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 * Copyright (c) 2012 by Delphix. All rights reserved
27 * Use is subject to license terms.
30 #pragma ident "%Z%%M% %I% %E% SMI"
33 * DTrace - Dynamic Tracing for Solaris
35 * This is the implementation of the Solaris Dynamic Tracing framework
36 * (DTrace). The user-visible interface to DTrace is described at length in
37 * the "Solaris Dynamic Tracing Guide". The interfaces between the libdtrace
38 * library, the in-kernel DTrace framework, and the DTrace providers are
39 * described in the block comments in the <sys/dtrace.h> header file. The
40 * internal architecture of DTrace is described in the block comments in the
41 * <sys/dtrace_impl.h> header file. The comments contained within the DTrace
42 * implementation very much assume mastery of all of these sources; if one has
43 * an unanswered question about the implementation, one should consult them
46 * The functions here are ordered roughly as follows:
48 * - Probe context functions
49 * - Probe hashing functions
50 * - Non-probe context utility functions
51 * - Matching functions
52 * - Provider-to-Framework API functions
53 * - Probe management functions
54 * - DIF object functions
56 * - Predicate functions
59 * - Enabling functions
61 * - Anonymous enabling functions
62 * - Consumer state functions
65 * - Driver cookbook functions
67 * Each group of functions begins with a block comment labelled the "DTrace
68 * [Group] Functions", allowing one to find each block by searching forward
69 * on capital-f functions.
71 #include <sys/errno.h>
76 #include <sys/modctl.h>
78 #include <sys/systm.h>
81 #include <sys/sunddi.h>
83 #include <sys/cpuvar.h>
86 #include <sys/strsubr.h>
88 #include <sys/sysmacros.h>
89 #include <sys/dtrace_impl.h>
90 #include <sys/atomic.h>
91 #include <sys/cmn_err.h>
93 #include <sys/mutex_impl.h>
94 #include <sys/rwlock_impl.h>
96 #include <sys/ctf_api.h>
98 #include <sys/panic.h>
99 #include <sys/priv_impl.h>
101 #include <sys/policy.h>
103 #include <sys/cred_impl.h>
104 #include <sys/procfs_isa.h>
106 #include <sys/taskq.h>
108 #include <sys/mkdev.h>
111 #include <sys/zone.h>
112 #include <sys/socket.h>
113 #include <netinet/in.h>
115 /* FreeBSD includes: */
117 #include <sys/callout.h>
118 #include <sys/ctype.h>
119 #include <sys/eventhandler.h>
120 #include <sys/limits.h>
122 #include <sys/kernel.h>
123 #include <sys/malloc.h>
124 #include <sys/sysctl.h>
125 #include <sys/lock.h>
126 #include <sys/mutex.h>
127 #include <sys/rwlock.h>
129 #include <sys/dtrace_bsd.h>
130 #include <netinet/in.h>
131 #include "dtrace_cddl.h"
132 #include "dtrace_debug.c"
136 * DTrace Tunable Variables
138 * The following variables may be tuned by adding a line to /etc/system that
139 * includes both the name of the DTrace module ("dtrace") and the name of the
140 * variable. For example:
142 * set dtrace:dtrace_destructive_disallow = 1
144 * In general, the only variables that one should be tuning this way are those
145 * that affect system-wide DTrace behavior, and for which the default behavior
146 * is undesirable. Most of these variables are tunable on a per-consumer
147 * basis using DTrace options, and need not be tuned on a system-wide basis.
148 * When tuning these variables, avoid pathological values; while some attempt
149 * is made to verify the integrity of these variables, they are not considered
150 * part of the supported interface to DTrace, and they are therefore not
151 * checked comprehensively. Further, these variables should not be tuned
152 * dynamically via "mdb -kw" or other means; they should only be tuned via
155 int dtrace_destructive_disallow = 0;
156 dtrace_optval_t dtrace_nonroot_maxsize = (16 * 1024 * 1024);
157 size_t dtrace_difo_maxsize = (256 * 1024);
158 dtrace_optval_t dtrace_dof_maxsize = (256 * 1024);
159 size_t dtrace_global_maxsize = (16 * 1024);
160 size_t dtrace_actions_max = (16 * 1024);
161 size_t dtrace_retain_max = 1024;
162 dtrace_optval_t dtrace_helper_actions_max = 128;
163 dtrace_optval_t dtrace_helper_providers_max = 32;
164 dtrace_optval_t dtrace_dstate_defsize = (1 * 1024 * 1024);
165 size_t dtrace_strsize_default = 256;
166 dtrace_optval_t dtrace_cleanrate_default = 9900990; /* 101 hz */
167 dtrace_optval_t dtrace_cleanrate_min = 200000; /* 5000 hz */
168 dtrace_optval_t dtrace_cleanrate_max = (uint64_t)60 * NANOSEC; /* 1/minute */
169 dtrace_optval_t dtrace_aggrate_default = NANOSEC; /* 1 hz */
170 dtrace_optval_t dtrace_statusrate_default = NANOSEC; /* 1 hz */
171 dtrace_optval_t dtrace_statusrate_max = (hrtime_t)10 * NANOSEC; /* 6/minute */
172 dtrace_optval_t dtrace_switchrate_default = NANOSEC; /* 1 hz */
173 dtrace_optval_t dtrace_nspec_default = 1;
174 dtrace_optval_t dtrace_specsize_default = 32 * 1024;
175 dtrace_optval_t dtrace_stackframes_default = 20;
176 dtrace_optval_t dtrace_ustackframes_default = 20;
177 dtrace_optval_t dtrace_jstackframes_default = 50;
178 dtrace_optval_t dtrace_jstackstrsize_default = 512;
179 int dtrace_msgdsize_max = 128;
180 hrtime_t dtrace_chill_max = 500 * (NANOSEC / MILLISEC); /* 500 ms */
181 hrtime_t dtrace_chill_interval = NANOSEC; /* 1000 ms */
182 int dtrace_devdepth_max = 32;
183 int dtrace_err_verbose;
184 hrtime_t dtrace_deadman_interval = NANOSEC;
185 hrtime_t dtrace_deadman_timeout = (hrtime_t)10 * NANOSEC;
186 hrtime_t dtrace_deadman_user = (hrtime_t)30 * NANOSEC;
187 hrtime_t dtrace_unregister_defunct_reap = (hrtime_t)60 * NANOSEC;
190 * DTrace External Variables
192 * As dtrace(7D) is a kernel module, any DTrace variables are obviously
193 * available to DTrace consumers via the backtick (`) syntax. One of these,
194 * dtrace_zero, is made deliberately so: it is provided as a source of
195 * well-known, zero-filled memory. While this variable is not documented,
196 * it is used by some translators as an implementation detail.
198 const char dtrace_zero[256] = { 0 }; /* zero-filled memory */
201 * DTrace Internal Variables
204 static dev_info_t *dtrace_devi; /* device info */
207 static vmem_t *dtrace_arena; /* probe ID arena */
208 static vmem_t *dtrace_minor; /* minor number arena */
210 static taskq_t *dtrace_taskq; /* task queue */
211 static struct unrhdr *dtrace_arena; /* Probe ID number. */
213 static dtrace_probe_t **dtrace_probes; /* array of all probes */
214 static int dtrace_nprobes; /* number of probes */
215 static dtrace_provider_t *dtrace_provider; /* provider list */
216 static dtrace_meta_t *dtrace_meta_pid; /* user-land meta provider */
217 static int dtrace_opens; /* number of opens */
218 static int dtrace_helpers; /* number of helpers */
220 static void *dtrace_softstate; /* softstate pointer */
222 static dtrace_hash_t *dtrace_bymod; /* probes hashed by module */
223 static dtrace_hash_t *dtrace_byfunc; /* probes hashed by function */
224 static dtrace_hash_t *dtrace_byname; /* probes hashed by name */
225 static dtrace_toxrange_t *dtrace_toxrange; /* toxic range array */
226 static int dtrace_toxranges; /* number of toxic ranges */
227 static int dtrace_toxranges_max; /* size of toxic range array */
228 static dtrace_anon_t dtrace_anon; /* anonymous enabling */
229 static kmem_cache_t *dtrace_state_cache; /* cache for dynamic state */
230 static uint64_t dtrace_vtime_references; /* number of vtimestamp refs */
231 static kthread_t *dtrace_panicked; /* panicking thread */
232 static dtrace_ecb_t *dtrace_ecb_create_cache; /* cached created ECB */
233 static dtrace_genid_t dtrace_probegen; /* current probe generation */
234 static dtrace_helpers_t *dtrace_deferred_pid; /* deferred helper list */
235 static dtrace_enabling_t *dtrace_retained; /* list of retained enablings */
236 static dtrace_dynvar_t dtrace_dynhash_sink; /* end of dynamic hash chains */
238 static struct mtx dtrace_unr_mtx;
239 MTX_SYSINIT(dtrace_unr_mtx, &dtrace_unr_mtx, "Unique resource identifier", MTX_DEF);
240 int dtrace_in_probe; /* non-zero if executing a probe */
241 #if defined(__i386__) || defined(__amd64__) || defined(__mips__) || defined(__powerpc__)
242 uintptr_t dtrace_in_probe_addr; /* Address of invop when already in probe */
244 static eventhandler_tag dtrace_kld_load_tag;
245 static eventhandler_tag dtrace_kld_unload_try_tag;
250 * DTrace is protected by three (relatively coarse-grained) locks:
252 * (1) dtrace_lock is required to manipulate essentially any DTrace state,
253 * including enabling state, probes, ECBs, consumer state, helper state,
254 * etc. Importantly, dtrace_lock is _not_ required when in probe context;
255 * probe context is lock-free -- synchronization is handled via the
256 * dtrace_sync() cross call mechanism.
258 * (2) dtrace_provider_lock is required when manipulating provider state, or
259 * when provider state must be held constant.
261 * (3) dtrace_meta_lock is required when manipulating meta provider state, or
262 * when meta provider state must be held constant.
264 * The lock ordering between these three locks is dtrace_meta_lock before
265 * dtrace_provider_lock before dtrace_lock. (In particular, there are
266 * several places where dtrace_provider_lock is held by the framework as it
267 * calls into the providers -- which then call back into the framework,
268 * grabbing dtrace_lock.)
270 * There are two other locks in the mix: mod_lock and cpu_lock. With respect
271 * to dtrace_provider_lock and dtrace_lock, cpu_lock continues its historical
272 * role as a coarse-grained lock; it is acquired before both of these locks.
273 * With respect to dtrace_meta_lock, its behavior is stranger: cpu_lock must
274 * be acquired _between_ dtrace_meta_lock and any other DTrace locks.
275 * mod_lock is similar with respect to dtrace_provider_lock in that it must be
276 * acquired _between_ dtrace_provider_lock and dtrace_lock.
278 static kmutex_t dtrace_lock; /* probe state lock */
279 static kmutex_t dtrace_provider_lock; /* provider state lock */
280 static kmutex_t dtrace_meta_lock; /* meta-provider state lock */
283 /* XXX FreeBSD hacks. */
284 #define cr_suid cr_svuid
285 #define cr_sgid cr_svgid
286 #define ipaddr_t in_addr_t
287 #define mod_modname pathname
288 #define vuprintf vprintf
289 #define ttoproc(_a) ((_a)->td_proc)
290 #define crgetzoneid(_a) 0
293 #define CPU_ON_INTR(_a) 0
295 #define PRIV_EFFECTIVE (1 << 0)
296 #define PRIV_DTRACE_KERNEL (1 << 1)
297 #define PRIV_DTRACE_PROC (1 << 2)
298 #define PRIV_DTRACE_USER (1 << 3)
299 #define PRIV_PROC_OWNER (1 << 4)
300 #define PRIV_PROC_ZONE (1 << 5)
303 SYSCTL_NODE(_debug, OID_AUTO, dtrace, CTLFLAG_RD, 0, "DTrace Information");
307 #define curcpu CPU->cpu_id
312 * DTrace Provider Variables
314 * These are the variables relating to DTrace as a provider (that is, the
315 * provider of the BEGIN, END, and ERROR probes).
317 static dtrace_pattr_t dtrace_provider_attr = {
318 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
319 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
320 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
321 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
322 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
329 static dtrace_pops_t dtrace_provider_ops = {
330 (void (*)(void *, dtrace_probedesc_t *))dtrace_nullop,
331 (void (*)(void *, modctl_t *))dtrace_nullop,
332 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
333 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
334 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
335 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
339 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop
342 static dtrace_id_t dtrace_probeid_begin; /* special BEGIN probe */
343 static dtrace_id_t dtrace_probeid_end; /* special END probe */
344 dtrace_id_t dtrace_probeid_error; /* special ERROR probe */
347 * DTrace Helper Tracing Variables
349 uint32_t dtrace_helptrace_next = 0;
350 uint32_t dtrace_helptrace_nlocals;
351 char *dtrace_helptrace_buffer;
352 int dtrace_helptrace_bufsize = 512 * 1024;
355 int dtrace_helptrace_enabled = 1;
357 int dtrace_helptrace_enabled = 0;
361 * DTrace Error Hashing
363 * On DEBUG kernels, DTrace will track the errors that has seen in a hash
364 * table. This is very useful for checking coverage of tests that are
365 * expected to induce DIF or DOF processing errors, and may be useful for
366 * debugging problems in the DIF code generator or in DOF generation . The
367 * error hash may be examined with the ::dtrace_errhash MDB dcmd.
370 static dtrace_errhash_t dtrace_errhash[DTRACE_ERRHASHSZ];
371 static const char *dtrace_errlast;
372 static kthread_t *dtrace_errthread;
373 static kmutex_t dtrace_errlock;
377 * DTrace Macros and Constants
379 * These are various macros that are useful in various spots in the
380 * implementation, along with a few random constants that have no meaning
381 * outside of the implementation. There is no real structure to this cpp
382 * mishmash -- but is there ever?
384 #define DTRACE_HASHSTR(hash, probe) \
385 dtrace_hash_str(*((char **)((uintptr_t)(probe) + (hash)->dth_stroffs)))
387 #define DTRACE_HASHNEXT(hash, probe) \
388 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_nextoffs)
390 #define DTRACE_HASHPREV(hash, probe) \
391 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_prevoffs)
393 #define DTRACE_HASHEQ(hash, lhs, rhs) \
394 (strcmp(*((char **)((uintptr_t)(lhs) + (hash)->dth_stroffs)), \
395 *((char **)((uintptr_t)(rhs) + (hash)->dth_stroffs))) == 0)
397 #define DTRACE_AGGHASHSIZE_SLEW 17
399 #define DTRACE_V4MAPPED_OFFSET (sizeof (uint32_t) * 3)
402 * The key for a thread-local variable consists of the lower 61 bits of the
403 * t_did, plus the 3 bits of the highest active interrupt above LOCK_LEVEL.
404 * We add DIF_VARIABLE_MAX to t_did to assure that the thread key is never
405 * equal to a variable identifier. This is necessary (but not sufficient) to
406 * assure that global associative arrays never collide with thread-local
407 * variables. To guarantee that they cannot collide, we must also define the
408 * order for keying dynamic variables. That order is:
410 * [ key0 ] ... [ keyn ] [ variable-key ] [ tls-key ]
412 * Because the variable-key and the tls-key are in orthogonal spaces, there is
413 * no way for a global variable key signature to match a thread-local key
417 #define DTRACE_TLS_THRKEY(where) { \
419 uint_t actv = CPU->cpu_intr_actv >> (LOCK_LEVEL + 1); \
420 for (; actv; actv >>= 1) \
422 ASSERT(intr < (1 << 3)); \
423 (where) = ((curthread->t_did + DIF_VARIABLE_MAX) & \
424 (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
427 #define DTRACE_TLS_THRKEY(where) { \
428 solaris_cpu_t *_c = &solaris_cpu[curcpu]; \
430 uint_t actv = _c->cpu_intr_actv; \
431 for (; actv; actv >>= 1) \
433 ASSERT(intr < (1 << 3)); \
434 (where) = ((curthread->td_tid + DIF_VARIABLE_MAX) & \
435 (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
439 #define DT_BSWAP_8(x) ((x) & 0xff)
440 #define DT_BSWAP_16(x) ((DT_BSWAP_8(x) << 8) | DT_BSWAP_8((x) >> 8))
441 #define DT_BSWAP_32(x) ((DT_BSWAP_16(x) << 16) | DT_BSWAP_16((x) >> 16))
442 #define DT_BSWAP_64(x) ((DT_BSWAP_32(x) << 32) | DT_BSWAP_32((x) >> 32))
444 #define DT_MASK_LO 0x00000000FFFFFFFFULL
446 #define DTRACE_STORE(type, tomax, offset, what) \
447 *((type *)((uintptr_t)(tomax) + (uintptr_t)offset)) = (type)(what);
450 #define DTRACE_ALIGNCHECK(addr, size, flags) \
451 if (addr & (size - 1)) { \
452 *flags |= CPU_DTRACE_BADALIGN; \
453 cpu_core[curcpu].cpuc_dtrace_illval = addr; \
457 #define DTRACE_ALIGNCHECK(addr, size, flags)
461 * Test whether a range of memory starting at testaddr of size testsz falls
462 * within the range of memory described by addr, sz. We take care to avoid
463 * problems with overflow and underflow of the unsigned quantities, and
464 * disallow all negative sizes. Ranges of size 0 are allowed.
466 #define DTRACE_INRANGE(testaddr, testsz, baseaddr, basesz) \
467 ((testaddr) - (baseaddr) < (basesz) && \
468 (testaddr) + (testsz) - (baseaddr) <= (basesz) && \
469 (testaddr) + (testsz) >= (testaddr))
472 * Test whether alloc_sz bytes will fit in the scratch region. We isolate
473 * alloc_sz on the righthand side of the comparison in order to avoid overflow
474 * or underflow in the comparison with it. This is simpler than the INRANGE
475 * check above, because we know that the dtms_scratch_ptr is valid in the
476 * range. Allocations of size zero are allowed.
478 #define DTRACE_INSCRATCH(mstate, alloc_sz) \
479 ((mstate)->dtms_scratch_base + (mstate)->dtms_scratch_size - \
480 (mstate)->dtms_scratch_ptr >= (alloc_sz))
482 #define DTRACE_LOADFUNC(bits) \
485 dtrace_load##bits(uintptr_t addr) \
487 size_t size = bits / NBBY; \
489 uint##bits##_t rval; \
491 volatile uint16_t *flags = (volatile uint16_t *) \
492 &cpu_core[curcpu].cpuc_dtrace_flags; \
494 DTRACE_ALIGNCHECK(addr, size, flags); \
496 for (i = 0; i < dtrace_toxranges; i++) { \
497 if (addr >= dtrace_toxrange[i].dtt_limit) \
500 if (addr + size <= dtrace_toxrange[i].dtt_base) \
504 * This address falls within a toxic region; return 0. \
506 *flags |= CPU_DTRACE_BADADDR; \
507 cpu_core[curcpu].cpuc_dtrace_illval = addr; \
511 *flags |= CPU_DTRACE_NOFAULT; \
513 rval = *((volatile uint##bits##_t *)addr); \
514 *flags &= ~CPU_DTRACE_NOFAULT; \
516 return (!(*flags & CPU_DTRACE_FAULT) ? rval : 0); \
520 #define dtrace_loadptr dtrace_load64
522 #define dtrace_loadptr dtrace_load32
525 #define DTRACE_DYNHASH_FREE 0
526 #define DTRACE_DYNHASH_SINK 1
527 #define DTRACE_DYNHASH_VALID 2
529 #define DTRACE_MATCH_NEXT 0
530 #define DTRACE_MATCH_DONE 1
531 #define DTRACE_ANCHORED(probe) ((probe)->dtpr_func[0] != '\0')
532 #define DTRACE_STATE_ALIGN 64
534 #define DTRACE_FLAGS2FLT(flags) \
535 (((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR : \
536 ((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP : \
537 ((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO : \
538 ((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV : \
539 ((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV : \
540 ((flags) & CPU_DTRACE_TUPOFLOW) ? DTRACEFLT_TUPOFLOW : \
541 ((flags) & CPU_DTRACE_BADALIGN) ? DTRACEFLT_BADALIGN : \
542 ((flags) & CPU_DTRACE_NOSCRATCH) ? DTRACEFLT_NOSCRATCH : \
543 ((flags) & CPU_DTRACE_BADSTACK) ? DTRACEFLT_BADSTACK : \
546 #define DTRACEACT_ISSTRING(act) \
547 ((act)->dta_kind == DTRACEACT_DIFEXPR && \
548 (act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING)
550 /* Function prototype definitions: */
551 static size_t dtrace_strlen(const char *, size_t);
552 static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id);
553 static void dtrace_enabling_provide(dtrace_provider_t *);
554 static int dtrace_enabling_match(dtrace_enabling_t *, int *);
555 static void dtrace_enabling_matchall(void);
556 static void dtrace_enabling_reap(void);
557 static dtrace_state_t *dtrace_anon_grab(void);
558 static uint64_t dtrace_helper(int, dtrace_mstate_t *,
559 dtrace_state_t *, uint64_t, uint64_t);
560 static dtrace_helpers_t *dtrace_helpers_create(proc_t *);
561 static void dtrace_buffer_drop(dtrace_buffer_t *);
562 static int dtrace_buffer_consumed(dtrace_buffer_t *, hrtime_t when);
563 static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t,
564 dtrace_state_t *, dtrace_mstate_t *);
565 static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t,
567 static int dtrace_ecb_create_enable(dtrace_probe_t *, void *);
568 static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);
569 uint16_t dtrace_load16(uintptr_t);
570 uint32_t dtrace_load32(uintptr_t);
571 uint64_t dtrace_load64(uintptr_t);
572 uint8_t dtrace_load8(uintptr_t);
573 void dtrace_dynvar_clean(dtrace_dstate_t *);
574 dtrace_dynvar_t *dtrace_dynvar(dtrace_dstate_t *, uint_t, dtrace_key_t *,
575 size_t, dtrace_dynvar_op_t, dtrace_mstate_t *, dtrace_vstate_t *);
576 uintptr_t dtrace_dif_varstr(uintptr_t, dtrace_state_t *, dtrace_mstate_t *);
579 * DTrace Probe Context Functions
581 * These functions are called from probe context. Because probe context is
582 * any context in which C may be called, arbitrarily locks may be held,
583 * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
584 * As a result, functions called from probe context may only call other DTrace
585 * support functions -- they may not interact at all with the system at large.
586 * (Note that the ASSERT macro is made probe-context safe by redefining it in
587 * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
588 * loads are to be performed from probe context, they _must_ be in terms of
589 * the safe dtrace_load*() variants.
591 * Some functions in this block are not actually called from probe context;
592 * for these functions, there will be a comment above the function reading
593 * "Note: not called from probe context."
596 dtrace_panic(const char *format, ...)
600 va_start(alist, format);
601 dtrace_vpanic(format, alist);
606 dtrace_assfail(const char *a, const char *f, int l)
608 dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);
611 * We just need something here that even the most clever compiler
612 * cannot optimize away.
614 return (a[(uintptr_t)f]);
618 * Atomically increment a specified error counter from probe context.
621 dtrace_error(uint32_t *counter)
624 * Most counters stored to in probe context are per-CPU counters.
625 * However, there are some error conditions that are sufficiently
626 * arcane that they don't merit per-CPU storage. If these counters
627 * are incremented concurrently on different CPUs, scalability will be
628 * adversely affected -- but we don't expect them to be white-hot in a
629 * correctly constructed enabling...
636 if ((nval = oval + 1) == 0) {
638 * If the counter would wrap, set it to 1 -- assuring
639 * that the counter is never zero when we have seen
640 * errors. (The counter must be 32-bits because we
641 * aren't guaranteed a 64-bit compare&swap operation.)
642 * To save this code both the infamy of being fingered
643 * by a priggish news story and the indignity of being
644 * the target of a neo-puritan witch trial, we're
645 * carefully avoiding any colorful description of the
646 * likelihood of this condition -- but suffice it to
647 * say that it is only slightly more likely than the
648 * overflow of predicate cache IDs, as discussed in
649 * dtrace_predicate_create().
653 } while (dtrace_cas32(counter, oval, nval) != oval);
657 * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
658 * uint8_t, a uint16_t, a uint32_t and a uint64_t.
666 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
668 if (dest < mstate->dtms_scratch_base)
671 if (dest + size < dest)
674 if (dest + size > mstate->dtms_scratch_ptr)
681 dtrace_canstore_statvar(uint64_t addr, size_t sz,
682 dtrace_statvar_t **svars, int nsvars)
686 for (i = 0; i < nsvars; i++) {
687 dtrace_statvar_t *svar = svars[i];
689 if (svar == NULL || svar->dtsv_size == 0)
692 if (DTRACE_INRANGE(addr, sz, svar->dtsv_data, svar->dtsv_size))
700 * Check to see if the address is within a memory region to which a store may
701 * be issued. This includes the DTrace scratch areas, and any DTrace variable
702 * region. The caller of dtrace_canstore() is responsible for performing any
703 * alignment checks that are needed before stores are actually executed.
706 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
707 dtrace_vstate_t *vstate)
710 * First, check to see if the address is in scratch space...
712 if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base,
713 mstate->dtms_scratch_size))
717 * Now check to see if it's a dynamic variable. This check will pick
718 * up both thread-local variables and any global dynamically-allocated
721 if (DTRACE_INRANGE(addr, sz, (uintptr_t)vstate->dtvs_dynvars.dtds_base,
722 vstate->dtvs_dynvars.dtds_size)) {
723 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
724 uintptr_t base = (uintptr_t)dstate->dtds_base +
725 (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t));
729 * Before we assume that we can store here, we need to make
730 * sure that it isn't in our metadata -- storing to our
731 * dynamic variable metadata would corrupt our state. For
732 * the range to not include any dynamic variable metadata,
735 * (1) Start above the hash table that is at the base of
736 * the dynamic variable space
738 * (2) Have a starting chunk offset that is beyond the
739 * dtrace_dynvar_t that is at the base of every chunk
741 * (3) Not span a chunk boundary
747 chunkoffs = (addr - base) % dstate->dtds_chunksize;
749 if (chunkoffs < sizeof (dtrace_dynvar_t))
752 if (chunkoffs + sz > dstate->dtds_chunksize)
759 * Finally, check the static local and global variables. These checks
760 * take the longest, so we perform them last.
762 if (dtrace_canstore_statvar(addr, sz,
763 vstate->dtvs_locals, vstate->dtvs_nlocals))
766 if (dtrace_canstore_statvar(addr, sz,
767 vstate->dtvs_globals, vstate->dtvs_nglobals))
775 * Convenience routine to check to see if the address is within a memory
776 * region in which a load may be issued given the user's privilege level;
777 * if not, it sets the appropriate error flags and loads 'addr' into the
778 * illegal value slot.
780 * DTrace subroutines (DIF_SUBR_*) should use this helper to implement
781 * appropriate memory access protection.
784 dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
785 dtrace_vstate_t *vstate)
787 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
790 * If we hold the privilege to read from kernel memory, then
791 * everything is readable.
793 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
797 * You can obviously read that which you can store.
799 if (dtrace_canstore(addr, sz, mstate, vstate))
803 * We're allowed to read from our own string table.
805 if (DTRACE_INRANGE(addr, sz, (uintptr_t)mstate->dtms_difo->dtdo_strtab,
806 mstate->dtms_difo->dtdo_strlen))
809 DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV);
815 * Convenience routine to check to see if a given string is within a memory
816 * region in which a load may be issued given the user's privilege level;
817 * this exists so that we don't need to issue unnecessary dtrace_strlen()
818 * calls in the event that the user has all privileges.
821 dtrace_strcanload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
822 dtrace_vstate_t *vstate)
827 * If we hold the privilege to read from kernel memory, then
828 * everything is readable.
830 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
833 strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr, sz);
834 if (dtrace_canload(addr, strsz, mstate, vstate))
841 * Convenience routine to check to see if a given variable is within a memory
842 * region in which a load may be issued given the user's privilege level.
845 dtrace_vcanload(void *src, dtrace_diftype_t *type, dtrace_mstate_t *mstate,
846 dtrace_vstate_t *vstate)
849 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
852 * If we hold the privilege to read from kernel memory, then
853 * everything is readable.
855 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
858 if (type->dtdt_kind == DIF_TYPE_STRING)
859 sz = dtrace_strlen(src,
860 vstate->dtvs_state->dts_options[DTRACEOPT_STRSIZE]) + 1;
862 sz = type->dtdt_size;
864 return (dtrace_canload((uintptr_t)src, sz, mstate, vstate));
868 * Compare two strings using safe loads.
871 dtrace_strncmp(char *s1, char *s2, size_t limit)
874 volatile uint16_t *flags;
876 if (s1 == s2 || limit == 0)
879 flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
885 c1 = dtrace_load8((uintptr_t)s1++);
891 c2 = dtrace_load8((uintptr_t)s2++);
896 } while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));
902 * Compute strlen(s) for a string using safe memory accesses. The additional
903 * len parameter is used to specify a maximum length to ensure completion.
906 dtrace_strlen(const char *s, size_t lim)
910 for (len = 0; len != lim; len++) {
911 if (dtrace_load8((uintptr_t)s++) == '\0')
919 * Check if an address falls within a toxic region.
922 dtrace_istoxic(uintptr_t kaddr, size_t size)
924 uintptr_t taddr, tsize;
927 for (i = 0; i < dtrace_toxranges; i++) {
928 taddr = dtrace_toxrange[i].dtt_base;
929 tsize = dtrace_toxrange[i].dtt_limit - taddr;
931 if (kaddr - taddr < tsize) {
932 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
933 cpu_core[curcpu].cpuc_dtrace_illval = kaddr;
937 if (taddr - kaddr < size) {
938 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
939 cpu_core[curcpu].cpuc_dtrace_illval = taddr;
948 * Copy src to dst using safe memory accesses. The src is assumed to be unsafe
949 * memory specified by the DIF program. The dst is assumed to be safe memory
950 * that we can store to directly because it is managed by DTrace. As with
951 * standard bcopy, overlapping copies are handled properly.
954 dtrace_bcopy(const void *src, void *dst, size_t len)
958 const uint8_t *s2 = src;
962 *s1++ = dtrace_load8((uintptr_t)s2++);
963 } while (--len != 0);
969 *--s1 = dtrace_load8((uintptr_t)--s2);
970 } while (--len != 0);
976 * Copy src to dst using safe memory accesses, up to either the specified
977 * length, or the point that a nul byte is encountered. The src is assumed to
978 * be unsafe memory specified by the DIF program. The dst is assumed to be
979 * safe memory that we can store to directly because it is managed by DTrace.
980 * Unlike dtrace_bcopy(), overlapping regions are not handled.
983 dtrace_strcpy(const void *src, void *dst, size_t len)
986 uint8_t *s1 = dst, c;
987 const uint8_t *s2 = src;
990 *s1++ = c = dtrace_load8((uintptr_t)s2++);
991 } while (--len != 0 && c != '\0');
996 * Copy src to dst, deriving the size and type from the specified (BYREF)
997 * variable type. The src is assumed to be unsafe memory specified by the DIF
998 * program. The dst is assumed to be DTrace variable memory that is of the
999 * specified type; we assume that we can store to directly.
1002 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type)
1004 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
1006 if (type->dtdt_kind == DIF_TYPE_STRING) {
1007 dtrace_strcpy(src, dst, type->dtdt_size);
1009 dtrace_bcopy(src, dst, type->dtdt_size);
1014 * Compare s1 to s2 using safe memory accesses. The s1 data is assumed to be
1015 * unsafe memory specified by the DIF program. The s2 data is assumed to be
1016 * safe memory that we can access directly because it is managed by DTrace.
1019 dtrace_bcmp(const void *s1, const void *s2, size_t len)
1021 volatile uint16_t *flags;
1023 flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
1028 if (s1 == NULL || s2 == NULL)
1031 if (s1 != s2 && len != 0) {
1032 const uint8_t *ps1 = s1;
1033 const uint8_t *ps2 = s2;
1036 if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
1038 } while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
1044 * Zero the specified region using a simple byte-by-byte loop. Note that this
1045 * is for safe DTrace-managed memory only.
1048 dtrace_bzero(void *dst, size_t len)
1052 for (cp = dst; len != 0; len--)
1057 dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum)
1061 result[0] = addend1[0] + addend2[0];
1062 result[1] = addend1[1] + addend2[1] +
1063 (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0);
1070 * Shift the 128-bit value in a by b. If b is positive, shift left.
1071 * If b is negative, shift right.
1074 dtrace_shift_128(uint64_t *a, int b)
1084 a[0] = a[1] >> (b - 64);
1088 mask = 1LL << (64 - b);
1090 a[0] |= ((a[1] & mask) << (64 - b));
1095 a[1] = a[0] << (b - 64);
1099 mask = a[0] >> (64 - b);
1107 * The basic idea is to break the 2 64-bit values into 4 32-bit values,
1108 * use native multiplication on those, and then re-combine into the
1109 * resulting 128-bit value.
1111 * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) =
1118 dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product)
1120 uint64_t hi1, hi2, lo1, lo2;
1123 hi1 = factor1 >> 32;
1124 hi2 = factor2 >> 32;
1126 lo1 = factor1 & DT_MASK_LO;
1127 lo2 = factor2 & DT_MASK_LO;
1129 product[0] = lo1 * lo2;
1130 product[1] = hi1 * hi2;
1134 dtrace_shift_128(tmp, 32);
1135 dtrace_add_128(product, tmp, product);
1139 dtrace_shift_128(tmp, 32);
1140 dtrace_add_128(product, tmp, product);
1144 * This privilege check should be used by actions and subroutines to
1145 * verify that the user credentials of the process that enabled the
1146 * invoking ECB match the target credentials
1149 dtrace_priv_proc_common_user(dtrace_state_t *state)
1151 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1154 * We should always have a non-NULL state cred here, since if cred
1155 * is null (anonymous tracing), we fast-path bypass this routine.
1157 ASSERT(s_cr != NULL);
1159 if ((cr = CRED()) != NULL &&
1160 s_cr->cr_uid == cr->cr_uid &&
1161 s_cr->cr_uid == cr->cr_ruid &&
1162 s_cr->cr_uid == cr->cr_suid &&
1163 s_cr->cr_gid == cr->cr_gid &&
1164 s_cr->cr_gid == cr->cr_rgid &&
1165 s_cr->cr_gid == cr->cr_sgid)
1172 * This privilege check should be used by actions and subroutines to
1173 * verify that the zone of the process that enabled the invoking ECB
1174 * matches the target credentials
1177 dtrace_priv_proc_common_zone(dtrace_state_t *state)
1180 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1183 * We should always have a non-NULL state cred here, since if cred
1184 * is null (anonymous tracing), we fast-path bypass this routine.
1186 ASSERT(s_cr != NULL);
1188 if ((cr = CRED()) != NULL &&
1189 s_cr->cr_zone == cr->cr_zone)
1199 * This privilege check should be used by actions and subroutines to
1200 * verify that the process has not setuid or changed credentials.
1203 dtrace_priv_proc_common_nocd(void)
1207 if ((proc = ttoproc(curthread)) != NULL &&
1208 !(proc->p_flag & SNOCD))
1215 dtrace_priv_proc_destructive(dtrace_state_t *state)
1217 int action = state->dts_cred.dcr_action;
1219 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
1220 dtrace_priv_proc_common_zone(state) == 0)
1223 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
1224 dtrace_priv_proc_common_user(state) == 0)
1227 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
1228 dtrace_priv_proc_common_nocd() == 0)
1234 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1240 dtrace_priv_proc_control(dtrace_state_t *state)
1242 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
1245 if (dtrace_priv_proc_common_zone(state) &&
1246 dtrace_priv_proc_common_user(state) &&
1247 dtrace_priv_proc_common_nocd())
1250 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1256 dtrace_priv_proc(dtrace_state_t *state)
1258 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC)
1261 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1267 dtrace_priv_kernel(dtrace_state_t *state)
1269 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
1272 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1278 dtrace_priv_kernel_destructive(dtrace_state_t *state)
1280 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
1283 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1289 * Note: not called from probe context. This function is called
1290 * asynchronously (and at a regular interval) from outside of probe context to
1291 * clean the dirty dynamic variable lists on all CPUs. Dynamic variable
1292 * cleaning is explained in detail in <sys/dtrace_impl.h>.
1295 dtrace_dynvar_clean(dtrace_dstate_t *dstate)
1297 dtrace_dynvar_t *dirty;
1298 dtrace_dstate_percpu_t *dcpu;
1301 for (i = 0; i < NCPU; i++) {
1302 dcpu = &dstate->dtds_percpu[i];
1304 ASSERT(dcpu->dtdsc_rinsing == NULL);
1307 * If the dirty list is NULL, there is no dirty work to do.
1309 if (dcpu->dtdsc_dirty == NULL)
1313 * If the clean list is non-NULL, then we're not going to do
1314 * any work for this CPU -- it means that there has not been
1315 * a dtrace_dynvar() allocation on this CPU (or from this CPU)
1316 * since the last time we cleaned house.
1318 if (dcpu->dtdsc_clean != NULL)
1324 * Atomically move the dirty list aside.
1327 dirty = dcpu->dtdsc_dirty;
1330 * Before we zap the dirty list, set the rinsing list.
1331 * (This allows for a potential assertion in
1332 * dtrace_dynvar(): if a free dynamic variable appears
1333 * on a hash chain, either the dirty list or the
1334 * rinsing list for some CPU must be non-NULL.)
1336 dcpu->dtdsc_rinsing = dirty;
1337 dtrace_membar_producer();
1338 } while (dtrace_casptr(&dcpu->dtdsc_dirty,
1339 dirty, NULL) != dirty);
1344 * We have no work to do; we can simply return.
1351 for (i = 0; i < NCPU; i++) {
1352 dcpu = &dstate->dtds_percpu[i];
1354 if (dcpu->dtdsc_rinsing == NULL)
1358 * We are now guaranteed that no hash chain contains a pointer
1359 * into this dirty list; we can make it clean.
1361 ASSERT(dcpu->dtdsc_clean == NULL);
1362 dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
1363 dcpu->dtdsc_rinsing = NULL;
1367 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
1368 * sure that all CPUs have seen all of the dtdsc_clean pointers.
1369 * This prevents a race whereby a CPU incorrectly decides that
1370 * the state should be something other than DTRACE_DSTATE_CLEAN
1371 * after dtrace_dynvar_clean() has completed.
1375 dstate->dtds_state = DTRACE_DSTATE_CLEAN;
1379 * Depending on the value of the op parameter, this function looks-up,
1380 * allocates or deallocates an arbitrarily-keyed dynamic variable. If an
1381 * allocation is requested, this function will return a pointer to a
1382 * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
1383 * variable can be allocated. If NULL is returned, the appropriate counter
1384 * will be incremented.
1387 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
1388 dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op,
1389 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1391 uint64_t hashval = DTRACE_DYNHASH_VALID;
1392 dtrace_dynhash_t *hash = dstate->dtds_hash;
1393 dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
1394 processorid_t me = curcpu, cpu = me;
1395 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
1396 size_t bucket, ksize;
1397 size_t chunksize = dstate->dtds_chunksize;
1398 uintptr_t kdata, lock, nstate;
1404 * Hash the key. As with aggregations, we use Jenkins' "One-at-a-time"
1405 * algorithm. For the by-value portions, we perform the algorithm in
1406 * 16-bit chunks (as opposed to 8-bit chunks). This speeds things up a
1407 * bit, and seems to have only a minute effect on distribution. For
1408 * the by-reference data, we perform "One-at-a-time" iterating (safely)
1409 * over each referenced byte. It's painful to do this, but it's much
1410 * better than pathological hash distribution. The efficacy of the
1411 * hashing algorithm (and a comparison with other algorithms) may be
1412 * found by running the ::dtrace_dynstat MDB dcmd.
1414 for (i = 0; i < nkeys; i++) {
1415 if (key[i].dttk_size == 0) {
1416 uint64_t val = key[i].dttk_value;
1418 hashval += (val >> 48) & 0xffff;
1419 hashval += (hashval << 10);
1420 hashval ^= (hashval >> 6);
1422 hashval += (val >> 32) & 0xffff;
1423 hashval += (hashval << 10);
1424 hashval ^= (hashval >> 6);
1426 hashval += (val >> 16) & 0xffff;
1427 hashval += (hashval << 10);
1428 hashval ^= (hashval >> 6);
1430 hashval += val & 0xffff;
1431 hashval += (hashval << 10);
1432 hashval ^= (hashval >> 6);
1435 * This is incredibly painful, but it beats the hell
1436 * out of the alternative.
1438 uint64_t j, size = key[i].dttk_size;
1439 uintptr_t base = (uintptr_t)key[i].dttk_value;
1441 if (!dtrace_canload(base, size, mstate, vstate))
1444 for (j = 0; j < size; j++) {
1445 hashval += dtrace_load8(base + j);
1446 hashval += (hashval << 10);
1447 hashval ^= (hashval >> 6);
1452 if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT))
1455 hashval += (hashval << 3);
1456 hashval ^= (hashval >> 11);
1457 hashval += (hashval << 15);
1460 * There is a remote chance (ideally, 1 in 2^31) that our hashval
1461 * comes out to be one of our two sentinel hash values. If this
1462 * actually happens, we set the hashval to be a value known to be a
1463 * non-sentinel value.
1465 if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
1466 hashval = DTRACE_DYNHASH_VALID;
1469 * Yes, it's painful to do a divide here. If the cycle count becomes
1470 * important here, tricks can be pulled to reduce it. (However, it's
1471 * critical that hash collisions be kept to an absolute minimum;
1472 * they're much more painful than a divide.) It's better to have a
1473 * solution that generates few collisions and still keeps things
1474 * relatively simple.
1476 bucket = hashval % dstate->dtds_hashsize;
1478 if (op == DTRACE_DYNVAR_DEALLOC) {
1479 volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;
1482 while ((lock = *lockp) & 1)
1485 if (dtrace_casptr((volatile void *)lockp,
1486 (volatile void *)lock, (volatile void *)(lock + 1)) == (void *)lock)
1490 dtrace_membar_producer();
1495 lock = hash[bucket].dtdh_lock;
1497 dtrace_membar_consumer();
1499 start = hash[bucket].dtdh_chain;
1500 ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
1501 start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
1502 op != DTRACE_DYNVAR_DEALLOC));
1504 for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
1505 dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
1506 dtrace_key_t *dkey = &dtuple->dtt_key[0];
1508 if (dvar->dtdv_hashval != hashval) {
1509 if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
1511 * We've reached the sink, and therefore the
1512 * end of the hash chain; we can kick out of
1513 * the loop knowing that we have seen a valid
1514 * snapshot of state.
1516 ASSERT(dvar->dtdv_next == NULL);
1517 ASSERT(dvar == &dtrace_dynhash_sink);
1521 if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
1523 * We've gone off the rails: somewhere along
1524 * the line, one of the members of this hash
1525 * chain was deleted. Note that we could also
1526 * detect this by simply letting this loop run
1527 * to completion, as we would eventually hit
1528 * the end of the dirty list. However, we
1529 * want to avoid running the length of the
1530 * dirty list unnecessarily (it might be quite
1531 * long), so we catch this as early as
1532 * possible by detecting the hash marker. In
1533 * this case, we simply set dvar to NULL and
1534 * break; the conditional after the loop will
1535 * send us back to top.
1544 if (dtuple->dtt_nkeys != nkeys)
1547 for (i = 0; i < nkeys; i++, dkey++) {
1548 if (dkey->dttk_size != key[i].dttk_size)
1549 goto next; /* size or type mismatch */
1551 if (dkey->dttk_size != 0) {
1553 (void *)(uintptr_t)key[i].dttk_value,
1554 (void *)(uintptr_t)dkey->dttk_value,
1558 if (dkey->dttk_value != key[i].dttk_value)
1563 if (op != DTRACE_DYNVAR_DEALLOC)
1566 ASSERT(dvar->dtdv_next == NULL ||
1567 dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);
1570 ASSERT(hash[bucket].dtdh_chain != dvar);
1571 ASSERT(start != dvar);
1572 ASSERT(prev->dtdv_next == dvar);
1573 prev->dtdv_next = dvar->dtdv_next;
1575 if (dtrace_casptr(&hash[bucket].dtdh_chain,
1576 start, dvar->dtdv_next) != start) {
1578 * We have failed to atomically swing the
1579 * hash table head pointer, presumably because
1580 * of a conflicting allocation on another CPU.
1581 * We need to reread the hash chain and try
1588 dtrace_membar_producer();
1591 * Now set the hash value to indicate that it's free.
1593 ASSERT(hash[bucket].dtdh_chain != dvar);
1594 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1596 dtrace_membar_producer();
1599 * Set the next pointer to point at the dirty list, and
1600 * atomically swing the dirty pointer to the newly freed dvar.
1603 next = dcpu->dtdsc_dirty;
1604 dvar->dtdv_next = next;
1605 } while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);
1608 * Finally, unlock this hash bucket.
1610 ASSERT(hash[bucket].dtdh_lock == lock);
1612 hash[bucket].dtdh_lock++;
1622 * If dvar is NULL, it is because we went off the rails:
1623 * one of the elements that we traversed in the hash chain
1624 * was deleted while we were traversing it. In this case,
1625 * we assert that we aren't doing a dealloc (deallocs lock
1626 * the hash bucket to prevent themselves from racing with
1627 * one another), and retry the hash chain traversal.
1629 ASSERT(op != DTRACE_DYNVAR_DEALLOC);
1633 if (op != DTRACE_DYNVAR_ALLOC) {
1635 * If we are not to allocate a new variable, we want to
1636 * return NULL now. Before we return, check that the value
1637 * of the lock word hasn't changed. If it has, we may have
1638 * seen an inconsistent snapshot.
1640 if (op == DTRACE_DYNVAR_NOALLOC) {
1641 if (hash[bucket].dtdh_lock != lock)
1644 ASSERT(op == DTRACE_DYNVAR_DEALLOC);
1645 ASSERT(hash[bucket].dtdh_lock == lock);
1647 hash[bucket].dtdh_lock++;
1654 * We need to allocate a new dynamic variable. The size we need is the
1655 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
1656 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
1657 * the size of any referred-to data (dsize). We then round the final
1658 * size up to the chunksize for allocation.
1660 for (ksize = 0, i = 0; i < nkeys; i++)
1661 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
1664 * This should be pretty much impossible, but could happen if, say,
1665 * strange DIF specified the tuple. Ideally, this should be an
1666 * assertion and not an error condition -- but that requires that the
1667 * chunksize calculation in dtrace_difo_chunksize() be absolutely
1668 * bullet-proof. (That is, it must not be able to be fooled by
1669 * malicious DIF.) Given the lack of backwards branches in DIF,
1670 * solving this would presumably not amount to solving the Halting
1671 * Problem -- but it still seems awfully hard.
1673 if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
1674 ksize + dsize > chunksize) {
1675 dcpu->dtdsc_drops++;
1679 nstate = DTRACE_DSTATE_EMPTY;
1683 free = dcpu->dtdsc_free;
1686 dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
1689 if (clean == NULL) {
1691 * We're out of dynamic variable space on
1692 * this CPU. Unless we have tried all CPUs,
1693 * we'll try to allocate from a different
1696 switch (dstate->dtds_state) {
1697 case DTRACE_DSTATE_CLEAN: {
1698 void *sp = &dstate->dtds_state;
1703 if (dcpu->dtdsc_dirty != NULL &&
1704 nstate == DTRACE_DSTATE_EMPTY)
1705 nstate = DTRACE_DSTATE_DIRTY;
1707 if (dcpu->dtdsc_rinsing != NULL)
1708 nstate = DTRACE_DSTATE_RINSING;
1710 dcpu = &dstate->dtds_percpu[cpu];
1715 (void) dtrace_cas32(sp,
1716 DTRACE_DSTATE_CLEAN, nstate);
1719 * To increment the correct bean
1720 * counter, take another lap.
1725 case DTRACE_DSTATE_DIRTY:
1726 dcpu->dtdsc_dirty_drops++;
1729 case DTRACE_DSTATE_RINSING:
1730 dcpu->dtdsc_rinsing_drops++;
1733 case DTRACE_DSTATE_EMPTY:
1734 dcpu->dtdsc_drops++;
1738 DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
1743 * The clean list appears to be non-empty. We want to
1744 * move the clean list to the free list; we start by
1745 * moving the clean pointer aside.
1747 if (dtrace_casptr(&dcpu->dtdsc_clean,
1748 clean, NULL) != clean) {
1750 * We are in one of two situations:
1752 * (a) The clean list was switched to the
1753 * free list by another CPU.
1755 * (b) The clean list was added to by the
1758 * In either of these situations, we can
1759 * just reattempt the free list allocation.
1764 ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);
1767 * Now we'll move the clean list to the free list.
1768 * It's impossible for this to fail: the only way
1769 * the free list can be updated is through this
1770 * code path, and only one CPU can own the clean list.
1771 * Thus, it would only be possible for this to fail if
1772 * this code were racing with dtrace_dynvar_clean().
1773 * (That is, if dtrace_dynvar_clean() updated the clean
1774 * list, and we ended up racing to update the free
1775 * list.) This race is prevented by the dtrace_sync()
1776 * in dtrace_dynvar_clean() -- which flushes the
1777 * owners of the clean lists out before resetting
1780 rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
1781 ASSERT(rval == NULL);
1786 new_free = dvar->dtdv_next;
1787 } while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);
1790 * We have now allocated a new chunk. We copy the tuple keys into the
1791 * tuple array and copy any referenced key data into the data space
1792 * following the tuple array. As we do this, we relocate dttk_value
1793 * in the final tuple to point to the key data address in the chunk.
