1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License, Version 1.0 only 6 * (the "License"). You may not use this file except in compliance 7 * with the License. 8 * 9 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 10 * or http://www.opensolaris.org/os/licensing. 11 * See the License for the specific language governing permissions 12 * and limitations under the License. 13 * 14 * When distributing Covered Code, include this CDDL HEADER in each 15 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 16 * If applicable, add the following below this CDDL HEADER, with the 17 * fields enclosed by brackets "[]" replaced with your own identifying 18 * information: Portions Copyright [yyyy] [name of copyright owner] 19 * 20 * CDDL HEADER END 21 */ 22 /* 23 * Copyright 2005 Sun Microsystems, Inc. All rights reserved. 24 * Use is subject to license terms. 25 */ 26 27 #pragma ident "%Z%%M% %I% %E% SMI" 28 29 /* 30 * DTrace - Dynamic Tracing for Solaris 31 * 32 * This is the implementation of the Solaris Dynamic Tracing framework 33 * (DTrace). The user-visible interface to DTrace is described at length in 34 * the "Solaris Dynamic Tracing Guide". The interfaces between the libdtrace 35 * library, the in-kernel DTrace framework, and the DTrace providers are 36 * described in the block comments in the <sys/dtrace.h> header file. The 37 * internal architecture of DTrace is described in the block comments in the 38 * <sys/dtrace_impl.h> header file. The comments contained within the DTrace 39 * implementation very much assume mastery of all of these sources; if one has 40 * an unanswered question about the implementation, one should consult them 41 * first. 42 * 43 * The functions here are ordered roughly as follows: 44 * 45 * - Probe context functions 46 * - Probe hashing functions 47 * - Non-probe context utility functions 48 * - Matching functions 49 * - Provider-to-Framework API functions 50 * - Probe management functions 51 * - DIF object functions 52 * - Format functions 53 * - Predicate functions 54 * - ECB functions 55 * - Buffer functions 56 * - Enabling functions 57 * - DOF functions 58 * - Anonymous enabling functions 59 * - Consumer state functions 60 * - Helper functions 61 * - Hook functions 62 * - Driver cookbook functions 63 * 64 * Each group of functions begins with a block comment labelled the "DTrace 65 * [Group] Functions", allowing one to find each block by searching forward 66 * on capital-f functions. 67 */ 68 #include <sys/errno.h> 69 #include <sys/stat.h> 70 #include <sys/modctl.h> 71 #include <sys/conf.h> 72 #include <sys/systm.h> 73 #include <sys/ddi.h> 74 #include <sys/sunddi.h> 75 #include <sys/cpuvar.h> 76 #include <sys/kmem.h> 77 #include <sys/strsubr.h> 78 #include <sys/sysmacros.h> 79 #include <sys/dtrace_impl.h> 80 #include <sys/atomic.h> 81 #include <sys/cmn_err.h> 82 #include <sys/mutex_impl.h> 83 #include <sys/rwlock_impl.h> 84 #include <sys/ctf_api.h> 85 #include <sys/panic.h> 86 #include <sys/priv_impl.h> 87 #include <sys/policy.h> 88 #include <sys/cred_impl.h> 89 #include <sys/procfs_isa.h> 90 #include <sys/taskq.h> 91 #include <sys/mkdev.h> 92 #include <sys/kdi.h> 93 #include <sys/zone.h> 94 95 /* 96 * DTrace Tunable Variables 97 * 98 * The following variables may be tuned by adding a line to /etc/system that 99 * includes both the name of the DTrace module ("dtrace") and the name of the 100 * variable. For example: 101 * 102 * set dtrace:dtrace_destructive_disallow = 1 103 * 104 * In general, the only variables that one should be tuning this way are those 105 * that affect system-wide DTrace behavior, and for which the default behavior 106 * is undesirable. Most of these variables are tunable on a per-consumer 107 * basis using DTrace options, and need not be tuned on a system-wide basis. 108 * When tuning these variables, avoid pathological values; while some attempt 109 * is made to verify the integrity of these variables, they are not considered 110 * part of the supported interface to DTrace, and they are therefore not 111 * checked comprehensively. Further, these variables should not be tuned 112 * dynamically via "mdb -kw" or other means; they should only be tuned via 113 * /etc/system. 