ops, /* driver ops */
16440 static struct modlinkage modlinkage = {
16449 return (mod_install(&modlinkage));
16453 _info(struct modinfo *modinfop)
16455 return (mod_info(&modlinkage, modinfop));
16461 return (mod_remove(&modlinkage));
16465 static d_ioctl_t dtrace_ioctl;
16466 static d_ioctl_t dtrace_ioctl_helper;
16467 static void dtrace_load(void *);
16468 static int dtrace_unload(void);
16469 #if __FreeBSD_version < 800039
16470 static void dtrace_clone(void *, struct ucred *, char *, int , struct cdev **);
16471 static struct clonedevs *dtrace_clones; /* Ptr to the array of cloned devices. */
16472 static eventhandler_tag eh_tag; /* Event handler tag. */
16474 static struct cdev *dtrace_dev;
16475 static struct cdev *helper_dev;
16478 void dtrace_invop_init(void);
16479 void dtrace_invop_uninit(void);
16481 static struct cdevsw dtrace_cdevsw = {
16482 .d_version = D_VERSION,
16483 .d_flags = D_TRACKCLOSE | D_NEEDMINOR,
16484 .d_close = dtrace_close,
16485 .d_ioctl = dtrace_ioctl,
16486 .d_open = dtrace_open,
16487 .d_name = "dtrace",
16490 static struct cdevsw helper_cdevsw = {
16491 .d_version = D_VERSION,
16492 .d_flags = D_TRACKCLOSE | D_NEEDMINOR,
16493 .d_ioctl = dtrace_ioctl_helper,
16494 .d_name = "helper",
16497 #include <dtrace_anon.c>
16498 #if __FreeBSD_version < 800039
16499 #include <dtrace_clone.c>
16501 #include <dtrace_ioctl.c>
16502 #include <dtrace_load.c>
16503 #include <dtrace_modevent.c>
16504 #include <dtrace_sysctl.c>
16505 #include <dtrace_unload.c>
16506 #include <dtrace_vtime.c>
16507 #include <dtrace_hacks.c>
16508 #include <dtrace_isa.c>
16510 SYSINIT(dtrace_load, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_load, NULL);
16511 SYSUNINIT(dtrace_unload, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_unload, NULL);
16512 SYSINIT(dtrace_anon_init, SI_SUB_DTRACE_ANON, SI_ORDER_FIRST, dtrace_anon_init, NULL);
16514 DEV_MODULE(dtrace, dtrace_modevent, NULL);
16515 MODULE_VERSION(dtrace, 1);
16516 MODULE_DEPEND(dtrace, cyclic, 1, 1, 1);
16517 MODULE_DEPEND(dtrace, opensolaris, 1, 1, 1);