1795 kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
1796 dvar->dtdv_data = (void *)(kdata + ksize);
1797 dvar->dtdv_tuple.dtt_nkeys = nkeys;
1799 for (i = 0; i < nkeys; i++) {
1800 dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
1801 size_t kesize = key[i].dttk_size;
1805 (const void *)(uintptr_t)key[i].dttk_value,
1806 (void *)kdata, kesize);
1807 dkey->dttk_value = kdata;
1808 kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
1810 dkey->dttk_value = key[i].dttk_value;
1813 dkey->dttk_size = kesize;
1816 ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
1817 dvar->dtdv_hashval = hashval;
1818 dvar->dtdv_next = start;
1820 if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
1824 * The cas has failed. Either another CPU is adding an element to
1825 * this hash chain, or another CPU is deleting an element from this
1826 * hash chain. The simplest way to deal with both of these cases
1827 * (though not necessarily the most efficient) is to free our
1828 * allocated block and tail-call ourselves. Note that the free is
1829 * to the dirty list and _not_ to the free list. This is to prevent
1830 * races with allocators, above.
1832 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1834 dtrace_membar_producer();
1837 free = dcpu->dtdsc_dirty;
1838 dvar->dtdv_next = free;
1839 } while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);
1841 return (dtrace_dynvar(dstate, nkeys, key, dsize, op, mstate, vstate));
1846 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
1848 if ((int64_t)nval < (int64_t)*oval)
1854 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
1856 if ((int64_t)nval > (int64_t)*oval)
1861 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
1863 int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
1864 int64_t val = (int64_t)nval;
1867 for (i = 0; i < zero; i++) {
1868 if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
1874 for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
1875 if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
1876 quanta[i - 1] += incr;
1881 quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
1889 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
1891 uint64_t arg = *lquanta++;
1892 int32_t base = DTRACE_LQUANTIZE_BASE(arg);
1893 uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
1894 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
1895 int32_t val = (int32_t)nval, level;
1898 ASSERT(levels != 0);
1902 * This is an underflow.
1908 level = (val - base) / step;
1910 if (level < levels) {
1911 lquanta[level + 1] += incr;
1916 * This is an overflow.
1918 lquanta[levels + 1] += incr;
1922 dtrace_aggregate_llquantize_bucket(uint16_t factor, uint16_t low,
1923 uint16_t high, uint16_t nsteps, int64_t value)
1925 int64_t this = 1, last, next;
1926 int base = 1, order;
1928 ASSERT(factor <= nsteps);
1929 ASSERT(nsteps % factor == 0);
1931 for (order = 0; order < low; order++)
1935 * If our value is less than our factor taken to the power of the
1936 * low order of magnitude, it goes into the zeroth bucket.
1938 if (value < (last = this))
1941 for (this *= factor; order <= high; order++) {
1942 int nbuckets = this > nsteps ? nsteps : this;
1944 if ((next = this * factor) < this) {
1946 * We should not generally get log/linear quantizations
1947 * with a high magnitude that allows 64-bits to
1948 * overflow, but we nonetheless protect against this
1949 * by explicitly checking for overflow, and clamping
1950 * our value accordingly.
1957 * If our value lies within this order of magnitude,
1958 * determine its position by taking the offset within
1959 * the order of magnitude, dividing by the bucket
1960 * width, and adding to our (accumulated) base.
1962 return (base + (value - last) / (this / nbuckets));
1965 base += nbuckets - (nbuckets / factor);
1971 * Our value is greater than or equal to our factor taken to the
1972 * power of one plus the high magnitude -- return the top bucket.
1978 dtrace_aggregate_llquantize(uint64_t *llquanta, uint64_t nval, uint64_t incr)
1980 uint64_t arg = *llquanta++;
1981 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(arg);
1982 uint16_t low = DTRACE_LLQUANTIZE_LOW(arg);
1983 uint16_t high = DTRACE_LLQUANTIZE_HIGH(arg);
1984 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(arg);
1986 llquanta[dtrace_aggregate_llquantize_bucket(factor,
1987 low, high, nsteps, nval)] += incr;
1992 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
2000 dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg)
2002 int64_t snval = (int64_t)nval;
2009 * What we want to say here is:
2011 * data[2] += nval * nval;
2013 * But given that nval is 64-bit, we could easily overflow, so
2014 * we do this as 128-bit arithmetic.
2019 dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp);
2020 dtrace_add_128(data + 2, tmp, data + 2);
2025 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
2032 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
2038 * Aggregate given the tuple in the principal data buffer, and the aggregating
2039 * action denoted by the specified dtrace_aggregation_t. The aggregation
2040 * buffer is specified as the buf parameter. This routine does not return
2041 * failure; if there is no space in the aggregation buffer, the data will be
2042 * dropped, and a corresponding counter incremented.
2045 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
2046 intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
2048 dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
2049 uint32_t i, ndx, size, fsize;
2050 uint32_t align = sizeof (uint64_t) - 1;
2051 dtrace_aggbuffer_t *agb;
2052 dtrace_aggkey_t *key;
2053 uint32_t hashval = 0, limit, isstr;
2054 caddr_t tomax, data, kdata;
2055 dtrace_actkind_t action;
2056 dtrace_action_t *act;
2062 if (!agg->dtag_hasarg) {
2064 * Currently, only quantize() and lquantize() take additional
2065 * arguments, and they have the same semantics: an increment
2066 * value that defaults to 1 when not present. If additional
2067 * aggregating actions take arguments, the setting of the
2068 * default argument value will presumably have to become more
2074 action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
2075 size = rec->dtrd_offset - agg->dtag_base;
2076 fsize = size + rec->dtrd_size;
2078 ASSERT(dbuf->dtb_tomax != NULL);
2079 data = dbuf->dtb_tomax + offset + agg->dtag_base;
2081 if ((tomax = buf->dtb_tomax) == NULL) {
2082 dtrace_buffer_drop(buf);
2087 * The metastructure is always at the bottom of the buffer.
2089 agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
2090 sizeof (dtrace_aggbuffer_t));
2092 if (buf->dtb_offset == 0) {
2094 * We just kludge up approximately 1/8th of the size to be
2095 * buckets. If this guess ends up being routinely
2096 * off-the-mark, we may need to dynamically readjust this
2097 * based on past performance.
2099 uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);
2101 if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
2102 (uintptr_t)tomax || hashsize == 0) {
2104 * We've been given a ludicrously small buffer;
2105 * increment our drop count and leave.
2107 dtrace_buffer_drop(buf);
2112 * And now, a pathetic attempt to try to get a an odd (or
2113 * perchance, a prime) hash size for better hash distribution.
2115 if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
2116 hashsize -= DTRACE_AGGHASHSIZE_SLEW;
2118 agb->dtagb_hashsize = hashsize;
2119 agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
2120 agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
2121 agb->dtagb_free = (uintptr_t)agb->dtagb_hash;
2123 for (i = 0; i < agb->dtagb_hashsize; i++)
2124 agb->dtagb_hash[i] = NULL;
2127 ASSERT(agg->dtag_first != NULL);
2128 ASSERT(agg->dtag_first->dta_intuple);
2131 * Calculate the hash value based on the key. Note that we _don't_
2132 * include the aggid in the hashing (but we will store it as part of
2133 * the key). The hashing algorithm is Bob Jenkins' "One-at-a-time"
2134 * algorithm: a simple, quick algorithm that has no known funnels, and
2135 * gets good distribution in practice. The efficacy of the hashing
2136 * algorithm (and a comparison with other algorithms) may be found by
2137 * running the ::dtrace_aggstat MDB dcmd.
2139 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2140 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2141 limit = i + act->dta_rec.dtrd_size;
2142 ASSERT(limit <= size);
2143 isstr = DTRACEACT_ISSTRING(act);
2145 for (; i < limit; i++) {
2147 hashval += (hashval << 10);
2148 hashval ^= (hashval >> 6);
2150 if (isstr && data[i] == '\0')
2155 hashval += (hashval << 3);
2156 hashval ^= (hashval >> 11);
2157 hashval += (hashval << 15);
2160 * Yes, the divide here is expensive -- but it's generally the least
2161 * of the performance issues given the amount of data that we iterate
2162 * over to compute hash values, compare data, etc.
2164 ndx = hashval % agb->dtagb_hashsize;
2166 for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
2167 ASSERT((caddr_t)key >= tomax);
2168 ASSERT((caddr_t)key < tomax + buf->dtb_size);
2170 if (hashval != key->dtak_hashval || key->dtak_size != size)
2173 kdata = key->dtak_data;
2174 ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);
2176 for (act = agg->dtag_first; act->dta_intuple;
2177 act = act->dta_next) {
2178 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2179 limit = i + act->dta_rec.dtrd_size;
2180 ASSERT(limit <= size);
2181 isstr = DTRACEACT_ISSTRING(act);
2183 for (; i < limit; i++) {
2184 if (kdata[i] != data[i])
2187 if (isstr && data[i] == '\0')
2192 if (action != key->dtak_action) {
2194 * We are aggregating on the same value in the same
2195 * aggregation with two different aggregating actions.
2196 * (This should have been picked up in the compiler,
2197 * so we may be dealing with errant or devious DIF.)
2198 * This is an error condition; we indicate as much,
2201 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2206 * This is a hit: we need to apply the aggregator to
2207 * the value at this key.
2209 agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
2216 * We didn't find it. We need to allocate some zero-filled space,
2217 * link it into the hash table appropriately, and apply the aggregator
2218 * to the (zero-filled) value.
2220 offs = buf->dtb_offset;
2221 while (offs & (align - 1))
2222 offs += sizeof (uint32_t);
2225 * If we don't have enough room to both allocate a new key _and_
2226 * its associated data, increment the drop count and return.
2228 if ((uintptr_t)tomax + offs + fsize >
2229 agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
2230 dtrace_buffer_drop(buf);
2235 ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
2236 key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
2237 agb->dtagb_free -= sizeof (dtrace_aggkey_t);
2239 key->dtak_data = kdata = tomax + offs;
2240 buf->dtb_offset = offs + fsize;
2243 * Now copy the data across.
2245 *((dtrace_aggid_t *)kdata) = agg->dtag_id;
2247 for (i = sizeof (dtrace_aggid_t); i < size; i++)
2251 * Because strings are not zeroed out by default, we need to iterate
2252 * looking for actions that store strings, and we need to explicitly
2253 * pad these strings out with zeroes.
2255 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2258 if (!DTRACEACT_ISSTRING(act))
2261 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2262 limit = i + act->dta_rec.dtrd_size;
2263 ASSERT(limit <= size);
2265 for (nul = 0; i < limit; i++) {
2271 if (data[i] != '\0')
2278 for (i = size; i < fsize; i++)
2281 key->dtak_hashval = hashval;
2282 key->dtak_size = size;
2283 key->dtak_action = action;
2284 key->dtak_next = agb->dtagb_hash[ndx];
2285 agb->dtagb_hash[ndx] = key;
2288 * Finally, apply the aggregator.
2290 *((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
2291 agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
2295 * Given consumer state, this routine finds a speculation in the INACTIVE
2296 * state and transitions it into the ACTIVE state. If there is no speculation
2297 * in the INACTIVE state, 0 is returned. In this case, no error counter is
2298 * incremented -- it is up to the caller to take appropriate action.
2301 dtrace_speculation(dtrace_state_t *state)
2304 dtrace_speculation_state_t current;
2305 uint32_t *stat = &state->dts_speculations_unavail, count;
2307 while (i < state->dts_nspeculations) {
2308 dtrace_speculation_t *spec = &state->dts_speculations[i];
2310 current = spec->dtsp_state;
2312 if (current != DTRACESPEC_INACTIVE) {
2313 if (current == DTRACESPEC_COMMITTINGMANY ||
2314 current == DTRACESPEC_COMMITTING ||
2315 current == DTRACESPEC_DISCARDING)
2316 stat = &state->dts_speculations_busy;
2321 if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2322 current, DTRACESPEC_ACTIVE) == current)
2327 * We couldn't find a speculation. If we found as much as a single
2328 * busy speculation buffer, we'll attribute this failure as "busy"
2329 * instead of "unavail".
2333 } while (dtrace_cas32(stat, count, count + 1) != count);
2339 * This routine commits an active speculation. If the specified speculation
2340 * is not in a valid state to perform a commit(), this routine will silently do
2341 * nothing. The state of the specified speculation is transitioned according
2342 * to the state transition diagram outlined in <sys/dtrace_impl.h>
2345 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
2346 dtrace_specid_t which)
2348 dtrace_speculation_t *spec;
2349 dtrace_buffer_t *src, *dest;
2350 uintptr_t daddr, saddr, dlimit, slimit;
2351 dtrace_speculation_state_t current, new = 0;
2358 if (which > state->dts_nspeculations) {
2359 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2363 spec = &state->dts_speculations[which - 1];
2364 src = &spec->dtsp_buffer[cpu];
2365 dest = &state->dts_buffer[cpu];
2368 current = spec->dtsp_state;
2370 if (current == DTRACESPEC_COMMITTINGMANY)
2374 case DTRACESPEC_INACTIVE:
2375 case DTRACESPEC_DISCARDING:
2378 case DTRACESPEC_COMMITTING:
2380 * This is only possible if we are (a) commit()'ing
2381 * without having done a prior speculate() on this CPU
2382 * and (b) racing with another commit() on a different
2383 * CPU. There's nothing to do -- we just assert that
2386 ASSERT(src->dtb_offset == 0);
2389 case DTRACESPEC_ACTIVE:
2390 new = DTRACESPEC_COMMITTING;
2393 case DTRACESPEC_ACTIVEONE:
2395 * This speculation is active on one CPU. If our
2396 * buffer offset is non-zero, we know that the one CPU
2397 * must be us. Otherwise, we are committing on a
2398 * different CPU from the speculate(), and we must
2399 * rely on being asynchronously cleaned.
2401 if (src->dtb_offset != 0) {
2402 new = DTRACESPEC_COMMITTING;
2407 case DTRACESPEC_ACTIVEMANY:
2408 new = DTRACESPEC_COMMITTINGMANY;
2414 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2415 current, new) != current);
2418 * We have set the state to indicate that we are committing this
2419 * speculation. Now reserve the necessary space in the destination
2422 if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
2423 sizeof (uint64_t), state, NULL)) < 0) {
2424 dtrace_buffer_drop(dest);
2429 * We have sufficient space to copy the speculative buffer into the
2430 * primary buffer. First, modify the speculative buffer, filling
2431 * in the timestamp of all entries with the current time. The data
2432 * must have the commit() time rather than the time it was traced,
2433 * so that all entries in the primary buffer are in timestamp order.
2435 timestamp = dtrace_gethrtime();
2436 saddr = (uintptr_t)src->dtb_tomax;
2437 slimit = saddr + src->dtb_offset;
2438 while (saddr < slimit) {
2440 dtrace_rechdr_t *dtrh = (dtrace_rechdr_t *)saddr;
2442 if (dtrh->dtrh_epid == DTRACE_EPIDNONE) {
2443 saddr += sizeof (dtrace_epid_t);
2446 ASSERT3U(dtrh->dtrh_epid, <=, state->dts_necbs);
2447 size = state->dts_ecbs[dtrh->dtrh_epid - 1]->dte_size;
2449 ASSERT3U(saddr + size, <=, slimit);
2450 ASSERT3U(size, >=, sizeof (dtrace_rechdr_t));
2451 ASSERT3U(DTRACE_RECORD_LOAD_TIMESTAMP(dtrh), ==, UINT64_MAX);
2453 DTRACE_RECORD_STORE_TIMESTAMP(dtrh, timestamp);
2459 * Copy the buffer across. (Note that this is a
2460 * highly subobtimal bcopy(); in the unlikely event that this becomes
2461 * a serious performance issue, a high-performance DTrace-specific
2462 * bcopy() should obviously be invented.)
2464 daddr = (uintptr_t)dest->dtb_tomax + offs;
2465 dlimit = daddr + src->dtb_offset;
2466 saddr = (uintptr_t)src->dtb_tomax;
2469 * First, the aligned portion.
2471 while (dlimit - daddr >= sizeof (uint64_t)) {
2472 *((uint64_t *)daddr) = *((uint64_t *)saddr);
2474 daddr += sizeof (uint64_t);
2475 saddr += sizeof (uint64_t);
2479 * Now any left-over bit...
2481 while (dlimit - daddr)
2482 *((uint8_t *)daddr++) = *((uint8_t *)saddr++);
2485 * Finally, commit the reserved space in the destination buffer.
2487 dest->dtb_offset = offs + src->dtb_offset;
2491 * If we're lucky enough to be the only active CPU on this speculation
2492 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
2494 if (current == DTRACESPEC_ACTIVE ||
2495 (current == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
2496 uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
2497 DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);
2499 ASSERT(rval == DTRACESPEC_COMMITTING);
2502 src->dtb_offset = 0;
2503 src->dtb_xamot_drops += src->dtb_drops;
2508 * This routine discards an active speculation. If the specified speculation
2509 * is not in a valid state to perform a discard(), this routine will silently
2510 * do nothing. The state of the specified speculation is transitioned
2511 * according to the state transition diagram outlined in <sys/dtrace_impl.h>
2514 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
2515 dtrace_specid_t which)
2517 dtrace_speculation_t *spec;
2518 dtrace_speculation_state_t current, new = 0;
2519 dtrace_buffer_t *buf;
2524 if (which > state->dts_nspeculations) {
2525 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2529 spec = &state->dts_speculations[which - 1];
2530 buf = &spec->dtsp_buffer[cpu];
2533 current = spec->dtsp_state;
2536 case DTRACESPEC_INACTIVE:
2537 case DTRACESPEC_COMMITTINGMANY:
2538 case DTRACESPEC_COMMITTING:
2539 case DTRACESPEC_DISCARDING:
2542 case DTRACESPEC_ACTIVE:
2543 case DTRACESPEC_ACTIVEMANY:
2544 new = DTRACESPEC_DISCARDING;
2547 case DTRACESPEC_ACTIVEONE:
2548 if (buf->dtb_offset != 0) {
2549 new = DTRACESPEC_INACTIVE;
2551 new = DTRACESPEC_DISCARDING;
2558 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2559 current, new) != current);
2561 buf->dtb_offset = 0;
2566 * Note: not called from probe context. This function is called
2567 * asynchronously from cross call context to clean any speculations that are
2568 * in the COMMITTINGMANY or DISCARDING states. These speculations may not be
2569 * transitioned back to the INACTIVE state until all CPUs have cleaned the
2573 dtrace_speculation_clean_here(dtrace_state_t *state)
2575 dtrace_icookie_t cookie;
2576 processorid_t cpu = curcpu;
2577 dtrace_buffer_t *dest = &state->dts_buffer[cpu];
2580 cookie = dtrace_interrupt_disable();
2582 if (dest->dtb_tomax == NULL) {
2583 dtrace_interrupt_enable(cookie);
2587 for (i = 0; i < state->dts_nspeculations; i++) {
2588 dtrace_speculation_t *spec = &state->dts_speculations[i];
2589 dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];
2591 if (src->dtb_tomax == NULL)
2594 if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
2595 src->dtb_offset = 0;
2599 if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2602 if (src->dtb_offset == 0)
2605 dtrace_speculation_commit(state, cpu, i + 1);
2608 dtrace_interrupt_enable(cookie);
2612 * Note: not called from probe context. This function is called
2613 * asynchronously (and at a regular interval) to clean any speculations that
2614 * are in the COMMITTINGMANY or DISCARDING states. If it discovers that there
2615 * is work to be done, it cross calls all CPUs to perform that work;
2616 * COMMITMANY and DISCARDING speculations may not be transitioned back to the
2617 * INACTIVE state until they have been cleaned by all CPUs.
2620 dtrace_speculation_clean(dtrace_state_t *state)
2625 for (i = 0; i < state->dts_nspeculations; i++) {
2626 dtrace_speculation_t *spec = &state->dts_speculations[i];
2628 ASSERT(!spec->dtsp_cleaning);
2630 if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
2631 spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2635 spec->dtsp_cleaning = 1;
2641 dtrace_xcall(DTRACE_CPUALL,
2642 (dtrace_xcall_t)dtrace_speculation_clean_here, state);
2645 * We now know that all CPUs have committed or discarded their
2646 * speculation buffers, as appropriate. We can now set the state
2649 for (i = 0; i < state->dts_nspeculations; i++) {
2650 dtrace_speculation_t *spec = &state->dts_speculations[i];
2651 dtrace_speculation_state_t current, new;
2653 if (!spec->dtsp_cleaning)
2656 current = spec->dtsp_state;
2657 ASSERT(current == DTRACESPEC_DISCARDING ||
2658 current == DTRACESPEC_COMMITTINGMANY);
2660 new = DTRACESPEC_INACTIVE;
2662 rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, current, new);
2663 ASSERT(rv == current);
2664 spec->dtsp_cleaning = 0;
2669 * Called as part of a speculate() to get the speculative buffer associated
2670 * with a given speculation. Returns NULL if the specified speculation is not
2671 * in an ACTIVE state. If the speculation is in the ACTIVEONE state -- and
2672 * the active CPU is not the specified CPU -- the speculation will be
2673 * atomically transitioned into the ACTIVEMANY state.
2675 static dtrace_buffer_t *
2676 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
2677 dtrace_specid_t which)
2679 dtrace_speculation_t *spec;
2680 dtrace_speculation_state_t current, new = 0;
2681 dtrace_buffer_t *buf;
2686 if (which > state->dts_nspeculations) {
2687 cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2691 spec = &state->dts_speculations[which - 1];
2692 buf = &spec->dtsp_buffer[cpuid];
2695 current = spec->dtsp_state;
2698 case DTRACESPEC_INACTIVE:
2699 case DTRACESPEC_COMMITTINGMANY:
2700 case DTRACESPEC_DISCARDING:
2703 case DTRACESPEC_COMMITTING:
2704 ASSERT(buf->dtb_offset == 0);
2707 case DTRACESPEC_ACTIVEONE:
2709 * This speculation is currently active on one CPU.
2710 * Check the offset in the buffer; if it's non-zero,
2711 * that CPU must be us (and we leave the state alone).
2712 * If it's zero, assume that we're starting on a new
2713 * CPU -- and change the state to indicate that the
2714 * speculation is active on more than one CPU.
2716 if (buf->dtb_offset != 0)
2719 new = DTRACESPEC_ACTIVEMANY;
2722 case DTRACESPEC_ACTIVEMANY:
2725 case DTRACESPEC_ACTIVE:
2726 new = DTRACESPEC_ACTIVEONE;
2732 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2733 current, new) != current);
2735 ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
2740 * Return a string. In the event that the user lacks the privilege to access
2741 * arbitrary kernel memory, we copy the string out to scratch memory so that we
2742 * don't fail access checking.
2744 * dtrace_dif_variable() uses this routine as a helper for various
2745 * builtin values such as 'execname' and 'probefunc.'
2748 dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state,
2749 dtrace_mstate_t *mstate)
2751 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
2756 * The easy case: this probe is allowed to read all of memory, so
2757 * we can just return this as a vanilla pointer.
2759 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
2763 * This is the tougher case: we copy the string in question from
2764 * kernel memory into scratch memory and return it that way: this
2765 * ensures that we won't trip up when access checking tests the
2766 * BYREF return value.
2768 strsz = dtrace_strlen((char *)addr, size) + 1;
2770 if (mstate->dtms_scratch_ptr + strsz >
2771 mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2772 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2776 dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2778 ret = mstate->dtms_scratch_ptr;
2779 mstate->dtms_scratch_ptr += strsz;
2784 * Return a string from a memoy address which is known to have one or
2785 * more concatenated, individually zero terminated, sub-strings.
2786 * In the event that the user lacks the privilege to access
2787 * arbitrary kernel memory, we copy the string out to scratch memory so that we
2788 * don't fail access checking.
2790 * dtrace_dif_variable() uses this routine as a helper for various
2791 * builtin values such as 'execargs'.
2794 dtrace_dif_varstrz(uintptr_t addr, size_t strsz, dtrace_state_t *state,
2795 dtrace_mstate_t *mstate)
2801 if (mstate->dtms_scratch_ptr + strsz >
2802 mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2803 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2807 dtrace_bcopy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2810 /* Replace sub-string termination characters with a space. */
2811 for (p = (char *) mstate->dtms_scratch_ptr, i = 0; i < strsz - 1;
2816 ret = mstate->dtms_scratch_ptr;
2817 mstate->dtms_scratch_ptr += strsz;
2822 * This function implements the DIF emulator's variable lookups. The emulator
2823 * passes a reserved variable identifier and optional built-in array index.
2826 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
2830 * If we're accessing one of the uncached arguments, we'll turn this
2831 * into a reference in the args array.
2833 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
2834 ndx = v - DIF_VAR_ARG0;
2840 ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
2841 if (ndx >= sizeof (mstate->dtms_arg) /
2842 sizeof (mstate->dtms_arg[0])) {
2843 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2844 dtrace_provider_t *pv;
2847 pv = mstate->dtms_probe->dtpr_provider;
2848 if (pv->dtpv_pops.dtps_getargval != NULL)
2849 val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
2850 mstate->dtms_probe->dtpr_id,
2851 mstate->dtms_probe->dtpr_arg, ndx, aframes);
2853 val = dtrace_getarg(ndx, aframes);
2856 * This is regrettably required to keep the compiler
2857 * from tail-optimizing the call to dtrace_getarg().
2858 * The condition always evaluates to true, but the
2859 * compiler has no way of figuring that out a priori.
2860 * (None of this would be necessary if the compiler
2861 * could be relied upon to _always_ tail-optimize
2862 * the call to dtrace_getarg() -- but it can't.)
2864 if (mstate->dtms_probe != NULL)
2870 return (mstate->dtms_arg[ndx]);
2873 case DIF_VAR_UREGS: {
2876 if (!dtrace_priv_proc(state))
2879 if ((lwp = curthread->t_lwp) == NULL) {
2880 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
2881 cpu_core[curcpu].cpuc_dtrace_illval = NULL;
2885 return (dtrace_getreg(lwp->lwp_regs, ndx));
2889 case DIF_VAR_UREGS: {
2890 struct trapframe *tframe;
2892 if (!dtrace_priv_proc(state))
2895 if ((tframe = curthread->td_frame) == NULL) {
2896 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
2897 cpu_core[curcpu].cpuc_dtrace_illval = 0;
2901 return (dtrace_getreg(tframe, ndx));
2905 case DIF_VAR_CURTHREAD:
2906 if (!dtrace_priv_kernel(state))
2908 return ((uint64_t)(uintptr_t)curthread);
2910 case DIF_VAR_TIMESTAMP:
2911 if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
2912 mstate->dtms_timestamp = dtrace_gethrtime();
2913 mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
2915 return (mstate->dtms_timestamp);
2917 case DIF_VAR_VTIMESTAMP:
2918 ASSERT(dtrace_vtime_references != 0);
2919 return (curthread->t_dtrace_vtime);
2921 case DIF_VAR_WALLTIMESTAMP:
2922 if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
2923 mstate->dtms_walltimestamp = dtrace_gethrestime();
2924 mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
2926 return (mstate->dtms_walltimestamp);
2930 if (!dtrace_priv_kernel(state))
2932 if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
2933 mstate->dtms_ipl = dtrace_getipl();
2934 mstate->dtms_present |= DTRACE_MSTATE_IPL;
2936 return (mstate->dtms_ipl);
2940 ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
2941 return (mstate->dtms_epid);
2944 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2945 return (mstate->dtms_probe->dtpr_id);
2947 case DIF_VAR_STACKDEPTH:
2948 if (!dtrace_priv_kernel(state))
2950 if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
2951 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2953 mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
2954 mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
2956 return (mstate->dtms_stackdepth);
2958 case DIF_VAR_USTACKDEPTH:
2959 if (!dtrace_priv_proc(state))
2961 if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
2963 * See comment in DIF_VAR_PID.
2965 if (DTRACE_ANCHORED(mstate->dtms_probe) &&
2967 mstate->dtms_ustackdepth = 0;
2969 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2970 mstate->dtms_ustackdepth =
2971 dtrace_getustackdepth();
2972 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2974 mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
2976 return (mstate->dtms_ustackdepth);
2978 case DIF_VAR_CALLER:
2979 if (!dtrace_priv_kernel(state))
2981 if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
2982 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2984 if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
2986 * If this is an unanchored probe, we are
2987 * required to go through the slow path:
2988 * dtrace_caller() only guarantees correct
2989 * results for anchored probes.
2991 pc_t caller[2] = {0, 0};
2993 dtrace_getpcstack(caller, 2, aframes,
2994 (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
2995 mstate->dtms_caller = caller[1];
2996 } else if ((mstate->dtms_caller =
2997 dtrace_caller(aframes)) == -1) {
2999 * We have failed to do this the quick way;
3000 * we must resort to the slower approach of
3001 * calling dtrace_getpcstack().
3005 dtrace_getpcstack(&caller, 1, aframes, NULL);
3006 mstate->dtms_caller = caller;
3009 mstate->dtms_present |= DTRACE_MSTATE_CALLER;
3011 return (mstate->dtms_caller);
3013 case DIF_VAR_UCALLER:
3014 if (!dtrace_priv_proc(state))
3017 if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
3021 * dtrace_getupcstack() fills in the first uint64_t
3022 * with the current PID. The second uint64_t will
3023 * be the program counter at user-level. The third
3024 * uint64_t will contain the caller, which is what
3028 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3029 dtrace_getupcstack(ustack, 3);
3030 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3031 mstate->dtms_ucaller = ustack[2];
3032 mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
3035 return (mstate->dtms_ucaller);
3037 case DIF_VAR_PROBEPROV:
3038 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3039 return (dtrace_dif_varstr(
3040 (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name,
3043 case DIF_VAR_PROBEMOD:
3044 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3045 return (dtrace_dif_varstr(
3046 (uintptr_t)mstate->dtms_probe->dtpr_mod,
3049 case DIF_VAR_PROBEFUNC:
3050 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3051 return (dtrace_dif_varstr(
3052 (uintptr_t)mstate->dtms_probe->dtpr_func,
3055 case DIF_VAR_PROBENAME:
3056 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3057 return (dtrace_dif_varstr(
3058 (uintptr_t)mstate->dtms_probe->dtpr_name,
3062 if (!dtrace_priv_proc(state))
3067 * Note that we are assuming that an unanchored probe is
3068 * always due to a high-level interrupt. (And we're assuming
3069 * that there is only a single high level interrupt.)
3071 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3072 return (pid0.pid_id);
3075 * It is always safe to dereference one's own t_procp pointer:
3076 * it always points to a valid, allocated proc structure.
3077 * Further, it is always safe to dereference the p_pidp member
3078 * of one's own proc structure. (These are truisms becuase
3079 * threads and processes don't clean up their own state --
3080 * they leave that task to whomever reaps them.)
3082 return ((uint64_t)curthread->t_procp->p_pidp->pid_id);
3084 return ((uint64_t)curproc->p_pid);
3088 if (!dtrace_priv_proc(state))
3093 * See comment in DIF_VAR_PID.
3095 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3096 return (pid0.pid_id);
3099 * It is always safe to dereference one's own t_procp pointer:
3100 * it always points to a valid, allocated proc structure.
3101 * (This is true because threads don't clean up their own
3102 * state -- they leave that task to whomever reaps them.)
3104 return ((uint64_t)curthread->t_procp->p_ppid);
3106 return ((uint64_t)curproc->p_pptr->p_pid);
3112 * See comment in DIF_VAR_PID.
3114 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3118 return ((uint64_t)curthread->t_tid);
3120 case DIF_VAR_EXECARGS: {
3121 struct pargs *p_args = curthread->td_proc->p_args;
3126 return (dtrace_dif_varstrz(
3127 (uintptr_t) p_args->ar_args, p_args->ar_length, state, mstate));
3130 case DIF_VAR_EXECNAME:
3132 if (!dtrace_priv_proc(state))
3136 * See comment in DIF_VAR_PID.
3138 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3139 return ((uint64_t)(uintptr_t)p0.p_user.u_comm);
3142 * It is always safe to dereference one's own t_procp pointer:
3143 * it always points to a valid, allocated proc structure.
3144 * (This is true because threads don't clean up their own
3145 * state -- they leave that task to whomever reaps them.)
3147 return (dtrace_dif_varstr(
3148 (uintptr_t)curthread->t_procp->p_user.u_comm,
3151 return (dtrace_dif_varstr(
3152 (uintptr_t) curthread->td_proc->p_comm, state, mstate));
3155 case DIF_VAR_ZONENAME:
3157 if (!dtrace_priv_proc(state))
3161 * See comment in DIF_VAR_PID.
3163 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3164 return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);
3167 * It is always safe to dereference one's own t_procp pointer:
3168 * it always points to a valid, allocated proc structure.
3169 * (This is true because threads don't clean up their own
3170 * state -- they leave that task to whomever reaps them.)
3172 return (dtrace_dif_varstr(
3173 (uintptr_t)curthread->t_procp->p_zone->zone_name,
3180 if (!dtrace_priv_proc(state))
3185 * See comment in DIF_VAR_PID.
3187 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3188 return ((uint64_t)p0.p_cred->cr_uid);
3192 * It is always safe to dereference one's own t_procp pointer:
3193 * it always points to a valid, allocated proc structure.
3194 * (This is true because threads don't clean up their own
3195 * state -- they leave that task to whomever reaps them.)
3197 * Additionally, it is safe to dereference one's own process
3198 * credential, since this is never NULL after process birth.
3200 return ((uint64_t)curthread->t_procp->p_cred->cr_uid);
3203 if (!dtrace_priv_proc(state))
3208 * See comment in DIF_VAR_PID.
3210 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3211 return ((uint64_t)p0.p_cred->cr_gid);
3215 * It is always safe to dereference one's own t_procp pointer:
3216 * it always points to a valid, allocated proc structure.
3217 * (This is true because threads don't clean up their own
3218 * state -- they leave that task to whomever reaps them.)
3220 * Additionally, it is safe to dereference one's own process
3221 * credential, since this is never NULL after process birth.
3223 return ((uint64_t)curthread->t_procp->p_cred->cr_gid);
3225 case DIF_VAR_ERRNO: {
3228 if (!dtrace_priv_proc(state))
3232 * See comment in DIF_VAR_PID.
3234 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3238 * It is always safe to dereference one's own t_lwp pointer in
3239 * the event that this pointer is non-NULL. (This is true
3240 * because threads and lwps don't clean up their own state --
3241 * they leave that task to whomever reaps them.)
3243 if ((lwp = curthread->t_lwp) == NULL)
3246 return ((uint64_t)lwp->lwp_errno);
3248 return (curthread->td_errno);
3257 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
3263 * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
3264 * Notice that we don't bother validating the proper number of arguments or
3265 * their types in the tuple stack. This isn't needed because all argument
3266 * interpretation is safe because of our load safety -- the worst that can
3267 * happen is that a bogus program can obtain bogus results.
3270 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
3271 dtrace_key_t *tupregs, int nargs,
3272 dtrace_mstate_t *mstate, dtrace_state_t *state)
3274 volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
3275 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
3276 dtrace_vstate_t *vstate = &state->dts_vstate;
3289 struct thread *lowner;
3291 struct lock_object *li;
3298 regs[rd] = (dtrace_gethrtime() * 2416 + 374441) % 1771875;
3302 case DIF_SUBR_MUTEX_OWNED:
3303 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3309 m.mx = dtrace_load64(tupregs[0].dttk_value);
3310 if (MUTEX_TYPE_ADAPTIVE(&m.mi))
3311 regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
3313 regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
3316 case DIF_SUBR_MUTEX_OWNER:
3317 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3323 m.mx = dtrace_load64(tupregs[0].dttk_value);
3324 if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
3325 MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
3326 regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
3331 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3332 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3338 m.mx = dtrace_load64(tupregs[0].dttk_value);
3339 regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
3342 case DIF_SUBR_MUTEX_TYPE_SPIN:
3343 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3349 m.mx = dtrace_load64(tupregs[0].dttk_value);
3350 regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
3353 case DIF_SUBR_RW_READ_HELD: {
3356 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3362 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3363 regs[rd] = _RW_READ_HELD(&r.ri, tmp);
3367 case DIF_SUBR_RW_WRITE_HELD:
3368 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3374 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3375 regs[rd] = _RW_WRITE_HELD(&r.ri);
3378 case DIF_SUBR_RW_ISWRITER:
3379 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3385 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3386 regs[rd] = _RW_ISWRITER(&r.ri);
3390 case DIF_SUBR_MUTEX_OWNED:
3391 if (!dtrace_canload(tupregs[0].dttk_value,
3392 sizeof (struct lock_object), mstate, vstate)) {
3396 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3397 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3400 case DIF_SUBR_MUTEX_OWNER:
3401 if (!dtrace_canload(tupregs[0].dttk_value,
3402 sizeof (struct lock_object), mstate, vstate)) {
3406 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3407 LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3408 regs[rd] = (uintptr_t)lowner;
3411 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3412 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
3417 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3418 /* XXX - should be only LC_SLEEPABLE? */
3419 regs[rd] = (LOCK_CLASS(l.li)->lc_flags &
3420 (LC_SLEEPLOCK | LC_SLEEPABLE)) != 0;
3423 case DIF_SUBR_MUTEX_TYPE_SPIN:
3424 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
3429 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3430 regs[rd] = (LOCK_CLASS(l.li)->lc_flags & LC_SPINLOCK) != 0;
3433 case DIF_SUBR_RW_READ_HELD:
3434 case DIF_SUBR_SX_SHARED_HELD:
3435 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3440 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3441 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
3445 case DIF_SUBR_RW_WRITE_HELD:
3446 case DIF_SUBR_SX_EXCLUSIVE_HELD:
3447 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3452 l.lx = dtrace_loadptr(tupregs[0].dttk_value);
3453 LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3454 regs[rd] = (lowner == curthread);
3457 case DIF_SUBR_RW_ISWRITER:
3458 case DIF_SUBR_SX_ISEXCLUSIVE:
3459 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3464 l.lx = dtrace_loadptr(tupregs[0].dttk_value);
3465 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
3468 #endif /* ! defined(sun) */
3470 case DIF_SUBR_BCOPY: {
3472 * We need to be sure that the destination is in the scratch
3473 * region -- no other region is allowed.
3475 uintptr_t src = tupregs[0].dttk_value;
3476 uintptr_t dest = tupregs[1].dttk_value;
3477 size_t size = tupregs[2].dttk_value;
3479 if (!dtrace_inscratch(dest, size, mstate)) {
3480 *flags |= CPU_DTRACE_BADADDR;
3485 if (!dtrace_canload(src, size, mstate, vstate)) {
3490 dtrace_bcopy((void *)src, (void *)dest, size);
3494 case DIF_SUBR_ALLOCA:
3495 case DIF_SUBR_COPYIN: {
3496 uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
3498 tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
3499 size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;
3502 * This action doesn't require any credential checks since
3503 * probes will not activate in user contexts to which the
3504 * enabling user does not have permissions.
3508 * Rounding up the user allocation size could have overflowed
3509 * a large, bogus allocation (like -1ULL) to 0.
3511 if (scratch_size < size ||
3512 !DTRACE_INSCRATCH(mstate, scratch_size)) {
3513 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3518 if (subr == DIF_SUBR_COPYIN) {
3519 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3520 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3521 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3524 mstate->dtms_scratch_ptr += scratch_size;
3529 case DIF_SUBR_COPYINTO: {
3530 uint64_t size = tupregs[1].dttk_value;
3531 uintptr_t dest = tupregs[2].dttk_value;
3534 * This action doesn't require any credential checks since
3535 * probes will not activate in user contexts to which the
3536 * enabling user does not have permissions.
3538 if (!dtrace_inscratch(dest, size, mstate)) {
3539 *flags |= CPU_DTRACE_BADADDR;
3544 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3545 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3546 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3550 case DIF_SUBR_COPYINSTR: {
3551 uintptr_t dest = mstate->dtms_scratch_ptr;
3552 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3554 if (nargs > 1 && tupregs[1].dttk_value < size)
3555 size = tupregs[1].dttk_value + 1;
3558 * This action doesn't require any credential checks since
3559 * probes will not activate in user contexts to which the
3560 * enabling user does not have permissions.
3562 if (!DTRACE_INSCRATCH(mstate, size)) {
3563 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3568 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3569 dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags);
3570 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3572 ((char *)dest)[size - 1] = '\0';
3573 mstate->dtms_scratch_ptr += size;
3579 case DIF_SUBR_MSGSIZE:
3580 case DIF_SUBR_MSGDSIZE: {
3581 uintptr_t baddr = tupregs[0].dttk_value, daddr;
3582 uintptr_t wptr, rptr;
3586 while (baddr != 0 && !(*flags & CPU_DTRACE_FAULT)) {
3588 if (!dtrace_canload(baddr, sizeof (mblk_t), mstate,
3594 wptr = dtrace_loadptr(baddr +
3595 offsetof(mblk_t, b_wptr));
3597 rptr = dtrace_loadptr(baddr +
3598 offsetof(mblk_t, b_rptr));
3601 *flags |= CPU_DTRACE_BADADDR;
3602 *illval = tupregs[0].dttk_value;
3606 daddr = dtrace_loadptr(baddr +
3607 offsetof(mblk_t, b_datap));
3609 baddr = dtrace_loadptr(baddr +
3610 offsetof(mblk_t, b_cont));
3613 * We want to prevent against denial-of-service here,
3614 * so we're only going to search the list for
3615 * dtrace_msgdsize_max mblks.
3617 if (cont++ > dtrace_msgdsize_max) {
3618 *flags |= CPU_DTRACE_ILLOP;
3622 if (subr == DIF_SUBR_MSGDSIZE) {
3623 if (dtrace_load8(daddr +
3624 offsetof(dblk_t, db_type)) != M_DATA)
3628 count += wptr - rptr;
3631 if (!(*flags & CPU_DTRACE_FAULT))
3638 case DIF_SUBR_PROGENYOF: {
3639 pid_t pid = tupregs[0].dttk_value;
3643 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3645 for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
3647 if (p->p_pidp->pid_id == pid) {
3649 if (p->p_pid == pid) {
3656 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3662 case DIF_SUBR_SPECULATION:
3663 regs[rd] = dtrace_speculation(state);
3666 case DIF_SUBR_COPYOUT: {
3667 uintptr_t kaddr = tupregs[0].dttk_value;
3668 uintptr_t uaddr = tupregs[1].dttk_value;
3669 uint64_t size = tupregs[2].dttk_value;
3671 if (!dtrace_destructive_disallow &&
3672 dtrace_priv_proc_control(state) &&
3673 !dtrace_istoxic(kaddr, size)) {
3674 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3675 dtrace_copyout(kaddr, uaddr, size, flags);
3676 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3681 case DIF_SUBR_COPYOUTSTR: {
3682 uintptr_t kaddr = tupregs[0].dttk_value;
3683 uintptr_t uaddr = tupregs[1].dttk_value;
3684 uint64_t size = tupregs[2].dttk_value;
3686 if (!dtrace_destructive_disallow &&
3687 dtrace_priv_proc_control(state) &&
3688 !dtrace_istoxic(kaddr, size)) {
3689 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3690 dtrace_copyoutstr(kaddr, uaddr, size, flags);
3691 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3696 case DIF_SUBR_STRLEN: {
3698 uintptr_t addr = (uintptr_t)tupregs[0].dttk_value;
3699 sz = dtrace_strlen((char *)addr,
3700 state->dts_options[DTRACEOPT_STRSIZE]);
3702 if (!dtrace_canload(addr, sz + 1, mstate, vstate)) {
3712 case DIF_SUBR_STRCHR:
3713 case DIF_SUBR_STRRCHR: {
3715 * We're going to iterate over the string looking for the
3716 * specified character. We will iterate until we have reached
3717 * the string length or we have found the character. If this
3718 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
3719 * of the specified character instead of the first.
3721 uintptr_t saddr = tupregs[0].dttk_value;
3722 uintptr_t addr = tupregs[0].dttk_value;
3723 uintptr_t limit = addr + state->dts_options[DTRACEOPT_STRSIZE];
3724 char c, target = (char)tupregs[1].dttk_value;
3726 for (regs[rd] = 0; addr < limit; addr++) {
3727 if ((c = dtrace_load8(addr)) == target) {
3730 if (subr == DIF_SUBR_STRCHR)
3738 if (!dtrace_canload(saddr, addr - saddr, mstate, vstate)) {
3746 case DIF_SUBR_STRSTR:
3747 case DIF_SUBR_INDEX:
3748 case DIF_SUBR_RINDEX: {
3750 * We're going to iterate over the string looking for the
3751 * specified string. We will iterate until we have reached
3752 * the string length or we have found the string. (Yes, this
3753 * is done in the most naive way possible -- but considering
3754 * that the string we're searching for is likely to be
3755 * relatively short, the complexity of Rabin-Karp or similar
3756 * hardly seems merited.)
3758 char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
3759 char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
3760 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3761 size_t len = dtrace_strlen(addr, size);
3762 size_t sublen = dtrace_strlen(substr, size);
3763 char *limit = addr + len, *orig = addr;
3764 int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
3767 regs[rd] = notfound;
3769 if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) {
3774 if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate,
3781 * strstr() and index()/rindex() have similar semantics if
3782 * both strings are the empty string: strstr() returns a
3783 * pointer to the (empty) string, and index() and rindex()
3784 * both return index 0 (regardless of any position argument).
3786 if (sublen == 0 && len == 0) {
3787 if (subr == DIF_SUBR_STRSTR)
3788 regs[rd] = (uintptr_t)addr;
3794 if (subr != DIF_SUBR_STRSTR) {
3795 if (subr == DIF_SUBR_RINDEX) {
3802 * Both index() and rindex() take an optional position
3803 * argument that denotes the starting position.
3806 int64_t pos = (int64_t)tupregs[2].dttk_value;
3809 * If the position argument to index() is
3810 * negative, Perl implicitly clamps it at
3811 * zero. This semantic is a little surprising
3812 * given the special meaning of negative
3813 * positions to similar Perl functions like
3814 * substr(), but it appears to reflect a
3815 * notion that index() can start from a
3816 * negative index and increment its way up to
3817 * the string. Given this notion, Perl's
3818 * rindex() is at least self-consistent in
3819 * that it implicitly clamps positions greater
3820 * than the string length to be the string
3821 * length. Where Perl completely loses
3822 * coherence, however, is when the specified
3823 * substring is the empty string (""). In
3824 * this case, even if the position is
3825 * negative, rindex() returns 0 -- and even if
3826 * the position is greater than the length,
3827 * index() returns the string length. These
3828 * semantics violate the notion that index()
3829 * should never return a value less than the
3830 * specified position and that rindex() should
3831 * never return a value greater than the
3832 * specified position. (One assumes that
3833 * these semantics are artifacts of Perl's
3834 * implementation and not the results of
3835 * deliberate design -- it beggars belief that
3836 * even Larry Wall could desire such oddness.)
3837 * While in the abstract one would wish for
3838 * consistent position semantics across
3839 * substr(), index() and rindex() -- or at the
3840 * very least self-consistent position
3841 * semantics for index() and rindex() -- we
3842 * instead opt to keep with the extant Perl
3843 * semantics, in all their broken glory. (Do
3844 * we have more desire to maintain Perl's
3845 * semantics than Perl does? Probably.)
3847 if (subr == DIF_SUBR_RINDEX) {
3871 for (regs[rd] = notfound; addr != limit; addr += inc) {
3872 if (dtrace_strncmp(addr, substr, sublen) == 0) {
3873 if (subr != DIF_SUBR_STRSTR) {
3875 * As D index() and rindex() are
3876 * modeled on Perl (and not on awk),
3877 * we return a zero-based (and not a
3878 * one-based) index. (For you Perl
3879 * weenies: no, we're not going to add
3880 * $[ -- and shouldn't you be at a con
3883 regs[rd] = (uintptr_t)(addr - orig);
3887 ASSERT(subr == DIF_SUBR_STRSTR);
3888 regs[rd] = (uintptr_t)addr;
3896 case DIF_SUBR_STRTOK: {
3897 uintptr_t addr = tupregs[0].dttk_value;
3898 uintptr_t tokaddr = tupregs[1].dttk_value;
3899 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3900 uintptr_t limit, toklimit = tokaddr + size;
3901 uint8_t c = 0, tokmap[32]; /* 256 / 8 */
3902 char *dest = (char *)mstate->dtms_scratch_ptr;
3906 * Check both the token buffer and (later) the input buffer,
3907 * since both could be non-scratch addresses.
3909 if (!dtrace_strcanload(tokaddr, size, mstate, vstate)) {
3914 if (!DTRACE_INSCRATCH(mstate, size)) {
3915 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3922 * If the address specified is NULL, we use our saved
3923 * strtok pointer from the mstate. Note that this
3924 * means that the saved strtok pointer is _only_
3925 * valid within multiple enablings of the same probe --
3926 * it behaves like an implicit clause-local variable.
3928 addr = mstate->dtms_strtok;
3931 * If the user-specified address is non-NULL we must
3932 * access check it. This is the only time we have
3933 * a chance to do so, since this address may reside
3934 * in the string table of this clause-- future calls
3935 * (when we fetch addr from mstate->dtms_strtok)
3936 * would fail this access check.
3938 if (!dtrace_strcanload(addr, size, mstate, vstate)) {
3945 * First, zero the token map, and then process the token
3946 * string -- setting a bit in the map for every character
3947 * found in the token string.
3949 for (i = 0; i < sizeof (tokmap); i++)
3952 for (; tokaddr < toklimit; tokaddr++) {
3953 if ((c = dtrace_load8(tokaddr)) == '\0')
3956 ASSERT((c >> 3) < sizeof (tokmap));
3957 tokmap[c >> 3] |= (1 << (c & 0x7));
3960 for (limit = addr + size; addr < limit; addr++) {
3962 * We're looking for a character that is _not_ contained
3963 * in the token string.
3965 if ((c = dtrace_load8(addr)) == '\0')
3968 if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
3974 * We reached the end of the string without finding
3975 * any character that was not in the token string.
3976 * We return NULL in this case, and we set the saved
3977 * address to NULL as well.
3980 mstate->dtms_strtok = 0;
3985 * From here on, we're copying into the destination string.
3987 for (i = 0; addr < limit && i < size - 1; addr++) {
3988 if ((c = dtrace_load8(addr)) == '\0')
3991 if (tokmap[c >> 3] & (1 << (c & 0x7)))
4000 regs[rd] = (uintptr_t)dest;
4001 mstate->dtms_scratch_ptr += size;
4002 mstate->dtms_strtok = addr;
4006 case DIF_SUBR_SUBSTR: {
4007 uintptr_t s = tupregs[0].dttk_value;
4008 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4009 char *d = (char *)mstate->dtms_scratch_ptr;
4010 int64_t index = (int64_t)tupregs[1].dttk_value;
4011 int64_t remaining = (int64_t)tupregs[2].dttk_value;
4012 size_t len = dtrace_strlen((char *)s, size);
4015 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4020 if (!DTRACE_INSCRATCH(mstate, size)) {
4021 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4027 remaining = (int64_t)size;
4032 if (index < 0 && index + remaining > 0) {
4038 if (index >= len || index < 0) {
4040 } else if (remaining < 0) {
4041 remaining += len - index;
4042 } else if (index + remaining > size) {
4043 remaining = size - index;
4046 for (i = 0; i < remaining; i++) {
4047 if ((d[i] = dtrace_load8(s + index + i)) == '\0')
4053 mstate->dtms_scratch_ptr += size;
4054 regs[rd] = (uintptr_t)d;
4058 case DIF_SUBR_TOUPPER:
4059 case DIF_SUBR_TOLOWER: {
4060 uintptr_t s = tupregs[0].dttk_value;
4061 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4062 char *dest = (char *)mstate->dtms_scratch_ptr, c;
4063 size_t len = dtrace_strlen((char *)s, size);
4064 char lower, upper, convert;
4067 if (subr == DIF_SUBR_TOUPPER) {
4077 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4082 if (!DTRACE_INSCRATCH(mstate, size)) {
4083 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4088 for (i = 0; i < size - 1; i++) {
4089 if ((c = dtrace_load8(s + i)) == '\0')
4092 if (c >= lower && c <= upper)
4093 c = convert + (c - lower);
4100 regs[rd] = (uintptr_t)dest;
4101 mstate->dtms_scratch_ptr += size;
4106 case DIF_SUBR_GETMAJOR:
4108 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
4110 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
4114 case DIF_SUBR_GETMINOR:
4116 regs[rd] = tupregs[0].dttk_value & MAXMIN64;
4118 regs[rd] = tupregs[0].dttk_value & MAXMIN;
4122 case DIF_SUBR_DDI_PATHNAME: {
4124 * This one is a galactic mess. We are going to roughly
4125 * emulate ddi_pathname(), but it's made more complicated
4126 * by the fact that we (a) want to include the minor name and
4127 * (b) must proceed iteratively instead of recursively.
4129 uintptr_t dest = mstate->dtms_scratch_ptr;
4130 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4131 char *start = (char *)dest, *end = start + size - 1;
4132 uintptr_t daddr = tupregs[0].dttk_value;
4133 int64_t minor = (int64_t)tupregs[1].dttk_value;
4135 int i, len, depth = 0;
4138 * Due to all the pointer jumping we do and context we must
4139 * rely upon, we just mandate that the user must have kernel
4140 * read privileges to use this routine.
4142 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
4143 *flags |= CPU_DTRACE_KPRIV;
4148 if (!DTRACE_INSCRATCH(mstate, size)) {
4149 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4157 * We want to have a name for the minor. In order to do this,
4158 * we need to walk the minor list from the devinfo. We want
4159 * to be sure that we don't infinitely walk a circular list,
4160 * so we check for circularity by sending a scout pointer
4161 * ahead two elements for every element that we iterate over;
4162 * if the list is circular, these will ultimately point to the
4163 * same element. You may recognize this little trick as the
4164 * answer to a stupid interview question -- one that always
4165 * seems to be asked by those who had to have it laboriously
4166 * explained to them, and who can't even concisely describe
4167 * the conditions under which one would be forced to resort to
4168 * this technique. Needless to say, those conditions are
4169 * found here -- and probably only here. Is this the only use
4170 * of this infamous trick in shipping, production code? If it
4171 * isn't, it probably should be...
4174 uintptr_t maddr = dtrace_loadptr(daddr +
4175 offsetof(struct dev_info, devi_minor));
4177 uintptr_t next = offsetof(struct ddi_minor_data, next);
4178 uintptr_t name = offsetof(struct ddi_minor_data,
4179 d_minor) + offsetof(struct ddi_minor, name);
4180 uintptr_t dev = offsetof(struct ddi_minor_data,
4181 d_minor) + offsetof(struct ddi_minor, dev);
4185 scout = dtrace_loadptr(maddr + next);
4187 while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4190 m = dtrace_load64(maddr + dev) & MAXMIN64;
4192 m = dtrace_load32(maddr + dev) & MAXMIN;
4195 maddr = dtrace_loadptr(maddr + next);
4200 scout = dtrace_loadptr(scout + next);
4205 scout = dtrace_loadptr(scout + next);
4210 if (scout == maddr) {
4211 *flags |= CPU_DTRACE_ILLOP;
4219 * We have the minor data. Now we need to
4220 * copy the minor's name into the end of the
4223 s = (char *)dtrace_loadptr(maddr + name);
4224 len = dtrace_strlen(s, size);
4226 if (*flags & CPU_DTRACE_FAULT)
4230 if ((end -= (len + 1)) < start)
4236 for (i = 1; i <= len; i++)
4237 end[i] = dtrace_load8((uintptr_t)s++);
4242 while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4243 ddi_node_state_t devi_state;
4245 devi_state = dtrace_load32(daddr +
4246 offsetof(struct dev_info, devi_node_state));
4248 if (*flags & CPU_DTRACE_FAULT)
4251 if (devi_state >= DS_INITIALIZED) {
4252 s = (char *)dtrace_loadptr(daddr +
4253 offsetof(struct dev_info, devi_addr));
4254 len = dtrace_strlen(s, size);
4256 if (*flags & CPU_DTRACE_FAULT)
4260 if ((end -= (len + 1)) < start)
4266 for (i = 1; i <= len; i++)
4267 end[i] = dtrace_load8((uintptr_t)s++);
4271 * Now for the node name...
4273 s = (char *)dtrace_loadptr(daddr +
4274 offsetof(struct dev_info, devi_node_name));
4276 daddr = dtrace_loadptr(daddr +
4277 offsetof(struct dev_info, devi_parent));
4280 * If our parent is NULL (that is, if we're the root
4281 * node), we're going to use the special path
4287 len = dtrace_strlen(s, size);
4288 if (*flags & CPU_DTRACE_FAULT)
4291 if ((end -= (len + 1)) < start)
4294 for (i = 1; i <= len; i++)
4295 end[i] = dtrace_load8((uintptr_t)s++);
4298 if (depth++ > dtrace_devdepth_max) {
4299 *flags |= CPU_DTRACE_ILLOP;
4305 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4308 regs[rd] = (uintptr_t)end;
4309 mstate->dtms_scratch_ptr += size;
4316 case DIF_SUBR_STRJOIN: {
4317 char *d = (char *)mstate->dtms_scratch_ptr;
4318 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4319 uintptr_t s1 = tupregs[0].dttk_value;
4320 uintptr_t s2 = tupregs[1].dttk_value;
4323 if (!dtrace_strcanload(s1, size, mstate, vstate) ||
4324 !dtrace_strcanload(s2, size, mstate, vstate)) {
4329 if (!DTRACE_INSCRATCH(mstate, size)) {
4330 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4337 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4342 if ((d[i++] = dtrace_load8(s1++)) == '\0') {
4350 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4355 if ((d[i++] = dtrace_load8(s2++)) == '\0')
4360 mstate->dtms_scratch_ptr += i;
4361 regs[rd] = (uintptr_t)d;
4367 case DIF_SUBR_LLTOSTR: {
4368 int64_t i = (int64_t)tupregs[0].dttk_value;
4369 uint64_t val, digit;
4370 uint64_t size = 65; /* enough room for 2^64 in binary */
4371 char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
4375 if ((base = tupregs[1].dttk_value) <= 1 ||
4376 base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
4377 *flags |= CPU_DTRACE_ILLOP;
4382 val = (base == 10 && i < 0) ? i * -1 : i;
4384 if (!DTRACE_INSCRATCH(mstate, size)) {
4385 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4390 for (*end-- = '\0'; val; val /= base) {
4391 if ((digit = val % base) <= '9' - '0') {
4392 *end-- = '0' + digit;
4394 *end-- = 'a' + (digit - ('9' - '0') - 1);
4398 if (i == 0 && base == 16)
4404 if (i == 0 || base == 8 || base == 16)
4407 if (i < 0 && base == 10)
4410 regs[rd] = (uintptr_t)end + 1;
4411 mstate->dtms_scratch_ptr += size;
4415 case DIF_SUBR_HTONS:
4416 case DIF_SUBR_NTOHS:
4417 #if BYTE_ORDER == BIG_ENDIAN
4418 regs[rd] = (uint16_t)tupregs[0].dttk_value;
4420 regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
4425 case DIF_SUBR_HTONL:
4426 case DIF_SUBR_NTOHL:
4427 #if BYTE_ORDER == BIG_ENDIAN
4428 regs[rd] = (uint32_t)tupregs[0].dttk_value;
4430 regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
4435 case DIF_SUBR_HTONLL:
4436 case DIF_SUBR_NTOHLL:
4437 #if BYTE_ORDER == BIG_ENDIAN
4438 regs[rd] = (uint64_t)tupregs[0].dttk_value;
4440 regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
4445 case DIF_SUBR_DIRNAME:
4446 case DIF_SUBR_BASENAME: {
4447 char *dest = (char *)mstate->dtms_scratch_ptr;
4448 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4449 uintptr_t src = tupregs[0].dttk_value;
4450 int i, j, len = dtrace_strlen((char *)src, size);
4451 int lastbase = -1, firstbase = -1, lastdir = -1;
4454 if (!dtrace_canload(src, len + 1, mstate, vstate)) {
4459 if (!DTRACE_INSCRATCH(mstate, size)) {
4460 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4466 * The basename and dirname for a zero-length string is
4471 src = (uintptr_t)".";
4475 * Start from the back of the string, moving back toward the
4476 * front until we see a character that isn't a slash. That
4477 * character is the last character in the basename.
4479 for (i = len - 1; i >= 0; i--) {
4480 if (dtrace_load8(src + i) != '/')
4488 * Starting from the last character in the basename, move
4489 * towards the front until we find a slash. The character
4490 * that we processed immediately before that is the first
4491 * character in the basename.
4493 for (; i >= 0; i--) {
4494 if (dtrace_load8(src + i) == '/')
4502 * Now keep going until we find a non-slash character. That
4503 * character is the last character in the dirname.
4505 for (; i >= 0; i--) {
4506 if (dtrace_load8(src + i) != '/')
4513 ASSERT(!(lastbase == -1 && firstbase != -1));
4514 ASSERT(!(firstbase == -1 && lastdir != -1));
4516 if (lastbase == -1) {
4518 * We didn't find a non-slash character. We know that
4519 * the length is non-zero, so the whole string must be
4520 * slashes. In either the dirname or the basename
4521 * case, we return '/'.
4523 ASSERT(firstbase == -1);
4524 firstbase = lastbase = lastdir = 0;
4527 if (firstbase == -1) {
4529 * The entire string consists only of a basename
4530 * component. If we're looking for dirname, we need
4531 * to change our string to be just "."; if we're
4532 * looking for a basename, we'll just set the first
4533 * character of the basename to be 0.
4535 if (subr == DIF_SUBR_DIRNAME) {
4536 ASSERT(lastdir == -1);
4537 src = (uintptr_t)".";
4544 if (subr == DIF_SUBR_DIRNAME) {
4545 if (lastdir == -1) {
4547 * We know that we have a slash in the name --
4548 * or lastdir would be set to 0, above. And
4549 * because lastdir is -1, we know that this
4550 * slash must be the first character. (That
4551 * is, the full string must be of the form
4552 * "/basename".) In this case, the last
4553 * character of the directory name is 0.
4561 ASSERT(subr == DIF_SUBR_BASENAME);
4562 ASSERT(firstbase != -1 && lastbase != -1);
4567 for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
4568 dest[j] = dtrace_load8(src + i);
4571 regs[rd] = (uintptr_t)dest;
4572 mstate->dtms_scratch_ptr += size;
4576 case DIF_SUBR_CLEANPATH: {
4577 char *dest = (char *)mstate->dtms_scratch_ptr, c;
4578 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4579 uintptr_t src = tupregs[0].dttk_value;
4582 if (!dtrace_strcanload(src, size, mstate, vstate)) {
4587 if (!DTRACE_INSCRATCH(mstate, size)) {
4588 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4594 * Move forward, loading each character.
4597 c = dtrace_load8(src + i++);
4599 if (j + 5 >= size) /* 5 = strlen("/..c\0") */
4607 c = dtrace_load8(src + i++);
4611 * We have two slashes -- we can just advance
4612 * to the next character.
4619 * This is not "." and it's not ".." -- we can
4620 * just store the "/" and this character and
4628 c = dtrace_load8(src + i++);
4632 * This is a "/./" component. We're not going
4633 * to store anything in the destination buffer;
4634 * we're just going to go to the next component.
4641 * This is not ".." -- we can just store the
4642 * "/." and this character and continue
4651 c = dtrace_load8(src + i++);
4653 if (c != '/' && c != '\0') {
4655 * This is not ".." -- it's "..[mumble]".
4656 * We'll store the "/.." and this character
4657 * and continue processing.
4667 * This is "/../" or "/..\0". We need to back up
4668 * our destination pointer until we find a "/".
4671 while (j != 0 && dest[--j] != '/')
4676 } while (c != '\0');
4679 regs[rd] = (uintptr_t)dest;
4680 mstate->dtms_scratch_ptr += size;
4684 case DIF_SUBR_INET_NTOA:
4685 case DIF_SUBR_INET_NTOA6:
4686 case DIF_SUBR_INET_NTOP: {
4691 if (subr == DIF_SUBR_INET_NTOP) {
4692 af = (int)tupregs[0].dttk_value;
4695 af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
4699 if (af == AF_INET) {
4704 * Safely load the IPv4 address.
4706 ip4 = dtrace_load32(tupregs[argi].dttk_value);
4709 * Check an IPv4 string will fit in scratch.
4711 size = INET_ADDRSTRLEN;
4712 if (!DTRACE_INSCRATCH(mstate, size)) {
4713 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4717 base = (char *)mstate->dtms_scratch_ptr;
4718 end = (char *)mstate->dtms_scratch_ptr + size - 1;
4721 * Stringify as a dotted decimal quad.
4724 ptr8 = (uint8_t *)&ip4;
4725 for (i = 3; i >= 0; i--) {
4731 for (; val; val /= 10) {
4732 *end-- = '0' + (val % 10);
4739 ASSERT(end + 1 >= base);
4741 } else if (af == AF_INET6) {
4742 struct in6_addr ip6;
4743 int firstzero, tryzero, numzero, v6end;
4745 const char digits[] = "0123456789abcdef";
4748 * Stringify using RFC 1884 convention 2 - 16 bit
4749 * hexadecimal values with a zero-run compression.
4750 * Lower case hexadecimal digits are used.
4751 * eg, fe80::214:4fff:fe0b:76c8.
4752 * The IPv4 embedded form is returned for inet_ntop,
4753 * just the IPv4 string is returned for inet_ntoa6.
4757 * Safely load the IPv6 address.
4760 (void *)(uintptr_t)tupregs[argi].dttk_value,
4761 (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));
4764 * Check an IPv6 string will fit in scratch.
4766 size = INET6_ADDRSTRLEN;
4767 if (!DTRACE_INSCRATCH(mstate, size)) {
4768 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4772 base = (char *)mstate->dtms_scratch_ptr;
4773 end = (char *)mstate->dtms_scratch_ptr + size - 1;
4777 * Find the longest run of 16 bit zero values
4778 * for the single allowed zero compression - "::".
4783 for (i = 0; i < sizeof (struct in6_addr); i++) {
4785 if (ip6._S6_un._S6_u8[i] == 0 &&
4787 if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
4789 tryzero == -1 && i % 2 == 0) {
4794 if (tryzero != -1 &&
4796 (ip6._S6_un._S6_u8[i] != 0 ||
4798 (ip6.__u6_addr.__u6_addr8[i] != 0 ||
4800 i == sizeof (struct in6_addr) - 1)) {
4802 if (i - tryzero <= numzero) {
4807 firstzero = tryzero;
4808 numzero = i - i % 2 - tryzero;
4812 if (ip6._S6_un._S6_u8[i] == 0 &&
4814 if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
4816 i == sizeof (struct in6_addr) - 1)
4820 ASSERT(firstzero + numzero <= sizeof (struct in6_addr));
4823 * Check for an IPv4 embedded address.
4825 v6end = sizeof (struct in6_addr) - 2;
4826 if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
4827 IN6_IS_ADDR_V4COMPAT(&ip6)) {
4828 for (i = sizeof (struct in6_addr) - 1;
4829 i >= DTRACE_V4MAPPED_OFFSET; i--) {
4830 ASSERT(end >= base);
4833 val = ip6._S6_un._S6_u8[i];
4835 val = ip6.__u6_addr.__u6_addr8[i];
4841 for (; val; val /= 10) {
4842 *end-- = '0' + val % 10;
4846 if (i > DTRACE_V4MAPPED_OFFSET)
4850 if (subr == DIF_SUBR_INET_NTOA6)
4854 * Set v6end to skip the IPv4 address that
4855 * we have already stringified.
4861 * Build the IPv6 string by working through the
4862 * address in reverse.
4864 for (i = v6end; i >= 0; i -= 2) {
4865 ASSERT(end >= base);
4867 if (i == firstzero + numzero - 2) {
4874 if (i < 14 && i != firstzero - 2)
4878 val = (ip6._S6_un._S6_u8[i] << 8) +
4879 ip6._S6_un._S6_u8[i + 1];
4881 val = (ip6.__u6_addr.__u6_addr8[i] << 8) +
4882 ip6.__u6_addr.__u6_addr8[i + 1];
4888 for (; val; val /= 16) {
4889 *end-- = digits[val % 16];
4893 ASSERT(end + 1 >= base);
4897 * The user didn't use AH_INET or AH_INET6.
4899 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
4904 inetout: regs[rd] = (uintptr_t)end + 1;
4905 mstate->dtms_scratch_ptr += size;
4909 case DIF_SUBR_MEMREF: {
4910 uintptr_t size = 2 * sizeof(uintptr_t);
4911 uintptr_t *memref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
4912 size_t scratch_size = ((uintptr_t) memref - mstate->dtms_scratch_ptr) + size;
4914 /* address and length */
4915 memref[0] = tupregs[0].dttk_value;
4916 memref[1] = tupregs[1].dttk_value;
4918 regs[rd] = (uintptr_t) memref;
4919 mstate->dtms_scratch_ptr += scratch_size;
4923 case DIF_SUBR_TYPEREF: {
4924 uintptr_t size = 4 * sizeof(uintptr_t);
4925 uintptr_t *typeref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
4926 size_t scratch_size = ((uintptr_t) typeref - mstate->dtms_scratch_ptr) + size;
4928 /* address, num_elements, type_str, type_len */
4929 typeref[0] = tupregs[0].dttk_value;
4930 typeref[1] = tupregs[1].dttk_value;
4931 typeref[2] = tupregs[2].dttk_value;
4932 typeref[3] = tupregs[3].dttk_value;
4934 regs[rd] = (uintptr_t) typeref;
4935 mstate->dtms_scratch_ptr += scratch_size;
4942 * Emulate the execution of DTrace IR instructions specified by the given
4943 * DIF object. This function is deliberately void of assertions as all of
4944 * the necessary checks are handled by a call to dtrace_difo_validate().
4947 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
4948 dtrace_vstate_t *vstate, dtrace_state_t *state)
4950 const dif_instr_t *text = difo->dtdo_buf;
4951 const uint_t textlen = difo->dtdo_len;
4952 const char *strtab = difo->dtdo_strtab;
4953 const uint64_t *inttab = difo->dtdo_inttab;
4956 dtrace_statvar_t *svar;
4957 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
4959 volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
4960 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
4962 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
4963 uint64_t regs[DIF_DIR_NREGS];
4966 uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
4968 uint_t pc = 0, id, opc = 0;
4974 * We stash the current DIF object into the machine state: we need it
4975 * for subsequent access checking.
4977 mstate->dtms_difo = difo;
4979 regs[DIF_REG_R0] = 0; /* %r0 is fixed at zero */
4981 while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
4985 r1 = DIF_INSTR_R1(instr);
4986 r2 = DIF_INSTR_R2(instr);
4987 rd = DIF_INSTR_RD(instr);
4989 switch (DIF_INSTR_OP(instr)) {
4991 regs[rd] = regs[r1] | regs[r2];
4994 regs[rd] = regs[r1] ^ regs[r2];
4997 regs[rd] = regs[r1] & regs[r2];
5000 regs[rd] = regs[r1] << regs[r2];
5003 regs[rd] = regs[r1] >> regs[r2];
5006 regs[rd] = regs[r1] - regs[r2];
5009 regs[rd] = regs[r1] + regs[r2];
5012 regs[rd] = regs[r1] * regs[r2];
5015 if (regs[r2] == 0) {
5017 *flags |= CPU_DTRACE_DIVZERO;
5019 regs[rd] = (int64_t)regs[r1] /
5025 if (regs[r2] == 0) {
5027 *flags |= CPU_DTRACE_DIVZERO;
5029 regs[rd] = regs[r1] / regs[r2];
5034 if (regs[r2] == 0) {
5036 *flags |= CPU_DTRACE_DIVZERO;
5038 regs[rd] = (int64_t)regs[r1] %
5044 if (regs[r2] == 0) {
5046 *flags |= CPU_DTRACE_DIVZERO;
5048 regs[rd] = regs[r1] % regs[r2];
5053 regs[rd] = ~regs[r1];
5056 regs[rd] = regs[r1];
5059 cc_r = regs[r1] - regs[r2];
5063 cc_c = regs[r1] < regs[r2];
5066 cc_n = cc_v = cc_c = 0;
5067 cc_z = regs[r1] == 0;
5070 pc = DIF_INSTR_LABEL(instr);
5074 pc = DIF_INSTR_LABEL(instr);
5078 pc = DIF_INSTR_LABEL(instr);
5081 if ((cc_z | (cc_n ^ cc_v)) == 0)
5082 pc = DIF_INSTR_LABEL(instr);
5085 if ((cc_c | cc_z) == 0)
5086 pc = DIF_INSTR_LABEL(instr);
5089 if ((cc_n ^ cc_v) == 0)
5090 pc = DIF_INSTR_LABEL(instr);
5094 pc = DIF_INSTR_LABEL(instr);
5098 pc = DIF_INSTR_LABEL(instr);
5102 pc = DIF_INSTR_LABEL(instr);
5105 if (cc_z | (cc_n ^ cc_v))
5106 pc = DIF_INSTR_LABEL(instr);
5110 pc = DIF_INSTR_LABEL(instr);
5113 if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
5114 *flags |= CPU_DTRACE_KPRIV;
5120 regs[rd] = (int8_t)dtrace_load8(regs[r1]);
5123 if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
5124 *flags |= CPU_DTRACE_KPRIV;
5130 regs[rd] = (int16_t)dtrace_load16(regs[r1]);
5133 if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
5134 *flags |= CPU_DTRACE_KPRIV;
5140 regs[rd] = (int32_t)dtrace_load32(regs[r1]);
5143 if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
5144 *flags |= CPU_DTRACE_KPRIV;
5150 regs[rd] = dtrace_load8(regs[r1]);
5153 if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
5154 *flags |= CPU_DTRACE_KPRIV;
5160 regs[rd] = dtrace_load16(regs[r1]);
5163 if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
5164 *flags |= CPU_DTRACE_KPRIV;
5170 regs[rd] = dtrace_load32(regs[r1]);
5173 if (!dtrace_canstore(regs[r1], 8, mstate, vstate)) {
5174 *flags |= CPU_DTRACE_KPRIV;
5180 regs[rd] = dtrace_load64(regs[r1]);
5184 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5187 regs[rd] = (int16_t)
5188 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5191 regs[rd] = (int32_t)
5192 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5196 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5200 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5204 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5208 dtrace_fuword64((void *)(uintptr_t)regs[r1]);
5217 regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
5220 regs[rd] = (uint64_t)(uintptr_t)
5221 (strtab + DIF_INSTR_STRING(instr));
5224 size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
5225 uintptr_t s1 = regs[r1];
5226 uintptr_t s2 = regs[r2];
5229 !dtrace_strcanload(s1, sz, mstate, vstate))
5232 !dtrace_strcanload(s2, sz, mstate, vstate))
5235 cc_r = dtrace_strncmp((char *)s1, (char *)s2, sz);
5243 regs[rd] = dtrace_dif_variable(mstate, state,
5247 id = DIF_INSTR_VAR(instr);
5249 if (id >= DIF_VAR_OTHER_UBASE) {
5252 id -= DIF_VAR_OTHER_UBASE;
5253 svar = vstate->dtvs_globals[id];
5254 ASSERT(svar != NULL);
5255 v = &svar->dtsv_var;
5257 if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
5258 regs[rd] = svar->dtsv_data;
5262 a = (uintptr_t)svar->dtsv_data;
5264 if (*(uint8_t *)a == UINT8_MAX) {
5266 * If the 0th byte is set to UINT8_MAX
5267 * then this is to be treated as a
5268 * reference to a NULL variable.
5272 regs[rd] = a + sizeof (uint64_t);
5278 regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
5282 id = DIF_INSTR_VAR(instr);
5284 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5285 id -= DIF_VAR_OTHER_UBASE;
5287 svar = vstate->dtvs_globals[id];
5288 ASSERT(svar != NULL);
5289 v = &svar->dtsv_var;
5291 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5292 uintptr_t a = (uintptr_t)svar->dtsv_data;
5295 ASSERT(svar->dtsv_size != 0);
5297 if (regs[rd] == 0) {
5298 *(uint8_t *)a = UINT8_MAX;
5302 a += sizeof (uint64_t);
5304 if (!dtrace_vcanload(
5305 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5309 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5310 (void *)a, &v->dtdv_type);
5314 svar->dtsv_data = regs[rd];
5319 * There are no DTrace built-in thread-local arrays at
5320 * present. This opcode is saved for future work.
5322 *flags |= CPU_DTRACE_ILLOP;
5327 id = DIF_INSTR_VAR(instr);
5329 if (id < DIF_VAR_OTHER_UBASE) {
5331 * For now, this has no meaning.
5337 id -= DIF_VAR_OTHER_UBASE;
5339 ASSERT(id < vstate->dtvs_nlocals);
5340 ASSERT(vstate->dtvs_locals != NULL);
5342 svar = vstate->dtvs_locals[id];
5343 ASSERT(svar != NULL);
5344 v = &svar->dtsv_var;
5346 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5347 uintptr_t a = (uintptr_t)svar->dtsv_data;
5348 size_t sz = v->dtdv_type.dtdt_size;
5350 sz += sizeof (uint64_t);
5351 ASSERT(svar->dtsv_size == NCPU * sz);
5354 if (*(uint8_t *)a == UINT8_MAX) {
5356 * If the 0th byte is set to UINT8_MAX
5357 * then this is to be treated as a
5358 * reference to a NULL variable.
5362 regs[rd] = a + sizeof (uint64_t);
5368 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5369 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5370 regs[rd] = tmp[curcpu];
5374 id = DIF_INSTR_VAR(instr);
5376 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5377 id -= DIF_VAR_OTHER_UBASE;
5378 ASSERT(id < vstate->dtvs_nlocals);
5380 ASSERT(vstate->dtvs_locals != NULL);
5381 svar = vstate->dtvs_locals[id];
5382 ASSERT(svar != NULL);
5383 v = &svar->dtsv_var;
5385 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5386 uintptr_t a = (uintptr_t)svar->dtsv_data;
5387 size_t sz = v->dtdv_type.dtdt_size;
5389 sz += sizeof (uint64_t);
5390 ASSERT(svar->dtsv_size == NCPU * sz);
5393 if (regs[rd] == 0) {
5394 *(uint8_t *)a = UINT8_MAX;
5398 a += sizeof (uint64_t);
5401 if (!dtrace_vcanload(
5402 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5406 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5407 (void *)a, &v->dtdv_type);
5411 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5412 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5413 tmp[curcpu] = regs[rd];
5417 dtrace_dynvar_t *dvar;
5420 id = DIF_INSTR_VAR(instr);
5421 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5422 id -= DIF_VAR_OTHER_UBASE;
5423 v = &vstate->dtvs_tlocals[id];
5425 key = &tupregs[DIF_DTR_NREGS];
5426 key[0].dttk_value = (uint64_t)id;
5427 key[0].dttk_size = 0;
5428 DTRACE_TLS_THRKEY(key[1].dttk_value);
5429 key[1].dttk_size = 0;
5431 dvar = dtrace_dynvar(dstate, 2, key,
5432 sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
5440 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5441 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5443 regs[rd] = *((uint64_t *)dvar->dtdv_data);
5450 dtrace_dynvar_t *dvar;
5453 id = DIF_INSTR_VAR(instr);
5454 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5455 id -= DIF_VAR_OTHER_UBASE;
5457 key = &tupregs[DIF_DTR_NREGS];
5458 key[0].dttk_value = (uint64_t)id;
5459 key[0].dttk_size = 0;
5460 DTRACE_TLS_THRKEY(key[1].dttk_value);
5461 key[1].dttk_size = 0;
5462 v = &vstate->dtvs_tlocals[id];
5464 dvar = dtrace_dynvar(dstate, 2, key,
5465 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5466 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5467 regs[rd] ? DTRACE_DYNVAR_ALLOC :
5468 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5471 * Given that we're storing to thread-local data,
5472 * we need to flush our predicate cache.
5474 curthread->t_predcache = 0;
5479 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5480 if (!dtrace_vcanload(
5481 (void *)(uintptr_t)regs[rd],
5482 &v->dtdv_type, mstate, vstate))
5485 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5486 dvar->dtdv_data, &v->dtdv_type);
5488 *((uint64_t *)dvar->dtdv_data) = regs[rd];
5495 regs[rd] = (int64_t)regs[r1] >> regs[r2];
5499 dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
5500 regs, tupregs, ttop, mstate, state);
5504 if (ttop == DIF_DTR_NREGS) {
5505 *flags |= CPU_DTRACE_TUPOFLOW;
5509 if (r1 == DIF_TYPE_STRING) {
5511 * If this is a string type and the size is 0,
5512 * we'll use the system-wide default string
5513 * size. Note that we are _not_ looking at
5514 * the value of the DTRACEOPT_STRSIZE option;
5515 * had this been set, we would expect to have
5516 * a non-zero size value in the "pushtr".
5518 tupregs[ttop].dttk_size =
5519 dtrace_strlen((char *)(uintptr_t)regs[rd],
5520 regs[r2] ? regs[r2] :
5521 dtrace_strsize_default) + 1;
5523 tupregs[ttop].dttk_size = regs[r2];
5526 tupregs[ttop++].dttk_value = regs[rd];
5530 if (ttop == DIF_DTR_NREGS) {
5531 *flags |= CPU_DTRACE_TUPOFLOW;
5535 tupregs[ttop].dttk_value = regs[rd];
5536 tupregs[ttop++].dttk_size = 0;
5544 case DIF_OP_FLUSHTS:
5549 case DIF_OP_LDTAA: {
5550 dtrace_dynvar_t *dvar;
5551 dtrace_key_t *key = tupregs;
5552 uint_t nkeys = ttop;
5554 id = DIF_INSTR_VAR(instr);
5555 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5556 id -= DIF_VAR_OTHER_UBASE;
5558 key[nkeys].dttk_value = (uint64_t)id;
5559 key[nkeys++].dttk_size = 0;
5561 if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
5562 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5563 key[nkeys++].dttk_size = 0;
5564 v = &vstate->dtvs_tlocals[id];
5566 v = &vstate->dtvs_globals[id]->dtsv_var;
5569 dvar = dtrace_dynvar(dstate, nkeys, key,
5570 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5571 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5572 DTRACE_DYNVAR_NOALLOC, mstate, vstate);
5579 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5580 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5582 regs[rd] = *((uint64_t *)dvar->dtdv_data);
5589 case DIF_OP_STTAA: {
5590 dtrace_dynvar_t *dvar;
5591 dtrace_key_t *key = tupregs;
5592 uint_t nkeys = ttop;
5594 id = DIF_INSTR_VAR(instr);
5595 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5596 id -= DIF_VAR_OTHER_UBASE;
5598 key[nkeys].dttk_value = (uint64_t)id;
5599 key[nkeys++].dttk_size = 0;
5601 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
5602 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5603 key[nkeys++].dttk_size = 0;
5604 v = &vstate->dtvs_tlocals[id];
5606 v = &vstate->dtvs_globals[id]->dtsv_var;
5609 dvar = dtrace_dynvar(dstate, nkeys, key,
5610 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5611 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5612 regs[rd] ? DTRACE_DYNVAR_ALLOC :
5613 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5618 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5619 if (!dtrace_vcanload(
5620 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5624 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5625 dvar->dtdv_data, &v->dtdv_type);
5627 *((uint64_t *)dvar->dtdv_data) = regs[rd];
5633 case DIF_OP_ALLOCS: {
5634 uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5635 size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
5638 * Rounding up the user allocation size could have
5639 * overflowed large, bogus allocations (like -1ULL) to
5642 if (size < regs[r1] ||
5643 !DTRACE_INSCRATCH(mstate, size)) {
5644 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5649 dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
5650 mstate->dtms_scratch_ptr += size;
5656 if (!dtrace_canstore(regs[rd], regs[r2],
5658 *flags |= CPU_DTRACE_BADADDR;
5663 if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
5666 dtrace_bcopy((void *)(uintptr_t)regs[r1],
5667 (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
5671 if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
5672 *flags |= CPU_DTRACE_BADADDR;
5676 *((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
5680 if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
5681 *flags |= CPU_DTRACE_BADADDR;
5686 *flags |= CPU_DTRACE_BADALIGN;
5690 *((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
5694 if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
5695 *flags |= CPU_DTRACE_BADADDR;
5700 *flags |= CPU_DTRACE_BADALIGN;
5704 *((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
5708 if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
5709 *flags |= CPU_DTRACE_BADADDR;
5714 *flags |= CPU_DTRACE_BADALIGN;
5718 *((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
5723 if (!(*flags & CPU_DTRACE_FAULT))
5726 mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
5727 mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
5733 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
5735 dtrace_probe_t *probe = ecb->dte_probe;
5736 dtrace_provider_t *prov = probe->dtpr_provider;
5737 char c[DTRACE_FULLNAMELEN + 80], *str;
5738 char *msg = "dtrace: breakpoint action at probe ";
5739 char *ecbmsg = " (ecb ";
5740 uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
5741 uintptr_t val = (uintptr_t)ecb;
5742 int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
5744 if (dtrace_destructive_disallow)
5748 * It's impossible to be taking action on the NULL probe.
5750 ASSERT(probe != NULL);
5753 * This is a poor man's (destitute man's?) sprintf(): we want to
5754 * print the provider name, module name, function name and name of
5755 * the probe, along with the hex address of the ECB with the breakpoint
5756 * action -- all of which we must place in the character buffer by
5759 while (*msg != '\0')
5762 for (str = prov->dtpv_name; *str != '\0'; str++)
5766 for (str = probe->dtpr_mod; *str != '\0'; str++)
5770 for (str = probe->dtpr_func; *str != '\0'; str++)
5774 for (str = probe->dtpr_name; *str != '\0'; str++)
5777 while (*ecbmsg != '\0')
5780 while (shift >= 0) {
5781 mask = (uintptr_t)0xf << shift;
5783 if (val >= ((uintptr_t)1 << shift))
5784 c[i++] = "0123456789abcdef"[(val & mask) >> shift];
5794 kdb_enter(KDB_WHY_DTRACE, "breakpoint action");
5799 dtrace_action_panic(dtrace_ecb_t *ecb)
5801 dtrace_probe_t *probe = ecb->dte_probe;
5804 * It's impossible to be taking action on the NULL probe.
5806 ASSERT(probe != NULL);
5808 if (dtrace_destructive_disallow)
5811 if (dtrace_panicked != NULL)
5814 if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
5818 * We won the right to panic. (We want to be sure that only one
5819 * thread calls panic() from dtrace_probe(), and that panic() is
5820 * called exactly once.)
5822 dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
5823 probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
5824 probe->dtpr_func, probe->dtpr_name, (void *)ecb);
5828 dtrace_action_raise(uint64_t sig)
5830 if (dtrace_destructive_disallow)
5834 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
5840 * raise() has a queue depth of 1 -- we ignore all subsequent
5841 * invocations of the raise() action.
5843 if (curthread->t_dtrace_sig == 0)
5844 curthread->t_dtrace_sig = (uint8_t)sig;
5846 curthread->t_sig_check = 1;
5849 struct proc *p = curproc;
5851 kern_psignal(p, sig);
5857 dtrace_action_stop(void)
5859 if (dtrace_destructive_disallow)
5863 if (!curthread->t_dtrace_stop) {
5864 curthread->t_dtrace_stop = 1;
5865 curthread->t_sig_check = 1;
5869 struct proc *p = curproc;
5871 kern_psignal(p, SIGSTOP);
5877 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
5880 volatile uint16_t *flags;
5884 cpu_t *cpu = &solaris_cpu[curcpu];
5887 if (dtrace_destructive_disallow)
5890 flags = (volatile uint16_t *)&cpu_core[cpu->cpu_id].cpuc_dtrace_flags;
5892 now = dtrace_gethrtime();
5894 if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
5896 * We need to advance the mark to the current time.
5898 cpu->cpu_dtrace_chillmark = now;
5899 cpu->cpu_dtrace_chilled = 0;
5903 * Now check to see if the requested chill time would take us over
5904 * the maximum amount of time allowed in the chill interval. (Or
5905 * worse, if the calculation itself induces overflow.)
5907 if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
5908 cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
5909 *flags |= CPU_DTRACE_ILLOP;
5913 while (dtrace_gethrtime() - now < val)
5917 * Normally, we assure that the value of the variable "timestamp" does
5918 * not change within an ECB. The presence of chill() represents an
5919 * exception to this rule, however.
5921 mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
5922 cpu->cpu_dtrace_chilled += val;
5926 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
5927 uint64_t *buf, uint64_t arg)
5929 int nframes = DTRACE_USTACK_NFRAMES(arg);
5930 int strsize = DTRACE_USTACK_STRSIZE(arg);
5931 uint64_t *pcs = &buf[1], *fps;
5932 char *str = (char *)&pcs[nframes];
5933 int size, offs = 0, i, j;
5934 uintptr_t old = mstate->dtms_scratch_ptr, saved;
5935 uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
5939 * Should be taking a faster path if string space has not been
5942 ASSERT(strsize != 0);
5945 * We will first allocate some temporary space for the frame pointers.
5947 fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5948 size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
5949 (nframes * sizeof (uint64_t));
5951 if (!DTRACE_INSCRATCH(mstate, size)) {
5953 * Not enough room for our frame pointers -- need to indicate
5954 * that we ran out of scratch space.
5956 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5960 mstate->dtms_scratch_ptr += size;
5961 saved = mstate->dtms_scratch_ptr;
5964 * Now get a stack with both program counters and frame pointers.
5966 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5967 dtrace_getufpstack(buf, fps, nframes + 1);
5968 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5971 * If that faulted, we're cooked.
5973 if (*flags & CPU_DTRACE_FAULT)
5977 * Now we want to walk up the stack, calling the USTACK helper. For
5978 * each iteration, we restore the scratch pointer.
5980 for (i = 0; i < nframes; i++) {
5981 mstate->dtms_scratch_ptr = saved;
5983 if (offs >= strsize)
5986 sym = (char *)(uintptr_t)dtrace_helper(
5987 DTRACE_HELPER_ACTION_USTACK,
5988 mstate, state, pcs[i], fps[i]);
5991 * If we faulted while running the helper, we're going to
5992 * clear the fault and null out the corresponding string.
5994 if (*flags & CPU_DTRACE_FAULT) {
5995 *flags &= ~CPU_DTRACE_FAULT;
6005 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6008 * Now copy in the string that the helper returned to us.
6010 for (j = 0; offs + j < strsize; j++) {
6011 if ((str[offs + j] = sym[j]) == '\0')
6015 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6020 if (offs >= strsize) {
6022 * If we didn't have room for all of the strings, we don't
6023 * abort processing -- this needn't be a fatal error -- but we
6024 * still want to increment a counter (dts_stkstroverflows) to
6025 * allow this condition to be warned about. (If this is from
6026 * a jstack() action, it is easily tuned via jstackstrsize.)
6028 dtrace_error(&state->dts_stkstroverflows);
6031 while (offs < strsize)
6035 mstate->dtms_scratch_ptr = old;
6039 * If you're looking for the epicenter of DTrace, you just found it. This
6040 * is the function called by the provider to fire a probe -- from which all
6041 * subsequent probe-context DTrace activity emanates.
6044 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
6045 uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
6047 processorid_t cpuid;
6048 dtrace_icookie_t cookie;
6049 dtrace_probe_t *probe;
6050 dtrace_mstate_t mstate;
6052 dtrace_action_t *act;
6056 volatile uint16_t *flags;
6059 if (panicstr != NULL)
6064 * Kick out immediately if this CPU is still being born (in which case
6065 * curthread will be set to -1) or the current thread can't allow
6066 * probes in its current context.
6068 if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
6072 cookie = dtrace_interrupt_disable();
6073 probe = dtrace_probes[id - 1];
6075 onintr = CPU_ON_INTR(CPU);
6077 if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
6078 probe->dtpr_predcache == curthread->t_predcache) {
6080 * We have hit in the predicate cache; we know that
6081 * this predicate would evaluate to be false.
6083 dtrace_interrupt_enable(cookie);
6088 if (panic_quiesce) {
6090 if (panicstr != NULL) {
6093 * We don't trace anything if we're panicking.
6095 dtrace_interrupt_enable(cookie);
6099 now = dtrace_gethrtime();
6100 vtime = dtrace_vtime_references != 0;
6102 if (vtime && curthread->t_dtrace_start)
6103 curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
6105 mstate.dtms_difo = NULL;
6106 mstate.dtms_probe = probe;
6107 mstate.dtms_strtok = 0;
6108 mstate.dtms_arg[0] = arg0;
6109 mstate.dtms_arg[1] = arg1;
6110 mstate.dtms_arg[2] = arg2;
6111 mstate.dtms_arg[3] = arg3;
6112 mstate.dtms_arg[4] = arg4;
6114 flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
6116 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
6117 dtrace_predicate_t *pred = ecb->dte_predicate;
6118 dtrace_state_t *state = ecb->dte_state;
6119 dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
6120 dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
6121 dtrace_vstate_t *vstate = &state->dts_vstate;
6122 dtrace_provider_t *prov = probe->dtpr_provider;
6123 uint64_t tracememsize = 0;
6128 * A little subtlety with the following (seemingly innocuous)
6129 * declaration of the automatic 'val': by looking at the
6130 * code, you might think that it could be declared in the
6131 * action processing loop, below. (That is, it's only used in
6132 * the action processing loop.) However, it must be declared
6133 * out of that scope because in the case of DIF expression
6134 * arguments to aggregating actions, one iteration of the
6135 * action loop will use the last iteration's value.
6139 mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
6140 *flags &= ~CPU_DTRACE_ERROR;
6142 if (prov == dtrace_provider) {
6144 * If dtrace itself is the provider of this probe,
6145 * we're only going to continue processing the ECB if
6146 * arg0 (the dtrace_state_t) is equal to the ECB's
6147 * creating state. (This prevents disjoint consumers
6148 * from seeing one another's metaprobes.)
6150 if (arg0 != (uint64_t)(uintptr_t)state)
6154 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
6156 * We're not currently active. If our provider isn't
6157 * the dtrace pseudo provider, we're not interested.
6159 if (prov != dtrace_provider)
6163 * Now we must further check if we are in the BEGIN
6164 * probe. If we are, we will only continue processing
6165 * if we're still in WARMUP -- if one BEGIN enabling
6166 * has invoked the exit() action, we don't want to
6167 * evaluate subsequent BEGIN enablings.
6169 if (probe->dtpr_id == dtrace_probeid_begin &&
6170 state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
6171 ASSERT(state->dts_activity ==
6172 DTRACE_ACTIVITY_DRAINING);
6177 if (ecb->dte_cond) {
6179 * If the dte_cond bits indicate that this
6180 * consumer is only allowed to see user-mode firings
6181 * of this probe, call the provider's dtps_usermode()
6182 * entry point to check that the probe was fired
6183 * while in a user context. Skip this ECB if that's
6186 if ((ecb->dte_cond & DTRACE_COND_USERMODE) &&
6187 prov->dtpv_pops.dtps_usermode(prov->dtpv_arg,
6188 probe->dtpr_id, probe->dtpr_arg) == 0)
6193 * This is more subtle than it looks. We have to be
6194 * absolutely certain that CRED() isn't going to
6195 * change out from under us so it's only legit to
6196 * examine that structure if we're in constrained
6197 * situations. Currently, the only times we'll this
6198 * check is if a non-super-user has enabled the
6199 * profile or syscall providers -- providers that
6200 * allow visibility of all processes. For the
6201 * profile case, the check above will ensure that
6202 * we're examining a user context.
6204 if (ecb->dte_cond & DTRACE_COND_OWNER) {
6207 ecb->dte_state->dts_cred.dcr_cred;
6210 ASSERT(s_cr != NULL);
6212 if ((cr = CRED()) == NULL ||
6213 s_cr->cr_uid != cr->cr_uid ||
6214 s_cr->cr_uid != cr->cr_ruid ||
6215 s_cr->cr_uid != cr->cr_suid ||
6216 s_cr->cr_gid != cr->cr_gid ||
6217 s_cr->cr_gid != cr->cr_rgid ||
6218 s_cr->cr_gid != cr->cr_sgid ||
6219 (proc = ttoproc(curthread)) == NULL ||
6220 (proc->p_flag & SNOCD))
6224 if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
6227 ecb->dte_state->dts_cred.dcr_cred;
6229 ASSERT(s_cr != NULL);
6231 if ((cr = CRED()) == NULL ||
6232 s_cr->cr_zone->zone_id !=
6233 cr->cr_zone->zone_id)
6239 if (now - state->dts_alive > dtrace_deadman_timeout) {
6241 * We seem to be dead. Unless we (a) have kernel
6242 * destructive permissions (b) have explicitly enabled
6243 * destructive actions and (c) destructive actions have
6244 * not been disabled, we're going to transition into
6245 * the KILLED state, from which no further processing
6246 * on this state will be performed.
6248 if (!dtrace_priv_kernel_destructive(state) ||
6249 !state->dts_cred.dcr_destructive ||
6250 dtrace_destructive_disallow) {
6251 void *activity = &state->dts_activity;
6252 dtrace_activity_t current;
6255 current = state->dts_activity;
6256 } while (dtrace_cas32(activity, current,
6257 DTRACE_ACTIVITY_KILLED) != current);
6263 if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
6264 ecb->dte_alignment, state, &mstate)) < 0)
6267 tomax = buf->dtb_tomax;
6268 ASSERT(tomax != NULL);
6270 if (ecb->dte_size != 0) {
6271 dtrace_rechdr_t dtrh;
6272 if (!(mstate.dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
6273 mstate.dtms_timestamp = dtrace_gethrtime();
6274 mstate.dtms_present |= DTRACE_MSTATE_TIMESTAMP;
6276 ASSERT3U(ecb->dte_size, >=, sizeof (dtrace_rechdr_t));
6277 dtrh.dtrh_epid = ecb->dte_epid;
6278 DTRACE_RECORD_STORE_TIMESTAMP(&dtrh,
6279 mstate.dtms_timestamp);
6280 *((dtrace_rechdr_t *)(tomax + offs)) = dtrh;
6283 mstate.dtms_epid = ecb->dte_epid;
6284 mstate.dtms_present |= DTRACE_MSTATE_EPID;
6286 if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
6287 mstate.dtms_access = DTRACE_ACCESS_KERNEL;
6289 mstate.dtms_access = 0;
6292 dtrace_difo_t *dp = pred->dtp_difo;
6295 rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
6297 if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
6298 dtrace_cacheid_t cid = probe->dtpr_predcache;
6300 if (cid != DTRACE_CACHEIDNONE && !onintr) {
6302 * Update the predicate cache...
6304 ASSERT(cid == pred->dtp_cacheid);
6305 curthread->t_predcache = cid;
6312 for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
6313 act != NULL; act = act->dta_next) {
6316 dtrace_recdesc_t *rec = &act->dta_rec;
6318 size = rec->dtrd_size;
6319 valoffs = offs + rec->dtrd_offset;
6321 if (DTRACEACT_ISAGG(act->dta_kind)) {
6323 dtrace_aggregation_t *agg;
6325 agg = (dtrace_aggregation_t *)act;
6327 if ((dp = act->dta_difo) != NULL)
6328 v = dtrace_dif_emulate(dp,
6329 &mstate, vstate, state);
6331 if (*flags & CPU_DTRACE_ERROR)
6335 * Note that we always pass the expression
6336 * value from the previous iteration of the
6337 * action loop. This value will only be used
6338 * if there is an expression argument to the
6339 * aggregating action, denoted by the
6340 * dtag_hasarg field.
6342 dtrace_aggregate(agg, buf,
6343 offs, aggbuf, v, val);
6347 switch (act->dta_kind) {
6348 case DTRACEACT_STOP:
6349 if (dtrace_priv_proc_destructive(state))
6350 dtrace_action_stop();
6353 case DTRACEACT_BREAKPOINT:
6354 if (dtrace_priv_kernel_destructive(state))
6355 dtrace_action_breakpoint(ecb);
6358 case DTRACEACT_PANIC:
6359 if (dtrace_priv_kernel_destructive(state))
6360 dtrace_action_panic(ecb);
6363 case DTRACEACT_STACK:
6364 if (!dtrace_priv_kernel(state))
6367 dtrace_getpcstack((pc_t *)(tomax + valoffs),
6368 size / sizeof (pc_t), probe->dtpr_aframes,
6369 DTRACE_ANCHORED(probe) ? NULL :
6373 case DTRACEACT_JSTACK:
6374 case DTRACEACT_USTACK:
6375 if (!dtrace_priv_proc(state))
6379 * See comment in DIF_VAR_PID.
6381 if (DTRACE_ANCHORED(mstate.dtms_probe) &&
6383 int depth = DTRACE_USTACK_NFRAMES(
6386 dtrace_bzero((void *)(tomax + valoffs),
6387 DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
6388 + depth * sizeof (uint64_t));
6393 if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
6394 curproc->p_dtrace_helpers != NULL) {
6396 * This is the slow path -- we have
6397 * allocated string space, and we're
6398 * getting the stack of a process that
6399 * has helpers. Call into a separate
6400 * routine to perform this processing.
6402 dtrace_action_ustack(&mstate, state,
6403 (uint64_t *)(tomax + valoffs),
6408 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6409 dtrace_getupcstack((uint64_t *)
6411 DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
6412 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6422 val = dtrace_dif_emulate(dp, &mstate, vstate, state);
6424 if (*flags & CPU_DTRACE_ERROR)
6427 switch (act->dta_kind) {
6428 case DTRACEACT_SPECULATE: {
6429 dtrace_rechdr_t *dtrh;
6431 ASSERT(buf == &state->dts_buffer[cpuid]);
6432 buf = dtrace_speculation_buffer(state,
6436 *flags |= CPU_DTRACE_DROP;
6440 offs = dtrace_buffer_reserve(buf,
6441 ecb->dte_needed, ecb->dte_alignment,
6445 *flags |= CPU_DTRACE_DROP;
6449 tomax = buf->dtb_tomax;
6450 ASSERT(tomax != NULL);
6452 if (ecb->dte_size == 0)
6455 ASSERT3U(ecb->dte_size, >=,
6456 sizeof (dtrace_rechdr_t));
6457 dtrh = ((void *)(tomax + offs));
6458 dtrh->dtrh_epid = ecb->dte_epid;
6460 * When the speculation is committed, all of
6461 * the records in the speculative buffer will
6462 * have their timestamps set to the commit
6463 * time. Until then, it is set to a sentinel
6464 * value, for debugability.
6466 DTRACE_RECORD_STORE_TIMESTAMP(dtrh, UINT64_MAX);
6470 case DTRACEACT_PRINTM: {
6471 /* The DIF returns a 'memref'. */
6472 uintptr_t *memref = (uintptr_t *)(uintptr_t) val;
6474 /* Get the size from the memref. */
6478 * Check if the size exceeds the allocated
6481 if (size + sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
6483 *flags |= CPU_DTRACE_DROP;
6487 /* Store the size in the buffer first. */
6488 DTRACE_STORE(uintptr_t, tomax,
6492 * Offset the buffer address to the start
6495 valoffs += sizeof(uintptr_t);
6498 * Reset to the memory address rather than
6499 * the memref array, then let the BYREF
6500 * code below do the work to store the
6501 * memory data in the buffer.
6507 case DTRACEACT_PRINTT: {
6508 /* The DIF returns a 'typeref'. */
6509 uintptr_t *typeref = (uintptr_t *)(uintptr_t) val;
6514 * Get the type string length and round it
6515 * up so that the data that follows is
6516 * aligned for easy access.
6518 size_t typs = strlen((char *) typeref[2]) + 1;
6519 typs = roundup(typs, sizeof(uintptr_t));
6522 *Get the size from the typeref using the
6523 * number of elements and the type size.
6525 size = typeref[1] * typeref[3];
6528 * Check if the size exceeds the allocated
6531 if (size + typs + 2 * sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
6533 *flags |= CPU_DTRACE_DROP;
6537 /* Store the size in the buffer first. */
6538 DTRACE_STORE(uintptr_t, tomax,
6540 valoffs += sizeof(uintptr_t);
6542 /* Store the type size in the buffer. */
6543 DTRACE_STORE(uintptr_t, tomax,
6544 valoffs, typeref[3]);
6545 valoffs += sizeof(uintptr_t);
6549 for (s = 0; s < typs; s++) {
6551 c = dtrace_load8(val++);
6553 DTRACE_STORE(uint8_t, tomax,
6558 * Reset to the memory address rather than
6559 * the typeref array, then let the BYREF
6560 * code below do the work to store the
6561 * memory data in the buffer.
6567 case DTRACEACT_CHILL:
6568 if (dtrace_priv_kernel_destructive(state))
6569 dtrace_action_chill(&mstate, val);
6572 case DTRACEACT_RAISE:
6573 if (dtrace_priv_proc_destructive(state))
6574 dtrace_action_raise(val);
6577 case DTRACEACT_COMMIT:
6581 * We need to commit our buffer state.
6584 buf->dtb_offset = offs + ecb->dte_size;
6585 buf = &state->dts_buffer[cpuid];
6586 dtrace_speculation_commit(state, cpuid, val);
6590 case DTRACEACT_DISCARD:
6591 dtrace_speculation_discard(state, cpuid, val);
6594 case DTRACEACT_DIFEXPR:
6595 case DTRACEACT_LIBACT:
6596 case DTRACEACT_PRINTF:
6597 case DTRACEACT_PRINTA:
6598 case DTRACEACT_SYSTEM:
6599 case DTRACEACT_FREOPEN:
6600 case DTRACEACT_TRACEMEM:
6603 case DTRACEACT_TRACEMEM_DYNSIZE:
6609 if (!dtrace_priv_kernel(state))
6613 case DTRACEACT_USYM:
6614 case DTRACEACT_UMOD:
6615 case DTRACEACT_UADDR: {
6617 struct pid *pid = curthread->t_procp->p_pidp;
6620 if (!dtrace_priv_proc(state))
6623 DTRACE_STORE(uint64_t, tomax,
6625 valoffs, (uint64_t)pid->pid_id);
6627 valoffs, (uint64_t) curproc->p_pid);
6629 DTRACE_STORE(uint64_t, tomax,
6630 valoffs + sizeof (uint64_t), val);
6635 case DTRACEACT_EXIT: {
6637 * For the exit action, we are going to attempt
6638 * to atomically set our activity to be
6639 * draining. If this fails (either because
6640 * another CPU has beat us to the exit action,
6641 * or because our current activity is something
6642 * other than ACTIVE or WARMUP), we will
6643 * continue. This assures that the exit action
6644 * can be successfully recorded at most once
6645 * when we're in the ACTIVE state. If we're
6646 * encountering the exit() action while in
6647 * COOLDOWN, however, we want to honor the new
6648 * status code. (We know that we're the only
6649 * thread in COOLDOWN, so there is no race.)
6651 void *activity = &state->dts_activity;
6652 dtrace_activity_t current = state->dts_activity;
6654 if (current == DTRACE_ACTIVITY_COOLDOWN)
6657 if (current != DTRACE_ACTIVITY_WARMUP)
6658 current = DTRACE_ACTIVITY_ACTIVE;
6660 if (dtrace_cas32(activity, current,
6661 DTRACE_ACTIVITY_DRAINING) != current) {
6662 *flags |= CPU_DTRACE_DROP;
6673 if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF) {
6674 uintptr_t end = valoffs + size;
6676 if (tracememsize != 0 &&
6677 valoffs + tracememsize < end) {
6678 end = valoffs + tracememsize;
6682 if (!dtrace_vcanload((void *)(uintptr_t)val,
6683 &dp->dtdo_rtype, &mstate, vstate))
6687 * If this is a string, we're going to only
6688 * load until we find the zero byte -- after
6689 * which we'll store zero bytes.
6691 if (dp->dtdo_rtype.dtdt_kind ==
6694 int intuple = act->dta_intuple;
6697 for (s = 0; s < size; s++) {
6699 c = dtrace_load8(val++);
6701 DTRACE_STORE(uint8_t, tomax,
6704 if (c == '\0' && intuple)
6711 while (valoffs < end) {
6712 DTRACE_STORE(uint8_t, tomax, valoffs++,
6713 dtrace_load8(val++));
6723 case sizeof (uint8_t):
6724 DTRACE_STORE(uint8_t, tomax, valoffs, val);
6726 case sizeof (uint16_t):
6727 DTRACE_STORE(uint16_t, tomax, valoffs, val);
6729 case sizeof (uint32_t):
6730 DTRACE_STORE(uint32_t, tomax, valoffs, val);
6732 case sizeof (uint64_t):
6733 DTRACE_STORE(uint64_t, tomax, valoffs, val);
6737 * Any other size should have been returned by
6738 * reference, not by value.
6745 if (*flags & CPU_DTRACE_DROP)
6748 if (*flags & CPU_DTRACE_FAULT) {
6750 dtrace_action_t *err;
6754 if (probe->dtpr_id == dtrace_probeid_error) {
6756 * There's nothing we can do -- we had an
6757 * error on the error probe. We bump an
6758 * error counter to at least indicate that
6759 * this condition happened.
6761 dtrace_error(&state->dts_dblerrors);
6767 * Before recursing on dtrace_probe(), we
6768 * need to explicitly clear out our start
6769 * time to prevent it from being accumulated
6770 * into t_dtrace_vtime.
6772 curthread->t_dtrace_start = 0;
6776 * Iterate over the actions to figure out which action
6777 * we were processing when we experienced the error.
6778 * Note that act points _past_ the faulting action; if
6779 * act is ecb->dte_action, the fault was in the
6780 * predicate, if it's ecb->dte_action->dta_next it's
6781 * in action #1, and so on.
6783 for (err = ecb->dte_action, ndx = 0;
6784 err != act; err = err->dta_next, ndx++)
6787 dtrace_probe_error(state, ecb->dte_epid, ndx,
6788 (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
6789 mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
6790 cpu_core[cpuid].cpuc_dtrace_illval);
6796 buf->dtb_offset = offs + ecb->dte_size;
6800 curthread->t_dtrace_start = dtrace_gethrtime();
6802 dtrace_interrupt_enable(cookie);
6806 * DTrace Probe Hashing Functions
6808 * The functions in this section (and indeed, the functions in remaining
6809 * sections) are not _called_ from probe context. (Any exceptions to this are
6810 * marked with a "Note:".) Rather, they are called from elsewhere in the
6811 * DTrace framework to look-up probes in, add probes to and remove probes from
6812 * the DTrace probe hashes. (Each probe is hashed by each element of the
6813 * probe tuple -- allowing for fast lookups, regardless of what was
6817 dtrace_hash_str(const char *p)
6823 hval = (hval << 4) + *p++;
6824 if ((g = (hval & 0xf0000000)) != 0)
6831 static dtrace_hash_t *
6832 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
6834 dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
6836 hash->dth_stroffs = stroffs;
6837 hash->dth_nextoffs = nextoffs;
6838 hash->dth_prevoffs = prevoffs;
6841 hash->dth_mask = hash->dth_size - 1;
6843 hash->dth_tab = kmem_zalloc(hash->dth_size *
6844 sizeof (dtrace_hashbucket_t *), KM_SLEEP);
6850 dtrace_hash_destroy(dtrace_hash_t *hash)
6855 for (i = 0; i < hash->dth_size; i++)
6856 ASSERT(hash->dth_tab[i] == NULL);
6859 kmem_free(hash->dth_tab,
6860 hash->dth_size * sizeof (dtrace_hashbucket_t *));
6861 kmem_free(hash, sizeof (dtrace_hash_t));
6865 dtrace_hash_resize(dtrace_hash_t *hash)
6867 int size = hash->dth_size, i, ndx;
6868 int new_size = hash->dth_size << 1;
6869 int new_mask = new_size - 1;
6870 dtrace_hashbucket_t **new_tab, *bucket, *next;
6872 ASSERT((new_size & new_mask) == 0);
6874 new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
6876 for (i = 0; i < size; i++) {
6877 for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
6878 dtrace_probe_t *probe = bucket->dthb_chain;
6880 ASSERT(probe != NULL);
6881 ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
6883 next = bucket->dthb_next;
6884 bucket->dthb_next = new_tab[ndx];
6885 new_tab[ndx] = bucket;
6889 kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
6890 hash->dth_tab = new_tab;
6891 hash->dth_size = new_size;
6892 hash->dth_mask = new_mask;
6896 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
6898 int hashval = DTRACE_HASHSTR(hash, new);
6899 int ndx = hashval & hash->dth_mask;
6900 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6901 dtrace_probe_t **nextp, **prevp;
6903 for (; bucket != NULL; bucket = bucket->dthb_next) {
6904 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
6908 if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
6909 dtrace_hash_resize(hash);
6910 dtrace_hash_add(hash, new);
6914 bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
6915 bucket->dthb_next = hash->dth_tab[ndx];
6916 hash->dth_tab[ndx] = bucket;
6917 hash->dth_nbuckets++;
6920 nextp = DTRACE_HASHNEXT(hash, new);
6921 ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
6922 *nextp = bucket->dthb_chain;
6924 if (bucket->dthb_chain != NULL) {
6925 prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
6926 ASSERT(*prevp == NULL);
6930 bucket->dthb_chain = new;
6934 static dtrace_probe_t *
6935 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
6937 int hashval = DTRACE_HASHSTR(hash, template);
6938 int ndx = hashval & hash->dth_mask;
6939 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6941 for (; bucket != NULL; bucket = bucket->dthb_next) {
6942 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6943 return (bucket->dthb_chain);
6950 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
6952 int hashval = DTRACE_HASHSTR(hash, template);
6953 int ndx = hashval & hash->dth_mask;
6954 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6956 for (; bucket != NULL; bucket = bucket->dthb_next) {
6957 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6958 return (bucket->dthb_len);
6965 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
6967 int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
6968 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6970 dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
6971 dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
6974 * Find the bucket that we're removing this probe from.
6976 for (; bucket != NULL; bucket = bucket->dthb_next) {
6977 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
6981 ASSERT(bucket != NULL);
6983 if (*prevp == NULL) {
6984 if (*nextp == NULL) {
6986 * The removed probe was the only probe on this
6987 * bucket; we need to remove the bucket.
6989 dtrace_hashbucket_t *b = hash->dth_tab[ndx];
6991 ASSERT(bucket->dthb_chain == probe);
6995 hash->dth_tab[ndx] = bucket->dthb_next;
6997 while (b->dthb_next != bucket)
6999 b->dthb_next = bucket->dthb_next;
7002 ASSERT(hash->dth_nbuckets > 0);
7003 hash->dth_nbuckets--;
7004 kmem_free(bucket, sizeof (dtrace_hashbucket_t));
7008 bucket->dthb_chain = *nextp;
7010 *(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
7014 *(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
7018 * DTrace Utility Functions
7020 * These are random utility functions that are _not_ called from probe context.
7023 dtrace_badattr(const dtrace_attribute_t *a)
7025 return (a->dtat_name > DTRACE_STABILITY_MAX ||
7026 a->dtat_data > DTRACE_STABILITY_MAX ||
7027 a->dtat_class > DTRACE_CLASS_MAX);
7031 * Return a duplicate copy of a string. If the specified string is NULL,
7032 * this function returns a zero-length string.
7035 dtrace_strdup(const char *str)
7037 char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
7040 (void) strcpy(new, str);
7045 #define DTRACE_ISALPHA(c) \
7046 (((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
7049 dtrace_badname(const char *s)
7053 if (s == NULL || (c = *s++) == '\0')
7056 if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
7059 while ((c = *s++) != '\0') {
7060 if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
7061 c != '-' && c != '_' && c != '.' && c != '`')
7069 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
7074 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
7076 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
7078 priv = DTRACE_PRIV_ALL;
7080 *uidp = crgetuid(cr);
7081 *zoneidp = crgetzoneid(cr);
7084 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
7085 priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
7086 else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
7087 priv |= DTRACE_PRIV_USER;
7088 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
7089 priv |= DTRACE_PRIV_PROC;
7090 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
7091 priv |= DTRACE_PRIV_OWNER;
7092 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
7093 priv |= DTRACE_PRIV_ZONEOWNER;
7096 priv = DTRACE_PRIV_ALL;
7102 #ifdef DTRACE_ERRDEBUG
7104 dtrace_errdebug(const char *str)
7106 int hval = dtrace_hash_str(str) % DTRACE_ERRHASHSZ;
7109 mutex_enter(&dtrace_errlock);
7110 dtrace_errlast = str;
7111 dtrace_errthread = curthread;
7113 while (occupied++ < DTRACE_ERRHASHSZ) {
7114 if (dtrace_errhash[hval].dter_msg == str) {
7115 dtrace_errhash[hval].dter_count++;
7119 if (dtrace_errhash[hval].dter_msg != NULL) {
7120 hval = (hval + 1) % DTRACE_ERRHASHSZ;
7124 dtrace_errhash[hval].dter_msg = str;
7125 dtrace_errhash[hval].dter_count = 1;
7129 panic("dtrace: undersized error hash");
7131 mutex_exit(&dtrace_errlock);
7136 * DTrace Matching Functions
7138 * These functions are used to match groups of probes, given some elements of
7139 * a probe tuple, or some globbed expressions for elements of a probe tuple.
7142 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
7145 if (priv != DTRACE_PRIV_ALL) {
7146 uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
7147 uint32_t match = priv & ppriv;
7150 * No PRIV_DTRACE_* privileges...
7152 if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
7153 DTRACE_PRIV_KERNEL)) == 0)
7157 * No matching bits, but there were bits to match...
7159 if (match == 0 && ppriv != 0)
7163 * Need to have permissions to the process, but don't...
7165 if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
7166 uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
7171 * Need to be in the same zone unless we possess the
7172 * privilege to examine all zones.
7174 if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
7175 zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
7184 * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
7185 * consists of input pattern strings and an ops-vector to evaluate them.
7186 * This function returns >0 for match, 0 for no match, and <0 for error.
7189 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
7190 uint32_t priv, uid_t uid, zoneid_t zoneid)
7192 dtrace_provider_t *pvp = prp->dtpr_provider;
7195 if (pvp->dtpv_defunct)
7198 if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
7201 if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
7204 if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
7207 if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
7210 if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
7217 * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
7218 * interface for matching a glob pattern 'p' to an input string 's'. Unlike
7219 * libc's version, the kernel version only applies to 8-bit ASCII strings.
7220 * In addition, all of the recursion cases except for '*' matching have been
7221 * unwound. For '*', we still implement recursive evaluation, but a depth
7222 * counter is maintained and matching is aborted if we recurse too deep.
7223 * The function returns 0 if no match, >0 if match, and <0 if recursion error.
7226 dtrace_match_glob(const char *s, const char *p, int depth)
7232 if (depth > DTRACE_PROBEKEY_MAXDEPTH)
7236 s = ""; /* treat NULL as empty string */
7245 if ((c = *p++) == '\0')
7246 return (s1 == '\0');
7250 int ok = 0, notflag = 0;
7261 if ((c = *p++) == '\0')
7265 if (c == '-' && lc != '\0' && *p != ']') {
7266 if ((c = *p++) == '\0')
7268 if (c == '\\' && (c = *p++) == '\0')
7272 if (s1 < lc || s1 > c)
7276 } else if (lc <= s1 && s1 <= c)
7279 } else if (c == '\\' && (c = *p++) == '\0')
7282 lc = c; /* save left-hand 'c' for next iteration */
7292 if ((c = *p++) == '\0')
7304 if ((c = *p++) == '\0')
7320 p++; /* consecutive *'s are identical to a single one */
7325 for (s = olds; *s != '\0'; s++) {
7326 if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
7336 dtrace_match_string(const char *s, const char *p, int depth)
7338 return (s != NULL && strcmp(s, p) == 0);
7343 dtrace_match_nul(const char *s, const char *p, int depth)
7345 return (1); /* always match the empty pattern */
7350 dtrace_match_nonzero(const char *s, const char *p, int depth)
7352 return (s != NULL && s[0] != '\0');
7356 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
7357 zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
7359 dtrace_probe_t template, *probe;
7360 dtrace_hash_t *hash = NULL;
7361 int len, best = INT_MAX, nmatched = 0;
7364 ASSERT(MUTEX_HELD(&dtrace_lock));
7367 * If the probe ID is specified in the key, just lookup by ID and
7368 * invoke the match callback once if a matching probe is found.
7370 if (pkp->dtpk_id != DTRACE_IDNONE) {
7371 if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
7372 dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
7373 (void) (*matched)(probe, arg);
7379 template.dtpr_mod = (char *)pkp->dtpk_mod;
7380 template.dtpr_func = (char *)pkp->dtpk_func;
7381 template.dtpr_name = (char *)pkp->dtpk_name;
7384 * We want to find the most distinct of the module name, function
7385 * name, and name. So for each one that is not a glob pattern or
7386 * empty string, we perform a lookup in the corresponding hash and
7387 * use the hash table with the fewest collisions to do our search.
7389 if (pkp->dtpk_mmatch == &dtrace_match_string &&
7390 (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
7392 hash = dtrace_bymod;
7395 if (pkp->dtpk_fmatch == &dtrace_match_string &&
7396 (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
7398 hash = dtrace_byfunc;
7401 if (pkp->dtpk_nmatch == &dtrace_match_string &&
7402 (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
7404 hash = dtrace_byname;
7408 * If we did not select a hash table, iterate over every probe and
7409 * invoke our callback for each one that matches our input probe key.
7412 for (i = 0; i < dtrace_nprobes; i++) {
7413 if ((probe = dtrace_probes[i]) == NULL ||
7414 dtrace_match_probe(probe, pkp, priv, uid,
7420 if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
7428 * If we selected a hash table, iterate over each probe of the same key
7429 * name and invoke the callback for every probe that matches the other
7430 * attributes of our input probe key.
7432 for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
7433 probe = *(DTRACE_HASHNEXT(hash, probe))) {
7435 if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
7440 if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
7448 * Return the function pointer dtrace_probecmp() should use to compare the
7449 * specified pattern with a string. For NULL or empty patterns, we select
7450 * dtrace_match_nul(). For glob pattern strings, we use dtrace_match_glob().
7451 * For non-empty non-glob strings, we use dtrace_match_string().
7453 static dtrace_probekey_f *
7454 dtrace_probekey_func(const char *p)
7458 if (p == NULL || *p == '\0')
7459 return (&dtrace_match_nul);
7461 while ((c = *p++) != '\0') {
7462 if (c == '[' || c == '?' || c == '*' || c == '\\')
7463 return (&dtrace_match_glob);
7466 return (&dtrace_match_string);
7470 * Build a probe comparison key for use with dtrace_match_probe() from the
7471 * given probe description. By convention, a null key only matches anchored
7472 * probes: if each field is the empty string, reset dtpk_fmatch to
7473 * dtrace_match_nonzero().
7476 dtrace_probekey(dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
7478 pkp->dtpk_prov = pdp->dtpd_provider;
7479 pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
7481 pkp->dtpk_mod = pdp->dtpd_mod;
7482 pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
7484 pkp->dtpk_func = pdp->dtpd_func;
7485 pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
7487 pkp->dtpk_name = pdp->dtpd_name;
7488 pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
7490 pkp->dtpk_id = pdp->dtpd_id;
7492 if (pkp->dtpk_id == DTRACE_IDNONE &&
7493 pkp->dtpk_pmatch == &dtrace_match_nul &&
7494 pkp->dtpk_mmatch == &dtrace_match_nul &&
7495 pkp->dtpk_fmatch == &dtrace_match_nul &&
7496 pkp->dtpk_nmatch == &dtrace_match_nul)
7497 pkp->dtpk_fmatch = &dtrace_match_nonzero;
7501 * DTrace Provider-to-Framework API Functions
7503 * These functions implement much of the Provider-to-Framework API, as
7504 * described in <sys/dtrace.h>. The parts of the API not in this section are
7505 * the functions in the API for probe management (found below), and
7506 * dtrace_probe() itself (found above).
7510 * Register the calling provider with the DTrace framework. This should
7511 * generally be called by DTrace providers in their attach(9E) entry point.
7514 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
7515 cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
7517 dtrace_provider_t *provider;
7519 if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
7520 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7521 "arguments", name ? name : "<NULL>");
7525 if (name[0] == '\0' || dtrace_badname(name)) {
7526 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7527 "provider name", name);
7531 if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
7532 pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
7533 pops->dtps_destroy == NULL ||
7534 ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
7535 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7536 "provider ops", name);
7540 if (dtrace_badattr(&pap->dtpa_provider) ||
7541 dtrace_badattr(&pap->dtpa_mod) ||
7542 dtrace_badattr(&pap->dtpa_func) ||
7543 dtrace_badattr(&pap->dtpa_name) ||
7544 dtrace_badattr(&pap->dtpa_args)) {
7545 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7546 "provider attributes", name);
7550 if (priv & ~DTRACE_PRIV_ALL) {
7551 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7552 "privilege attributes", name);
7556 if ((priv & DTRACE_PRIV_KERNEL) &&
7557 (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
7558 pops->dtps_usermode == NULL) {
7559 cmn_err(CE_WARN, "failed to register provider '%s': need "
7560 "dtps_usermode() op for given privilege attributes", name);
7564 provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
7565 provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
7566 (void) strcpy(provider->dtpv_name, name);
7568 provider->dtpv_attr = *pap;
7569 provider->dtpv_priv.dtpp_flags = priv;
7571 provider->dtpv_priv.dtpp_uid = crgetuid(cr);
7572 provider->dtpv_priv.dtpp_zoneid = crgetzoneid(cr);
7574 provider->dtpv_pops = *pops;
7576 if (pops->dtps_provide == NULL) {
7577 ASSERT(pops->dtps_provide_module != NULL);
7578 provider->dtpv_pops.dtps_provide =
7579 (void (*)(void *, dtrace_probedesc_t *))dtrace_nullop;
7582 if (pops->dtps_provide_module == NULL) {
7583 ASSERT(pops->dtps_provide != NULL);
7584 provider->dtpv_pops.dtps_provide_module =
7585 (void (*)(void *, modctl_t *))dtrace_nullop;
7588 if (pops->dtps_suspend == NULL) {
7589 ASSERT(pops->dtps_resume == NULL);
7590 provider->dtpv_pops.dtps_suspend =
7591 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7592 provider->dtpv_pops.dtps_resume =
7593 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7596 provider->dtpv_arg = arg;
7597 *idp = (dtrace_provider_id_t)provider;
7599 if (pops == &dtrace_provider_ops) {
7600 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7601 ASSERT(MUTEX_HELD(&dtrace_lock));
7602 ASSERT(dtrace_anon.dta_enabling == NULL);
7605 * We make sure that the DTrace provider is at the head of
7606 * the provider chain.
7608 provider->dtpv_next = dtrace_provider;
7609 dtrace_provider = provider;
7613 mutex_enter(&dtrace_provider_lock);
7614 mutex_enter(&dtrace_lock);
7617 * If there is at least one provider registered, we'll add this
7618 * provider after the first provider.
7620 if (dtrace_provider != NULL) {
7621 provider->dtpv_next = dtrace_provider->dtpv_next;
7622 dtrace_provider->dtpv_next = provider;
7624 dtrace_provider = provider;
7627 if (dtrace_retained != NULL) {
7628 dtrace_enabling_provide(provider);
7631 * Now we need to call dtrace_enabling_matchall() -- which
7632 * will acquire cpu_lock and dtrace_lock. We therefore need
7633 * to drop all of our locks before calling into it...
7635 mutex_exit(&dtrace_lock);
7636 mutex_exit(&dtrace_provider_lock);
7637 dtrace_enabling_matchall();
7642 mutex_exit(&dtrace_lock);
7643 mutex_exit(&dtrace_provider_lock);
7649 * Unregister the specified provider from the DTrace framework. This should
7650 * generally be called by DTrace providers in their detach(9E) entry point.
7653 dtrace_unregister(dtrace_provider_id_t id)
7655 dtrace_provider_t *old = (dtrace_provider_t *)id;
7656 dtrace_provider_t *prev = NULL;
7657 int i, self = 0, noreap = 0;
7658 dtrace_probe_t *probe, *first = NULL;
7660 if (old->dtpv_pops.dtps_enable ==
7661 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop) {
7663 * If DTrace itself is the provider, we're called with locks
7666 ASSERT(old == dtrace_provider);
7668 ASSERT(dtrace_devi != NULL);
7670 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7671 ASSERT(MUTEX_HELD(&dtrace_lock));
7674 if (dtrace_provider->dtpv_next != NULL) {
7676 * There's another provider here; return failure.
7681 mutex_enter(&dtrace_provider_lock);
7683 mutex_enter(&mod_lock);
7685 mutex_enter(&dtrace_lock);
7689 * If anyone has /dev/dtrace open, or if there are anonymous enabled
7690 * probes, we refuse to let providers slither away, unless this
7691 * provider has already been explicitly invalidated.
7693 if (!old->dtpv_defunct &&
7694 (dtrace_opens || (dtrace_anon.dta_state != NULL &&
7695 dtrace_anon.dta_state->dts_necbs > 0))) {
7697 mutex_exit(&dtrace_lock);
7699 mutex_exit(&mod_lock);
7701 mutex_exit(&dtrace_provider_lock);
7707 * Attempt to destroy the probes associated with this provider.
7709 for (i = 0; i < dtrace_nprobes; i++) {
7710 if ((probe = dtrace_probes[i]) == NULL)
7713 if (probe->dtpr_provider != old)
7716 if (probe->dtpr_ecb == NULL)
7720 * If we are trying to unregister a defunct provider, and the
7721 * provider was made defunct within the interval dictated by
7722 * dtrace_unregister_defunct_reap, we'll (asynchronously)
7723 * attempt to reap our enablings. To denote that the provider
7724 * should reattempt to unregister itself at some point in the
7725 * future, we will return a differentiable error code (EAGAIN
7726 * instead of EBUSY) in this case.
7728 if (dtrace_gethrtime() - old->dtpv_defunct >
7729 dtrace_unregister_defunct_reap)
7733 mutex_exit(&dtrace_lock);
7735 mutex_exit(&mod_lock);
7737 mutex_exit(&dtrace_provider_lock);
7743 (void) taskq_dispatch(dtrace_taskq,
7744 (task_func_t *)dtrace_enabling_reap, NULL, TQ_SLEEP);
7750 * All of the probes for this provider are disabled; we can safely
7751 * remove all of them from their hash chains and from the probe array.
7753 for (i = 0; i < dtrace_nprobes; i++) {
7754 if ((probe = dtrace_probes[i]) == NULL)
7757 if (probe->dtpr_provider != old)
7760 dtrace_probes[i] = NULL;
7762 dtrace_hash_remove(dtrace_bymod, probe);
7763 dtrace_hash_remove(dtrace_byfunc, probe);
7764 dtrace_hash_remove(dtrace_byname, probe);
7766 if (first == NULL) {
7768 probe->dtpr_nextmod = NULL;
7770 probe->dtpr_nextmod = first;
7776 * The provider's probes have been removed from the hash chains and
7777 * from the probe array. Now issue a dtrace_sync() to be sure that
7778 * everyone has cleared out from any probe array processing.
7782 for (probe = first; probe != NULL; probe = first) {
7783 first = probe->dtpr_nextmod;
7785 old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
7787 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7788 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7789 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7791 vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
7793 free_unr(dtrace_arena, probe->dtpr_id);
7795 kmem_free(probe, sizeof (dtrace_probe_t));
7798 if ((prev = dtrace_provider) == old) {
7800 ASSERT(self || dtrace_devi == NULL);
7801 ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
7803 dtrace_provider = old->dtpv_next;
7805 while (prev != NULL && prev->dtpv_next != old)
7806 prev = prev->dtpv_next;
7809 panic("attempt to unregister non-existent "
7810 "dtrace provider %p\n", (void *)id);
7813 prev->dtpv_next = old->dtpv_next;
7817 mutex_exit(&dtrace_lock);
7819 mutex_exit(&mod_lock);
7821 mutex_exit(&dtrace_provider_lock);
7824 kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
7825 kmem_free(old, sizeof (dtrace_provider_t));
7831 * Invalidate the specified provider. All subsequent probe lookups for the
7832 * specified provider will fail, but its probes will not be removed.
7835 dtrace_invalidate(dtrace_provider_id_t id)
7837 dtrace_provider_t *pvp = (dtrace_provider_t *)id;
7839 ASSERT(pvp->dtpv_pops.dtps_enable !=
7840 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
7842 mutex_enter(&dtrace_provider_lock);
7843 mutex_enter(&dtrace_lock);
7845 pvp->dtpv_defunct = dtrace_gethrtime();
7847 mutex_exit(&dtrace_lock);
7848 mutex_exit(&dtrace_provider_lock);
7852 * Indicate whether or not DTrace has attached.
7855 dtrace_attached(void)
7858 * dtrace_provider will be non-NULL iff the DTrace driver has
7859 * attached. (It's non-NULL because DTrace is always itself a
7862 return (dtrace_provider != NULL);
7866 * Remove all the unenabled probes for the given provider. This function is
7867 * not unlike dtrace_unregister(), except that it doesn't remove the provider
7868 * -- just as many of its associated probes as it can.
7871 dtrace_condense(dtrace_provider_id_t id)
7873 dtrace_provider_t *prov = (dtrace_provider_t *)id;
7875 dtrace_probe_t *probe;
7878 * Make sure this isn't the dtrace provider itself.
7880 ASSERT(prov->dtpv_pops.dtps_enable !=
7881 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
7883 mutex_enter(&dtrace_provider_lock);
7884 mutex_enter(&dtrace_lock);
7887 * Attempt to destroy the probes associated with this provider.
7889 for (i = 0; i < dtrace_nprobes; i++) {
7890 if ((probe = dtrace_probes[i]) == NULL)
7893 if (probe->dtpr_provider != prov)
7896 if (probe->dtpr_ecb != NULL)
7899 dtrace_probes[i] = NULL;
7901 dtrace_hash_remove(dtrace_bymod, probe);
7902 dtrace_hash_remove(dtrace_byfunc, probe);
7903 dtrace_hash_remove(dtrace_byname, probe);
7905 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
7907 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7908 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7909 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7910 kmem_free(probe, sizeof (dtrace_probe_t));
7912 vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
7914 free_unr(dtrace_arena, i + 1);
7918 mutex_exit(&dtrace_lock);
7919 mutex_exit(&dtrace_provider_lock);
7925 * DTrace Probe Management Functions
7927 * The functions in this section perform the DTrace probe management,
7928 * including functions to create probes, look-up probes, and call into the
7929 * providers to request that probes be provided. Some of these functions are
7930 * in the Provider-to-Framework API; these functions can be identified by the
7931 * fact that they are not declared "static".
7935 * Create a probe with the specified module name, function name, and name.
7938 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
7939 const char *func, const char *name, int aframes, void *arg)
7941 dtrace_probe_t *probe, **probes;
7942 dtrace_provider_t *provider = (dtrace_provider_t *)prov;
7945 if (provider == dtrace_provider) {
7946 ASSERT(MUTEX_HELD(&dtrace_lock));
7948 mutex_enter(&dtrace_lock);
7952 id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
7953 VM_BESTFIT | VM_SLEEP);
7955 id = alloc_unr(dtrace_arena);
7957 probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
7959 probe->dtpr_id = id;
7960 probe->dtpr_gen = dtrace_probegen++;
7961 probe->dtpr_mod = dtrace_strdup(mod);
7962 probe->dtpr_func = dtrace_strdup(func);
7963 probe->dtpr_name = dtrace_strdup(name);
7964 probe->dtpr_arg = arg;
7965 probe->dtpr_aframes = aframes;
7966 probe->dtpr_provider = provider;
7968 dtrace_hash_add(dtrace_bymod, probe);
7969 dtrace_hash_add(dtrace_byfunc, probe);
7970 dtrace_hash_add(dtrace_byname, probe);
7972 if (id - 1 >= dtrace_nprobes) {
7973 size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
7974 size_t nsize = osize << 1;
7978 ASSERT(dtrace_probes == NULL);
7979 nsize = sizeof (dtrace_probe_t *);
7982 probes = kmem_zalloc(nsize, KM_SLEEP);
7984 if (dtrace_probes == NULL) {
7986 dtrace_probes = probes;
7989 dtrace_probe_t **oprobes = dtrace_probes;
7991 bcopy(oprobes, probes, osize);
7992 dtrace_membar_producer();
7993 dtrace_probes = probes;
7998 * All CPUs are now seeing the new probes array; we can
7999 * safely free the old array.
8001 kmem_free(oprobes, osize);
8002 dtrace_nprobes <<= 1;
8005 ASSERT(id - 1 < dtrace_nprobes);
8008 ASSERT(dtrace_probes[id - 1] == NULL);
8009 dtrace_probes[id - 1] = probe;
8011 if (provider != dtrace_provider)
8012 mutex_exit(&dtrace_lock);
8017 static dtrace_probe_t *
8018 dtrace_probe_lookup_id(dtrace_id_t id)
8020 ASSERT(MUTEX_HELD(&dtrace_lock));
8022 if (id == 0 || id > dtrace_nprobes)
8025 return (dtrace_probes[id - 1]);
8029 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
8031 *((dtrace_id_t *)arg) = probe->dtpr_id;
8033 return (DTRACE_MATCH_DONE);
8037 * Look up a probe based on provider and one or more of module name, function
8038 * name and probe name.
8041 dtrace_probe_lookup(dtrace_provider_id_t prid, char *mod,
8042 char *func, char *name)
8044 dtrace_probekey_t pkey;
8048 pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
8049 pkey.dtpk_pmatch = &dtrace_match_string;
8050 pkey.dtpk_mod = mod;
8051 pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
8052 pkey.dtpk_func = func;
8053 pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
8054 pkey.dtpk_name = name;
8055 pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
8056 pkey.dtpk_id = DTRACE_IDNONE;
8058 mutex_enter(&dtrace_lock);
8059 match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
8060 dtrace_probe_lookup_match, &id);
8061 mutex_exit(&dtrace_lock);
8063 ASSERT(match == 1 || match == 0);
8064 return (match ? id : 0);
8068 * Returns the probe argument associated with the specified probe.
8071 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
8073 dtrace_probe_t *probe;
8076 mutex_enter(&dtrace_lock);
8078 if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
8079 probe->dtpr_provider == (dtrace_provider_t *)id)
8080 rval = probe->dtpr_arg;
8082 mutex_exit(&dtrace_lock);
8088 * Copy a probe into a probe description.
8091 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
8093 bzero(pdp, sizeof (dtrace_probedesc_t));
8094 pdp->dtpd_id = prp->dtpr_id;
8096 (void) strncpy(pdp->dtpd_provider,
8097 prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
8099 (void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
8100 (void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
8101 (void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
8105 * Called to indicate that a probe -- or probes -- should be provided by a
8106 * specfied provider. If the specified description is NULL, the provider will
8107 * be told to provide all of its probes. (This is done whenever a new
8108 * consumer comes along, or whenever a retained enabling is to be matched.) If
8109 * the specified description is non-NULL, the provider is given the
8110 * opportunity to dynamically provide the specified probe, allowing providers
8111 * to support the creation of probes on-the-fly. (So-called _autocreated_
8112 * probes.) If the provider is NULL, the operations will be applied to all
8113 * providers; if the provider is non-NULL the operations will only be applied
8114 * to the specified provider. The dtrace_provider_lock must be held, and the
8115 * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
8116 * will need to grab the dtrace_lock when it reenters the framework through
8117 * dtrace_probe_lookup(), dtrace_probe_create(), etc.
8120 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
8127 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8131 prv = dtrace_provider;
8136 * First, call the blanket provide operation.
8138 prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
8142 * Now call the per-module provide operation. We will grab
8143 * mod_lock to prevent the list from being modified. Note
8144 * that this also prevents the mod_busy bits from changing.
8145 * (mod_busy can only be changed with mod_lock held.)
8147 mutex_enter(&mod_lock);
8151 if (ctl->mod_busy || ctl->mod_mp == NULL)
8154 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
8156 } while ((ctl = ctl->mod_next) != &modules);
8158 mutex_exit(&mod_lock);
8160 } while (all && (prv = prv->dtpv_next) != NULL);
8165 * Iterate over each probe, and call the Framework-to-Provider API function
8169 dtrace_probe_foreach(uintptr_t offs)
8171 dtrace_provider_t *prov;
8172 void (*func)(void *, dtrace_id_t, void *);
8173 dtrace_probe_t *probe;
8174 dtrace_icookie_t cookie;
8178 * We disable interrupts to walk through the probe array. This is
8179 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
8180 * won't see stale data.
8182 cookie = dtrace_interrupt_disable();
8184 for (i = 0; i < dtrace_nprobes; i++) {
8185 if ((probe = dtrace_probes[i]) == NULL)
8188 if (probe->dtpr_ecb == NULL) {
8190 * This probe isn't enabled -- don't call the function.
8195 prov = probe->dtpr_provider;
8196 func = *((void(**)(void *, dtrace_id_t, void *))
8197 ((uintptr_t)&prov->dtpv_pops + offs));
8199 func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
8202 dtrace_interrupt_enable(cookie);
8207 dtrace_probe_enable(dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
8209 dtrace_probekey_t pkey;
8214 ASSERT(MUTEX_HELD(&dtrace_lock));
8215 dtrace_ecb_create_cache = NULL;
8219 * If we're passed a NULL description, we're being asked to
8220 * create an ECB with a NULL probe.
8222 (void) dtrace_ecb_create_enable(NULL, enab);
8226 dtrace_probekey(desc, &pkey);
8227 dtrace_cred2priv(enab->dten_vstate->dtvs_state->dts_cred.dcr_cred,
8228 &priv, &uid, &zoneid);
8230 return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
8235 * DTrace Helper Provider Functions
8238 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
8240 attr->dtat_name = DOF_ATTR_NAME(dofattr);
8241 attr->dtat_data = DOF_ATTR_DATA(dofattr);
8242 attr->dtat_class = DOF_ATTR_CLASS(dofattr);
8246 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
8247 const dof_provider_t *dofprov, char *strtab)
8249 hprov->dthpv_provname = strtab + dofprov->dofpv_name;
8250 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
8251 dofprov->dofpv_provattr);
8252 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
8253 dofprov->dofpv_modattr);
8254 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
8255 dofprov->dofpv_funcattr);
8256 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
8257 dofprov->dofpv_nameattr);
8258 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
8259 dofprov->dofpv_argsattr);
8263 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8265 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8266 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8267 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
8268 dof_provider_t *provider;
8270 uint32_t *off, *enoff;
8274 dtrace_helper_provdesc_t dhpv;
8275 dtrace_helper_probedesc_t dhpb;
8276 dtrace_meta_t *meta = dtrace_meta_pid;
8277 dtrace_mops_t *mops = &meta->dtm_mops;
8280 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8281 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8282 provider->dofpv_strtab * dof->dofh_secsize);
8283 prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8284 provider->dofpv_probes * dof->dofh_secsize);
8285 arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8286 provider->dofpv_prargs * dof->dofh_secsize);
8287 off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8288 provider->dofpv_proffs * dof->dofh_secsize);
8290 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8291 off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
8292 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
8296 * See dtrace_helper_provider_validate().
8298 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
8299 provider->dofpv_prenoffs != DOF_SECT_NONE) {
8300 enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8301 provider->dofpv_prenoffs * dof->dofh_secsize);
8302 enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
8305 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
8308 * Create the provider.
8310 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8312 if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
8318 * Create the probes.
8320 for (i = 0; i < nprobes; i++) {
8321 probe = (dof_probe_t *)(uintptr_t)(daddr +
8322 prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
8324 dhpb.dthpb_mod = dhp->dofhp_mod;
8325 dhpb.dthpb_func = strtab + probe->dofpr_func;
8326 dhpb.dthpb_name = strtab + probe->dofpr_name;
8327 dhpb.dthpb_base = probe->dofpr_addr;
8328 dhpb.dthpb_offs = off + probe->dofpr_offidx;
8329 dhpb.dthpb_noffs = probe->dofpr_noffs;
8330 if (enoff != NULL) {
8331 dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
8332 dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
8334 dhpb.dthpb_enoffs = NULL;
8335 dhpb.dthpb_nenoffs = 0;
8337 dhpb.dthpb_args = arg + probe->dofpr_argidx;
8338 dhpb.dthpb_nargc = probe->dofpr_nargc;
8339 dhpb.dthpb_xargc = probe->dofpr_xargc;
8340 dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
8341 dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
8343 mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
8348 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
8350 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8351 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8354 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8356 for (i = 0; i < dof->dofh_secnum; i++) {
8357 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8358 dof->dofh_secoff + i * dof->dofh_secsize);
8360 if (sec->dofs_type != DOF_SECT_PROVIDER)
8363 dtrace_helper_provide_one(dhp, sec, pid);
8367 * We may have just created probes, so we must now rematch against
8368 * any retained enablings. Note that this call will acquire both
8369 * cpu_lock and dtrace_lock; the fact that we are holding
8370 * dtrace_meta_lock now is what defines the ordering with respect to
8371 * these three locks.
8373 dtrace_enabling_matchall();
8377 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8379 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8380 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8382 dof_provider_t *provider;
8384 dtrace_helper_provdesc_t dhpv;
8385 dtrace_meta_t *meta = dtrace_meta_pid;
8386 dtrace_mops_t *mops = &meta->dtm_mops;
8388 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8389 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8390 provider->dofpv_strtab * dof->dofh_secsize);
8392 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8395 * Create the provider.
8397 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8399 mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
8405 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
8407 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8408 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8411 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8413 for (i = 0; i < dof->dofh_secnum; i++) {
8414 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8415 dof->dofh_secoff + i * dof->dofh_secsize);
8417 if (sec->dofs_type != DOF_SECT_PROVIDER)
8420 dtrace_helper_provider_remove_one(dhp, sec, pid);
8425 * DTrace Meta Provider-to-Framework API Functions
8427 * These functions implement the Meta Provider-to-Framework API, as described
8428 * in <sys/dtrace.h>.
8431 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
8432 dtrace_meta_provider_id_t *idp)
8434 dtrace_meta_t *meta;
8435 dtrace_helpers_t *help, *next;
8438 *idp = DTRACE_METAPROVNONE;
8441 * We strictly don't need the name, but we hold onto it for
8442 * debuggability. All hail error queues!
8445 cmn_err(CE_WARN, "failed to register meta-provider: "
8451 mops->dtms_create_probe == NULL ||
8452 mops->dtms_provide_pid == NULL ||
8453 mops->dtms_remove_pid == NULL) {
8454 cmn_err(CE_WARN, "failed to register meta-register %s: "
8455 "invalid ops", name);
8459 meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
8460 meta->dtm_mops = *mops;
8461 meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
8462 (void) strcpy(meta->dtm_name, name);
8463 meta->dtm_arg = arg;
8465 mutex_enter(&dtrace_meta_lock);
8466 mutex_enter(&dtrace_lock);
8468 if (dtrace_meta_pid != NULL) {
8469 mutex_exit(&dtrace_lock);
8470 mutex_exit(&dtrace_meta_lock);
8471 cmn_err(CE_WARN, "failed to register meta-register %s: "
8472 "user-land meta-provider exists", name);
8473 kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
8474 kmem_free(meta, sizeof (dtrace_meta_t));
8478 dtrace_meta_pid = meta;
8479 *idp = (dtrace_meta_provider_id_t)meta;
8482 * If there are providers and probes ready to go, pass them
8483 * off to the new meta provider now.
8486 help = dtrace_deferred_pid;
8487 dtrace_deferred_pid = NULL;
8489 mutex_exit(&dtrace_lock);
8491 while (help != NULL) {
8492 for (i = 0; i < help->dthps_nprovs; i++) {
8493 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
8497 next = help->dthps_next;
8498 help->dthps_next = NULL;
8499 help->dthps_prev = NULL;
8500 help->dthps_deferred = 0;
8504 mutex_exit(&dtrace_meta_lock);
8510 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
8512 dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
8514 mutex_enter(&dtrace_meta_lock);
8515 mutex_enter(&dtrace_lock);
8517 if (old == dtrace_meta_pid) {
8518 pp = &dtrace_meta_pid;
8520 panic("attempt to unregister non-existent "
8521 "dtrace meta-provider %p\n", (void *)old);
8524 if (old->dtm_count != 0) {
8525 mutex_exit(&dtrace_lock);
8526 mutex_exit(&dtrace_meta_lock);
8532 mutex_exit(&dtrace_lock);
8533 mutex_exit(&dtrace_meta_lock);
8535 kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
8536 kmem_free(old, sizeof (dtrace_meta_t));
8543 * DTrace DIF Object Functions
8546 dtrace_difo_err(uint_t pc, const char *format, ...)
8548 if (dtrace_err_verbose) {
8551 (void) uprintf("dtrace DIF object error: [%u]: ", pc);
8552 va_start(alist, format);
8553 (void) vuprintf(format, alist);
8557 #ifdef DTRACE_ERRDEBUG
8558 dtrace_errdebug(format);
8564 * Validate a DTrace DIF object by checking the IR instructions. The following
8565 * rules are currently enforced by dtrace_difo_validate():
8567 * 1. Each instruction must have a valid opcode
8568 * 2. Each register, string, variable, or subroutine reference must be valid
8569 * 3. No instruction can modify register %r0 (must be zero)
8570 * 4. All instruction reserved bits must be set to zero
8571 * 5. The last instruction must be a "ret" instruction
8572 * 6. All branch targets must reference a valid instruction _after_ the branch
8575 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
8579 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8583 kcheckload = cr == NULL ||
8584 (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
8586 dp->dtdo_destructive = 0;
8588 for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
8589 dif_instr_t instr = dp->dtdo_buf[pc];
8591 uint_t r1 = DIF_INSTR_R1(instr);
8592 uint_t r2 = DIF_INSTR_R2(instr);
8593 uint_t rd = DIF_INSTR_RD(instr);
8594 uint_t rs = DIF_INSTR_RS(instr);
8595 uint_t label = DIF_INSTR_LABEL(instr);
8596 uint_t v = DIF_INSTR_VAR(instr);
8597 uint_t subr = DIF_INSTR_SUBR(instr);
8598 uint_t type = DIF_INSTR_TYPE(instr);
8599 uint_t op = DIF_INSTR_OP(instr);
8617 err += efunc(pc, "invalid register %u\n", r1);
8619 err += efunc(pc, "invalid register %u\n", r2);
8621 err += efunc(pc, "invalid register %u\n", rd);
8623 err += efunc(pc, "cannot write to %r0\n");
8629 err += efunc(pc, "invalid register %u\n", r1);
8631 err += efunc(pc, "non-zero reserved bits\n");
8633 err += efunc(pc, "invalid register %u\n", rd);
8635 err += efunc(pc, "cannot write to %r0\n");
8645 err += efunc(pc, "invalid register %u\n", r1);
8647 err += efunc(pc, "non-zero reserved bits\n");
8649 err += efunc(pc, "invalid register %u\n", rd);
8651 err += efunc(pc, "cannot write to %r0\n");
8653 dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
8654 DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
8664 err += efunc(pc, "invalid register %u\n", r1);
8666 err += efunc(pc, "non-zero reserved bits\n");
8668 err += efunc(pc, "invalid register %u\n", rd);
8670 err += efunc(pc, "cannot write to %r0\n");
8680 err += efunc(pc, "invalid register %u\n", r1);
8682 err += efunc(pc, "non-zero reserved bits\n");
8684 err += efunc(pc, "invalid register %u\n", rd);
8686 err += efunc(pc, "cannot write to %r0\n");
8693 err += efunc(pc, "invalid register %u\n", r1);
8695 err += efunc(pc, "non-zero reserved bits\n");
8697 err += efunc(pc, "invalid register %u\n", rd);
8699 err += efunc(pc, "cannot write to 0 address\n");
8704 err += efunc(pc, "invalid register %u\n", r1);
8706 err += efunc(pc, "invalid register %u\n", r2);
8708 err += efunc(pc, "non-zero reserved bits\n");
8712 err += efunc(pc, "invalid register %u\n", r1);
8713 if (r2 != 0 || rd != 0)
8714 err += efunc(pc, "non-zero reserved bits\n");
8727 if (label >= dp->dtdo_len) {
8728 err += efunc(pc, "invalid branch target %u\n",
8732 err += efunc(pc, "backward branch to %u\n",
8737 if (r1 != 0 || r2 != 0)
8738 err += efunc(pc, "non-zero reserved bits\n");
8740 err += efunc(pc, "invalid register %u\n", rd);
8744 case DIF_OP_FLUSHTS:
8745 if (r1 != 0 || r2 != 0 || rd != 0)
8746 err += efunc(pc, "non-zero reserved bits\n");
8749 if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
8750 err += efunc(pc, "invalid integer ref %u\n",
8751 DIF_INSTR_INTEGER(instr));
8754 err += efunc(pc, "invalid register %u\n", rd);
8756 err += efunc(pc, "cannot write to %r0\n");
8759 if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
8760 err += efunc(pc, "invalid string ref %u\n",
8761 DIF_INSTR_STRING(instr));
8764 err += efunc(pc, "invalid register %u\n", rd);
8766 err += efunc(pc, "cannot write to %r0\n");
8770 if (r1 > DIF_VAR_ARRAY_MAX)
8771 err += efunc(pc, "invalid array %u\n", r1);
8773 err += efunc(pc, "invalid register %u\n", r2);
8775 err += efunc(pc, "invalid register %u\n", rd);
8777 err += efunc(pc, "cannot write to %r0\n");
8784 if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
8785 err += efunc(pc, "invalid variable %u\n", v);
8787 err += efunc(pc, "invalid register %u\n", rd);
8789 err += efunc(pc, "cannot write to %r0\n");
8796 if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
8797 err += efunc(pc, "invalid variable %u\n", v);
8799 err += efunc(pc, "invalid register %u\n", rd);
8802 if (subr > DIF_SUBR_MAX)
8803 err += efunc(pc, "invalid subr %u\n", subr);
8805 err += efunc(pc, "invalid register %u\n", rd);
8807 err += efunc(pc, "cannot write to %r0\n");
8809 if (subr == DIF_SUBR_COPYOUT ||
8810 subr == DIF_SUBR_COPYOUTSTR) {
8811 dp->dtdo_destructive = 1;
8815 if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
8816 err += efunc(pc, "invalid ref type %u\n", type);
8818 err += efunc(pc, "invalid register %u\n", r2);
8820 err += efunc(pc, "invalid register %u\n", rs);
8823 if (type != DIF_TYPE_CTF)
8824 err += efunc(pc, "invalid val type %u\n", type);
8826 err += efunc(pc, "invalid register %u\n", r2);
8828 err += efunc(pc, "invalid register %u\n", rs);
8831 err += efunc(pc, "invalid opcode %u\n",
8832 DIF_INSTR_OP(instr));
8836 if (dp->dtdo_len != 0 &&
8837 DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
8838 err += efunc(dp->dtdo_len - 1,
8839 "expected 'ret' as last DIF instruction\n");
8842 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) {
8844 * If we're not returning by reference, the size must be either
8845 * 0 or the size of one of the base types.
8847 switch (dp->dtdo_rtype.dtdt_size) {
8849 case sizeof (uint8_t):
8850 case sizeof (uint16_t):
8851 case sizeof (uint32_t):
8852 case sizeof (uint64_t):
8856 err += efunc(dp->dtdo_len - 1, "bad return size");
8860 for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
8861 dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
8862 dtrace_diftype_t *vt, *et;
8865 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
8866 v->dtdv_scope != DIFV_SCOPE_THREAD &&
8867 v->dtdv_scope != DIFV_SCOPE_LOCAL) {
8868 err += efunc(i, "unrecognized variable scope %d\n",
8873 if (v->dtdv_kind != DIFV_KIND_ARRAY &&
8874 v->dtdv_kind != DIFV_KIND_SCALAR) {
8875 err += efunc(i, "unrecognized variable type %d\n",
8880 if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
8881 err += efunc(i, "%d exceeds variable id limit\n", id);
8885 if (id < DIF_VAR_OTHER_UBASE)
8889 * For user-defined variables, we need to check that this
8890 * definition is identical to any previous definition that we
8893 ndx = id - DIF_VAR_OTHER_UBASE;
8895 switch (v->dtdv_scope) {
8896 case DIFV_SCOPE_GLOBAL:
8897 if (ndx < vstate->dtvs_nglobals) {
8898 dtrace_statvar_t *svar;
8900 if ((svar = vstate->dtvs_globals[ndx]) != NULL)
8901 existing = &svar->dtsv_var;
8906 case DIFV_SCOPE_THREAD:
8907 if (ndx < vstate->dtvs_ntlocals)
8908 existing = &vstate->dtvs_tlocals[ndx];
8911 case DIFV_SCOPE_LOCAL:
8912 if (ndx < vstate->dtvs_nlocals) {
8913 dtrace_statvar_t *svar;
8915 if ((svar = vstate->dtvs_locals[ndx]) != NULL)
8916 existing = &svar->dtsv_var;
8924 if (vt->dtdt_flags & DIF_TF_BYREF) {
8925 if (vt->dtdt_size == 0) {
8926 err += efunc(i, "zero-sized variable\n");
8930 if (v->dtdv_scope == DIFV_SCOPE_GLOBAL &&
8931 vt->dtdt_size > dtrace_global_maxsize) {
8932 err += efunc(i, "oversized by-ref global\n");
8937 if (existing == NULL || existing->dtdv_id == 0)
8940 ASSERT(existing->dtdv_id == v->dtdv_id);
8941 ASSERT(existing->dtdv_scope == v->dtdv_scope);
8943 if (existing->dtdv_kind != v->dtdv_kind)
8944 err += efunc(i, "%d changed variable kind\n", id);
8946 et = &existing->dtdv_type;
8948 if (vt->dtdt_flags != et->dtdt_flags) {
8949 err += efunc(i, "%d changed variable type flags\n", id);
8953 if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
8954 err += efunc(i, "%d changed variable type size\n", id);
8963 * Validate a DTrace DIF object that it is to be used as a helper. Helpers
8964 * are much more constrained than normal DIFOs. Specifically, they may
8967 * 1. Make calls to subroutines other than copyin(), copyinstr() or
8968 * miscellaneous string routines
8969 * 2. Access DTrace variables other than the args[] array, and the
8970 * curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
8971 * 3. Have thread-local variables.
8972 * 4. Have dynamic variables.
8975 dtrace_difo_validate_helper(dtrace_difo_t *dp)
8977 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8981 for (pc = 0; pc < dp->dtdo_len; pc++) {
8982 dif_instr_t instr = dp->dtdo_buf[pc];
8984 uint_t v = DIF_INSTR_VAR(instr);
8985 uint_t subr = DIF_INSTR_SUBR(instr);
8986 uint_t op = DIF_INSTR_OP(instr);
9041 case DIF_OP_FLUSHTS:
9053 if (v >= DIF_VAR_OTHER_UBASE)
9056 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
9059 if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
9060 v == DIF_VAR_PPID || v == DIF_VAR_TID ||
9061 v == DIF_VAR_EXECARGS ||
9062 v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
9063 v == DIF_VAR_UID || v == DIF_VAR_GID)
9066 err += efunc(pc, "illegal variable %u\n", v);
9073 err += efunc(pc, "illegal dynamic variable load\n");
9079 err += efunc(pc, "illegal dynamic variable store\n");
9083 if (subr == DIF_SUBR_ALLOCA ||
9084 subr == DIF_SUBR_BCOPY ||
9085 subr == DIF_SUBR_COPYIN ||
9086 subr == DIF_SUBR_COPYINTO ||
9087 subr == DIF_SUBR_COPYINSTR ||
9088 subr == DIF_SUBR_INDEX ||
9089 subr == DIF_SUBR_INET_NTOA ||
9090 subr == DIF_SUBR_INET_NTOA6 ||
9091 subr == DIF_SUBR_INET_NTOP ||
9092 subr == DIF_SUBR_LLTOSTR ||
9093 subr == DIF_SUBR_RINDEX ||
9094 subr == DIF_SUBR_STRCHR ||
9095 subr == DIF_SUBR_STRJOIN ||
9096 subr == DIF_SUBR_STRRCHR ||
9097 subr == DIF_SUBR_STRSTR ||
9098 subr == DIF_SUBR_HTONS ||
9099 subr == DIF_SUBR_HTONL ||
9100 subr == DIF_SUBR_HTONLL ||
9101 subr == DIF_SUBR_NTOHS ||
9102 subr == DIF_SUBR_NTOHL ||
9103 subr == DIF_SUBR_NTOHLL ||
9104 subr == DIF_SUBR_MEMREF ||
9105 subr == DIF_SUBR_TYPEREF)
9108 err += efunc(pc, "invalid subr %u\n", subr);
9112 err += efunc(pc, "invalid opcode %u\n",
9113 DIF_INSTR_OP(instr));
9121 * Returns 1 if the expression in the DIF object can be cached on a per-thread
9125 dtrace_difo_cacheable(dtrace_difo_t *dp)
9132 for (i = 0; i < dp->dtdo_varlen; i++) {
9133 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9135 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
9138 switch (v->dtdv_id) {
9139 case DIF_VAR_CURTHREAD:
9142 case DIF_VAR_EXECARGS:
9143 case DIF_VAR_EXECNAME:
9144 case DIF_VAR_ZONENAME:
9153 * This DIF object may be cacheable. Now we need to look for any
9154 * array loading instructions, any memory loading instructions, or
9155 * any stores to thread-local variables.
9157 for (i = 0; i < dp->dtdo_len; i++) {
9158 uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
9160 if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
9161 (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
9162 (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
9163 op == DIF_OP_LDGA || op == DIF_OP_STTS)
9171 dtrace_difo_hold(dtrace_difo_t *dp)
9175 ASSERT(MUTEX_HELD(&dtrace_lock));
9178 ASSERT(dp->dtdo_refcnt != 0);
9181 * We need to check this DIF object for references to the variable
9182 * DIF_VAR_VTIMESTAMP.
9184 for (i = 0; i < dp->dtdo_varlen; i++) {
9185 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9187 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9190 if (dtrace_vtime_references++ == 0)
9191 dtrace_vtime_enable();
9196 * This routine calculates the dynamic variable chunksize for a given DIF
9197 * object. The calculation is not fool-proof, and can probably be tricked by
9198 * malicious DIF -- but it works for all compiler-generated DIF. Because this
9199 * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
9200 * if a dynamic variable size exceeds the chunksize.
9203 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9206 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
9207 const dif_instr_t *text = dp->dtdo_buf;
9213 for (pc = 0; pc < dp->dtdo_len; pc++) {
9214 dif_instr_t instr = text[pc];
9215 uint_t op = DIF_INSTR_OP(instr);
9216 uint_t rd = DIF_INSTR_RD(instr);
9217 uint_t r1 = DIF_INSTR_R1(instr);
9221 dtrace_key_t *key = tupregs;
9225 sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
9230 key = &tupregs[DIF_DTR_NREGS];
9231 key[0].dttk_size = 0;
9232 key[1].dttk_size = 0;
9234 scope = DIFV_SCOPE_THREAD;
9241 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
9242 key[nkeys++].dttk_size = 0;
9244 key[nkeys++].dttk_size = 0;
9246 if (op == DIF_OP_STTAA) {
9247 scope = DIFV_SCOPE_THREAD;
9249 scope = DIFV_SCOPE_GLOBAL;
9255 if (ttop == DIF_DTR_NREGS)
9258 if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
9260 * If the register for the size of the "pushtr"
9261 * is %r0 (or the value is 0) and the type is
9262 * a string, we'll use the system-wide default
9265 tupregs[ttop++].dttk_size =
9266 dtrace_strsize_default;
9271 tupregs[ttop++].dttk_size = sval;
9277 if (ttop == DIF_DTR_NREGS)
9280 tupregs[ttop++].dttk_size = 0;
9283 case DIF_OP_FLUSHTS:
9300 * We have a dynamic variable allocation; calculate its size.
9302 for (ksize = 0, i = 0; i < nkeys; i++)
9303 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
9305 size = sizeof (dtrace_dynvar_t);
9306 size += sizeof (dtrace_key_t) * (nkeys - 1);
9310 * Now we need to determine the size of the stored data.
9312 id = DIF_INSTR_VAR(instr);
9314 for (i = 0; i < dp->dtdo_varlen; i++) {
9315 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9317 if (v->dtdv_id == id && v->dtdv_scope == scope) {
9318 size += v->dtdv_type.dtdt_size;
9323 if (i == dp->dtdo_varlen)
9327 * We have the size. If this is larger than the chunk size
9328 * for our dynamic variable state, reset the chunk size.
9330 size = P2ROUNDUP(size, sizeof (uint64_t));
9332 if (size > vstate->dtvs_dynvars.dtds_chunksize)
9333 vstate->dtvs_dynvars.dtds_chunksize = size;
9338 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9340 int i, oldsvars, osz, nsz, otlocals, ntlocals;
9343 ASSERT(MUTEX_HELD(&dtrace_lock));
9344 ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
9346 for (i = 0; i < dp->dtdo_varlen; i++) {
9347 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9348 dtrace_statvar_t *svar, ***svarp = NULL;
9350 uint8_t scope = v->dtdv_scope;
9353 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9356 id -= DIF_VAR_OTHER_UBASE;
9359 case DIFV_SCOPE_THREAD:
9360 while (id >= (otlocals = vstate->dtvs_ntlocals)) {
9361 dtrace_difv_t *tlocals;
9363 if ((ntlocals = (otlocals << 1)) == 0)
9366 osz = otlocals * sizeof (dtrace_difv_t);
9367 nsz = ntlocals * sizeof (dtrace_difv_t);
9369 tlocals = kmem_zalloc(nsz, KM_SLEEP);
9372 bcopy(vstate->dtvs_tlocals,
9374 kmem_free(vstate->dtvs_tlocals, osz);
9377 vstate->dtvs_tlocals = tlocals;
9378 vstate->dtvs_ntlocals = ntlocals;
9381 vstate->dtvs_tlocals[id] = *v;
9384 case DIFV_SCOPE_LOCAL:
9385 np = &vstate->dtvs_nlocals;
9386 svarp = &vstate->dtvs_locals;
9388 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9389 dsize = NCPU * (v->dtdv_type.dtdt_size +
9392 dsize = NCPU * sizeof (uint64_t);
9396 case DIFV_SCOPE_GLOBAL:
9397 np = &vstate->dtvs_nglobals;
9398 svarp = &vstate->dtvs_globals;
9400 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9401 dsize = v->dtdv_type.dtdt_size +
9410 while (id >= (oldsvars = *np)) {
9411 dtrace_statvar_t **statics;
9412 int newsvars, oldsize, newsize;
9414 if ((newsvars = (oldsvars << 1)) == 0)
9417 oldsize = oldsvars * sizeof (dtrace_statvar_t *);
9418 newsize = newsvars * sizeof (dtrace_statvar_t *);
9420 statics = kmem_zalloc(newsize, KM_SLEEP);
9423 bcopy(*svarp, statics, oldsize);
9424 kmem_free(*svarp, oldsize);
9431 if ((svar = (*svarp)[id]) == NULL) {
9432 svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
9433 svar->dtsv_var = *v;
9435 if ((svar->dtsv_size = dsize) != 0) {
9436 svar->dtsv_data = (uint64_t)(uintptr_t)
9437 kmem_zalloc(dsize, KM_SLEEP);
9440 (*svarp)[id] = svar;
9443 svar->dtsv_refcnt++;
9446 dtrace_difo_chunksize(dp, vstate);
9447 dtrace_difo_hold(dp);
9450 static dtrace_difo_t *
9451 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9456 ASSERT(dp->dtdo_buf != NULL);
9457 ASSERT(dp->dtdo_refcnt != 0);
9459 new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
9461 ASSERT(dp->dtdo_buf != NULL);
9462 sz = dp->dtdo_len * sizeof (dif_instr_t);
9463 new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
9464 bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
9465 new->dtdo_len = dp->dtdo_len;
9467 if (dp->dtdo_strtab != NULL) {
9468 ASSERT(dp->dtdo_strlen != 0);
9469 new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
9470 bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
9471 new->dtdo_strlen = dp->dtdo_strlen;
9474 if (dp->dtdo_inttab != NULL) {
9475 ASSERT(dp->dtdo_intlen != 0);
9476 sz = dp->dtdo_intlen * sizeof (uint64_t);
9477 new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
9478 bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
9479 new->dtdo_intlen = dp->dtdo_intlen;
9482 if (dp->dtdo_vartab != NULL) {
9483 ASSERT(dp->dtdo_varlen != 0);
9484 sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
9485 new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
9486 bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
9487 new->dtdo_varlen = dp->dtdo_varlen;
9490 dtrace_difo_init(new, vstate);
9495 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9499 ASSERT(dp->dtdo_refcnt == 0);
9501 for (i = 0; i < dp->dtdo_varlen; i++) {
9502 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9503 dtrace_statvar_t *svar, **svarp = NULL;
9505 uint8_t scope = v->dtdv_scope;
9509 case DIFV_SCOPE_THREAD:
9512 case DIFV_SCOPE_LOCAL:
9513 np = &vstate->dtvs_nlocals;
9514 svarp = vstate->dtvs_locals;
9517 case DIFV_SCOPE_GLOBAL:
9518 np = &vstate->dtvs_nglobals;
9519 svarp = vstate->dtvs_globals;
9526 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9529 id -= DIF_VAR_OTHER_UBASE;
9533 ASSERT(svar != NULL);
9534 ASSERT(svar->dtsv_refcnt > 0);
9536 if (--svar->dtsv_refcnt > 0)
9539 if (svar->dtsv_size != 0) {
9540 ASSERT(svar->dtsv_data != 0);
9541 kmem_free((void *)(uintptr_t)svar->dtsv_data,
9545 kmem_free(svar, sizeof (dtrace_statvar_t));
9549 if (dp->dtdo_buf != NULL)
9550 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
9551 if (dp->dtdo_inttab != NULL)
9552 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
9553 if (dp->dtdo_strtab != NULL)
9554 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
9555 if (dp->dtdo_vartab != NULL)
9556 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
9558 kmem_free(dp, sizeof (dtrace_difo_t));
9562 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9566 ASSERT(MUTEX_HELD(&dtrace_lock));
9567 ASSERT(dp->dtdo_refcnt != 0);
9569 for (i = 0; i < dp->dtdo_varlen; i++) {
9570 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9572 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9575 ASSERT(dtrace_vtime_references > 0);
9576 if (--dtrace_vtime_references == 0)
9577 dtrace_vtime_disable();
9580 if (--dp->dtdo_refcnt == 0)
9581 dtrace_difo_destroy(dp, vstate);
9585 * DTrace Format Functions
9588 dtrace_format_add(dtrace_state_t *state, char *str)
9591 uint16_t ndx, len = strlen(str) + 1;
9593 fmt = kmem_zalloc(len, KM_SLEEP);
9594 bcopy(str, fmt, len);
9596 for (ndx = 0; ndx < state->dts_nformats; ndx++) {
9597 if (state->dts_formats[ndx] == NULL) {
9598 state->dts_formats[ndx] = fmt;
9603 if (state->dts_nformats == USHRT_MAX) {
9605 * This is only likely if a denial-of-service attack is being
9606 * attempted. As such, it's okay to fail silently here.
9608 kmem_free(fmt, len);
9613 * For simplicity, we always resize the formats array to be exactly the
9614 * number of formats.
9616 ndx = state->dts_nformats++;
9617 new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
9619 if (state->dts_formats != NULL) {
9621 bcopy(state->dts_formats, new, ndx * sizeof (char *));
9622 kmem_free(state->dts_formats, ndx * sizeof (char *));
9625 state->dts_formats = new;
9626 state->dts_formats[ndx] = fmt;
9632 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
9636 ASSERT(state->dts_formats != NULL);
9637 ASSERT(format <= state->dts_nformats);
9638 ASSERT(state->dts_formats[format - 1] != NULL);
9640 fmt = state->dts_formats[format - 1];
9641 kmem_free(fmt, strlen(fmt) + 1);
9642 state->dts_formats[format - 1] = NULL;
9646 dtrace_format_destroy(dtrace_state_t *state)
9650 if (state->dts_nformats == 0) {
9651 ASSERT(state->dts_formats == NULL);
9655 ASSERT(state->dts_formats != NULL);
9657 for (i = 0; i < state->dts_nformats; i++) {
9658 char *fmt = state->dts_formats[i];
9663 kmem_free(fmt, strlen(fmt) + 1);
9666 kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
9667 state->dts_nformats = 0;
9668 state->dts_formats = NULL;
9672 * DTrace Predicate Functions
9674 static dtrace_predicate_t *
9675 dtrace_predicate_create(dtrace_difo_t *dp)
9677 dtrace_predicate_t *pred;
9679 ASSERT(MUTEX_HELD(&dtrace_lock));
9680 ASSERT(dp->dtdo_refcnt != 0);
9682 pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
9683 pred->dtp_difo = dp;
9684 pred->dtp_refcnt = 1;
9686 if (!dtrace_difo_cacheable(dp))
9689 if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
9691 * This is only theoretically possible -- we have had 2^32
9692 * cacheable predicates on this machine. We cannot allow any
9693 * more predicates to become cacheable: as unlikely as it is,
9694 * there may be a thread caching a (now stale) predicate cache
9695 * ID. (N.B.: the temptation is being successfully resisted to
9696 * have this cmn_err() "Holy shit -- we executed this code!")
9701 pred->dtp_cacheid = dtrace_predcache_id++;
9707 dtrace_predicate_hold(dtrace_predicate_t *pred)
9709 ASSERT(MUTEX_HELD(&dtrace_lock));
9710 ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
9711 ASSERT(pred->dtp_refcnt > 0);
9717 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
9719 dtrace_difo_t *dp = pred->dtp_difo;
9721 ASSERT(MUTEX_HELD(&dtrace_lock));
9722 ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
9723 ASSERT(pred->dtp_refcnt > 0);
9725 if (--pred->dtp_refcnt == 0) {
9726 dtrace_difo_release(pred->dtp_difo, vstate);
9727 kmem_free(pred, sizeof (dtrace_predicate_t));
9732 * DTrace Action Description Functions
9734 static dtrace_actdesc_t *
9735 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
9736 uint64_t uarg, uint64_t arg)
9738 dtrace_actdesc_t *act;
9741 ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
9742 arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
9745 act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
9746 act->dtad_kind = kind;
9747 act->dtad_ntuple = ntuple;
9748 act->dtad_uarg = uarg;
9749 act->dtad_arg = arg;
9750 act->dtad_refcnt = 1;
9756 dtrace_actdesc_hold(dtrace_actdesc_t *act)
9758 ASSERT(act->dtad_refcnt >= 1);
9763 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
9765 dtrace_actkind_t kind = act->dtad_kind;
9768 ASSERT(act->dtad_refcnt >= 1);
9770 if (--act->dtad_refcnt != 0)
9773 if ((dp = act->dtad_difo) != NULL)
9774 dtrace_difo_release(dp, vstate);
9776 if (DTRACEACT_ISPRINTFLIKE(kind)) {
9777 char *str = (char *)(uintptr_t)act->dtad_arg;
9780 ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
9781 (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
9785 kmem_free(str, strlen(str) + 1);
9788 kmem_free(act, sizeof (dtrace_actdesc_t));
9792 * DTrace ECB Functions
9794 static dtrace_ecb_t *
9795 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
9800 ASSERT(MUTEX_HELD(&dtrace_lock));
9802 ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
9803 ecb->dte_predicate = NULL;
9804 ecb->dte_probe = probe;
9807 * The default size is the size of the default action: recording
9810 ecb->dte_size = ecb->dte_needed = sizeof (dtrace_rechdr_t);
9811 ecb->dte_alignment = sizeof (dtrace_epid_t);
9813 epid = state->dts_epid++;
9815 if (epid - 1 >= state->dts_necbs) {
9816 dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
9817 int necbs = state->dts_necbs << 1;
9819 ASSERT(epid == state->dts_necbs + 1);
9822 ASSERT(oecbs == NULL);
9826 ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
9829 bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
9831 dtrace_membar_producer();
9832 state->dts_ecbs = ecbs;
9834 if (oecbs != NULL) {
9836 * If this state is active, we must dtrace_sync()
9837 * before we can free the old dts_ecbs array: we're
9838 * coming in hot, and there may be active ring
9839 * buffer processing (which indexes into the dts_ecbs
9840 * array) on another CPU.
9842 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
9845 kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
9848 dtrace_membar_producer();
9849 state->dts_necbs = necbs;
9852 ecb->dte_state = state;
9854 ASSERT(state->dts_ecbs[epid - 1] == NULL);
9855 dtrace_membar_producer();
9856 state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
9862 dtrace_ecb_enable(dtrace_ecb_t *ecb)
9864 dtrace_probe_t *probe = ecb->dte_probe;
9866 ASSERT(MUTEX_HELD(&cpu_lock));
9867 ASSERT(MUTEX_HELD(&dtrace_lock));
9868 ASSERT(ecb->dte_next == NULL);
9870 if (probe == NULL) {
9872 * This is the NULL probe -- there's nothing to do.
9877 if (probe->dtpr_ecb == NULL) {
9878 dtrace_provider_t *prov = probe->dtpr_provider;
9881 * We're the first ECB on this probe.
9883 probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
9885 if (ecb->dte_predicate != NULL)
9886 probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
9888 prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
9889 probe->dtpr_id, probe->dtpr_arg);
9892 * This probe is already active. Swing the last pointer to
9893 * point to the new ECB, and issue a dtrace_sync() to assure
9894 * that all CPUs have seen the change.
9896 ASSERT(probe->dtpr_ecb_last != NULL);
9897 probe->dtpr_ecb_last->dte_next = ecb;
9898 probe->dtpr_ecb_last = ecb;
9899 probe->dtpr_predcache = 0;
9906 dtrace_ecb_resize(dtrace_ecb_t *ecb)
9908 dtrace_action_t *act;
9909 uint32_t curneeded = UINT32_MAX;
9910 uint32_t aggbase = UINT32_MAX;
9913 * If we record anything, we always record the dtrace_rechdr_t. (And
9914 * we always record it first.)
9916 ecb->dte_size = sizeof (dtrace_rechdr_t);
9917 ecb->dte_alignment = sizeof (dtrace_epid_t);
9919 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
9920 dtrace_recdesc_t *rec = &act->dta_rec;
9921 ASSERT(rec->dtrd_size > 0 || rec->dtrd_alignment == 1);
9923 ecb->dte_alignment = MAX(ecb->dte_alignment,
9924 rec->dtrd_alignment);
9926 if (DTRACEACT_ISAGG(act->dta_kind)) {
9927 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
9929 ASSERT(rec->dtrd_size != 0);
9930 ASSERT(agg->dtag_first != NULL);
9931 ASSERT(act->dta_prev->dta_intuple);
9932 ASSERT(aggbase != UINT32_MAX);
9933 ASSERT(curneeded != UINT32_MAX);
9935 agg->dtag_base = aggbase;
9937 curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
9938 rec->dtrd_offset = curneeded;
9939 curneeded += rec->dtrd_size;
9940 ecb->dte_needed = MAX(ecb->dte_needed, curneeded);
9942 aggbase = UINT32_MAX;
9943 curneeded = UINT32_MAX;
9944 } else if (act->dta_intuple) {
9945 if (curneeded == UINT32_MAX) {
9947 * This is the first record in a tuple. Align
9948 * curneeded to be at offset 4 in an 8-byte
9951 ASSERT(act->dta_prev == NULL ||
9952 !act->dta_prev->dta_intuple);
9953 ASSERT3U(aggbase, ==, UINT32_MAX);
9954 curneeded = P2PHASEUP(ecb->dte_size,
9955 sizeof (uint64_t), sizeof (dtrace_aggid_t));
9957 aggbase = curneeded - sizeof (dtrace_aggid_t);
9958 ASSERT(IS_P2ALIGNED(aggbase,
9959 sizeof (uint64_t)));
9961 curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
9962 rec->dtrd_offset = curneeded;
9963 curneeded += rec->dtrd_size;
9965 /* tuples must be followed by an aggregation */
9966 ASSERT(act->dta_prev == NULL ||
9967 !act->dta_prev->dta_intuple);
9969 ecb->dte_size = P2ROUNDUP(ecb->dte_size,
9970 rec->dtrd_alignment);
9971 rec->dtrd_offset = ecb->dte_size;
9972 ecb->dte_size += rec->dtrd_size;
9973 ecb->dte_needed = MAX(ecb->dte_needed, ecb->dte_size);
9977 if ((act = ecb->dte_action) != NULL &&
9978 !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
9979 ecb->dte_size == sizeof (dtrace_rechdr_t)) {
9981 * If the size is still sizeof (dtrace_rechdr_t), then all
9982 * actions store no data; set the size to 0.
9987 ecb->dte_size = P2ROUNDUP(ecb->dte_size, sizeof (dtrace_epid_t));
9988 ecb->dte_needed = P2ROUNDUP(ecb->dte_needed, (sizeof (dtrace_epid_t)));
9989 ecb->dte_state->dts_needed = MAX(ecb->dte_state->dts_needed,
9993 static dtrace_action_t *
9994 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
9996 dtrace_aggregation_t *agg;
9997 size_t size = sizeof (uint64_t);
9998 int ntuple = desc->dtad_ntuple;
9999 dtrace_action_t *act;
10000 dtrace_recdesc_t *frec;
10001 dtrace_aggid_t aggid;
10002 dtrace_state_t *state = ecb->dte_state;
10004 agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
10005 agg->dtag_ecb = ecb;
10007 ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
10009 switch (desc->dtad_kind) {
10010 case DTRACEAGG_MIN:
10011 agg->dtag_initial = INT64_MAX;
10012 agg->dtag_aggregate = dtrace_aggregate_min;
10015 case DTRACEAGG_MAX:
10016 agg->dtag_initial = INT64_MIN;
10017 agg->dtag_aggregate = dtrace_aggregate_max;
10020 case DTRACEAGG_COUNT:
10021 agg->dtag_aggregate = dtrace_aggregate_count;
10024 case DTRACEAGG_QUANTIZE:
10025 agg->dtag_aggregate = dtrace_aggregate_quantize;
10026 size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
10030 case DTRACEAGG_LQUANTIZE: {
10031 uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
10032 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
10034 agg->dtag_initial = desc->dtad_arg;
10035 agg->dtag_aggregate = dtrace_aggregate_lquantize;
10037 if (step == 0 || levels == 0)
10040 size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
10044 case DTRACEAGG_LLQUANTIZE: {
10045 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(desc->dtad_arg);
10046 uint16_t low = DTRACE_LLQUANTIZE_LOW(desc->dtad_arg);
10047 uint16_t high = DTRACE_LLQUANTIZE_HIGH(desc->dtad_arg);
10048 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(desc->dtad_arg);
10051 agg->dtag_initial = desc->dtad_arg;
10052 agg->dtag_aggregate = dtrace_aggregate_llquantize;
10054 if (factor < 2 || low >= high || nsteps < factor)
10058 * Now check that the number of steps evenly divides a power
10059 * of the factor. (This assures both integer bucket size and
10060 * linearity within each magnitude.)
10062 for (v = factor; v < nsteps; v *= factor)
10065 if ((v % nsteps) || (nsteps % factor))
10068 size = (dtrace_aggregate_llquantize_bucket(factor,
10069 low, high, nsteps, INT64_MAX) + 2) * sizeof (uint64_t);
10073 case DTRACEAGG_AVG:
10074 agg->dtag_aggregate = dtrace_aggregate_avg;
10075 size = sizeof (uint64_t) * 2;
10078 case DTRACEAGG_STDDEV:
10079 agg->dtag_aggregate = dtrace_aggregate_stddev;
10080 size = sizeof (uint64_t) * 4;
10083 case DTRACEAGG_SUM:
10084 agg->dtag_aggregate = dtrace_aggregate_sum;
10091 agg->dtag_action.dta_rec.dtrd_size = size;
10097 * We must make sure that we have enough actions for the n-tuple.
10099 for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
10100 if (DTRACEACT_ISAGG(act->dta_kind))
10103 if (--ntuple == 0) {
10105 * This is the action with which our n-tuple begins.
10107 agg->dtag_first = act;
10113 * This n-tuple is short by ntuple elements. Return failure.
10115 ASSERT(ntuple != 0);
10117 kmem_free(agg, sizeof (dtrace_aggregation_t));
10122 * If the last action in the tuple has a size of zero, it's actually
10123 * an expression argument for the aggregating action.
10125 ASSERT(ecb->dte_action_last != NULL);
10126 act = ecb->dte_action_last;
10128 if (act->dta_kind == DTRACEACT_DIFEXPR) {
10129 ASSERT(act->dta_difo != NULL);
10131 if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
10132 agg->dtag_hasarg = 1;
10136 * We need to allocate an id for this aggregation.
10139 aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
10140 VM_BESTFIT | VM_SLEEP);
10142 aggid = alloc_unr(state->dts_aggid_arena);
10145 if (aggid - 1 >= state->dts_naggregations) {
10146 dtrace_aggregation_t **oaggs = state->dts_aggregations;
10147 dtrace_aggregation_t **aggs;
10148 int naggs = state->dts_naggregations << 1;
10149 int onaggs = state->dts_naggregations;
10151 ASSERT(aggid == state->dts_naggregations + 1);
10154 ASSERT(oaggs == NULL);
10158 aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
10160 if (oaggs != NULL) {
10161 bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
10162 kmem_free(oaggs, onaggs * sizeof (*aggs));
10165 state->dts_aggregations = aggs;
10166 state->dts_naggregations = naggs;
10169 ASSERT(state->dts_aggregations[aggid - 1] == NULL);
10170 state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
10172 frec = &agg->dtag_first->dta_rec;
10173 if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
10174 frec->dtrd_alignment = sizeof (dtrace_aggid_t);
10176 for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
10177 ASSERT(!act->dta_intuple);
10178 act->dta_intuple = 1;
10181 return (&agg->dtag_action);
10185 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
10187 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
10188 dtrace_state_t *state = ecb->dte_state;
10189 dtrace_aggid_t aggid = agg->dtag_id;
10191 ASSERT(DTRACEACT_ISAGG(act->dta_kind));
10193 vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
10195 free_unr(state->dts_aggid_arena, aggid);
10198 ASSERT(state->dts_aggregations[aggid - 1] == agg);
10199 state->dts_aggregations[aggid - 1] = NULL;
10201 kmem_free(agg, sizeof (dtrace_aggregation_t));
10205 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
10207 dtrace_action_t *action, *last;
10208 dtrace_difo_t *dp = desc->dtad_difo;
10209 uint32_t size = 0, align = sizeof (uint8_t), mask;
10210 uint16_t format = 0;
10211 dtrace_recdesc_t *rec;
10212 dtrace_state_t *state = ecb->dte_state;
10213 dtrace_optval_t *opt = state->dts_options, nframes = 0, strsize;
10214 uint64_t arg = desc->dtad_arg;
10216 ASSERT(MUTEX_HELD(&dtrace_lock));
10217 ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
10219 if (DTRACEACT_ISAGG(desc->dtad_kind)) {
10221 * If this is an aggregating action, there must be neither
10222 * a speculate nor a commit on the action chain.
10224 dtrace_action_t *act;
10226 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
10227 if (act->dta_kind == DTRACEACT_COMMIT)
10230 if (act->dta_kind == DTRACEACT_SPECULATE)
10234 action = dtrace_ecb_aggregation_create(ecb, desc);
10236 if (action == NULL)
10239 if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
10240 (desc->dtad_kind == DTRACEACT_DIFEXPR &&
10241 dp != NULL && dp->dtdo_destructive)) {
10242 state->dts_destructive = 1;
10245 switch (desc->dtad_kind) {
10246 case DTRACEACT_PRINTF:
10247 case DTRACEACT_PRINTA:
10248 case DTRACEACT_SYSTEM:
10249 case DTRACEACT_FREOPEN:
10250 case DTRACEACT_DIFEXPR:
10252 * We know that our arg is a string -- turn it into a
10256 ASSERT(desc->dtad_kind == DTRACEACT_PRINTA ||
10257 desc->dtad_kind == DTRACEACT_DIFEXPR);
10262 ASSERT(arg > KERNELBASE);
10264 format = dtrace_format_add(state,
10265 (char *)(uintptr_t)arg);
10269 case DTRACEACT_LIBACT:
10270 case DTRACEACT_TRACEMEM:
10271 case DTRACEACT_TRACEMEM_DYNSIZE:
10275 if ((size = dp->dtdo_rtype.dtdt_size) != 0)
10278 if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
10279 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10282 size = opt[DTRACEOPT_STRSIZE];
10287 case DTRACEACT_STACK:
10288 if ((nframes = arg) == 0) {
10289 nframes = opt[DTRACEOPT_STACKFRAMES];
10290 ASSERT(nframes > 0);
10294 size = nframes * sizeof (pc_t);
10297 case DTRACEACT_JSTACK:
10298 if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
10299 strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
10301 if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
10302 nframes = opt[DTRACEOPT_JSTACKFRAMES];
10304 arg = DTRACE_USTACK_ARG(nframes, strsize);
10307 case DTRACEACT_USTACK:
10308 if (desc->dtad_kind != DTRACEACT_JSTACK &&
10309 (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
10310 strsize = DTRACE_USTACK_STRSIZE(arg);
10311 nframes = opt[DTRACEOPT_USTACKFRAMES];
10312 ASSERT(nframes > 0);
10313 arg = DTRACE_USTACK_ARG(nframes, strsize);
10317 * Save a slot for the pid.
10319 size = (nframes + 1) * sizeof (uint64_t);
10320 size += DTRACE_USTACK_STRSIZE(arg);
10321 size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
10325 case DTRACEACT_SYM:
10326 case DTRACEACT_MOD:
10327 if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
10328 sizeof (uint64_t)) ||
10329 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10333 case DTRACEACT_USYM:
10334 case DTRACEACT_UMOD:
10335 case DTRACEACT_UADDR:
10337 (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
10338 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10342 * We have a slot for the pid, plus a slot for the
10343 * argument. To keep things simple (aligned with
10344 * bitness-neutral sizing), we store each as a 64-bit
10347 size = 2 * sizeof (uint64_t);
10350 case DTRACEACT_STOP:
10351 case DTRACEACT_BREAKPOINT:
10352 case DTRACEACT_PANIC:
10355 case DTRACEACT_CHILL:
10356 case DTRACEACT_DISCARD:
10357 case DTRACEACT_RAISE:
10362 case DTRACEACT_EXIT:
10364 (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
10365 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10369 case DTRACEACT_SPECULATE:
10370 if (ecb->dte_size > sizeof (dtrace_rechdr_t))
10376 state->dts_speculates = 1;
10379 case DTRACEACT_PRINTM:
10380 size = dp->dtdo_rtype.dtdt_size;
10383 case DTRACEACT_PRINTT:
10384 size = dp->dtdo_rtype.dtdt_size;
10387 case DTRACEACT_COMMIT: {
10388 dtrace_action_t *act = ecb->dte_action;
10390 for (; act != NULL; act = act->dta_next) {
10391 if (act->dta_kind == DTRACEACT_COMMIT)
10404 if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
10406 * If this is a data-storing action or a speculate,
10407 * we must be sure that there isn't a commit on the
10410 dtrace_action_t *act = ecb->dte_action;
10412 for (; act != NULL; act = act->dta_next) {
10413 if (act->dta_kind == DTRACEACT_COMMIT)
10418 action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
10419 action->dta_rec.dtrd_size = size;
10422 action->dta_refcnt = 1;
10423 rec = &action->dta_rec;
10424 size = rec->dtrd_size;
10426 for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
10427 if (!(size & mask)) {
10433 action->dta_kind = desc->dtad_kind;
10435 if ((action->dta_difo = dp) != NULL)
10436 dtrace_difo_hold(dp);
10438 rec->dtrd_action = action->dta_kind;
10439 rec->dtrd_arg = arg;
10440 rec->dtrd_uarg = desc->dtad_uarg;
10441 rec->dtrd_alignment = (uint16_t)align;
10442 rec->dtrd_format = format;
10444 if ((last = ecb->dte_action_last) != NULL) {
10445 ASSERT(ecb->dte_action != NULL);
10446 action->dta_prev = last;
10447 last->dta_next = action;
10449 ASSERT(ecb->dte_action == NULL);
10450 ecb->dte_action = action;
10453 ecb->dte_action_last = action;
10459 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
10461 dtrace_action_t *act = ecb->dte_action, *next;
10462 dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
10466 if (act != NULL && act->dta_refcnt > 1) {
10467 ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
10470 for (; act != NULL; act = next) {
10471 next = act->dta_next;
10472 ASSERT(next != NULL || act == ecb->dte_action_last);
10473 ASSERT(act->dta_refcnt == 1);
10475 if ((format = act->dta_rec.dtrd_format) != 0)
10476 dtrace_format_remove(ecb->dte_state, format);
10478 if ((dp = act->dta_difo) != NULL)
10479 dtrace_difo_release(dp, vstate);
10481 if (DTRACEACT_ISAGG(act->dta_kind)) {
10482 dtrace_ecb_aggregation_destroy(ecb, act);
10484 kmem_free(act, sizeof (dtrace_action_t));
10489 ecb->dte_action = NULL;
10490 ecb->dte_action_last = NULL;
10495 dtrace_ecb_disable(dtrace_ecb_t *ecb)
10498 * We disable the ECB by removing it from its probe.
10500 dtrace_ecb_t *pecb, *prev = NULL;
10501 dtrace_probe_t *probe = ecb->dte_probe;
10503 ASSERT(MUTEX_HELD(&dtrace_lock));
10505 if (probe == NULL) {
10507 * This is the NULL probe; there is nothing to disable.
10512 for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
10518 ASSERT(pecb != NULL);
10520 if (prev == NULL) {
10521 probe->dtpr_ecb = ecb->dte_next;
10523 prev->dte_next = ecb->dte_next;
10526 if (ecb == probe->dtpr_ecb_last) {
10527 ASSERT(ecb->dte_next == NULL);
10528 probe->dtpr_ecb_last = prev;
10532 * The ECB has been disconnected from the probe; now sync to assure
10533 * that all CPUs have seen the change before returning.
10537 if (probe->dtpr_ecb == NULL) {
10539 * That was the last ECB on the probe; clear the predicate
10540 * cache ID for the probe, disable it and sync one more time
10541 * to assure that we'll never hit it again.
10543 dtrace_provider_t *prov = probe->dtpr_provider;
10545 ASSERT(ecb->dte_next == NULL);
10546 ASSERT(probe->dtpr_ecb_last == NULL);
10547 probe->dtpr_predcache = DTRACE_CACHEIDNONE;
10548 prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
10549 probe->dtpr_id, probe->dtpr_arg);
10553 * There is at least one ECB remaining on the probe. If there
10554 * is _exactly_ one, set the probe's predicate cache ID to be
10555 * the predicate cache ID of the remaining ECB.
10557 ASSERT(probe->dtpr_ecb_last != NULL);
10558 ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
10560 if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
10561 dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
10563 ASSERT(probe->dtpr_ecb->dte_next == NULL);
10566 probe->dtpr_predcache = p->dtp_cacheid;
10569 ecb->dte_next = NULL;
10574 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
10576 dtrace_state_t *state = ecb->dte_state;
10577 dtrace_vstate_t *vstate = &state->dts_vstate;
10578 dtrace_predicate_t *pred;
10579 dtrace_epid_t epid = ecb->dte_epid;
10581 ASSERT(MUTEX_HELD(&dtrace_lock));
10582 ASSERT(ecb->dte_next == NULL);
10583 ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
10585 if ((pred = ecb->dte_predicate) != NULL)
10586 dtrace_predicate_release(pred, vstate);
10588 dtrace_ecb_action_remove(ecb);
10590 ASSERT(state->dts_ecbs[epid - 1] == ecb);
10591 state->dts_ecbs[epid - 1] = NULL;
10593 kmem_free(ecb, sizeof (dtrace_ecb_t));
10596 static dtrace_ecb_t *
10597 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
10598 dtrace_enabling_t *enab)
10601 dtrace_predicate_t *pred;
10602 dtrace_actdesc_t *act;
10603 dtrace_provider_t *prov;
10604 dtrace_ecbdesc_t *desc = enab->dten_current;
10606 ASSERT(MUTEX_HELD(&dtrace_lock));
10607 ASSERT(state != NULL);
10609 ecb = dtrace_ecb_add(state, probe);
10610 ecb->dte_uarg = desc->dted_uarg;
10612 if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
10613 dtrace_predicate_hold(pred);
10614 ecb->dte_predicate = pred;
10617 if (probe != NULL) {
10619 * If the provider shows more leg than the consumer is old
10620 * enough to see, we need to enable the appropriate implicit
10621 * predicate bits to prevent the ecb from activating at
10624 * Providers specifying DTRACE_PRIV_USER at register time
10625 * are stating that they need the /proc-style privilege
10626 * model to be enforced, and this is what DTRACE_COND_OWNER
10627 * and DTRACE_COND_ZONEOWNER will then do at probe time.
10629 prov = probe->dtpr_provider;
10630 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
10631 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10632 ecb->dte_cond |= DTRACE_COND_OWNER;
10634 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
10635 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10636 ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
10639 * If the provider shows us kernel innards and the user
10640 * is lacking sufficient privilege, enable the
10641 * DTRACE_COND_USERMODE implicit predicate.
10643 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
10644 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
10645 ecb->dte_cond |= DTRACE_COND_USERMODE;
10648 if (dtrace_ecb_create_cache != NULL) {
10650 * If we have a cached ecb, we'll use its action list instead
10651 * of creating our own (saving both time and space).
10653 dtrace_ecb_t *cached = dtrace_ecb_create_cache;
10654 dtrace_action_t *act = cached->dte_action;
10657 ASSERT(act->dta_refcnt > 0);
10659 ecb->dte_action = act;
10660 ecb->dte_action_last = cached->dte_action_last;
10661 ecb->dte_needed = cached->dte_needed;
10662 ecb->dte_size = cached->dte_size;
10663 ecb->dte_alignment = cached->dte_alignment;
10669 for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
10670 if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
10671 dtrace_ecb_destroy(ecb);
10676 dtrace_ecb_resize(ecb);
10678 return (dtrace_ecb_create_cache = ecb);
10682 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
10685 dtrace_enabling_t *enab = arg;
10686 dtrace_state_t *state = enab->dten_vstate->dtvs_state;
10688 ASSERT(state != NULL);
10690 if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
10692 * This probe was created in a generation for which this
10693 * enabling has previously created ECBs; we don't want to
10694 * enable it again, so just kick out.
10696 return (DTRACE_MATCH_NEXT);
10699 if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
10700 return (DTRACE_MATCH_DONE);
10702 dtrace_ecb_enable(ecb);
10703 return (DTRACE_MATCH_NEXT);
10706 static dtrace_ecb_t *
10707 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
10711 ASSERT(MUTEX_HELD(&dtrace_lock));
10713 if (id == 0 || id > state->dts_necbs)
10716 ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
10717 ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
10719 return (state->dts_ecbs[id - 1]);
10722 static dtrace_aggregation_t *
10723 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
10725 dtrace_aggregation_t *agg;
10727 ASSERT(MUTEX_HELD(&dtrace_lock));
10729 if (id == 0 || id > state->dts_naggregations)
10732 ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
10733 ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
10734 agg->dtag_id == id);
10736 return (state->dts_aggregations[id - 1]);
10740 * DTrace Buffer Functions
10742 * The following functions manipulate DTrace buffers. Most of these functions
10743 * are called in the context of establishing or processing consumer state;
10744 * exceptions are explicitly noted.
10748 * Note: called from cross call context. This function switches the two
10749 * buffers on a given CPU. The atomicity of this operation is assured by
10750 * disabling interrupts while the actual switch takes place; the disabling of
10751 * interrupts serializes the execution with any execution of dtrace_probe() on
10755 dtrace_buffer_switch(dtrace_buffer_t *buf)
10757 caddr_t tomax = buf->dtb_tomax;
10758 caddr_t xamot = buf->dtb_xamot;
10759 dtrace_icookie_t cookie;
10762 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
10763 ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
10765 cookie = dtrace_interrupt_disable();
10766 now = dtrace_gethrtime();
10767 buf->dtb_tomax = xamot;
10768 buf->dtb_xamot = tomax;
10769 buf->dtb_xamot_drops = buf->dtb_drops;
10770 buf->dtb_xamot_offset = buf->dtb_offset;
10771 buf->dtb_xamot_errors = buf->dtb_errors;
10772 buf->dtb_xamot_flags = buf->dtb_flags;
10773 buf->dtb_offset = 0;
10774 buf->dtb_drops = 0;
10775 buf->dtb_errors = 0;
10776 buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
10777 buf->dtb_interval = now - buf->dtb_switched;
10778 buf->dtb_switched = now;
10779 dtrace_interrupt_enable(cookie);
10783 * Note: called from cross call context. This function activates a buffer
10784 * on a CPU. As with dtrace_buffer_switch(), the atomicity of the operation
10785 * is guaranteed by the disabling of interrupts.
10788 dtrace_buffer_activate(dtrace_state_t *state)
10790 dtrace_buffer_t *buf;
10791 dtrace_icookie_t cookie = dtrace_interrupt_disable();
10793 buf = &state->dts_buffer[curcpu];
10795 if (buf->dtb_tomax != NULL) {
10797 * We might like to assert that the buffer is marked inactive,
10798 * but this isn't necessarily true: the buffer for the CPU
10799 * that processes the BEGIN probe has its buffer activated
10800 * manually. In this case, we take the (harmless) action
10801 * re-clearing the bit INACTIVE bit.
10803 buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
10806 dtrace_interrupt_enable(cookie);
10810 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
10816 dtrace_buffer_t *buf;
10819 ASSERT(MUTEX_HELD(&cpu_lock));
10820 ASSERT(MUTEX_HELD(&dtrace_lock));
10822 if (size > dtrace_nonroot_maxsize &&
10823 !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
10829 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10832 buf = &bufs[cp->cpu_id];
10835 * If there is already a buffer allocated for this CPU, it
10836 * is only possible that this is a DR event. In this case,
10838 if (buf->dtb_tomax != NULL) {
10839 ASSERT(buf->dtb_size == size);
10843 ASSERT(buf->dtb_xamot == NULL);
10845 if ((buf->dtb_tomax = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10848 buf->dtb_size = size;
10849 buf->dtb_flags = flags;
10850 buf->dtb_offset = 0;
10851 buf->dtb_drops = 0;
10853 if (flags & DTRACEBUF_NOSWITCH)
10856 if ((buf->dtb_xamot = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10858 } while ((cp = cp->cpu_next) != cpu_list);
10866 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10869 buf = &bufs[cp->cpu_id];
10871 if (buf->dtb_xamot != NULL) {
10872 ASSERT(buf->dtb_tomax != NULL);
10873 ASSERT(buf->dtb_size == size);
10874 kmem_free(buf->dtb_xamot, size);
10877 if (buf->dtb_tomax != NULL) {
10878 ASSERT(buf->dtb_size == size);
10879 kmem_free(buf->dtb_tomax, size);
10882 buf->dtb_tomax = NULL;
10883 buf->dtb_xamot = NULL;
10885 } while ((cp = cp->cpu_next) != cpu_list);
10891 #if defined(__amd64__) || defined(__mips__) || defined(__powerpc__)
10893 * FreeBSD isn't good at limiting the amount of memory we
10894 * ask to malloc, so let's place a limit here before trying
10895 * to do something that might well end in tears at bedtime.
10897 if (size > physmem * PAGE_SIZE / (128 * (mp_maxid + 1)))
10901 ASSERT(MUTEX_HELD(&dtrace_lock));
10903 if (cpu != DTRACE_CPUALL && cpu != i)
10909 * If there is already a buffer allocated for this CPU, it
10910 * is only possible that this is a DR event. In this case,
10911 * the buffer size must match our specified size.
10913 if (buf->dtb_tomax != NULL) {
10914 ASSERT(buf->dtb_size == size);
10918 ASSERT(buf->dtb_xamot == NULL);
10920 if ((buf->dtb_tomax = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10923 buf->dtb_size = size;
10924 buf->dtb_flags = flags;
10925 buf->dtb_offset = 0;
10926 buf->dtb_drops = 0;
10928 if (flags & DTRACEBUF_NOSWITCH)
10931 if ((buf->dtb_xamot = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10939 * Error allocating memory, so free the buffers that were
10940 * allocated before the failed allocation.
10943 if (cpu != DTRACE_CPUALL && cpu != i)
10948 if (buf->dtb_xamot != NULL) {
10949 ASSERT(buf->dtb_tomax != NULL);
10950 ASSERT(buf->dtb_size == size);
10951 kmem_free(buf->dtb_xamot, size);
10954 if (buf->dtb_tomax != NULL) {
10955 ASSERT(buf->dtb_size == size);
10956 kmem_free(buf->dtb_tomax, size);
10959 buf->dtb_tomax = NULL;
10960 buf->dtb_xamot = NULL;
10970 * Note: called from probe context. This function just increments the drop
10971 * count on a buffer. It has been made a function to allow for the
10972 * possibility of understanding the source of mysterious drop counts. (A
10973 * problem for which one may be particularly disappointed that DTrace cannot
10974 * be used to understand DTrace.)
10977 dtrace_buffer_drop(dtrace_buffer_t *buf)
10983 * Note: called from probe context. This function is called to reserve space
10984 * in a buffer. If mstate is non-NULL, sets the scratch base and size in the
10985 * mstate. Returns the new offset in the buffer, or a negative value if an
10986 * error has occurred.
10989 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
10990 dtrace_state_t *state, dtrace_mstate_t *mstate)
10992 intptr_t offs = buf->dtb_offset, soffs;
10997 if (buf->dtb_flags & DTRACEBUF_INACTIVE)
11000 if ((tomax = buf->dtb_tomax) == NULL) {
11001 dtrace_buffer_drop(buf);
11005 if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
11006 while (offs & (align - 1)) {
11008 * Assert that our alignment is off by a number which
11009 * is itself sizeof (uint32_t) aligned.
11011 ASSERT(!((align - (offs & (align - 1))) &
11012 (sizeof (uint32_t) - 1)));
11013 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
11014 offs += sizeof (uint32_t);
11017 if ((soffs = offs + needed) > buf->dtb_size) {
11018 dtrace_buffer_drop(buf);
11022 if (mstate == NULL)
11025 mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
11026 mstate->dtms_scratch_size = buf->dtb_size - soffs;
11027 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11032 if (buf->dtb_flags & DTRACEBUF_FILL) {
11033 if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
11034 (buf->dtb_flags & DTRACEBUF_FULL))
11039 total = needed + (offs & (align - 1));
11042 * For a ring buffer, life is quite a bit more complicated. Before
11043 * we can store any padding, we need to adjust our wrapping offset.
11044 * (If we've never before wrapped or we're not about to, no adjustment
11047 if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
11048 offs + total > buf->dtb_size) {
11049 woffs = buf->dtb_xamot_offset;
11051 if (offs + total > buf->dtb_size) {
11053 * We can't fit in the end of the buffer. First, a
11054 * sanity check that we can fit in the buffer at all.
11056 if (total > buf->dtb_size) {
11057 dtrace_buffer_drop(buf);
11062 * We're going to be storing at the top of the buffer,
11063 * so now we need to deal with the wrapped offset. We
11064 * only reset our wrapped offset to 0 if it is
11065 * currently greater than the current offset. If it
11066 * is less than the current offset, it is because a
11067 * previous allocation induced a wrap -- but the
11068 * allocation didn't subsequently take the space due
11069 * to an error or false predicate evaluation. In this
11070 * case, we'll just leave the wrapped offset alone: if
11071 * the wrapped offset hasn't been advanced far enough
11072 * for this allocation, it will be adjusted in the
11075 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
11083 * Now we know that we're going to be storing to the
11084 * top of the buffer and that there is room for us
11085 * there. We need to clear the buffer from the current
11086 * offset to the end (there may be old gunk there).
11088 while (offs < buf->dtb_size)
11092 * We need to set our offset to zero. And because we
11093 * are wrapping, we need to set the bit indicating as
11094 * much. We can also adjust our needed space back
11095 * down to the space required by the ECB -- we know
11096 * that the top of the buffer is aligned.
11100 buf->dtb_flags |= DTRACEBUF_WRAPPED;
11103 * There is room for us in the buffer, so we simply
11104 * need to check the wrapped offset.
11106 if (woffs < offs) {
11108 * The wrapped offset is less than the offset.
11109 * This can happen if we allocated buffer space
11110 * that induced a wrap, but then we didn't
11111 * subsequently take the space due to an error
11112 * or false predicate evaluation. This is
11113 * okay; we know that _this_ allocation isn't
11114 * going to induce a wrap. We still can't
11115 * reset the wrapped offset to be zero,
11116 * however: the space may have been trashed in
11117 * the previous failed probe attempt. But at
11118 * least the wrapped offset doesn't need to
11119 * be adjusted at all...
11125 while (offs + total > woffs) {
11126 dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
11129 if (epid == DTRACE_EPIDNONE) {
11130 size = sizeof (uint32_t);
11132 ASSERT3U(epid, <=, state->dts_necbs);
11133 ASSERT(state->dts_ecbs[epid - 1] != NULL);
11135 size = state->dts_ecbs[epid - 1]->dte_size;
11138 ASSERT(woffs + size <= buf->dtb_size);
11141 if (woffs + size == buf->dtb_size) {
11143 * We've reached the end of the buffer; we want
11144 * to set the wrapped offset to 0 and break
11145 * out. However, if the offs is 0, then we're
11146 * in a strange edge-condition: the amount of
11147 * space that we want to reserve plus the size
11148 * of the record that we're overwriting is
11149 * greater than the size of the buffer. This
11150 * is problematic because if we reserve the
11151 * space but subsequently don't consume it (due
11152 * to a failed predicate or error) the wrapped
11153 * offset will be 0 -- yet the EPID at offset 0
11154 * will not be committed. This situation is
11155 * relatively easy to deal with: if we're in
11156 * this case, the buffer is indistinguishable
11157 * from one that hasn't wrapped; we need only
11158 * finish the job by clearing the wrapped bit,
11159 * explicitly setting the offset to be 0, and
11160 * zero'ing out the old data in the buffer.
11163 buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
11164 buf->dtb_offset = 0;
11167 while (woffs < buf->dtb_size)
11168 tomax[woffs++] = 0;
11179 * We have a wrapped offset. It may be that the wrapped offset
11180 * has become zero -- that's okay.
11182 buf->dtb_xamot_offset = woffs;
11187 * Now we can plow the buffer with any necessary padding.
11189 while (offs & (align - 1)) {
11191 * Assert that our alignment is off by a number which
11192 * is itself sizeof (uint32_t) aligned.
11194 ASSERT(!((align - (offs & (align - 1))) &
11195 (sizeof (uint32_t) - 1)));
11196 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
11197 offs += sizeof (uint32_t);
11200 if (buf->dtb_flags & DTRACEBUF_FILL) {
11201 if (offs + needed > buf->dtb_size - state->dts_reserve) {
11202 buf->dtb_flags |= DTRACEBUF_FULL;
11207 if (mstate == NULL)
11211 * For ring buffers and fill buffers, the scratch space is always
11212 * the inactive buffer.
11214 mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
11215 mstate->dtms_scratch_size = buf->dtb_size;
11216 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11222 dtrace_buffer_polish(dtrace_buffer_t *buf)
11224 ASSERT(buf->dtb_flags & DTRACEBUF_RING);
11225 ASSERT(MUTEX_HELD(&dtrace_lock));
11227 if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
11231 * We need to polish the ring buffer. There are three cases:
11233 * - The first (and presumably most common) is that there is no gap
11234 * between the buffer offset and the wrapped offset. In this case,
11235 * there is nothing in the buffer that isn't valid data; we can
11236 * mark the buffer as polished and return.
11238 * - The second (less common than the first but still more common
11239 * than the third) is that there is a gap between the buffer offset
11240 * and the wrapped offset, and the wrapped offset is larger than the
11241 * buffer offset. This can happen because of an alignment issue, or
11242 * can happen because of a call to dtrace_buffer_reserve() that
11243 * didn't subsequently consume the buffer space. In this case,
11244 * we need to zero the data from the buffer offset to the wrapped
11247 * - The third (and least common) is that there is a gap between the
11248 * buffer offset and the wrapped offset, but the wrapped offset is
11249 * _less_ than the buffer offset. This can only happen because a
11250 * call to dtrace_buffer_reserve() induced a wrap, but the space
11251 * was not subsequently consumed. In this case, we need to zero the
11252 * space from the offset to the end of the buffer _and_ from the
11253 * top of the buffer to the wrapped offset.
11255 if (buf->dtb_offset < buf->dtb_xamot_offset) {
11256 bzero(buf->dtb_tomax + buf->dtb_offset,
11257 buf->dtb_xamot_offset - buf->dtb_offset);
11260 if (buf->dtb_offset > buf->dtb_xamot_offset) {
11261 bzero(buf->dtb_tomax + buf->dtb_offset,
11262 buf->dtb_size - buf->dtb_offset);
11263 bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
11268 * This routine determines if data generated at the specified time has likely
11269 * been entirely consumed at user-level. This routine is called to determine
11270 * if an ECB on a defunct probe (but for an active enabling) can be safely
11271 * disabled and destroyed.
11274 dtrace_buffer_consumed(dtrace_buffer_t *bufs, hrtime_t when)
11278 for (i = 0; i < NCPU; i++) {
11279 dtrace_buffer_t *buf = &bufs[i];
11281 if (buf->dtb_size == 0)
11284 if (buf->dtb_flags & DTRACEBUF_RING)
11287 if (!buf->dtb_switched && buf->dtb_offset != 0)
11290 if (buf->dtb_switched - buf->dtb_interval < when)
11298 dtrace_buffer_free(dtrace_buffer_t *bufs)
11302 for (i = 0; i < NCPU; i++) {
11303 dtrace_buffer_t *buf = &bufs[i];
11305 if (buf->dtb_tomax == NULL) {
11306 ASSERT(buf->dtb_xamot == NULL);
11307 ASSERT(buf->dtb_size == 0);
11311 if (buf->dtb_xamot != NULL) {
11312 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
11313 kmem_free(buf->dtb_xamot, buf->dtb_size);
11316 kmem_free(buf->dtb_tomax, buf->dtb_size);
11318 buf->dtb_tomax = NULL;
11319 buf->dtb_xamot = NULL;
11324 * DTrace Enabling Functions
11326 static dtrace_enabling_t *
11327 dtrace_enabling_create(dtrace_vstate_t *vstate)
11329 dtrace_enabling_t *enab;
11331 enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
11332 enab->dten_vstate = vstate;
11338 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
11340 dtrace_ecbdesc_t **ndesc;
11341 size_t osize, nsize;
11344 * We can't add to enablings after we've enabled them, or after we've
11347 ASSERT(enab->dten_probegen == 0);
11348 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11350 if (enab->dten_ndesc < enab->dten_maxdesc) {
11351 enab->dten_desc[enab->dten_ndesc++] = ecb;
11355 osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11357 if (enab->dten_maxdesc == 0) {
11358 enab->dten_maxdesc = 1;
11360 enab->dten_maxdesc <<= 1;
11363 ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
11365 nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11366 ndesc = kmem_zalloc(nsize, KM_SLEEP);
11367 bcopy(enab->dten_desc, ndesc, osize);
11368 if (enab->dten_desc != NULL)
11369 kmem_free(enab->dten_desc, osize);
11371 enab->dten_desc = ndesc;
11372 enab->dten_desc[enab->dten_ndesc++] = ecb;
11376 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
11377 dtrace_probedesc_t *pd)
11379 dtrace_ecbdesc_t *new;
11380 dtrace_predicate_t *pred;
11381 dtrace_actdesc_t *act;
11384 * We're going to create a new ECB description that matches the
11385 * specified ECB in every way, but has the specified probe description.
11387 new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
11389 if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
11390 dtrace_predicate_hold(pred);
11392 for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
11393 dtrace_actdesc_hold(act);
11395 new->dted_action = ecb->dted_action;
11396 new->dted_pred = ecb->dted_pred;
11397 new->dted_probe = *pd;
11398 new->dted_uarg = ecb->dted_uarg;
11400 dtrace_enabling_add(enab, new);
11404 dtrace_enabling_dump(dtrace_enabling_t *enab)
11408 for (i = 0; i < enab->dten_ndesc; i++) {
11409 dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
11411 cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
11412 desc->dtpd_provider, desc->dtpd_mod,
11413 desc->dtpd_func, desc->dtpd_name);
11418 dtrace_enabling_destroy(dtrace_enabling_t *enab)
11421 dtrace_ecbdesc_t *ep;
11422 dtrace_vstate_t *vstate = enab->dten_vstate;
11424 ASSERT(MUTEX_HELD(&dtrace_lock));
11426 for (i = 0; i < enab->dten_ndesc; i++) {
11427 dtrace_actdesc_t *act, *next;
11428 dtrace_predicate_t *pred;
11430 ep = enab->dten_desc[i];
11432 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
11433 dtrace_predicate_release(pred, vstate);
11435 for (act = ep->dted_action; act != NULL; act = next) {
11436 next = act->dtad_next;
11437 dtrace_actdesc_release(act, vstate);
11440 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
11443 if (enab->dten_desc != NULL)
11444 kmem_free(enab->dten_desc,
11445 enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
11448 * If this was a retained enabling, decrement the dts_nretained count
11449 * and take it off of the dtrace_retained list.
11451 if (enab->dten_prev != NULL || enab->dten_next != NULL ||
11452 dtrace_retained == enab) {
11453 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11454 ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
11455 enab->dten_vstate->dtvs_state->dts_nretained--;
11458 if (enab->dten_prev == NULL) {
11459 if (dtrace_retained == enab) {
11460 dtrace_retained = enab->dten_next;
11462 if (dtrace_retained != NULL)
11463 dtrace_retained->dten_prev = NULL;
11466 ASSERT(enab != dtrace_retained);
11467 ASSERT(dtrace_retained != NULL);
11468 enab->dten_prev->dten_next = enab->dten_next;
11471 if (enab->dten_next != NULL) {
11472 ASSERT(dtrace_retained != NULL);
11473 enab->dten_next->dten_prev = enab->dten_prev;
11476 kmem_free(enab, sizeof (dtrace_enabling_t));
11480 dtrace_enabling_retain(dtrace_enabling_t *enab)
11482 dtrace_state_t *state;
11484 ASSERT(MUTEX_HELD(&dtrace_lock));
11485 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11486 ASSERT(enab->dten_vstate != NULL);
11488 state = enab->dten_vstate->dtvs_state;
11489 ASSERT(state != NULL);
11492 * We only allow each state to retain dtrace_retain_max enablings.
11494 if (state->dts_nretained >= dtrace_retain_max)
11497 state->dts_nretained++;
11499 if (dtrace_retained == NULL) {
11500 dtrace_retained = enab;
11504 enab->dten_next = dtrace_retained;
11505 dtrace_retained->dten_prev = enab;
11506 dtrace_retained = enab;
11512 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
11513 dtrace_probedesc_t *create)
11515 dtrace_enabling_t *new, *enab;
11516 int found = 0, err = ENOENT;
11518 ASSERT(MUTEX_HELD(&dtrace_lock));
11519 ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
11520 ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
11521 ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
11522 ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
11524 new = dtrace_enabling_create(&state->dts_vstate);
11527 * Iterate over all retained enablings, looking for enablings that
11528 * match the specified state.
11530 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11534 * dtvs_state can only be NULL for helper enablings -- and
11535 * helper enablings can't be retained.
11537 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11539 if (enab->dten_vstate->dtvs_state != state)
11543 * Now iterate over each probe description; we're looking for
11544 * an exact match to the specified probe description.
11546 for (i = 0; i < enab->dten_ndesc; i++) {
11547 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11548 dtrace_probedesc_t *pd = &ep->dted_probe;
11550 if (strcmp(pd->dtpd_provider, match->dtpd_provider))
11553 if (strcmp(pd->dtpd_mod, match->dtpd_mod))
11556 if (strcmp(pd->dtpd_func, match->dtpd_func))
11559 if (strcmp(pd->dtpd_name, match->dtpd_name))
11563 * We have a winning probe! Add it to our growing
11567 dtrace_enabling_addlike(new, ep, create);
11571 if (!found || (err = dtrace_enabling_retain(new)) != 0) {
11572 dtrace_enabling_destroy(new);
11580 dtrace_enabling_retract(dtrace_state_t *state)
11582 dtrace_enabling_t *enab, *next;
11584 ASSERT(MUTEX_HELD(&dtrace_lock));
11587 * Iterate over all retained enablings, destroy the enablings retained
11588 * for the specified state.
11590 for (enab = dtrace_retained; enab != NULL; enab = next) {
11591 next = enab->dten_next;
11594 * dtvs_state can only be NULL for helper enablings -- and
11595 * helper enablings can't be retained.
11597 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11599 if (enab->dten_vstate->dtvs_state == state) {
11600 ASSERT(state->dts_nretained > 0);
11601 dtrace_enabling_destroy(enab);
11605 ASSERT(state->dts_nretained == 0);
11609 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
11614 ASSERT(MUTEX_HELD(&cpu_lock));
11615 ASSERT(MUTEX_HELD(&dtrace_lock));
11617 for (i = 0; i < enab->dten_ndesc; i++) {
11618 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11620 enab->dten_current = ep;
11621 enab->dten_error = 0;
11623 matched += dtrace_probe_enable(&ep->dted_probe, enab);
11625 if (enab->dten_error != 0) {
11627 * If we get an error half-way through enabling the
11628 * probes, we kick out -- perhaps with some number of
11629 * them enabled. Leaving enabled probes enabled may
11630 * be slightly confusing for user-level, but we expect
11631 * that no one will attempt to actually drive on in
11632 * the face of such errors. If this is an anonymous
11633 * enabling (indicated with a NULL nmatched pointer),
11634 * we cmn_err() a message. We aren't expecting to
11635 * get such an error -- such as it can exist at all,
11636 * it would be a result of corrupted DOF in the driver
11639 if (nmatched == NULL) {
11640 cmn_err(CE_WARN, "dtrace_enabling_match() "
11641 "error on %p: %d", (void *)ep,
11645 return (enab->dten_error);
11649 enab->dten_probegen = dtrace_probegen;
11650 if (nmatched != NULL)
11651 *nmatched = matched;
11657 dtrace_enabling_matchall(void)
11659 dtrace_enabling_t *enab;
11661 mutex_enter(&cpu_lock);
11662 mutex_enter(&dtrace_lock);
11665 * Iterate over all retained enablings to see if any probes match
11666 * against them. We only perform this operation on enablings for which
11667 * we have sufficient permissions by virtue of being in the global zone
11668 * or in the same zone as the DTrace client. Because we can be called
11669 * after dtrace_detach() has been called, we cannot assert that there
11670 * are retained enablings. We can safely load from dtrace_retained,
11671 * however: the taskq_destroy() at the end of dtrace_detach() will
11672 * block pending our completion.
11674 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11676 cred_t *cr = enab->dten_vstate->dtvs_state->dts_cred.dcr_cred;
11678 if (INGLOBALZONE(curproc) || getzoneid() == crgetzoneid(cr))
11680 (void) dtrace_enabling_match(enab, NULL);
11683 mutex_exit(&dtrace_lock);
11684 mutex_exit(&cpu_lock);
11688 * If an enabling is to be enabled without having matched probes (that is, if
11689 * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
11690 * enabling must be _primed_ by creating an ECB for every ECB description.
11691 * This must be done to assure that we know the number of speculations, the
11692 * number of aggregations, the minimum buffer size needed, etc. before we
11693 * transition out of DTRACE_ACTIVITY_INACTIVE. To do this without actually
11694 * enabling any probes, we create ECBs for every ECB decription, but with a
11695 * NULL probe -- which is exactly what this function does.
11698 dtrace_enabling_prime(dtrace_state_t *state)
11700 dtrace_enabling_t *enab;
11703 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11704 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11706 if (enab->dten_vstate->dtvs_state != state)
11710 * We don't want to prime an enabling more than once, lest
11711 * we allow a malicious user to induce resource exhaustion.
11712 * (The ECBs that result from priming an enabling aren't
11713 * leaked -- but they also aren't deallocated until the
11714 * consumer state is destroyed.)
11716 if (enab->dten_primed)
11719 for (i = 0; i < enab->dten_ndesc; i++) {
11720 enab->dten_current = enab->dten_desc[i];
11721 (void) dtrace_probe_enable(NULL, enab);
11724 enab->dten_primed = 1;
11729 * Called to indicate that probes should be provided due to retained
11730 * enablings. This is implemented in terms of dtrace_probe_provide(), but it
11731 * must take an initial lap through the enabling calling the dtps_provide()
11732 * entry point explicitly to allow for autocreated probes.
11735 dtrace_enabling_provide(dtrace_provider_t *prv)
11738 dtrace_probedesc_t desc;
11740 ASSERT(MUTEX_HELD(&dtrace_lock));
11741 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
11745 prv = dtrace_provider;
11749 dtrace_enabling_t *enab = dtrace_retained;
11750 void *parg = prv->dtpv_arg;
11752 for (; enab != NULL; enab = enab->dten_next) {
11753 for (i = 0; i < enab->dten_ndesc; i++) {
11754 desc = enab->dten_desc[i]->dted_probe;
11755 mutex_exit(&dtrace_lock);
11756 prv->dtpv_pops.dtps_provide(parg, &desc);
11757 mutex_enter(&dtrace_lock);
11760 } while (all && (prv = prv->dtpv_next) != NULL);
11762 mutex_exit(&dtrace_lock);
11763 dtrace_probe_provide(NULL, all ? NULL : prv);
11764 mutex_enter(&dtrace_lock);
11768 * Called to reap ECBs that are attached to probes from defunct providers.
11771 dtrace_enabling_reap(void)
11773 dtrace_provider_t *prov;
11774 dtrace_probe_t *probe;
11779 mutex_enter(&cpu_lock);
11780 mutex_enter(&dtrace_lock);
11782 for (i = 0; i < dtrace_nprobes; i++) {
11783 if ((probe = dtrace_probes[i]) == NULL)
11786 if (probe->dtpr_ecb == NULL)
11789 prov = probe->dtpr_provider;
11791 if ((when = prov->dtpv_defunct) == 0)
11795 * We have ECBs on a defunct provider: we want to reap these
11796 * ECBs to allow the provider to unregister. The destruction
11797 * of these ECBs must be done carefully: if we destroy the ECB
11798 * and the consumer later wishes to consume an EPID that
11799 * corresponds to the destroyed ECB (and if the EPID metadata
11800 * has not been previously consumed), the consumer will abort
11801 * processing on the unknown EPID. To reduce (but not, sadly,
11802 * eliminate) the possibility of this, we will only destroy an
11803 * ECB for a defunct provider if, for the state that
11804 * corresponds to the ECB:
11806 * (a) There is no speculative tracing (which can effectively
11807 * cache an EPID for an arbitrary amount of time).
11809 * (b) The principal buffers have been switched twice since the
11810 * provider became defunct.
11812 * (c) The aggregation buffers are of zero size or have been
11813 * switched twice since the provider became defunct.
11815 * We use dts_speculates to determine (a) and call a function
11816 * (dtrace_buffer_consumed()) to determine (b) and (c). Note
11817 * that as soon as we've been unable to destroy one of the ECBs
11818 * associated with the probe, we quit trying -- reaping is only
11819 * fruitful in as much as we can destroy all ECBs associated
11820 * with the defunct provider's probes.
11822 while ((ecb = probe->dtpr_ecb) != NULL) {
11823 dtrace_state_t *state = ecb->dte_state;
11824 dtrace_buffer_t *buf = state->dts_buffer;
11825 dtrace_buffer_t *aggbuf = state->dts_aggbuffer;
11827 if (state->dts_speculates)
11830 if (!dtrace_buffer_consumed(buf, when))
11833 if (!dtrace_buffer_consumed(aggbuf, when))
11836 dtrace_ecb_disable(ecb);
11837 ASSERT(probe->dtpr_ecb != ecb);
11838 dtrace_ecb_destroy(ecb);
11842 mutex_exit(&dtrace_lock);
11843 mutex_exit(&cpu_lock);
11847 * DTrace DOF Functions
11851 dtrace_dof_error(dof_hdr_t *dof, const char *str)
11853 if (dtrace_err_verbose)
11854 cmn_err(CE_WARN, "failed to process DOF: %s", str);
11856 #ifdef DTRACE_ERRDEBUG
11857 dtrace_errdebug(str);
11862 * Create DOF out of a currently enabled state. Right now, we only create
11863 * DOF containing the run-time options -- but this could be expanded to create
11864 * complete DOF representing the enabled state.
11867 dtrace_dof_create(dtrace_state_t *state)
11871 dof_optdesc_t *opt;
11872 int i, len = sizeof (dof_hdr_t) +
11873 roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
11874 sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11876 ASSERT(MUTEX_HELD(&dtrace_lock));
11878 dof = kmem_zalloc(len, KM_SLEEP);
11879 dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
11880 dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
11881 dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
11882 dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
11884 dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
11885 dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
11886 dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
11887 dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
11888 dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
11889 dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
11891 dof->dofh_flags = 0;
11892 dof->dofh_hdrsize = sizeof (dof_hdr_t);
11893 dof->dofh_secsize = sizeof (dof_sec_t);
11894 dof->dofh_secnum = 1; /* only DOF_SECT_OPTDESC */
11895 dof->dofh_secoff = sizeof (dof_hdr_t);
11896 dof->dofh_loadsz = len;
11897 dof->dofh_filesz = len;
11901 * Fill in the option section header...
11903 sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
11904 sec->dofs_type = DOF_SECT_OPTDESC;
11905 sec->dofs_align = sizeof (uint64_t);
11906 sec->dofs_flags = DOF_SECF_LOAD;
11907 sec->dofs_entsize = sizeof (dof_optdesc_t);
11909 opt = (dof_optdesc_t *)((uintptr_t)sec +
11910 roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
11912 sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
11913 sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11915 for (i = 0; i < DTRACEOPT_MAX; i++) {
11916 opt[i].dofo_option = i;
11917 opt[i].dofo_strtab = DOF_SECIDX_NONE;
11918 opt[i].dofo_value = state->dts_options[i];
11925 dtrace_dof_copyin(uintptr_t uarg, int *errp)
11927 dof_hdr_t hdr, *dof;
11929 ASSERT(!MUTEX_HELD(&dtrace_lock));
11932 * First, we're going to copyin() the sizeof (dof_hdr_t).
11934 if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
11935 dtrace_dof_error(NULL, "failed to copyin DOF header");
11941 * Now we'll allocate the entire DOF and copy it in -- provided
11942 * that the length isn't outrageous.
11944 if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
11945 dtrace_dof_error(&hdr, "load size exceeds maximum");
11950 if (hdr.dofh_loadsz < sizeof (hdr)) {
11951 dtrace_dof_error(&hdr, "invalid load size");
11956 dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
11958 if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0) {
11959 kmem_free(dof, hdr.dofh_loadsz);
11968 static __inline uchar_t
11969 dtrace_dof_char(char c) {
11988 return (c - 'A' + 10);
11995 return (c - 'a' + 10);
11997 /* Should not reach here. */
12003 dtrace_dof_property(const char *name)
12007 unsigned int len, i;
12012 * Unfortunately, array of values in .conf files are always (and
12013 * only) interpreted to be integer arrays. We must read our DOF
12014 * as an integer array, and then squeeze it into a byte array.
12016 if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
12017 (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
12020 for (i = 0; i < len; i++)
12021 buf[i] = (uchar_t)(((int *)buf)[i]);
12023 if (len < sizeof (dof_hdr_t)) {
12024 ddi_prop_free(buf);
12025 dtrace_dof_error(NULL, "truncated header");
12029 if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
12030 ddi_prop_free(buf);
12031 dtrace_dof_error(NULL, "truncated DOF");
12035 if (loadsz >= dtrace_dof_maxsize) {
12036 ddi_prop_free(buf);
12037 dtrace_dof_error(NULL, "oversized DOF");
12041 dof = kmem_alloc(loadsz, KM_SLEEP);
12042 bcopy(buf, dof, loadsz);
12043 ddi_prop_free(buf);
12048 if ((p_env = getenv(name)) == NULL)
12051 len = strlen(p_env) / 2;
12053 buf = kmem_alloc(len, KM_SLEEP);
12055 dof = (dof_hdr_t *) buf;
12059 for (i = 0; i < len; i++) {
12060 buf[i] = (dtrace_dof_char(p[0]) << 4) |
12061 dtrace_dof_char(p[1]);
12067 if (len < sizeof (dof_hdr_t)) {
12069 dtrace_dof_error(NULL, "truncated header");
12073 if (len < (loadsz = dof->dofh_loadsz)) {
12075 dtrace_dof_error(NULL, "truncated DOF");
12079 if (loadsz >= dtrace_dof_maxsize) {
12081 dtrace_dof_error(NULL, "oversized DOF");
12090 dtrace_dof_destroy(dof_hdr_t *dof)
12092 kmem_free(dof, dof->dofh_loadsz);
12096 * Return the dof_sec_t pointer corresponding to a given section index. If the
12097 * index is not valid, dtrace_dof_error() is called and NULL is returned. If
12098 * a type other than DOF_SECT_NONE is specified, the header is checked against
12099 * this type and NULL is returned if the types do not match.
12102 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
12104 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
12105 ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
12107 if (i >= dof->dofh_secnum) {
12108 dtrace_dof_error(dof, "referenced section index is invalid");
12112 if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
12113 dtrace_dof_error(dof, "referenced section is not loadable");
12117 if (type != DOF_SECT_NONE && type != sec->dofs_type) {
12118 dtrace_dof_error(dof, "referenced section is the wrong type");
12125 static dtrace_probedesc_t *
12126 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
12128 dof_probedesc_t *probe;
12130 uintptr_t daddr = (uintptr_t)dof;
12134 if (sec->dofs_type != DOF_SECT_PROBEDESC) {
12135 dtrace_dof_error(dof, "invalid probe section");
12139 if (sec->dofs_align != sizeof (dof_secidx_t)) {
12140 dtrace_dof_error(dof, "bad alignment in probe description");
12144 if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
12145 dtrace_dof_error(dof, "truncated probe description");
12149 probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
12150 strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
12152 if (strtab == NULL)
12155 str = daddr + strtab->dofs_offset;
12156 size = strtab->dofs_size;
12158 if (probe->dofp_provider >= strtab->dofs_size) {
12159 dtrace_dof_error(dof, "corrupt probe provider");
12163 (void) strncpy(desc->dtpd_provider,
12164 (char *)(str + probe->dofp_provider),
12165 MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
12167 if (probe->dofp_mod >= strtab->dofs_size) {
12168 dtrace_dof_error(dof, "corrupt probe module");
12172 (void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
12173 MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
12175 if (probe->dofp_func >= strtab->dofs_size) {
12176 dtrace_dof_error(dof, "corrupt probe function");
12180 (void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
12181 MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
12183 if (probe->dofp_name >= strtab->dofs_size) {
12184 dtrace_dof_error(dof, "corrupt probe name");
12188 (void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
12189 MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
12194 static dtrace_difo_t *
12195 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12200 dof_difohdr_t *dofd;
12201 uintptr_t daddr = (uintptr_t)dof;
12202 size_t max = dtrace_difo_maxsize;
12205 static const struct {
12213 { DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
12214 offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
12215 sizeof (dif_instr_t), "multiple DIF sections" },
12217 { DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
12218 offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
12219 sizeof (uint64_t), "multiple integer tables" },
12221 { DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
12222 offsetof(dtrace_difo_t, dtdo_strlen), 0,
12223 sizeof (char), "multiple string tables" },
12225 { DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
12226 offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
12227 sizeof (uint_t), "multiple variable tables" },
12229 { DOF_SECT_NONE, 0, 0, 0, 0, NULL }
12232 if (sec->dofs_type != DOF_SECT_DIFOHDR) {
12233 dtrace_dof_error(dof, "invalid DIFO header section");
12237 if (sec->dofs_align != sizeof (dof_secidx_t)) {
12238 dtrace_dof_error(dof, "bad alignment in DIFO header");
12242 if (sec->dofs_size < sizeof (dof_difohdr_t) ||
12243 sec->dofs_size % sizeof (dof_secidx_t)) {
12244 dtrace_dof_error(dof, "bad size in DIFO header");
12248 dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12249 n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
12251 dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
12252 dp->dtdo_rtype = dofd->dofd_rtype;
12254 for (l = 0; l < n; l++) {
12259 if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
12260 dofd->dofd_links[l])) == NULL)
12261 goto err; /* invalid section link */
12263 if (ttl + subsec->dofs_size > max) {
12264 dtrace_dof_error(dof, "exceeds maximum size");
12268 ttl += subsec->dofs_size;
12270 for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
12271 if (subsec->dofs_type != difo[i].section)
12274 if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
12275 dtrace_dof_error(dof, "section not loaded");
12279 if (subsec->dofs_align != difo[i].align) {
12280 dtrace_dof_error(dof, "bad alignment");
12284 bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
12285 lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
12287 if (*bufp != NULL) {
12288 dtrace_dof_error(dof, difo[i].msg);
12292 if (difo[i].entsize != subsec->dofs_entsize) {
12293 dtrace_dof_error(dof, "entry size mismatch");
12297 if (subsec->dofs_entsize != 0 &&
12298 (subsec->dofs_size % subsec->dofs_entsize) != 0) {
12299 dtrace_dof_error(dof, "corrupt entry size");
12303 *lenp = subsec->dofs_size;
12304 *bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
12305 bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
12306 *bufp, subsec->dofs_size);
12308 if (subsec->dofs_entsize != 0)
12309 *lenp /= subsec->dofs_entsize;
12315 * If we encounter a loadable DIFO sub-section that is not
12316 * known to us, assume this is a broken program and fail.
12318 if (difo[i].section == DOF_SECT_NONE &&
12319 (subsec->dofs_flags & DOF_SECF_LOAD)) {
12320 dtrace_dof_error(dof, "unrecognized DIFO subsection");
12325 if (dp->dtdo_buf == NULL) {
12327 * We can't have a DIF object without DIF text.
12329 dtrace_dof_error(dof, "missing DIF text");
12334 * Before we validate the DIF object, run through the variable table
12335 * looking for the strings -- if any of their size are under, we'll set
12336 * their size to be the system-wide default string size. Note that
12337 * this should _not_ happen if the "strsize" option has been set --
12338 * in this case, the compiler should have set the size to reflect the
12339 * setting of the option.
12341 for (i = 0; i < dp->dtdo_varlen; i++) {
12342 dtrace_difv_t *v = &dp->dtdo_vartab[i];
12343 dtrace_diftype_t *t = &v->dtdv_type;
12345 if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
12348 if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
12349 t->dtdt_size = dtrace_strsize_default;
12352 if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
12355 dtrace_difo_init(dp, vstate);
12359 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
12360 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
12361 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
12362 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
12364 kmem_free(dp, sizeof (dtrace_difo_t));
12368 static dtrace_predicate_t *
12369 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12374 if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
12377 return (dtrace_predicate_create(dp));
12380 static dtrace_actdesc_t *
12381 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12384 dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
12385 dof_actdesc_t *desc;
12386 dof_sec_t *difosec;
12388 uintptr_t daddr = (uintptr_t)dof;
12390 dtrace_actkind_t kind;
12392 if (sec->dofs_type != DOF_SECT_ACTDESC) {
12393 dtrace_dof_error(dof, "invalid action section");
12397 if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
12398 dtrace_dof_error(dof, "truncated action description");
12402 if (sec->dofs_align != sizeof (uint64_t)) {
12403 dtrace_dof_error(dof, "bad alignment in action description");
12407 if (sec->dofs_size < sec->dofs_entsize) {
12408 dtrace_dof_error(dof, "section entry size exceeds total size");
12412 if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
12413 dtrace_dof_error(dof, "bad entry size in action description");
12417 if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
12418 dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
12422 for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
12423 desc = (dof_actdesc_t *)(daddr +
12424 (uintptr_t)sec->dofs_offset + offs);
12425 kind = (dtrace_actkind_t)desc->dofa_kind;
12427 if ((DTRACEACT_ISPRINTFLIKE(kind) &&
12428 (kind != DTRACEACT_PRINTA ||
12429 desc->dofa_strtab != DOF_SECIDX_NONE)) ||
12430 (kind == DTRACEACT_DIFEXPR &&
12431 desc->dofa_strtab != DOF_SECIDX_NONE)) {
12437 * The argument to these actions is an index into the
12438 * DOF string table. For printf()-like actions, this
12439 * is the format string. For print(), this is the
12440 * CTF type of the expression result.
12442 if ((strtab = dtrace_dof_sect(dof,
12443 DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
12446 str = (char *)((uintptr_t)dof +
12447 (uintptr_t)strtab->dofs_offset);
12449 for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
12450 if (str[i] == '\0')
12454 if (i >= strtab->dofs_size) {
12455 dtrace_dof_error(dof, "bogus format string");
12459 if (i == desc->dofa_arg) {
12460 dtrace_dof_error(dof, "empty format string");
12464 i -= desc->dofa_arg;
12465 fmt = kmem_alloc(i + 1, KM_SLEEP);
12466 bcopy(&str[desc->dofa_arg], fmt, i + 1);
12467 arg = (uint64_t)(uintptr_t)fmt;
12469 if (kind == DTRACEACT_PRINTA) {
12470 ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
12473 arg = desc->dofa_arg;
12477 act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
12478 desc->dofa_uarg, arg);
12480 if (last != NULL) {
12481 last->dtad_next = act;
12488 if (desc->dofa_difo == DOF_SECIDX_NONE)
12491 if ((difosec = dtrace_dof_sect(dof,
12492 DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
12495 act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
12497 if (act->dtad_difo == NULL)
12501 ASSERT(first != NULL);
12505 for (act = first; act != NULL; act = next) {
12506 next = act->dtad_next;
12507 dtrace_actdesc_release(act, vstate);
12513 static dtrace_ecbdesc_t *
12514 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12517 dtrace_ecbdesc_t *ep;
12518 dof_ecbdesc_t *ecb;
12519 dtrace_probedesc_t *desc;
12520 dtrace_predicate_t *pred = NULL;
12522 if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
12523 dtrace_dof_error(dof, "truncated ECB description");
12527 if (sec->dofs_align != sizeof (uint64_t)) {
12528 dtrace_dof_error(dof, "bad alignment in ECB description");
12532 ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
12533 sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
12538 ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
12539 ep->dted_uarg = ecb->dofe_uarg;
12540 desc = &ep->dted_probe;
12542 if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
12545 if (ecb->dofe_pred != DOF_SECIDX_NONE) {
12546 if ((sec = dtrace_dof_sect(dof,
12547 DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
12550 if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
12553 ep->dted_pred.dtpdd_predicate = pred;
12556 if (ecb->dofe_actions != DOF_SECIDX_NONE) {
12557 if ((sec = dtrace_dof_sect(dof,
12558 DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
12561 ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
12563 if (ep->dted_action == NULL)
12571 dtrace_predicate_release(pred, vstate);
12572 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
12577 * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
12578 * specified DOF. At present, this amounts to simply adding 'ubase' to the
12579 * site of any user SETX relocations to account for load object base address.
12580 * In the future, if we need other relocations, this function can be extended.
12583 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase)
12585 uintptr_t daddr = (uintptr_t)dof;
12586 dof_relohdr_t *dofr =
12587 (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12588 dof_sec_t *ss, *rs, *ts;
12592 if (sec->dofs_size < sizeof (dof_relohdr_t) ||
12593 sec->dofs_align != sizeof (dof_secidx_t)) {
12594 dtrace_dof_error(dof, "invalid relocation header");
12598 ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
12599 rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
12600 ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
12602 if (ss == NULL || rs == NULL || ts == NULL)
12603 return (-1); /* dtrace_dof_error() has been called already */
12605 if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
12606 rs->dofs_align != sizeof (uint64_t)) {
12607 dtrace_dof_error(dof, "invalid relocation section");
12611 r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
12612 n = rs->dofs_size / rs->dofs_entsize;
12614 for (i = 0; i < n; i++) {
12615 uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
12617 switch (r->dofr_type) {
12618 case DOF_RELO_NONE:
12620 case DOF_RELO_SETX:
12621 if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
12622 sizeof (uint64_t) > ts->dofs_size) {
12623 dtrace_dof_error(dof, "bad relocation offset");
12627 if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
12628 dtrace_dof_error(dof, "misaligned setx relo");
12632 *(uint64_t *)taddr += ubase;
12635 dtrace_dof_error(dof, "invalid relocation type");
12639 r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
12646 * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
12647 * header: it should be at the front of a memory region that is at least
12648 * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
12649 * size. It need not be validated in any other way.
12652 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
12653 dtrace_enabling_t **enabp, uint64_t ubase, int noprobes)
12655 uint64_t len = dof->dofh_loadsz, seclen;
12656 uintptr_t daddr = (uintptr_t)dof;
12657 dtrace_ecbdesc_t *ep;
12658 dtrace_enabling_t *enab;
12661 ASSERT(MUTEX_HELD(&dtrace_lock));
12662 ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
12665 * Check the DOF header identification bytes. In addition to checking
12666 * valid settings, we also verify that unused bits/bytes are zeroed so
12667 * we can use them later without fear of regressing existing binaries.
12669 if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
12670 DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
12671 dtrace_dof_error(dof, "DOF magic string mismatch");
12675 if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
12676 dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
12677 dtrace_dof_error(dof, "DOF has invalid data model");
12681 if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
12682 dtrace_dof_error(dof, "DOF encoding mismatch");
12686 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
12687 dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
12688 dtrace_dof_error(dof, "DOF version mismatch");
12692 if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
12693 dtrace_dof_error(dof, "DOF uses unsupported instruction set");
12697 if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
12698 dtrace_dof_error(dof, "DOF uses too many integer registers");
12702 if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
12703 dtrace_dof_error(dof, "DOF uses too many tuple registers");
12707 for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
12708 if (dof->dofh_ident[i] != 0) {
12709 dtrace_dof_error(dof, "DOF has invalid ident byte set");
12714 if (dof->dofh_flags & ~DOF_FL_VALID) {
12715 dtrace_dof_error(dof, "DOF has invalid flag bits set");
12719 if (dof->dofh_secsize == 0) {
12720 dtrace_dof_error(dof, "zero section header size");
12725 * Check that the section headers don't exceed the amount of DOF
12726 * data. Note that we cast the section size and number of sections
12727 * to uint64_t's to prevent possible overflow in the multiplication.
12729 seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
12731 if (dof->dofh_secoff > len || seclen > len ||
12732 dof->dofh_secoff + seclen > len) {
12733 dtrace_dof_error(dof, "truncated section headers");
12737 if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
12738 dtrace_dof_error(dof, "misaligned section headers");
12742 if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
12743 dtrace_dof_error(dof, "misaligned section size");
12748 * Take an initial pass through the section headers to be sure that
12749 * the headers don't have stray offsets. If the 'noprobes' flag is
12750 * set, do not permit sections relating to providers, probes, or args.
12752 for (i = 0; i < dof->dofh_secnum; i++) {
12753 dof_sec_t *sec = (dof_sec_t *)(daddr +
12754 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12757 switch (sec->dofs_type) {
12758 case DOF_SECT_PROVIDER:
12759 case DOF_SECT_PROBES:
12760 case DOF_SECT_PRARGS:
12761 case DOF_SECT_PROFFS:
12762 dtrace_dof_error(dof, "illegal sections "
12768 if (!(sec->dofs_flags & DOF_SECF_LOAD))
12769 continue; /* just ignore non-loadable sections */
12771 if (sec->dofs_align & (sec->dofs_align - 1)) {
12772 dtrace_dof_error(dof, "bad section alignment");
12776 if (sec->dofs_offset & (sec->dofs_align - 1)) {
12777 dtrace_dof_error(dof, "misaligned section");
12781 if (sec->dofs_offset > len || sec->dofs_size > len ||
12782 sec->dofs_offset + sec->dofs_size > len) {
12783 dtrace_dof_error(dof, "corrupt section header");
12787 if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
12788 sec->dofs_offset + sec->dofs_size - 1) != '\0') {
12789 dtrace_dof_error(dof, "non-terminating string table");
12795 * Take a second pass through the sections and locate and perform any
12796 * relocations that are present. We do this after the first pass to
12797 * be sure that all sections have had their headers validated.
12799 for (i = 0; i < dof->dofh_secnum; i++) {
12800 dof_sec_t *sec = (dof_sec_t *)(daddr +
12801 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12803 if (!(sec->dofs_flags & DOF_SECF_LOAD))
12804 continue; /* skip sections that are not loadable */
12806 switch (sec->dofs_type) {
12807 case DOF_SECT_URELHDR:
12808 if (dtrace_dof_relocate(dof, sec, ubase) != 0)
12814 if ((enab = *enabp) == NULL)
12815 enab = *enabp = dtrace_enabling_create(vstate);
12817 for (i = 0; i < dof->dofh_secnum; i++) {
12818 dof_sec_t *sec = (dof_sec_t *)(daddr +
12819 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12821 if (sec->dofs_type != DOF_SECT_ECBDESC)
12824 if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
12825 dtrace_enabling_destroy(enab);
12830 dtrace_enabling_add(enab, ep);
12837 * Process DOF for any options. This routine assumes that the DOF has been
12838 * at least processed by dtrace_dof_slurp().
12841 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
12846 dof_optdesc_t *desc;
12848 for (i = 0; i < dof->dofh_secnum; i++) {
12849 dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
12850 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12852 if (sec->dofs_type != DOF_SECT_OPTDESC)
12855 if (sec->dofs_align != sizeof (uint64_t)) {
12856 dtrace_dof_error(dof, "bad alignment in "
12857 "option description");
12861 if ((entsize = sec->dofs_entsize) == 0) {
12862 dtrace_dof_error(dof, "zeroed option entry size");
12866 if (entsize < sizeof (dof_optdesc_t)) {
12867 dtrace_dof_error(dof, "bad option entry size");
12871 for (offs = 0; offs < sec->dofs_size; offs += entsize) {
12872 desc = (dof_optdesc_t *)((uintptr_t)dof +
12873 (uintptr_t)sec->dofs_offset + offs);
12875 if (desc->dofo_strtab != DOF_SECIDX_NONE) {
12876 dtrace_dof_error(dof, "non-zero option string");
12880 if (desc->dofo_value == DTRACEOPT_UNSET) {
12881 dtrace_dof_error(dof, "unset option");
12885 if ((rval = dtrace_state_option(state,
12886 desc->dofo_option, desc->dofo_value)) != 0) {
12887 dtrace_dof_error(dof, "rejected option");
12897 * DTrace Consumer State Functions
12900 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
12902 size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
12905 dtrace_dynvar_t *dvar, *next, *start;
12908 ASSERT(MUTEX_HELD(&dtrace_lock));
12909 ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
12911 bzero(dstate, sizeof (dtrace_dstate_t));
12913 if ((dstate->dtds_chunksize = chunksize) == 0)
12914 dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
12916 if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
12919 if ((base = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
12922 dstate->dtds_size = size;
12923 dstate->dtds_base = base;
12924 dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
12925 bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
12927 hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
12929 if (hashsize != 1 && (hashsize & 1))
12932 dstate->dtds_hashsize = hashsize;
12933 dstate->dtds_hash = dstate->dtds_base;
12936 * Set all of our hash buckets to point to the single sink, and (if
12937 * it hasn't already been set), set the sink's hash value to be the
12938 * sink sentinel value. The sink is needed for dynamic variable
12939 * lookups to know that they have iterated over an entire, valid hash
12942 for (i = 0; i < hashsize; i++)
12943 dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
12945 if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
12946 dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
12949 * Determine number of active CPUs. Divide free list evenly among
12952 start = (dtrace_dynvar_t *)
12953 ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
12954 limit = (uintptr_t)base + size;
12956 maxper = (limit - (uintptr_t)start) / NCPU;
12957 maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
12962 for (i = 0; i < NCPU; i++) {
12964 dstate->dtds_percpu[i].dtdsc_free = dvar = start;
12967 * If we don't even have enough chunks to make it once through
12968 * NCPUs, we're just going to allocate everything to the first
12969 * CPU. And if we're on the last CPU, we're going to allocate
12970 * whatever is left over. In either case, we set the limit to
12971 * be the limit of the dynamic variable space.
12973 if (maxper == 0 || i == NCPU - 1) {
12974 limit = (uintptr_t)base + size;
12977 limit = (uintptr_t)start + maxper;
12978 start = (dtrace_dynvar_t *)limit;
12981 ASSERT(limit <= (uintptr_t)base + size);
12984 next = (dtrace_dynvar_t *)((uintptr_t)dvar +
12985 dstate->dtds_chunksize);
12987 if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
12990 dvar->dtdv_next = next;
13002 dtrace_dstate_fini(dtrace_dstate_t *dstate)
13004 ASSERT(MUTEX_HELD(&cpu_lock));
13006 if (dstate->dtds_base == NULL)
13009 kmem_free(dstate->dtds_base, dstate->dtds_size);
13010 kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
13014 dtrace_vstate_fini(dtrace_vstate_t *vstate)
13017 * Logical XOR, where are you?
13019 ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
13021 if (vstate->dtvs_nglobals > 0) {
13022 kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
13023 sizeof (dtrace_statvar_t *));
13026 if (vstate->dtvs_ntlocals > 0) {
13027 kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
13028 sizeof (dtrace_difv_t));
13031 ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
13033 if (vstate->dtvs_nlocals > 0) {
13034 kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
13035 sizeof (dtrace_statvar_t *));
13041 dtrace_state_clean(dtrace_state_t *state)
13043 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
13046 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
13047 dtrace_speculation_clean(state);
13051 dtrace_state_deadman(dtrace_state_t *state)
13057 now = dtrace_gethrtime();
13059 if (state != dtrace_anon.dta_state &&
13060 now - state->dts_laststatus >= dtrace_deadman_user)
13064 * We must be sure that dts_alive never appears to be less than the
13065 * value upon entry to dtrace_state_deadman(), and because we lack a
13066 * dtrace_cas64(), we cannot store to it atomically. We thus instead
13067 * store INT64_MAX to it, followed by a memory barrier, followed by
13068 * the new value. This assures that dts_alive never appears to be
13069 * less than its true value, regardless of the order in which the
13070 * stores to the underlying storage are issued.
13072 state->dts_alive = INT64_MAX;
13073 dtrace_membar_producer();
13074 state->dts_alive = now;
13078 dtrace_state_clean(void *arg)
13080 dtrace_state_t *state = arg;
13081 dtrace_optval_t *opt = state->dts_options;
13083 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
13086 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
13087 dtrace_speculation_clean(state);
13089 callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
13090 dtrace_state_clean, state);
13094 dtrace_state_deadman(void *arg)
13096 dtrace_state_t *state = arg;
13101 dtrace_debug_output();
13103 now = dtrace_gethrtime();
13105 if (state != dtrace_anon.dta_state &&
13106 now - state->dts_laststatus >= dtrace_deadman_user)
13110 * We must be sure that dts_alive never appears to be less than the
13111 * value upon entry to dtrace_state_deadman(), and because we lack a
13112 * dtrace_cas64(), we cannot store to it atomically. We thus instead
13113 * store INT64_MAX to it, followed by a memory barrier, followed by
13114 * the new value. This assures that dts_alive never appears to be
13115 * less than its true value, regardless of the order in which the
13116 * stores to the underlying storage are issued.
13118 state->dts_alive = INT64_MAX;
13119 dtrace_membar_producer();
13120 state->dts_alive = now;
13122 callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
13123 dtrace_state_deadman, state);
13127 static dtrace_state_t *
13129 dtrace_state_create(dev_t *devp, cred_t *cr)
13131 dtrace_state_create(struct cdev *dev)
13142 dtrace_state_t *state;
13143 dtrace_optval_t *opt;
13144 int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
13146 ASSERT(MUTEX_HELD(&dtrace_lock));
13147 ASSERT(MUTEX_HELD(&cpu_lock));
13150 minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
13151 VM_BESTFIT | VM_SLEEP);
13153 if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
13154 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
13158 state = ddi_get_soft_state(dtrace_softstate, minor);
13165 /* Allocate memory for the state. */
13166 state = kmem_zalloc(sizeof(dtrace_state_t), KM_SLEEP);
13169 state->dts_epid = DTRACE_EPIDNONE + 1;
13171 (void) snprintf(c, sizeof (c), "dtrace_aggid_%d", m);
13173 state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
13174 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
13176 if (devp != NULL) {
13177 major = getemajor(*devp);
13179 major = ddi_driver_major(dtrace_devi);
13182 state->dts_dev = makedevice(major, minor);
13185 *devp = state->dts_dev;
13187 state->dts_aggid_arena = new_unrhdr(1, INT_MAX, &dtrace_unr_mtx);
13188 state->dts_dev = dev;
13192 * We allocate NCPU buffers. On the one hand, this can be quite
13193 * a bit of memory per instance (nearly 36K on a Starcat). On the
13194 * other hand, it saves an additional memory reference in the probe
13197 state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
13198 state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
13201 state->dts_cleaner = CYCLIC_NONE;
13202 state->dts_deadman = CYCLIC_NONE;
13204 callout_init(&state->dts_cleaner, CALLOUT_MPSAFE);
13205 callout_init(&state->dts_deadman, CALLOUT_MPSAFE);
13207 state->dts_vstate.dtvs_state = state;
13209 for (i = 0; i < DTRACEOPT_MAX; i++)
13210 state->dts_options[i] = DTRACEOPT_UNSET;
13213 * Set the default options.
13215 opt = state->dts_options;
13216 opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
13217 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
13218 opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
13219 opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
13220 opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
13221 opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
13222 opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
13223 opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
13224 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
13225 opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
13226 opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
13227 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
13228 opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
13229 opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
13231 state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
13234 * Depending on the user credentials, we set flag bits which alter probe
13235 * visibility or the amount of destructiveness allowed. In the case of
13236 * actual anonymous tracing, or the possession of all privileges, all of
13237 * the normal checks are bypassed.
13239 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
13240 state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
13241 state->dts_cred.dcr_action = DTRACE_CRA_ALL;
13244 * Set up the credentials for this instantiation. We take a
13245 * hold on the credential to prevent it from disappearing on
13246 * us; this in turn prevents the zone_t referenced by this
13247 * credential from disappearing. This means that we can
13248 * examine the credential and the zone from probe context.
13251 state->dts_cred.dcr_cred = cr;
13254 * CRA_PROC means "we have *some* privilege for dtrace" and
13255 * unlocks the use of variables like pid, zonename, etc.
13257 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
13258 PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13259 state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
13263 * dtrace_user allows use of syscall and profile providers.
13264 * If the user also has proc_owner and/or proc_zone, we
13265 * extend the scope to include additional visibility and
13266 * destructive power.
13268 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
13269 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
13270 state->dts_cred.dcr_visible |=
13271 DTRACE_CRV_ALLPROC;
13273 state->dts_cred.dcr_action |=
13274 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13277 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
13278 state->dts_cred.dcr_visible |=
13279 DTRACE_CRV_ALLZONE;
13281 state->dts_cred.dcr_action |=
13282 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13286 * If we have all privs in whatever zone this is,
13287 * we can do destructive things to processes which
13288 * have altered credentials.
13291 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13292 cr->cr_zone->zone_privset)) {
13293 state->dts_cred.dcr_action |=
13294 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13300 * Holding the dtrace_kernel privilege also implies that
13301 * the user has the dtrace_user privilege from a visibility
13302 * perspective. But without further privileges, some
13303 * destructive actions are not available.
13305 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
13307 * Make all probes in all zones visible. However,
13308 * this doesn't mean that all actions become available
13311 state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
13312 DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
13314 state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
13317 * Holding proc_owner means that destructive actions
13318 * for *this* zone are allowed.
13320 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13321 state->dts_cred.dcr_action |=
13322 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13325 * Holding proc_zone means that destructive actions
13326 * for this user/group ID in all zones is allowed.
13328 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13329 state->dts_cred.dcr_action |=
13330 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13334 * If we have all privs in whatever zone this is,
13335 * we can do destructive things to processes which
13336 * have altered credentials.
13338 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13339 cr->cr_zone->zone_privset)) {
13340 state->dts_cred.dcr_action |=
13341 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13347 * Holding the dtrace_proc privilege gives control over fasttrap
13348 * and pid providers. We need to grant wider destructive
13349 * privileges in the event that the user has proc_owner and/or
13352 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13353 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13354 state->dts_cred.dcr_action |=
13355 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13357 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13358 state->dts_cred.dcr_action |=
13359 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13367 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
13369 dtrace_optval_t *opt = state->dts_options, size;
13370 processorid_t cpu = 0;;
13371 int flags = 0, rval;
13373 ASSERT(MUTEX_HELD(&dtrace_lock));
13374 ASSERT(MUTEX_HELD(&cpu_lock));
13375 ASSERT(which < DTRACEOPT_MAX);
13376 ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
13377 (state == dtrace_anon.dta_state &&
13378 state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
13380 if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
13383 if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
13384 cpu = opt[DTRACEOPT_CPU];
13386 if (which == DTRACEOPT_SPECSIZE)
13387 flags |= DTRACEBUF_NOSWITCH;
13389 if (which == DTRACEOPT_BUFSIZE) {
13390 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
13391 flags |= DTRACEBUF_RING;
13393 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
13394 flags |= DTRACEBUF_FILL;
13396 if (state != dtrace_anon.dta_state ||
13397 state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
13398 flags |= DTRACEBUF_INACTIVE;
13401 for (size = opt[which]; size >= sizeof (uint64_t); size >>= 1) {
13403 * The size must be 8-byte aligned. If the size is not 8-byte
13404 * aligned, drop it down by the difference.
13406 if (size & (sizeof (uint64_t) - 1))
13407 size -= size & (sizeof (uint64_t) - 1);
13409 if (size < state->dts_reserve) {
13411 * Buffers always must be large enough to accommodate
13412 * their prereserved space. We return E2BIG instead
13413 * of ENOMEM in this case to allow for user-level
13414 * software to differentiate the cases.
13419 rval = dtrace_buffer_alloc(buf, size, flags, cpu);
13421 if (rval != ENOMEM) {
13426 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13434 dtrace_state_buffers(dtrace_state_t *state)
13436 dtrace_speculation_t *spec = state->dts_speculations;
13439 if ((rval = dtrace_state_buffer(state, state->dts_buffer,
13440 DTRACEOPT_BUFSIZE)) != 0)
13443 if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
13444 DTRACEOPT_AGGSIZE)) != 0)
13447 for (i = 0; i < state->dts_nspeculations; i++) {
13448 if ((rval = dtrace_state_buffer(state,
13449 spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
13457 dtrace_state_prereserve(dtrace_state_t *state)
13460 dtrace_probe_t *probe;
13462 state->dts_reserve = 0;
13464 if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
13468 * If our buffer policy is a "fill" buffer policy, we need to set the
13469 * prereserved space to be the space required by the END probes.
13471 probe = dtrace_probes[dtrace_probeid_end - 1];
13472 ASSERT(probe != NULL);
13474 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
13475 if (ecb->dte_state != state)
13478 state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
13483 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
13485 dtrace_optval_t *opt = state->dts_options, sz, nspec;
13486 dtrace_speculation_t *spec;
13487 dtrace_buffer_t *buf;
13489 cyc_handler_t hdlr;
13492 int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13493 dtrace_icookie_t cookie;
13495 mutex_enter(&cpu_lock);
13496 mutex_enter(&dtrace_lock);
13498 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
13504 * Before we can perform any checks, we must prime all of the
13505 * retained enablings that correspond to this state.
13507 dtrace_enabling_prime(state);
13509 if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
13514 dtrace_state_prereserve(state);
13517 * Now we want to do is try to allocate our speculations.
13518 * We do not automatically resize the number of speculations; if
13519 * this fails, we will fail the operation.
13521 nspec = opt[DTRACEOPT_NSPEC];
13522 ASSERT(nspec != DTRACEOPT_UNSET);
13524 if (nspec > INT_MAX) {
13529 spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t), KM_NOSLEEP);
13531 if (spec == NULL) {
13536 state->dts_speculations = spec;
13537 state->dts_nspeculations = (int)nspec;
13539 for (i = 0; i < nspec; i++) {
13540 if ((buf = kmem_zalloc(bufsize, KM_NOSLEEP)) == NULL) {
13545 spec[i].dtsp_buffer = buf;
13548 if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
13549 if (dtrace_anon.dta_state == NULL) {
13554 if (state->dts_necbs != 0) {
13559 state->dts_anon = dtrace_anon_grab();
13560 ASSERT(state->dts_anon != NULL);
13561 state = state->dts_anon;
13564 * We want "grabanon" to be set in the grabbed state, so we'll
13565 * copy that option value from the grabbing state into the
13568 state->dts_options[DTRACEOPT_GRABANON] =
13569 opt[DTRACEOPT_GRABANON];
13571 *cpu = dtrace_anon.dta_beganon;
13574 * If the anonymous state is active (as it almost certainly
13575 * is if the anonymous enabling ultimately matched anything),
13576 * we don't allow any further option processing -- but we
13577 * don't return failure.
13579 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13583 if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
13584 opt[DTRACEOPT_AGGSIZE] != 0) {
13585 if (state->dts_aggregations == NULL) {
13587 * We're not going to create an aggregation buffer
13588 * because we don't have any ECBs that contain
13589 * aggregations -- set this option to 0.
13591 opt[DTRACEOPT_AGGSIZE] = 0;
13594 * If we have an aggregation buffer, we must also have
13595 * a buffer to use as scratch.
13597 if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
13598 opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
13599 opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
13604 if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
13605 opt[DTRACEOPT_SPECSIZE] != 0) {
13606 if (!state->dts_speculates) {
13608 * We're not going to create speculation buffers
13609 * because we don't have any ECBs that actually
13610 * speculate -- set the speculation size to 0.
13612 opt[DTRACEOPT_SPECSIZE] = 0;
13617 * The bare minimum size for any buffer that we're actually going to
13618 * do anything to is sizeof (uint64_t).
13620 sz = sizeof (uint64_t);
13622 if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
13623 (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
13624 (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
13626 * A buffer size has been explicitly set to 0 (or to a size
13627 * that will be adjusted to 0) and we need the space -- we
13628 * need to return failure. We return ENOSPC to differentiate
13629 * it from failing to allocate a buffer due to failure to meet
13630 * the reserve (for which we return E2BIG).
13636 if ((rval = dtrace_state_buffers(state)) != 0)
13639 if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
13640 sz = dtrace_dstate_defsize;
13643 rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
13648 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13650 } while (sz >>= 1);
13652 opt[DTRACEOPT_DYNVARSIZE] = sz;
13657 if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
13658 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
13660 if (opt[DTRACEOPT_CLEANRATE] == 0)
13661 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13663 if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
13664 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
13666 if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
13667 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13669 state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
13671 hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
13672 hdlr.cyh_arg = state;
13673 hdlr.cyh_level = CY_LOW_LEVEL;
13676 when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
13678 state->dts_cleaner = cyclic_add(&hdlr, &when);
13680 hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
13681 hdlr.cyh_arg = state;
13682 hdlr.cyh_level = CY_LOW_LEVEL;
13685 when.cyt_interval = dtrace_deadman_interval;
13687 state->dts_deadman = cyclic_add(&hdlr, &when);
13689 callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
13690 dtrace_state_clean, state);
13691 callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
13692 dtrace_state_deadman, state);
13695 state->dts_activity = DTRACE_ACTIVITY_WARMUP;
13698 * Now it's time to actually fire the BEGIN probe. We need to disable
13699 * interrupts here both to record the CPU on which we fired the BEGIN
13700 * probe (the data from this CPU will be processed first at user
13701 * level) and to manually activate the buffer for this CPU.
13703 cookie = dtrace_interrupt_disable();
13705 ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
13706 state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
13708 dtrace_probe(dtrace_probeid_begin,
13709 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13710 dtrace_interrupt_enable(cookie);
13712 * We may have had an exit action from a BEGIN probe; only change our
13713 * state to ACTIVE if we're still in WARMUP.
13715 ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
13716 state->dts_activity == DTRACE_ACTIVITY_DRAINING);
13718 if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
13719 state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
13722 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
13723 * want each CPU to transition its principal buffer out of the
13724 * INACTIVE state. Doing this assures that no CPU will suddenly begin
13725 * processing an ECB halfway down a probe's ECB chain; all CPUs will
13726 * atomically transition from processing none of a state's ECBs to
13727 * processing all of them.
13729 dtrace_xcall(DTRACE_CPUALL,
13730 (dtrace_xcall_t)dtrace_buffer_activate, state);
13734 dtrace_buffer_free(state->dts_buffer);
13735 dtrace_buffer_free(state->dts_aggbuffer);
13737 if ((nspec = state->dts_nspeculations) == 0) {
13738 ASSERT(state->dts_speculations == NULL);
13742 spec = state->dts_speculations;
13743 ASSERT(spec != NULL);
13745 for (i = 0; i < state->dts_nspeculations; i++) {
13746 if ((buf = spec[i].dtsp_buffer) == NULL)
13749 dtrace_buffer_free(buf);
13750 kmem_free(buf, bufsize);
13753 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13754 state->dts_nspeculations = 0;
13755 state->dts_speculations = NULL;
13758 mutex_exit(&dtrace_lock);
13759 mutex_exit(&cpu_lock);
13765 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
13767 dtrace_icookie_t cookie;
13769 ASSERT(MUTEX_HELD(&dtrace_lock));
13771 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
13772 state->dts_activity != DTRACE_ACTIVITY_DRAINING)
13776 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
13777 * to be sure that every CPU has seen it. See below for the details
13778 * on why this is done.
13780 state->dts_activity = DTRACE_ACTIVITY_DRAINING;
13784 * By this point, it is impossible for any CPU to be still processing
13785 * with DTRACE_ACTIVITY_ACTIVE. We can thus set our activity to
13786 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
13787 * other CPU in dtrace_buffer_reserve(). This allows dtrace_probe()
13788 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
13789 * iff we're in the END probe.
13791 state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
13793 ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
13796 * Finally, we can release the reserve and call the END probe. We
13797 * disable interrupts across calling the END probe to allow us to
13798 * return the CPU on which we actually called the END probe. This
13799 * allows user-land to be sure that this CPU's principal buffer is
13802 state->dts_reserve = 0;
13804 cookie = dtrace_interrupt_disable();
13806 dtrace_probe(dtrace_probeid_end,
13807 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13808 dtrace_interrupt_enable(cookie);
13810 state->dts_activity = DTRACE_ACTIVITY_STOPPED;
13817 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
13818 dtrace_optval_t val)
13820 ASSERT(MUTEX_HELD(&dtrace_lock));
13822 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13825 if (option >= DTRACEOPT_MAX)
13828 if (option != DTRACEOPT_CPU && val < 0)
13832 case DTRACEOPT_DESTRUCTIVE:
13833 if (dtrace_destructive_disallow)
13836 state->dts_cred.dcr_destructive = 1;
13839 case DTRACEOPT_BUFSIZE:
13840 case DTRACEOPT_DYNVARSIZE:
13841 case DTRACEOPT_AGGSIZE:
13842 case DTRACEOPT_SPECSIZE:
13843 case DTRACEOPT_STRSIZE:
13847 if (val >= LONG_MAX) {
13849 * If this is an otherwise negative value, set it to
13850 * the highest multiple of 128m less than LONG_MAX.
13851 * Technically, we're adjusting the size without
13852 * regard to the buffer resizing policy, but in fact,
13853 * this has no effect -- if we set the buffer size to
13854 * ~LONG_MAX and the buffer policy is ultimately set to
13855 * be "manual", the buffer allocation is guaranteed to
13856 * fail, if only because the allocation requires two
13857 * buffers. (We set the the size to the highest
13858 * multiple of 128m because it ensures that the size
13859 * will remain a multiple of a megabyte when
13860 * repeatedly halved -- all the way down to 15m.)
13862 val = LONG_MAX - (1 << 27) + 1;
13866 state->dts_options[option] = val;
13872 dtrace_state_destroy(dtrace_state_t *state)
13875 dtrace_vstate_t *vstate = &state->dts_vstate;
13877 minor_t minor = getminor(state->dts_dev);
13879 int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13880 dtrace_speculation_t *spec = state->dts_speculations;
13881 int nspec = state->dts_nspeculations;
13884 ASSERT(MUTEX_HELD(&dtrace_lock));
13885 ASSERT(MUTEX_HELD(&cpu_lock));
13888 * First, retract any retained enablings for this state.
13890 dtrace_enabling_retract(state);
13891 ASSERT(state->dts_nretained == 0);
13893 if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
13894 state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
13896 * We have managed to come into dtrace_state_destroy() on a
13897 * hot enabling -- almost certainly because of a disorderly
13898 * shutdown of a consumer. (That is, a consumer that is
13899 * exiting without having called dtrace_stop().) In this case,
13900 * we're going to set our activity to be KILLED, and then
13901 * issue a sync to be sure that everyone is out of probe
13902 * context before we start blowing away ECBs.
13904 state->dts_activity = DTRACE_ACTIVITY_KILLED;
13909 * Release the credential hold we took in dtrace_state_create().
13911 if (state->dts_cred.dcr_cred != NULL)
13912 crfree(state->dts_cred.dcr_cred);
13915 * Now we can safely disable and destroy any enabled probes. Because
13916 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
13917 * (especially if they're all enabled), we take two passes through the
13918 * ECBs: in the first, we disable just DTRACE_PRIV_KERNEL probes, and
13919 * in the second we disable whatever is left over.
13921 for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
13922 for (i = 0; i < state->dts_necbs; i++) {
13923 if ((ecb = state->dts_ecbs[i]) == NULL)
13926 if (match && ecb->dte_probe != NULL) {
13927 dtrace_probe_t *probe = ecb->dte_probe;
13928 dtrace_provider_t *prov = probe->dtpr_provider;
13930 if (!(prov->dtpv_priv.dtpp_flags & match))
13934 dtrace_ecb_disable(ecb);
13935 dtrace_ecb_destroy(ecb);
13943 * Before we free the buffers, perform one more sync to assure that
13944 * every CPU is out of probe context.
13948 dtrace_buffer_free(state->dts_buffer);
13949 dtrace_buffer_free(state->dts_aggbuffer);
13951 for (i = 0; i < nspec; i++)
13952 dtrace_buffer_free(spec[i].dtsp_buffer);
13955 if (state->dts_cleaner != CYCLIC_NONE)
13956 cyclic_remove(state->dts_cleaner);
13958 if (state->dts_deadman != CYCLIC_NONE)
13959 cyclic_remove(state->dts_deadman);
13961 callout_stop(&state->dts_cleaner);
13962 callout_drain(&state->dts_cleaner);
13963 callout_stop(&state->dts_deadman);
13964 callout_drain(&state->dts_deadman);
13967 dtrace_dstate_fini(&vstate->dtvs_dynvars);
13968 dtrace_vstate_fini(vstate);
13969 if (state->dts_ecbs != NULL)
13970 kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
13972 if (state->dts_aggregations != NULL) {
13974 for (i = 0; i < state->dts_naggregations; i++)
13975 ASSERT(state->dts_aggregations[i] == NULL);
13977 ASSERT(state->dts_naggregations > 0);
13978 kmem_free(state->dts_aggregations,
13979 state->dts_naggregations * sizeof (dtrace_aggregation_t *));
13982 kmem_free(state->dts_buffer, bufsize);
13983 kmem_free(state->dts_aggbuffer, bufsize);
13985 for (i = 0; i < nspec; i++)
13986 kmem_free(spec[i].dtsp_buffer, bufsize);
13989 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13991 dtrace_format_destroy(state);
13993 if (state->dts_aggid_arena != NULL) {
13995 vmem_destroy(state->dts_aggid_arena);
13997 delete_unrhdr(state->dts_aggid_arena);
13999 state->dts_aggid_arena = NULL;
14002 ddi_soft_state_free(dtrace_softstate, minor);
14003 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
14008 * DTrace Anonymous Enabling Functions
14010 static dtrace_state_t *
14011 dtrace_anon_grab(void)
14013 dtrace_state_t *state;
14015 ASSERT(MUTEX_HELD(&dtrace_lock));
14017 if ((state = dtrace_anon.dta_state) == NULL) {
14018 ASSERT(dtrace_anon.dta_enabling == NULL);
14022 ASSERT(dtrace_anon.dta_enabling != NULL);
14023 ASSERT(dtrace_retained != NULL);
14025 dtrace_enabling_destroy(dtrace_anon.dta_enabling);
14026 dtrace_anon.dta_enabling = NULL;
14027 dtrace_anon.dta_state = NULL;
14033 dtrace_anon_property(void)
14036 dtrace_state_t *state;
14038 char c[32]; /* enough for "dof-data-" + digits */
14040 ASSERT(MUTEX_HELD(&dtrace_lock));
14041 ASSERT(MUTEX_HELD(&cpu_lock));
14043 for (i = 0; ; i++) {
14044 (void) snprintf(c, sizeof (c), "dof-data-%d", i);
14046 dtrace_err_verbose = 1;
14048 if ((dof = dtrace_dof_property(c)) == NULL) {
14049 dtrace_err_verbose = 0;
14055 * We want to create anonymous state, so we need to transition
14056 * the kernel debugger to indicate that DTrace is active. If
14057 * this fails (e.g. because the debugger has modified text in
14058 * some way), we won't continue with the processing.
14060 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
14061 cmn_err(CE_NOTE, "kernel debugger active; anonymous "
14062 "enabling ignored.");
14063 dtrace_dof_destroy(dof);
14069 * If we haven't allocated an anonymous state, we'll do so now.
14071 if ((state = dtrace_anon.dta_state) == NULL) {
14073 state = dtrace_state_create(NULL, NULL);
14075 state = dtrace_state_create(NULL);
14077 dtrace_anon.dta_state = state;
14079 if (state == NULL) {
14081 * This basically shouldn't happen: the only
14082 * failure mode from dtrace_state_create() is a
14083 * failure of ddi_soft_state_zalloc() that
14084 * itself should never happen. Still, the
14085 * interface allows for a failure mode, and
14086 * we want to fail as gracefully as possible:
14087 * we'll emit an error message and cease
14088 * processing anonymous state in this case.
14090 cmn_err(CE_WARN, "failed to create "
14091 "anonymous state");
14092 dtrace_dof_destroy(dof);
14097 rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
14098 &dtrace_anon.dta_enabling, 0, B_TRUE);
14101 rv = dtrace_dof_options(dof, state);
14103 dtrace_err_verbose = 0;
14104 dtrace_dof_destroy(dof);
14108 * This is malformed DOF; chuck any anonymous state
14111 ASSERT(dtrace_anon.dta_enabling == NULL);
14112 dtrace_state_destroy(state);
14113 dtrace_anon.dta_state = NULL;
14117 ASSERT(dtrace_anon.dta_enabling != NULL);
14120 if (dtrace_anon.dta_enabling != NULL) {
14124 * dtrace_enabling_retain() can only fail because we are
14125 * trying to retain more enablings than are allowed -- but
14126 * we only have one anonymous enabling, and we are guaranteed
14127 * to be allowed at least one retained enabling; we assert
14128 * that dtrace_enabling_retain() returns success.
14130 rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
14133 dtrace_enabling_dump(dtrace_anon.dta_enabling);
14138 * DTrace Helper Functions
14141 dtrace_helper_trace(dtrace_helper_action_t *helper,
14142 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
14144 uint32_t size, next, nnext, i;
14145 dtrace_helptrace_t *ent;
14146 uint16_t flags = cpu_core[curcpu].cpuc_dtrace_flags;
14148 if (!dtrace_helptrace_enabled)
14151 ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
14154 * What would a tracing framework be without its own tracing
14155 * framework? (Well, a hell of a lot simpler, for starters...)
14157 size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
14158 sizeof (uint64_t) - sizeof (uint64_t);
14161 * Iterate until we can allocate a slot in the trace buffer.
14164 next = dtrace_helptrace_next;
14166 if (next + size < dtrace_helptrace_bufsize) {
14167 nnext = next + size;
14171 } while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
14174 * We have our slot; fill it in.
14179 ent = (dtrace_helptrace_t *)&dtrace_helptrace_buffer[next];
14180 ent->dtht_helper = helper;
14181 ent->dtht_where = where;
14182 ent->dtht_nlocals = vstate->dtvs_nlocals;
14184 ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
14185 mstate->dtms_fltoffs : -1;
14186 ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
14187 ent->dtht_illval = cpu_core[curcpu].cpuc_dtrace_illval;
14189 for (i = 0; i < vstate->dtvs_nlocals; i++) {
14190 dtrace_statvar_t *svar;
14192 if ((svar = vstate->dtvs_locals[i]) == NULL)
14195 ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
14196 ent->dtht_locals[i] =
14197 ((uint64_t *)(uintptr_t)svar->dtsv_data)[curcpu];
14202 dtrace_helper(int which, dtrace_mstate_t *mstate,
14203 dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
14205 uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
14206 uint64_t sarg0 = mstate->dtms_arg[0];
14207 uint64_t sarg1 = mstate->dtms_arg[1];
14209 dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
14210 dtrace_helper_action_t *helper;
14211 dtrace_vstate_t *vstate;
14212 dtrace_difo_t *pred;
14213 int i, trace = dtrace_helptrace_enabled;
14215 ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
14217 if (helpers == NULL)
14220 if ((helper = helpers->dthps_actions[which]) == NULL)
14223 vstate = &helpers->dthps_vstate;
14224 mstate->dtms_arg[0] = arg0;
14225 mstate->dtms_arg[1] = arg1;
14228 * Now iterate over each helper. If its predicate evaluates to 'true',
14229 * we'll call the corresponding actions. Note that the below calls
14230 * to dtrace_dif_emulate() may set faults in machine state. This is
14231 * okay: our caller (the outer dtrace_dif_emulate()) will simply plow
14232 * the stored DIF offset with its own (which is the desired behavior).
14233 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
14234 * from machine state; this is okay, too.
14236 for (; helper != NULL; helper = helper->dtha_next) {
14237 if ((pred = helper->dtha_predicate) != NULL) {
14239 dtrace_helper_trace(helper, mstate, vstate, 0);
14241 if (!dtrace_dif_emulate(pred, mstate, vstate, state))
14244 if (*flags & CPU_DTRACE_FAULT)
14248 for (i = 0; i < helper->dtha_nactions; i++) {
14250 dtrace_helper_trace(helper,
14251 mstate, vstate, i + 1);
14253 rval = dtrace_dif_emulate(helper->dtha_actions[i],
14254 mstate, vstate, state);
14256 if (*flags & CPU_DTRACE_FAULT)
14262 dtrace_helper_trace(helper, mstate, vstate,
14263 DTRACE_HELPTRACE_NEXT);
14267 dtrace_helper_trace(helper, mstate, vstate,
14268 DTRACE_HELPTRACE_DONE);
14271 * Restore the arg0 that we saved upon entry.
14273 mstate->dtms_arg[0] = sarg0;
14274 mstate->dtms_arg[1] = sarg1;
14280 dtrace_helper_trace(helper, mstate, vstate,
14281 DTRACE_HELPTRACE_ERR);
14284 * Restore the arg0 that we saved upon entry.
14286 mstate->dtms_arg[0] = sarg0;
14287 mstate->dtms_arg[1] = sarg1;
14293 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
14294 dtrace_vstate_t *vstate)
14298 if (helper->dtha_predicate != NULL)
14299 dtrace_difo_release(helper->dtha_predicate, vstate);
14301 for (i = 0; i < helper->dtha_nactions; i++) {
14302 ASSERT(helper->dtha_actions[i] != NULL);
14303 dtrace_difo_release(helper->dtha_actions[i], vstate);
14306 kmem_free(helper->dtha_actions,
14307 helper->dtha_nactions * sizeof (dtrace_difo_t *));
14308 kmem_free(helper, sizeof (dtrace_helper_action_t));
14312 dtrace_helper_destroygen(int gen)
14314 proc_t *p = curproc;
14315 dtrace_helpers_t *help = p->p_dtrace_helpers;
14316 dtrace_vstate_t *vstate;
14319 ASSERT(MUTEX_HELD(&dtrace_lock));
14321 if (help == NULL || gen > help->dthps_generation)
14324 vstate = &help->dthps_vstate;
14326 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14327 dtrace_helper_action_t *last = NULL, *h, *next;
14329 for (h = help->dthps_actions[i]; h != NULL; h = next) {
14330 next = h->dtha_next;
14332 if (h->dtha_generation == gen) {
14333 if (last != NULL) {
14334 last->dtha_next = next;
14336 help->dthps_actions[i] = next;
14339 dtrace_helper_action_destroy(h, vstate);
14347 * Interate until we've cleared out all helper providers with the
14348 * given generation number.
14351 dtrace_helper_provider_t *prov;
14354 * Look for a helper provider with the right generation. We
14355 * have to start back at the beginning of the list each time
14356 * because we drop dtrace_lock. It's unlikely that we'll make
14357 * more than two passes.
14359 for (i = 0; i < help->dthps_nprovs; i++) {
14360 prov = help->dthps_provs[i];
14362 if (prov->dthp_generation == gen)
14367 * If there were no matches, we're done.
14369 if (i == help->dthps_nprovs)
14373 * Move the last helper provider into this slot.
14375 help->dthps_nprovs--;
14376 help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
14377 help->dthps_provs[help->dthps_nprovs] = NULL;
14379 mutex_exit(&dtrace_lock);
14382 * If we have a meta provider, remove this helper provider.
14384 mutex_enter(&dtrace_meta_lock);
14385 if (dtrace_meta_pid != NULL) {
14386 ASSERT(dtrace_deferred_pid == NULL);
14387 dtrace_helper_provider_remove(&prov->dthp_prov,
14390 mutex_exit(&dtrace_meta_lock);
14392 dtrace_helper_provider_destroy(prov);
14394 mutex_enter(&dtrace_lock);
14401 dtrace_helper_validate(dtrace_helper_action_t *helper)
14406 if ((dp = helper->dtha_predicate) != NULL)
14407 err += dtrace_difo_validate_helper(dp);
14409 for (i = 0; i < helper->dtha_nactions; i++)
14410 err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
14416 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep)
14418 dtrace_helpers_t *help;
14419 dtrace_helper_action_t *helper, *last;
14420 dtrace_actdesc_t *act;
14421 dtrace_vstate_t *vstate;
14422 dtrace_predicate_t *pred;
14423 int count = 0, nactions = 0, i;
14425 if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
14428 help = curproc->p_dtrace_helpers;
14429 last = help->dthps_actions[which];
14430 vstate = &help->dthps_vstate;
14432 for (count = 0; last != NULL; last = last->dtha_next) {
14434 if (last->dtha_next == NULL)
14439 * If we already have dtrace_helper_actions_max helper actions for this
14440 * helper action type, we'll refuse to add a new one.
14442 if (count >= dtrace_helper_actions_max)
14445 helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
14446 helper->dtha_generation = help->dthps_generation;
14448 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
14449 ASSERT(pred->dtp_difo != NULL);
14450 dtrace_difo_hold(pred->dtp_difo);
14451 helper->dtha_predicate = pred->dtp_difo;
14454 for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
14455 if (act->dtad_kind != DTRACEACT_DIFEXPR)
14458 if (act->dtad_difo == NULL)
14464 helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
14465 (helper->dtha_nactions = nactions), KM_SLEEP);
14467 for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
14468 dtrace_difo_hold(act->dtad_difo);
14469 helper->dtha_actions[i++] = act->dtad_difo;
14472 if (!dtrace_helper_validate(helper))
14475 if (last == NULL) {
14476 help->dthps_actions[which] = helper;
14478 last->dtha_next = helper;
14481 if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
14482 dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
14483 dtrace_helptrace_next = 0;
14488 dtrace_helper_action_destroy(helper, vstate);
14493 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
14494 dof_helper_t *dofhp)
14496 ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
14498 mutex_enter(&dtrace_meta_lock);
14499 mutex_enter(&dtrace_lock);
14501 if (!dtrace_attached() || dtrace_meta_pid == NULL) {
14503 * If the dtrace module is loaded but not attached, or if
14504 * there aren't isn't a meta provider registered to deal with
14505 * these provider descriptions, we need to postpone creating
14506 * the actual providers until later.
14509 if (help->dthps_next == NULL && help->dthps_prev == NULL &&
14510 dtrace_deferred_pid != help) {
14511 help->dthps_deferred = 1;
14512 help->dthps_pid = p->p_pid;
14513 help->dthps_next = dtrace_deferred_pid;
14514 help->dthps_prev = NULL;
14515 if (dtrace_deferred_pid != NULL)
14516 dtrace_deferred_pid->dthps_prev = help;
14517 dtrace_deferred_pid = help;
14520 mutex_exit(&dtrace_lock);
14522 } else if (dofhp != NULL) {
14524 * If the dtrace module is loaded and we have a particular
14525 * helper provider description, pass that off to the
14529 mutex_exit(&dtrace_lock);
14531 dtrace_helper_provide(dofhp, p->p_pid);
14535 * Otherwise, just pass all the helper provider descriptions
14536 * off to the meta provider.
14540 mutex_exit(&dtrace_lock);
14542 for (i = 0; i < help->dthps_nprovs; i++) {
14543 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
14548 mutex_exit(&dtrace_meta_lock);
14552 dtrace_helper_provider_add(dof_helper_t *dofhp, int gen)
14554 dtrace_helpers_t *help;
14555 dtrace_helper_provider_t *hprov, **tmp_provs;
14556 uint_t tmp_maxprovs, i;
14558 ASSERT(MUTEX_HELD(&dtrace_lock));
14560 help = curproc->p_dtrace_helpers;
14561 ASSERT(help != NULL);
14564 * If we already have dtrace_helper_providers_max helper providers,
14565 * we're refuse to add a new one.
14567 if (help->dthps_nprovs >= dtrace_helper_providers_max)
14571 * Check to make sure this isn't a duplicate.
14573 for (i = 0; i < help->dthps_nprovs; i++) {
14574 if (dofhp->dofhp_addr ==
14575 help->dthps_provs[i]->dthp_prov.dofhp_addr)
14579 hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
14580 hprov->dthp_prov = *dofhp;
14581 hprov->dthp_ref = 1;
14582 hprov->dthp_generation = gen;
14585 * Allocate a bigger table for helper providers if it's already full.
14587 if (help->dthps_maxprovs == help->dthps_nprovs) {
14588 tmp_maxprovs = help->dthps_maxprovs;
14589 tmp_provs = help->dthps_provs;
14591 if (help->dthps_maxprovs == 0)
14592 help->dthps_maxprovs = 2;
14594 help->dthps_maxprovs *= 2;
14595 if (help->dthps_maxprovs > dtrace_helper_providers_max)
14596 help->dthps_maxprovs = dtrace_helper_providers_max;
14598 ASSERT(tmp_maxprovs < help->dthps_maxprovs);
14600 help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
14601 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
14603 if (tmp_provs != NULL) {
14604 bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
14605 sizeof (dtrace_helper_provider_t *));
14606 kmem_free(tmp_provs, tmp_maxprovs *
14607 sizeof (dtrace_helper_provider_t *));
14611 help->dthps_provs[help->dthps_nprovs] = hprov;
14612 help->dthps_nprovs++;
14618 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
14620 mutex_enter(&dtrace_lock);
14622 if (--hprov->dthp_ref == 0) {
14624 mutex_exit(&dtrace_lock);
14625 dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
14626 dtrace_dof_destroy(dof);
14627 kmem_free(hprov, sizeof (dtrace_helper_provider_t));
14629 mutex_exit(&dtrace_lock);
14634 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
14636 uintptr_t daddr = (uintptr_t)dof;
14637 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
14638 dof_provider_t *provider;
14639 dof_probe_t *probe;
14641 char *strtab, *typestr;
14642 dof_stridx_t typeidx;
14644 uint_t nprobes, j, k;
14646 ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
14648 if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
14649 dtrace_dof_error(dof, "misaligned section offset");
14654 * The section needs to be large enough to contain the DOF provider
14655 * structure appropriate for the given version.
14657 if (sec->dofs_size <
14658 ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
14659 offsetof(dof_provider_t, dofpv_prenoffs) :
14660 sizeof (dof_provider_t))) {
14661 dtrace_dof_error(dof, "provider section too small");
14665 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
14666 str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
14667 prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
14668 arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
14669 off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
14671 if (str_sec == NULL || prb_sec == NULL ||
14672 arg_sec == NULL || off_sec == NULL)
14677 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
14678 provider->dofpv_prenoffs != DOF_SECT_NONE &&
14679 (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
14680 provider->dofpv_prenoffs)) == NULL)
14683 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
14685 if (provider->dofpv_name >= str_sec->dofs_size ||
14686 strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
14687 dtrace_dof_error(dof, "invalid provider name");
14691 if (prb_sec->dofs_entsize == 0 ||
14692 prb_sec->dofs_entsize > prb_sec->dofs_size) {
14693 dtrace_dof_error(dof, "invalid entry size");
14697 if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
14698 dtrace_dof_error(dof, "misaligned entry size");
14702 if (off_sec->dofs_entsize != sizeof (uint32_t)) {
14703 dtrace_dof_error(dof, "invalid entry size");
14707 if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
14708 dtrace_dof_error(dof, "misaligned section offset");
14712 if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
14713 dtrace_dof_error(dof, "invalid entry size");
14717 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
14719 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
14722 * Take a pass through the probes to check for errors.
14724 for (j = 0; j < nprobes; j++) {
14725 probe = (dof_probe_t *)(uintptr_t)(daddr +
14726 prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
14728 if (probe->dofpr_func >= str_sec->dofs_size) {
14729 dtrace_dof_error(dof, "invalid function name");
14733 if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
14734 dtrace_dof_error(dof, "function name too long");
14738 if (probe->dofpr_name >= str_sec->dofs_size ||
14739 strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
14740 dtrace_dof_error(dof, "invalid probe name");
14745 * The offset count must not wrap the index, and the offsets
14746 * must also not overflow the section's data.
14748 if (probe->dofpr_offidx + probe->dofpr_noffs <
14749 probe->dofpr_offidx ||
14750 (probe->dofpr_offidx + probe->dofpr_noffs) *
14751 off_sec->dofs_entsize > off_sec->dofs_size) {
14752 dtrace_dof_error(dof, "invalid probe offset");
14756 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
14758 * If there's no is-enabled offset section, make sure
14759 * there aren't any is-enabled offsets. Otherwise
14760 * perform the same checks as for probe offsets
14761 * (immediately above).
14763 if (enoff_sec == NULL) {
14764 if (probe->dofpr_enoffidx != 0 ||
14765 probe->dofpr_nenoffs != 0) {
14766 dtrace_dof_error(dof, "is-enabled "
14767 "offsets with null section");
14770 } else if (probe->dofpr_enoffidx +
14771 probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
14772 (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
14773 enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
14774 dtrace_dof_error(dof, "invalid is-enabled "
14779 if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
14780 dtrace_dof_error(dof, "zero probe and "
14781 "is-enabled offsets");
14784 } else if (probe->dofpr_noffs == 0) {
14785 dtrace_dof_error(dof, "zero probe offsets");
14789 if (probe->dofpr_argidx + probe->dofpr_xargc <
14790 probe->dofpr_argidx ||
14791 (probe->dofpr_argidx + probe->dofpr_xargc) *
14792 arg_sec->dofs_entsize > arg_sec->dofs_size) {
14793 dtrace_dof_error(dof, "invalid args");
14797 typeidx = probe->dofpr_nargv;
14798 typestr = strtab + probe->dofpr_nargv;
14799 for (k = 0; k < probe->dofpr_nargc; k++) {
14800 if (typeidx >= str_sec->dofs_size) {
14801 dtrace_dof_error(dof, "bad "
14802 "native argument type");
14806 typesz = strlen(typestr) + 1;
14807 if (typesz > DTRACE_ARGTYPELEN) {
14808 dtrace_dof_error(dof, "native "
14809 "argument type too long");
14816 typeidx = probe->dofpr_xargv;
14817 typestr = strtab + probe->dofpr_xargv;
14818 for (k = 0; k < probe->dofpr_xargc; k++) {
14819 if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
14820 dtrace_dof_error(dof, "bad "
14821 "native argument index");
14825 if (typeidx >= str_sec->dofs_size) {
14826 dtrace_dof_error(dof, "bad "
14827 "translated argument type");
14831 typesz = strlen(typestr) + 1;
14832 if (typesz > DTRACE_ARGTYPELEN) {
14833 dtrace_dof_error(dof, "translated argument "
14847 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp)
14849 dtrace_helpers_t *help;
14850 dtrace_vstate_t *vstate;
14851 dtrace_enabling_t *enab = NULL;
14852 int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
14853 uintptr_t daddr = (uintptr_t)dof;
14855 ASSERT(MUTEX_HELD(&dtrace_lock));
14857 if ((help = curproc->p_dtrace_helpers) == NULL)
14858 help = dtrace_helpers_create(curproc);
14860 vstate = &help->dthps_vstate;
14862 if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab,
14863 dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) {
14864 dtrace_dof_destroy(dof);
14869 * Look for helper providers and validate their descriptions.
14872 for (i = 0; i < dof->dofh_secnum; i++) {
14873 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
14874 dof->dofh_secoff + i * dof->dofh_secsize);
14876 if (sec->dofs_type != DOF_SECT_PROVIDER)
14879 if (dtrace_helper_provider_validate(dof, sec) != 0) {
14880 dtrace_enabling_destroy(enab);
14881 dtrace_dof_destroy(dof);
14890 * Now we need to walk through the ECB descriptions in the enabling.
14892 for (i = 0; i < enab->dten_ndesc; i++) {
14893 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
14894 dtrace_probedesc_t *desc = &ep->dted_probe;
14896 if (strcmp(desc->dtpd_provider, "dtrace") != 0)
14899 if (strcmp(desc->dtpd_mod, "helper") != 0)
14902 if (strcmp(desc->dtpd_func, "ustack") != 0)
14905 if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
14908 * Adding this helper action failed -- we are now going
14909 * to rip out the entire generation and return failure.
14911 (void) dtrace_helper_destroygen(help->dthps_generation);
14912 dtrace_enabling_destroy(enab);
14913 dtrace_dof_destroy(dof);
14920 if (nhelpers < enab->dten_ndesc)
14921 dtrace_dof_error(dof, "unmatched helpers");
14923 gen = help->dthps_generation++;
14924 dtrace_enabling_destroy(enab);
14926 if (dhp != NULL && nprovs > 0) {
14927 dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
14928 if (dtrace_helper_provider_add(dhp, gen) == 0) {
14929 mutex_exit(&dtrace_lock);
14930 dtrace_helper_provider_register(curproc, help, dhp);
14931 mutex_enter(&dtrace_lock);
14938 dtrace_dof_destroy(dof);
14943 static dtrace_helpers_t *
14944 dtrace_helpers_create(proc_t *p)
14946 dtrace_helpers_t *help;
14948 ASSERT(MUTEX_HELD(&dtrace_lock));
14949 ASSERT(p->p_dtrace_helpers == NULL);
14951 help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
14952 help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
14953 DTRACE_NHELPER_ACTIONS, KM_SLEEP);
14955 p->p_dtrace_helpers = help;
14965 dtrace_helpers_destroy(proc_t *p)
14967 dtrace_helpers_t *help;
14968 dtrace_vstate_t *vstate;
14970 proc_t *p = curproc;
14974 mutex_enter(&dtrace_lock);
14976 ASSERT(p->p_dtrace_helpers != NULL);
14977 ASSERT(dtrace_helpers > 0);
14979 help = p->p_dtrace_helpers;
14980 vstate = &help->dthps_vstate;
14983 * We're now going to lose the help from this process.
14985 p->p_dtrace_helpers = NULL;
14989 * Destory the helper actions.
14991 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14992 dtrace_helper_action_t *h, *next;
14994 for (h = help->dthps_actions[i]; h != NULL; h = next) {
14995 next = h->dtha_next;
14996 dtrace_helper_action_destroy(h, vstate);
15001 mutex_exit(&dtrace_lock);
15004 * Destroy the helper providers.
15006 if (help->dthps_maxprovs > 0) {
15007 mutex_enter(&dtrace_meta_lock);
15008 if (dtrace_meta_pid != NULL) {
15009 ASSERT(dtrace_deferred_pid == NULL);
15011 for (i = 0; i < help->dthps_nprovs; i++) {
15012 dtrace_helper_provider_remove(
15013 &help->dthps_provs[i]->dthp_prov, p->p_pid);
15016 mutex_enter(&dtrace_lock);
15017 ASSERT(help->dthps_deferred == 0 ||
15018 help->dthps_next != NULL ||
15019 help->dthps_prev != NULL ||
15020 help == dtrace_deferred_pid);
15023 * Remove the helper from the deferred list.
15025 if (help->dthps_next != NULL)
15026 help->dthps_next->dthps_prev = help->dthps_prev;
15027 if (help->dthps_prev != NULL)
15028 help->dthps_prev->dthps_next = help->dthps_next;
15029 if (dtrace_deferred_pid == help) {
15030 dtrace_deferred_pid = help->dthps_next;
15031 ASSERT(help->dthps_prev == NULL);
15034 mutex_exit(&dtrace_lock);
15037 mutex_exit(&dtrace_meta_lock);
15039 for (i = 0; i < help->dthps_nprovs; i++) {
15040 dtrace_helper_provider_destroy(help->dthps_provs[i]);
15043 kmem_free(help->dthps_provs, help->dthps_maxprovs *
15044 sizeof (dtrace_helper_provider_t *));
15047 mutex_enter(&dtrace_lock);
15049 dtrace_vstate_fini(&help->dthps_vstate);
15050 kmem_free(help->dthps_actions,
15051 sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
15052 kmem_free(help, sizeof (dtrace_helpers_t));
15055 mutex_exit(&dtrace_lock);
15062 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
15064 dtrace_helpers_t *help, *newhelp;
15065 dtrace_helper_action_t *helper, *new, *last;
15067 dtrace_vstate_t *vstate;
15068 int i, j, sz, hasprovs = 0;
15070 mutex_enter(&dtrace_lock);
15071 ASSERT(from->p_dtrace_helpers != NULL);
15072 ASSERT(dtrace_helpers > 0);
15074 help = from->p_dtrace_helpers;
15075 newhelp = dtrace_helpers_create(to);
15076 ASSERT(to->p_dtrace_helpers != NULL);
15078 newhelp->dthps_generation = help->dthps_generation;
15079 vstate = &newhelp->dthps_vstate;
15082 * Duplicate the helper actions.
15084 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15085 if ((helper = help->dthps_actions[i]) == NULL)
15088 for (last = NULL; helper != NULL; helper = helper->dtha_next) {
15089 new = kmem_zalloc(sizeof (dtrace_helper_action_t),
15091 new->dtha_generation = helper->dtha_generation;
15093 if ((dp = helper->dtha_predicate) != NULL) {
15094 dp = dtrace_difo_duplicate(dp, vstate);
15095 new->dtha_predicate = dp;
15098 new->dtha_nactions = helper->dtha_nactions;
15099 sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
15100 new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
15102 for (j = 0; j < new->dtha_nactions; j++) {
15103 dtrace_difo_t *dp = helper->dtha_actions[j];
15105 ASSERT(dp != NULL);
15106 dp = dtrace_difo_duplicate(dp, vstate);
15107 new->dtha_actions[j] = dp;
15110 if (last != NULL) {
15111 last->dtha_next = new;
15113 newhelp->dthps_actions[i] = new;
15121 * Duplicate the helper providers and register them with the
15122 * DTrace framework.
15124 if (help->dthps_nprovs > 0) {
15125 newhelp->dthps_nprovs = help->dthps_nprovs;
15126 newhelp->dthps_maxprovs = help->dthps_nprovs;
15127 newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
15128 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
15129 for (i = 0; i < newhelp->dthps_nprovs; i++) {
15130 newhelp->dthps_provs[i] = help->dthps_provs[i];
15131 newhelp->dthps_provs[i]->dthp_ref++;
15137 mutex_exit(&dtrace_lock);
15140 dtrace_helper_provider_register(to, newhelp, NULL);
15144 * DTrace Hook Functions
15147 dtrace_module_loaded(modctl_t *ctl)
15149 dtrace_provider_t *prv;
15151 mutex_enter(&dtrace_provider_lock);
15153 mutex_enter(&mod_lock);
15157 ASSERT(ctl->mod_busy);
15161 * We're going to call each providers per-module provide operation
15162 * specifying only this module.
15164 for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
15165 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
15168 mutex_exit(&mod_lock);
15170 mutex_exit(&dtrace_provider_lock);
15173 * If we have any retained enablings, we need to match against them.
15174 * Enabling probes requires that cpu_lock be held, and we cannot hold
15175 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
15176 * module. (In particular, this happens when loading scheduling
15177 * classes.) So if we have any retained enablings, we need to dispatch
15178 * our task queue to do the match for us.
15180 mutex_enter(&dtrace_lock);
15182 if (dtrace_retained == NULL) {
15183 mutex_exit(&dtrace_lock);
15187 (void) taskq_dispatch(dtrace_taskq,
15188 (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
15190 mutex_exit(&dtrace_lock);
15193 * And now, for a little heuristic sleaze: in general, we want to
15194 * match modules as soon as they load. However, we cannot guarantee
15195 * this, because it would lead us to the lock ordering violation
15196 * outlined above. The common case, of course, is that cpu_lock is
15197 * _not_ held -- so we delay here for a clock tick, hoping that that's
15198 * long enough for the task queue to do its work. If it's not, it's
15199 * not a serious problem -- it just means that the module that we
15200 * just loaded may not be immediately instrumentable.
15207 dtrace_module_unloaded(modctl_t *ctl)
15209 dtrace_module_unloaded(modctl_t *ctl, int *error)
15212 dtrace_probe_t template, *probe, *first, *next;
15213 dtrace_provider_t *prov;
15215 char modname[DTRACE_MODNAMELEN];
15220 template.dtpr_mod = ctl->mod_modname;
15222 /* Handle the fact that ctl->filename may end in ".ko". */
15223 strlcpy(modname, ctl->filename, sizeof(modname));
15224 len = strlen(ctl->filename);
15225 if (len > 3 && strcmp(modname + len - 3, ".ko") == 0)
15226 modname[len - 3] = '\0';
15227 template.dtpr_mod = modname;
15230 mutex_enter(&dtrace_provider_lock);
15232 mutex_enter(&mod_lock);
15234 mutex_enter(&dtrace_lock);
15237 if (ctl->nenabled > 0) {
15238 /* Don't allow unloads if a probe is enabled. */
15239 mutex_exit(&dtrace_provider_lock);
15240 mutex_exit(&dtrace_lock);
15243 "kldunload: attempt to unload module that has DTrace probes enabled\n");
15248 if (dtrace_bymod == NULL) {
15250 * The DTrace module is loaded (obviously) but not attached;
15251 * we don't have any work to do.
15253 mutex_exit(&dtrace_provider_lock);
15255 mutex_exit(&mod_lock);
15257 mutex_exit(&dtrace_lock);
15261 for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
15262 probe != NULL; probe = probe->dtpr_nextmod) {
15263 if (probe->dtpr_ecb != NULL) {
15264 mutex_exit(&dtrace_provider_lock);
15266 mutex_exit(&mod_lock);
15268 mutex_exit(&dtrace_lock);
15271 * This shouldn't _actually_ be possible -- we're
15272 * unloading a module that has an enabled probe in it.
15273 * (It's normally up to the provider to make sure that
15274 * this can't happen.) However, because dtps_enable()
15275 * doesn't have a failure mode, there can be an
15276 * enable/unload race. Upshot: we don't want to
15277 * assert, but we're not going to disable the
15280 if (dtrace_err_verbose) {
15282 cmn_err(CE_WARN, "unloaded module '%s' had "
15283 "enabled probes", ctl->mod_modname);
15285 cmn_err(CE_WARN, "unloaded module '%s' had "
15286 "enabled probes", modname);
15296 for (first = NULL; probe != NULL; probe = next) {
15297 ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
15299 dtrace_probes[probe->dtpr_id - 1] = NULL;
15301 next = probe->dtpr_nextmod;
15302 dtrace_hash_remove(dtrace_bymod, probe);
15303 dtrace_hash_remove(dtrace_byfunc, probe);
15304 dtrace_hash_remove(dtrace_byname, probe);
15306 if (first == NULL) {
15308 probe->dtpr_nextmod = NULL;
15310 probe->dtpr_nextmod = first;
15316 * We've removed all of the module's probes from the hash chains and
15317 * from the probe array. Now issue a dtrace_sync() to be sure that
15318 * everyone has cleared out from any probe array processing.
15322 for (probe = first; probe != NULL; probe = first) {
15323 first = probe->dtpr_nextmod;
15324 prov = probe->dtpr_provider;
15325 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
15327 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
15328 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
15329 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
15331 vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
15333 free_unr(dtrace_arena, probe->dtpr_id);
15335 kmem_free(probe, sizeof (dtrace_probe_t));
15338 mutex_exit(&dtrace_lock);
15340 mutex_exit(&mod_lock);
15342 mutex_exit(&dtrace_provider_lock);
15347 dtrace_kld_load(void *arg __unused, linker_file_t lf)
15350 dtrace_module_loaded(lf);
15354 dtrace_kld_unload_try(void *arg __unused, linker_file_t lf, int *error)
15358 /* We already have an error, so don't do anything. */
15360 dtrace_module_unloaded(lf, error);
15366 dtrace_suspend(void)
15368 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
15372 dtrace_resume(void)
15374 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
15379 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
15381 ASSERT(MUTEX_HELD(&cpu_lock));
15382 mutex_enter(&dtrace_lock);
15386 dtrace_state_t *state;
15387 dtrace_optval_t *opt, rs, c;
15390 * For now, we only allocate a new buffer for anonymous state.
15392 if ((state = dtrace_anon.dta_state) == NULL)
15395 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
15398 opt = state->dts_options;
15399 c = opt[DTRACEOPT_CPU];
15401 if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
15405 * Regardless of what the actual policy is, we're going to
15406 * temporarily set our resize policy to be manual. We're
15407 * also going to temporarily set our CPU option to denote
15408 * the newly configured CPU.
15410 rs = opt[DTRACEOPT_BUFRESIZE];
15411 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
15412 opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
15414 (void) dtrace_state_buffers(state);
15416 opt[DTRACEOPT_BUFRESIZE] = rs;
15417 opt[DTRACEOPT_CPU] = c;
15424 * We don't free the buffer in the CPU_UNCONFIG case. (The
15425 * buffer will be freed when the consumer exits.)
15433 mutex_exit(&dtrace_lock);
15439 dtrace_cpu_setup_initial(processorid_t cpu)
15441 (void) dtrace_cpu_setup(CPU_CONFIG, cpu);
15446 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
15448 if (dtrace_toxranges >= dtrace_toxranges_max) {
15450 dtrace_toxrange_t *range;
15452 osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15455 ASSERT(dtrace_toxrange == NULL);
15456 ASSERT(dtrace_toxranges_max == 0);
15457 dtrace_toxranges_max = 1;
15459 dtrace_toxranges_max <<= 1;
15462 nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15463 range = kmem_zalloc(nsize, KM_SLEEP);
15465 if (dtrace_toxrange != NULL) {
15466 ASSERT(osize != 0);
15467 bcopy(dtrace_toxrange, range, osize);
15468 kmem_free(dtrace_toxrange, osize);
15471 dtrace_toxrange = range;
15474 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == 0);
15475 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == 0);
15477 dtrace_toxrange[dtrace_toxranges].dtt_base = base;
15478 dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
15479 dtrace_toxranges++;
15483 * DTrace Driver Cookbook Functions
15488 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
15490 dtrace_provider_id_t id;
15491 dtrace_state_t *state = NULL;
15492 dtrace_enabling_t *enab;
15494 mutex_enter(&cpu_lock);
15495 mutex_enter(&dtrace_provider_lock);
15496 mutex_enter(&dtrace_lock);
15498 if (ddi_soft_state_init(&dtrace_softstate,
15499 sizeof (dtrace_state_t), 0) != 0) {
15500 cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
15501 mutex_exit(&cpu_lock);
15502 mutex_exit(&dtrace_provider_lock);
15503 mutex_exit(&dtrace_lock);
15504 return (DDI_FAILURE);
15507 if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
15508 DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
15509 ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
15510 DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
15511 cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
15512 ddi_remove_minor_node(devi, NULL);
15513 ddi_soft_state_fini(&dtrace_softstate);
15514 mutex_exit(&cpu_lock);
15515 mutex_exit(&dtrace_provider_lock);
15516 mutex_exit(&dtrace_lock);
15517 return (DDI_FAILURE);
15520 ddi_report_dev(devi);
15521 dtrace_devi = devi;
15523 dtrace_modload = dtrace_module_loaded;
15524 dtrace_modunload = dtrace_module_unloaded;
15525 dtrace_cpu_init = dtrace_cpu_setup_initial;
15526 dtrace_helpers_cleanup = dtrace_helpers_destroy;
15527 dtrace_helpers_fork = dtrace_helpers_duplicate;
15528 dtrace_cpustart_init = dtrace_suspend;
15529 dtrace_cpustart_fini = dtrace_resume;
15530 dtrace_debugger_init = dtrace_suspend;
15531 dtrace_debugger_fini = dtrace_resume;
15533 register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
15535 ASSERT(MUTEX_HELD(&cpu_lock));
15537 dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
15538 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
15539 dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
15540 UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
15541 VM_SLEEP | VMC_IDENTIFIER);
15542 dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
15545 dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
15546 sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
15547 NULL, NULL, NULL, NULL, NULL, 0);
15549 ASSERT(MUTEX_HELD(&cpu_lock));
15550 dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
15551 offsetof(dtrace_probe_t, dtpr_nextmod),
15552 offsetof(dtrace_probe_t, dtpr_prevmod));
15554 dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
15555 offsetof(dtrace_probe_t, dtpr_nextfunc),
15556 offsetof(dtrace_probe_t, dtpr_prevfunc));
15558 dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
15559 offsetof(dtrace_probe_t, dtpr_nextname),
15560 offsetof(dtrace_probe_t, dtpr_prevname));
15562 if (dtrace_retain_max < 1) {
15563 cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; "
15564 "setting to 1", dtrace_retain_max);
15565 dtrace_retain_max = 1;
15569 * Now discover our toxic ranges.
15571 dtrace_toxic_ranges(dtrace_toxrange_add);
15574 * Before we register ourselves as a provider to our own framework,
15575 * we would like to assert that dtrace_provider is NULL -- but that's
15576 * not true if we were loaded as a dependency of a DTrace provider.
15577 * Once we've registered, we can assert that dtrace_provider is our
15580 (void) dtrace_register("dtrace", &dtrace_provider_attr,
15581 DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
15583 ASSERT(dtrace_provider != NULL);
15584 ASSERT((dtrace_provider_id_t)dtrace_provider == id);
15586 dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
15587 dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
15588 dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
15589 dtrace_provider, NULL, NULL, "END", 0, NULL);
15590 dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
15591 dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
15593 dtrace_anon_property();
15594 mutex_exit(&cpu_lock);
15597 * If DTrace helper tracing is enabled, we need to allocate the
15598 * trace buffer and initialize the values.
15600 if (dtrace_helptrace_enabled) {
15601 ASSERT(dtrace_helptrace_buffer == NULL);
15602 dtrace_helptrace_buffer =
15603 kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
15604 dtrace_helptrace_next = 0;
15608 * If there are already providers, we must ask them to provide their
15609 * probes, and then match any anonymous enabling against them. Note
15610 * that there should be no other retained enablings at this time:
15611 * the only retained enablings at this time should be the anonymous
15614 if (dtrace_anon.dta_enabling != NULL) {
15615 ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
15617 dtrace_enabling_provide(NULL);
15618 state = dtrace_anon.dta_state;
15621 * We couldn't hold cpu_lock across the above call to
15622 * dtrace_enabling_provide(), but we must hold it to actually
15623 * enable the probes. We have to drop all of our locks, pick
15624 * up cpu_lock, and regain our locks before matching the
15625 * retained anonymous enabling.
15627 mutex_exit(&dtrace_lock);
15628 mutex_exit(&dtrace_provider_lock);
15630 mutex_enter(&cpu_lock);
15631 mutex_enter(&dtrace_provider_lock);
15632 mutex_enter(&dtrace_lock);
15634 if ((enab = dtrace_anon.dta_enabling) != NULL)
15635 (void) dtrace_enabling_match(enab, NULL);
15637 mutex_exit(&cpu_lock);
15640 mutex_exit(&dtrace_lock);
15641 mutex_exit(&dtrace_provider_lock);
15643 if (state != NULL) {
15645 * If we created any anonymous state, set it going now.
15647 (void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
15650 return (DDI_SUCCESS);
15655 #if __FreeBSD_version >= 800039
15656 static void dtrace_dtr(void *);
15663 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
15665 dtrace_open(struct cdev *dev, int oflags, int devtype, struct thread *td)
15668 dtrace_state_t *state;
15674 if (getminor(*devp) == DTRACEMNRN_HELPER)
15678 * If this wasn't an open with the "helper" minor, then it must be
15679 * the "dtrace" minor.
15681 ASSERT(getminor(*devp) == DTRACEMNRN_DTRACE);
15683 cred_t *cred_p = NULL;
15685 #if __FreeBSD_version < 800039
15687 * The first minor device is the one that is cloned so there is
15688 * nothing more to do here.
15690 if (dev2unit(dev) == 0)
15694 * Devices are cloned, so if the DTrace state has already
15695 * been allocated, that means this device belongs to a
15696 * different client. Each client should open '/dev/dtrace'
15697 * to get a cloned device.
15699 if (dev->si_drv1 != NULL)
15703 cred_p = dev->si_cred;
15707 * If no DTRACE_PRIV_* bits are set in the credential, then the
15708 * caller lacks sufficient permission to do anything with DTrace.
15710 dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
15711 if (priv == DTRACE_PRIV_NONE) {
15713 #if __FreeBSD_version < 800039
15714 /* Destroy the cloned device. */
15723 * Ask all providers to provide all their probes.
15725 mutex_enter(&dtrace_provider_lock);
15726 dtrace_probe_provide(NULL, NULL);
15727 mutex_exit(&dtrace_provider_lock);
15729 mutex_enter(&cpu_lock);
15730 mutex_enter(&dtrace_lock);
15732 dtrace_membar_producer();
15736 * If the kernel debugger is active (that is, if the kernel debugger
15737 * modified text in some way), we won't allow the open.
15739 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
15741 mutex_exit(&cpu_lock);
15742 mutex_exit(&dtrace_lock);
15746 state = dtrace_state_create(devp, cred_p);
15748 state = dtrace_state_create(dev);
15749 #if __FreeBSD_version < 800039
15750 dev->si_drv1 = state;
15752 devfs_set_cdevpriv(state, dtrace_dtr);
15756 mutex_exit(&cpu_lock);
15758 if (state == NULL) {
15760 if (--dtrace_opens == 0)
15761 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15765 mutex_exit(&dtrace_lock);
15767 #if __FreeBSD_version < 800039
15768 /* Destroy the cloned device. */
15775 mutex_exit(&dtrace_lock);
15783 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
15784 #elif __FreeBSD_version < 800039
15786 dtrace_close(struct cdev *dev, int flags, int fmt __unused, struct thread *td)
15789 dtrace_dtr(void *data)
15793 minor_t minor = getminor(dev);
15794 dtrace_state_t *state;
15796 if (minor == DTRACEMNRN_HELPER)
15799 state = ddi_get_soft_state(dtrace_softstate, minor);
15801 #if __FreeBSD_version < 800039
15802 dtrace_state_t *state = dev->si_drv1;
15804 /* Check if this is not a cloned device. */
15805 if (dev2unit(dev) == 0)
15808 dtrace_state_t *state = data;
15813 mutex_enter(&cpu_lock);
15814 mutex_enter(&dtrace_lock);
15816 if (state != NULL) {
15817 if (state->dts_anon) {
15819 * There is anonymous state. Destroy that first.
15821 ASSERT(dtrace_anon.dta_state == NULL);
15822 dtrace_state_destroy(state->dts_anon);
15825 dtrace_state_destroy(state);
15828 kmem_free(state, 0);
15829 #if __FreeBSD_version < 800039
15830 dev->si_drv1 = NULL;
15835 ASSERT(dtrace_opens > 0);
15837 if (--dtrace_opens == 0)
15838 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15843 mutex_exit(&dtrace_lock);
15844 mutex_exit(&cpu_lock);
15846 #if __FreeBSD_version < 800039
15847 /* Schedule this cloned device to be destroyed. */
15848 destroy_dev_sched(dev);
15851 #if defined(sun) || __FreeBSD_version < 800039
15859 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
15862 dof_helper_t help, *dhp = NULL;
15865 case DTRACEHIOC_ADDDOF:
15866 if (copyin((void *)arg, &help, sizeof (help)) != 0) {
15867 dtrace_dof_error(NULL, "failed to copyin DOF helper");
15872 arg = (intptr_t)help.dofhp_dof;
15875 case DTRACEHIOC_ADD: {
15876 dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
15881 mutex_enter(&dtrace_lock);
15884 * dtrace_helper_slurp() takes responsibility for the dof --
15885 * it may free it now or it may save it and free it later.
15887 if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
15894 mutex_exit(&dtrace_lock);
15898 case DTRACEHIOC_REMOVE: {
15899 mutex_enter(&dtrace_lock);
15900 rval = dtrace_helper_destroygen(arg);
15901 mutex_exit(&dtrace_lock);
15915 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
15917 minor_t minor = getminor(dev);
15918 dtrace_state_t *state;
15921 if (minor == DTRACEMNRN_HELPER)
15922 return (dtrace_ioctl_helper(cmd, arg, rv));
15924 state = ddi_get_soft_state(dtrace_softstate, minor);
15926 if (state->dts_anon) {
15927 ASSERT(dtrace_anon.dta_state == NULL);
15928 state = state->dts_anon;
15932 case DTRACEIOC_PROVIDER: {
15933 dtrace_providerdesc_t pvd;
15934 dtrace_provider_t *pvp;
15936 if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
15939 pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
15940 mutex_enter(&dtrace_provider_lock);
15942 for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
15943 if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
15947 mutex_exit(&dtrace_provider_lock);
15952 bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
15953 bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
15955 if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
15961 case DTRACEIOC_EPROBE: {
15962 dtrace_eprobedesc_t epdesc;
15964 dtrace_action_t *act;
15970 if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
15973 mutex_enter(&dtrace_lock);
15975 if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
15976 mutex_exit(&dtrace_lock);
15980 if (ecb->dte_probe == NULL) {
15981 mutex_exit(&dtrace_lock);
15985 epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
15986 epdesc.dtepd_uarg = ecb->dte_uarg;
15987 epdesc.dtepd_size = ecb->dte_size;
15989 nrecs = epdesc.dtepd_nrecs;
15990 epdesc.dtepd_nrecs = 0;
15991 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15992 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15995 epdesc.dtepd_nrecs++;
15999 * Now that we have the size, we need to allocate a temporary
16000 * buffer in which to store the complete description. We need
16001 * the temporary buffer to be able to drop dtrace_lock()
16002 * across the copyout(), below.
16004 size = sizeof (dtrace_eprobedesc_t) +
16005 (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
16007 buf = kmem_alloc(size, KM_SLEEP);
16008 dest = (uintptr_t)buf;
16010 bcopy(&epdesc, (void *)dest, sizeof (epdesc));
16011 dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
16013 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
16014 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
16020 bcopy(&act->dta_rec, (void *)dest,
16021 sizeof (dtrace_recdesc_t));
16022 dest += sizeof (dtrace_recdesc_t);
16025 mutex_exit(&dtrace_lock);
16027 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
16028 kmem_free(buf, size);
16032 kmem_free(buf, size);
16036 case DTRACEIOC_AGGDESC: {
16037 dtrace_aggdesc_t aggdesc;
16038 dtrace_action_t *act;
16039 dtrace_aggregation_t *agg;
16042 dtrace_recdesc_t *lrec;
16047 if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
16050 mutex_enter(&dtrace_lock);
16052 if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
16053 mutex_exit(&dtrace_lock);
16057 aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
16059 nrecs = aggdesc.dtagd_nrecs;
16060 aggdesc.dtagd_nrecs = 0;
16062 offs = agg->dtag_base;
16063 lrec = &agg->dtag_action.dta_rec;
16064 aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
16066 for (act = agg->dtag_first; ; act = act->dta_next) {
16067 ASSERT(act->dta_intuple ||
16068 DTRACEACT_ISAGG(act->dta_kind));
16071 * If this action has a record size of zero, it
16072 * denotes an argument to the aggregating action.
16073 * Because the presence of this record doesn't (or
16074 * shouldn't) affect the way the data is interpreted,
16075 * we don't copy it out to save user-level the
16076 * confusion of dealing with a zero-length record.
16078 if (act->dta_rec.dtrd_size == 0) {
16079 ASSERT(agg->dtag_hasarg);
16083 aggdesc.dtagd_nrecs++;
16085 if (act == &agg->dtag_action)
16090 * Now that we have the size, we need to allocate a temporary
16091 * buffer in which to store the complete description. We need
16092 * the temporary buffer to be able to drop dtrace_lock()
16093 * across the copyout(), below.
16095 size = sizeof (dtrace_aggdesc_t) +
16096 (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
16098 buf = kmem_alloc(size, KM_SLEEP);
16099 dest = (uintptr_t)buf;
16101 bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
16102 dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
16104 for (act = agg->dtag_first; ; act = act->dta_next) {
16105 dtrace_recdesc_t rec = act->dta_rec;
16108 * See the comment in the above loop for why we pass
16109 * over zero-length records.
16111 if (rec.dtrd_size == 0) {
16112 ASSERT(agg->dtag_hasarg);
16119 rec.dtrd_offset -= offs;
16120 bcopy(&rec, (void *)dest, sizeof (rec));
16121 dest += sizeof (dtrace_recdesc_t);
16123 if (act == &agg->dtag_action)
16127 mutex_exit(&dtrace_lock);
16129 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
16130 kmem_free(buf, size);
16134 kmem_free(buf, size);
16138 case DTRACEIOC_ENABLE: {
16140 dtrace_enabling_t *enab = NULL;
16141 dtrace_vstate_t *vstate;
16147 * If a NULL argument has been passed, we take this as our
16148 * cue to reevaluate our enablings.
16151 dtrace_enabling_matchall();
16156 if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
16159 mutex_enter(&cpu_lock);
16160 mutex_enter(&dtrace_lock);
16161 vstate = &state->dts_vstate;
16163 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
16164 mutex_exit(&dtrace_lock);
16165 mutex_exit(&cpu_lock);
16166 dtrace_dof_destroy(dof);
16170 if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
16171 mutex_exit(&dtrace_lock);
16172 mutex_exit(&cpu_lock);
16173 dtrace_dof_destroy(dof);
16177 if ((rval = dtrace_dof_options(dof, state)) != 0) {
16178 dtrace_enabling_destroy(enab);
16179 mutex_exit(&dtrace_lock);
16180 mutex_exit(&cpu_lock);
16181 dtrace_dof_destroy(dof);
16185 if ((err = dtrace_enabling_match(enab, rv)) == 0) {
16186 err = dtrace_enabling_retain(enab);
16188 dtrace_enabling_destroy(enab);
16191 mutex_exit(&cpu_lock);
16192 mutex_exit(&dtrace_lock);
16193 dtrace_dof_destroy(dof);
16198 case DTRACEIOC_REPLICATE: {
16199 dtrace_repldesc_t desc;
16200 dtrace_probedesc_t *match = &desc.dtrpd_match;
16201 dtrace_probedesc_t *create = &desc.dtrpd_create;
16204 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16207 match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16208 match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16209 match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16210 match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16212 create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16213 create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16214 create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16215 create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16217 mutex_enter(&dtrace_lock);
16218 err = dtrace_enabling_replicate(state, match, create);
16219 mutex_exit(&dtrace_lock);
16224 case DTRACEIOC_PROBEMATCH:
16225 case DTRACEIOC_PROBES: {
16226 dtrace_probe_t *probe = NULL;
16227 dtrace_probedesc_t desc;
16228 dtrace_probekey_t pkey;
16235 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16238 desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16239 desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16240 desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16241 desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16244 * Before we attempt to match this probe, we want to give
16245 * all providers the opportunity to provide it.
16247 if (desc.dtpd_id == DTRACE_IDNONE) {
16248 mutex_enter(&dtrace_provider_lock);
16249 dtrace_probe_provide(&desc, NULL);
16250 mutex_exit(&dtrace_provider_lock);
16254 if (cmd == DTRACEIOC_PROBEMATCH) {
16255 dtrace_probekey(&desc, &pkey);
16256 pkey.dtpk_id = DTRACE_IDNONE;
16259 dtrace_cred2priv(cr, &priv, &uid, &zoneid);
16261 mutex_enter(&dtrace_lock);
16263 if (cmd == DTRACEIOC_PROBEMATCH) {
16264 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
16265 if ((probe = dtrace_probes[i - 1]) != NULL &&
16266 (m = dtrace_match_probe(probe, &pkey,
16267 priv, uid, zoneid)) != 0)
16272 mutex_exit(&dtrace_lock);
16277 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
16278 if ((probe = dtrace_probes[i - 1]) != NULL &&
16279 dtrace_match_priv(probe, priv, uid, zoneid))
16284 if (probe == NULL) {
16285 mutex_exit(&dtrace_lock);
16289 dtrace_probe_description(probe, &desc);
16290 mutex_exit(&dtrace_lock);
16292 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16298 case DTRACEIOC_PROBEARG: {
16299 dtrace_argdesc_t desc;
16300 dtrace_probe_t *probe;
16301 dtrace_provider_t *prov;
16303 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16306 if (desc.dtargd_id == DTRACE_IDNONE)
16309 if (desc.dtargd_ndx == DTRACE_ARGNONE)
16312 mutex_enter(&dtrace_provider_lock);
16313 mutex_enter(&mod_lock);
16314 mutex_enter(&dtrace_lock);
16316 if (desc.dtargd_id > dtrace_nprobes) {
16317 mutex_exit(&dtrace_lock);
16318 mutex_exit(&mod_lock);
16319 mutex_exit(&dtrace_provider_lock);
16323 if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
16324 mutex_exit(&dtrace_lock);
16325 mutex_exit(&mod_lock);
16326 mutex_exit(&dtrace_provider_lock);
16330 mutex_exit(&dtrace_lock);
16332 prov = probe->dtpr_provider;
16334 if (prov->dtpv_pops.dtps_getargdesc == NULL) {
16336 * There isn't any typed information for this probe.
16337 * Set the argument number to DTRACE_ARGNONE.
16339 desc.dtargd_ndx = DTRACE_ARGNONE;
16341 desc.dtargd_native[0] = '\0';
16342 desc.dtargd_xlate[0] = '\0';
16343 desc.dtargd_mapping = desc.dtargd_ndx;
16345 prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
16346 probe->dtpr_id, probe->dtpr_arg, &desc);
16349 mutex_exit(&mod_lock);
16350 mutex_exit(&dtrace_provider_lock);
16352 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16358 case DTRACEIOC_GO: {
16359 processorid_t cpuid;
16360 rval = dtrace_state_go(state, &cpuid);
16365 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16371 case DTRACEIOC_STOP: {
16372 processorid_t cpuid;
16374 mutex_enter(&dtrace_lock);
16375 rval = dtrace_state_stop(state, &cpuid);
16376 mutex_exit(&dtrace_lock);
16381 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16387 case DTRACEIOC_DOFGET: {
16388 dof_hdr_t hdr, *dof;
16391 if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
16394 mutex_enter(&dtrace_lock);
16395 dof = dtrace_dof_create(state);
16396 mutex_exit(&dtrace_lock);
16398 len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
16399 rval = copyout(dof, (void *)arg, len);
16400 dtrace_dof_destroy(dof);
16402 return (rval == 0 ? 0 : EFAULT);
16405 case DTRACEIOC_AGGSNAP:
16406 case DTRACEIOC_BUFSNAP: {
16407 dtrace_bufdesc_t desc;
16409 dtrace_buffer_t *buf;
16411 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16414 if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
16417 mutex_enter(&dtrace_lock);
16419 if (cmd == DTRACEIOC_BUFSNAP) {
16420 buf = &state->dts_buffer[desc.dtbd_cpu];
16422 buf = &state->dts_aggbuffer[desc.dtbd_cpu];
16425 if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
16426 size_t sz = buf->dtb_offset;
16428 if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
16429 mutex_exit(&dtrace_lock);
16434 * If this buffer has already been consumed, we're
16435 * going to indicate that there's nothing left here
16438 if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
16439 mutex_exit(&dtrace_lock);
16441 desc.dtbd_size = 0;
16442 desc.dtbd_drops = 0;
16443 desc.dtbd_errors = 0;
16444 desc.dtbd_oldest = 0;
16445 sz = sizeof (desc);
16447 if (copyout(&desc, (void *)arg, sz) != 0)
16454 * If this is a ring buffer that has wrapped, we want
16455 * to copy the whole thing out.
16457 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
16458 dtrace_buffer_polish(buf);
16459 sz = buf->dtb_size;
16462 if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
16463 mutex_exit(&dtrace_lock);
16467 desc.dtbd_size = sz;
16468 desc.dtbd_drops = buf->dtb_drops;
16469 desc.dtbd_errors = buf->dtb_errors;
16470 desc.dtbd_oldest = buf->dtb_xamot_offset;
16471 desc.dtbd_timestamp = dtrace_gethrtime();
16473 mutex_exit(&dtrace_lock);
16475 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16478 buf->dtb_flags |= DTRACEBUF_CONSUMED;
16483 if (buf->dtb_tomax == NULL) {
16484 ASSERT(buf->dtb_xamot == NULL);
16485 mutex_exit(&dtrace_lock);
16489 cached = buf->dtb_tomax;
16490 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
16492 dtrace_xcall(desc.dtbd_cpu,
16493 (dtrace_xcall_t)dtrace_buffer_switch, buf);
16495 state->dts_errors += buf->dtb_xamot_errors;
16498 * If the buffers did not actually switch, then the cross call
16499 * did not take place -- presumably because the given CPU is
16500 * not in the ready set. If this is the case, we'll return
16503 if (buf->dtb_tomax == cached) {
16504 ASSERT(buf->dtb_xamot != cached);
16505 mutex_exit(&dtrace_lock);
16509 ASSERT(cached == buf->dtb_xamot);
16512 * We have our snapshot; now copy it out.
16514 if (copyout(buf->dtb_xamot, desc.dtbd_data,
16515 buf->dtb_xamot_offset) != 0) {
16516 mutex_exit(&dtrace_lock);
16520 desc.dtbd_size = buf->dtb_xamot_offset;
16521 desc.dtbd_drops = buf->dtb_xamot_drops;
16522 desc.dtbd_errors = buf->dtb_xamot_errors;
16523 desc.dtbd_oldest = 0;
16524 desc.dtbd_timestamp = buf->dtb_switched;
16526 mutex_exit(&dtrace_lock);
16529 * Finally, copy out the buffer description.
16531 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16537 case DTRACEIOC_CONF: {
16538 dtrace_conf_t conf;
16540 bzero(&conf, sizeof (conf));
16541 conf.dtc_difversion = DIF_VERSION;
16542 conf.dtc_difintregs = DIF_DIR_NREGS;
16543 conf.dtc_diftupregs = DIF_DTR_NREGS;
16544 conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
16546 if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
16552 case DTRACEIOC_STATUS: {
16553 dtrace_status_t stat;
16554 dtrace_dstate_t *dstate;
16559 * See the comment in dtrace_state_deadman() for the reason
16560 * for setting dts_laststatus to INT64_MAX before setting
16561 * it to the correct value.
16563 state->dts_laststatus = INT64_MAX;
16564 dtrace_membar_producer();
16565 state->dts_laststatus = dtrace_gethrtime();
16567 bzero(&stat, sizeof (stat));
16569 mutex_enter(&dtrace_lock);
16571 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
16572 mutex_exit(&dtrace_lock);
16576 if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
16577 stat.dtst_exiting = 1;
16579 nerrs = state->dts_errors;
16580 dstate = &state->dts_vstate.dtvs_dynvars;
16582 for (i = 0; i < NCPU; i++) {
16583 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
16585 stat.dtst_dyndrops += dcpu->dtdsc_drops;
16586 stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
16587 stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
16589 if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
16590 stat.dtst_filled++;
16592 nerrs += state->dts_buffer[i].dtb_errors;
16594 for (j = 0; j < state->dts_nspeculations; j++) {
16595 dtrace_speculation_t *spec;
16596 dtrace_buffer_t *buf;
16598 spec = &state->dts_speculations[j];
16599 buf = &spec->dtsp_buffer[i];
16600 stat.dtst_specdrops += buf->dtb_xamot_drops;
16604 stat.dtst_specdrops_busy = state->dts_speculations_busy;
16605 stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
16606 stat.dtst_stkstroverflows = state->dts_stkstroverflows;
16607 stat.dtst_dblerrors = state->dts_dblerrors;
16609 (state->dts_activity == DTRACE_ACTIVITY_KILLED);
16610 stat.dtst_errors = nerrs;
16612 mutex_exit(&dtrace_lock);
16614 if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
16620 case DTRACEIOC_FORMAT: {
16621 dtrace_fmtdesc_t fmt;
16625 if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
16628 mutex_enter(&dtrace_lock);
16630 if (fmt.dtfd_format == 0 ||
16631 fmt.dtfd_format > state->dts_nformats) {
16632 mutex_exit(&dtrace_lock);
16637 * Format strings are allocated contiguously and they are
16638 * never freed; if a format index is less than the number
16639 * of formats, we can assert that the format map is non-NULL
16640 * and that the format for the specified index is non-NULL.
16642 ASSERT(state->dts_formats != NULL);
16643 str = state->dts_formats[fmt.dtfd_format - 1];
16644 ASSERT(str != NULL);
16646 len = strlen(str) + 1;
16648 if (len > fmt.dtfd_length) {
16649 fmt.dtfd_length = len;
16651 if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
16652 mutex_exit(&dtrace_lock);
16656 if (copyout(str, fmt.dtfd_string, len) != 0) {
16657 mutex_exit(&dtrace_lock);
16662 mutex_exit(&dtrace_lock);
16675 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
16677 dtrace_state_t *state;
16684 return (DDI_SUCCESS);
16687 return (DDI_FAILURE);
16690 mutex_enter(&cpu_lock);
16691 mutex_enter(&dtrace_provider_lock);
16692 mutex_enter(&dtrace_lock);
16694 ASSERT(dtrace_opens == 0);
16696 if (dtrace_helpers > 0) {
16697 mutex_exit(&dtrace_provider_lock);
16698 mutex_exit(&dtrace_lock);
16699 mutex_exit(&cpu_lock);
16700 return (DDI_FAILURE);
16703 if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
16704 mutex_exit(&dtrace_provider_lock);
16705 mutex_exit(&dtrace_lock);
16706 mutex_exit(&cpu_lock);
16707 return (DDI_FAILURE);
16710 dtrace_provider = NULL;
16712 if ((state = dtrace_anon_grab()) != NULL) {
16714 * If there were ECBs on this state, the provider should
16715 * have not been allowed to detach; assert that there is
16718 ASSERT(state->dts_necbs == 0);
16719 dtrace_state_destroy(state);
16722 * If we're being detached with anonymous state, we need to
16723 * indicate to the kernel debugger that DTrace is now inactive.
16725 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
16728 bzero(&dtrace_anon, sizeof (dtrace_anon_t));
16729 unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
16730 dtrace_cpu_init = NULL;
16731 dtrace_helpers_cleanup = NULL;
16732 dtrace_helpers_fork = NULL;
16733 dtrace_cpustart_init = NULL;
16734 dtrace_cpustart_fini = NULL;
16735 dtrace_debugger_init = NULL;
16736 dtrace_debugger_fini = NULL;
16737 dtrace_modload = NULL;
16738 dtrace_modunload = NULL;
16740 mutex_exit(&cpu_lock);
16742 if (dtrace_helptrace_enabled) {
16743 kmem_free(dtrace_helptrace_buffer, dtrace_helptrace_bufsize);
16744 dtrace_helptrace_buffer = NULL;
16747 kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
16748 dtrace_probes = NULL;
16749 dtrace_nprobes = 0;
16751 dtrace_hash_destroy(dtrace_bymod);
16752 dtrace_hash_destroy(dtrace_byfunc);
16753 dtrace_hash_destroy(dtrace_byname);
16754 dtrace_bymod = NULL;
16755 dtrace_byfunc = NULL;
16756 dtrace_byname = NULL;
16758 kmem_cache_destroy(dtrace_state_cache);
16759 vmem_destroy(dtrace_minor);
16760 vmem_destroy(dtrace_arena);
16762 if (dtrace_toxrange != NULL) {
16763 kmem_free(dtrace_toxrange,
16764 dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
16765 dtrace_toxrange = NULL;
16766 dtrace_toxranges = 0;
16767 dtrace_toxranges_max = 0;
16770 ddi_remove_minor_node(dtrace_devi, NULL);
16771 dtrace_devi = NULL;
16773 ddi_soft_state_fini(&dtrace_softstate);
16775 ASSERT(dtrace_vtime_references == 0);
16776 ASSERT(dtrace_opens == 0);
16777 ASSERT(dtrace_retained == NULL);
16779 mutex_exit(&dtrace_lock);
16780 mutex_exit(&dtrace_provider_lock);
16783 * We don't destroy the task queue until after we have dropped our
16784 * locks (taskq_destroy() may block on running tasks). To prevent
16785 * attempting to do work after we have effectively detached but before
16786 * the task queue has been destroyed, all tasks dispatched via the
16787 * task queue must check that DTrace is still attached before
16788 * performing any operation.
16790 taskq_destroy(dtrace_taskq);
16791 dtrace_taskq = NULL;
16793 return (DDI_SUCCESS);
16800 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
16805 case DDI_INFO_DEVT2DEVINFO:
16806 *result = (void *)dtrace_devi;
16807 error = DDI_SUCCESS;
16809 case DDI_INFO_DEVT2INSTANCE:
16810 *result = (void *)0;
16811 error = DDI_SUCCESS;
16814 error = DDI_FAILURE;
16821 static struct cb_ops dtrace_cb_ops = {
16822 dtrace_open, /* open */
16823 dtrace_close, /* close */
16824 nulldev, /* strategy */
16825 nulldev, /* print */
16829 dtrace_ioctl, /* ioctl */
16830 nodev, /* devmap */
16832 nodev, /* segmap */
16833 nochpoll, /* poll */
16834 ddi_prop_op, /* cb_prop_op */
16836 D_NEW | D_MP /* Driver compatibility flag */
16839 static struct dev_ops dtrace_ops = {
16840 DEVO_REV, /* devo_rev */
16842 dtrace_info, /* get_dev_info */
16843 nulldev, /* identify */
16844 nulldev, /* probe */
16845 dtrace_attach, /* attach */
16846 dtrace_detach, /* detach */
16848 &dtrace_cb_ops, /* driver operations */
16849 NULL, /* bus operations */
16850 nodev /* dev power */
16853 static struct modldrv modldrv = {
16854 &mod_driverops, /* module type (this is a pseudo driver) */
16855 "Dynamic Tracing", /* name of module */
16856 &dtrace_ops, /* driver ops */
16859 static struct modlinkage modlinkage = {
16868 return (mod_install(&modlinkage));
16872 _info(struct modinfo *modinfop)
16874 return (mod_info(&modlinkage, modinfop));
16880 return (mod_remove(&modlinkage));
16884 static d_ioctl_t dtrace_ioctl;
16885 static d_ioctl_t dtrace_ioctl_helper;
16886 static void dtrace_load(void *);
16887 static int dtrace_unload(void);
16888 #if __FreeBSD_version < 800039
16889 static void dtrace_clone(void *, struct ucred *, char *, int , struct cdev **);
16890 static struct clonedevs *dtrace_clones; /* Ptr to the array of cloned devices. */
16891 static eventhandler_tag eh_tag; /* Event handler tag. */
16893 static struct cdev *dtrace_dev;
16894 static struct cdev *helper_dev;
16897 void dtrace_invop_init(void);
16898 void dtrace_invop_uninit(void);
16900 static struct cdevsw dtrace_cdevsw = {
16901 .d_version = D_VERSION,
16902 #if __FreeBSD_version < 800039
16903 .d_flags = D_TRACKCLOSE | D_NEEDMINOR,
16904 .d_close = dtrace_close,
16906 .d_ioctl = dtrace_ioctl,
16907 .d_open = dtrace_open,
16908 .d_name = "dtrace",
16911 static struct cdevsw helper_cdevsw = {
16912 .d_version = D_VERSION,
16913 .d_ioctl = dtrace_ioctl_helper,
16914 .d_name = "helper",
16917 #include <dtrace_anon.c>
16918 #if __FreeBSD_version < 800039
16919 #include <dtrace_clone.c>
16921 #include <dtrace_ioctl.c>
16922 #include <dtrace_load.c>
16923 #include <dtrace_modevent.c>
16924 #include <dtrace_sysctl.c>
16925 #include <dtrace_unload.c>
16926 #include <dtrace_vtime.c>
16927 #include <dtrace_hacks.c>
16928 #include <dtrace_isa.c>
16930 SYSINIT(dtrace_load, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_load, NULL);
16931 SYSUNINIT(dtrace_unload, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_unload, NULL);
16932 SYSINIT(dtrace_anon_init, SI_SUB_DTRACE_ANON, SI_ORDER_FIRST, dtrace_anon_init, NULL);
16934 DEV_MODULE(dtrace, dtrace_modevent, NULL);
16935 MODULE_VERSION(dtrace, 1);
16936 MODULE_DEPEND(dtrace, cyclic, 1, 1, 1);
16937 MODULE_DEPEND(dtrace, opensolaris, 1, 1, 1);