114 */ 115 int dtrace_destructive_disallow = 0; 116 dtrace_optval_t dtrace_nonroot_maxsize = (16 * 1024 * 1024); 117 size_t dtrace_difo_maxsize = (256 * 1024); 118 dtrace_optval_t dtrace_dof_maxsize = (256 * 1024); 119 size_t dtrace_global_maxsize = (16 * 1024); 120 size_t dtrace_actions_max = (16 * 1024); 121 size_t dtrace_retain_max = 1024; 122 dtrace_optval_t dtrace_helper_actions_max = 32; 123 dtrace_optval_t dtrace_helper_providers_max = 32; 124 dtrace_optval_t dtrace_dstate_defsize = (1 * 1024 * 1024); 125 size_t dtrace_strsize_default = 256; 126 dtrace_optval_t dtrace_cleanrate_default = 9900990; /* 101 hz */ 127 dtrace_optval_t dtrace_cleanrate_min = 200000; /* 5000 hz */ 128 dtrace_optval_t dtrace_cleanrate_max = (uint64_t)60 * NANOSEC; /* 1/minute */ 129 dtrace_optval_t dtrace_aggrate_default = NANOSEC; /* 1 hz */ 130 dtrace_optval_t dtrace_statusrate_default = NANOSEC; /* 1 hz */ 131 dtrace_optval_t dtrace_statusrate_max = (hrtime_t)10 * NANOSEC; /* 6/minute */ 132 dtrace_optval_t dtrace_switchrate_default = NANOSEC; /* 1 hz */ 133 dtrace_optval_t dtrace_nspec_default = 1; 134 dtrace_optval_t dtrace_specsize_default = 32 * 1024; 135 dtrace_optval_t dtrace_stackframes_default = 20; 136 dtrace_optval_t dtrace_ustackframes_default = 20; 137 dtrace_optval_t dtrace_jstackframes_default = 50; 138 dtrace_optval_t dtrace_jstackstrsize_default = 512; 139 int dtrace_msgdsize_max = 128; 140 hrtime_t dtrace_chill_max = 500 * (NANOSEC / MILLISEC); /* 500 ms */ 141 hrtime_t dtrace_chill_interval = NANOSEC; /* 1000 ms */ 142 int dtrace_devdepth_max = 32; 143 int dtrace_err_verbose; 144 hrtime_t dtrace_deadman_interval = NANOSEC; 145 hrtime_t dtrace_deadman_timeout = (hrtime_t)10 * NANOSEC; 146 hrtime_t dtrace_deadman_user = (hrtime_t)30 * NANOSEC; 147 148 /* 149 * DTrace External Variables 150 * 151 * As dtrace(7D) is a kernel module, any DTrace variables are obviously 152 * available to DTrace consumers via the backtick (`) syntax. One of these, 153 * dtrace_zero, is made deliberately so: it is provided as a source of 154 * well-known, zero-filled memory. While this variable is not documented, 155 * it is used by some translators as an implementation detail. 156 */ 157 const char dtrace_zero[256] = { 0 }; /* zero-filled memory */ 158 159 /* 160 * DTrace Internal Variables 161 */ 162 static dev_info_t *dtrace_devi; /* device info */ 163 static vmem_t *dtrace_arena; /* probe ID arena */ 164 static vmem_t *dtrace_minor; /* minor number arena */ 165 static taskq_t *dtrace_taskq; /* task queue */ 166 static dtrace_probe_t **dtrace_probes; /* array of all probes */ 167 static int dtrace_nprobes; /* number of probes */ 168 static dtrace_provider_t *dtrace_provider; /* provider list */ 169 static dtrace_meta_t *dtrace_meta_pid; /* user-land meta provider */ 170 static int dtrace_opens; /* number of opens */ 171 static int dtrace_helpers; /* number of helpers */ 172 static void *dtrace_softstate; /* softstate pointer */ 173 static dtrace_hash_t *dtrace_bymod; /* probes hashed by module */ 174 static dtrace_hash_t *dtrace_byfunc; /* probes hashed by function */ 175 static dtrace_hash_t *dtrace_byname; /* probes hashed by name */ 176 static dtrace_toxrange_t *dtrace_toxrange; /* toxic range array */ 177 static int dtrace_toxranges; /* number of toxic ranges */ 178 static int dtrace_toxranges_max; /* size of toxic range array */ 179 static dtrace_anon_t dtrace_anon; /* anonymous enabling */ 180 static kmem_cache_t *dtrace_state_cache; /* cache for dynamic state */ 181 static uint64_t dtrace_vtime_references; /* number of vtimestamp refs */ 182 static kthread_t *dtrace_panicked; /* panicking thread */ 183 static dtrace_ecb_t *dtrace_ecb_create_cache; /* cached created ECB */ 184 static dtrace_genid_t dtrace_probegen; /* current probe generation */ 185 static dtrace_helpers_t *dtrace_deferred_pid; /* deferred helper list */ 186 static dtrace_enabling_t *dtrace_retained; /* list of retained enablings */ 187 static dtrace_state_t *dtrace_state; /* temporary variable */ 188 static int dtrace_err; /* temporary variable */ 189 190 /* 191 * DTrace Locking 192 * DTrace is protected by three (relatively coarse-grained) locks: 193 * 194 * (1) dtrace_lock is required to manipulate essentially any DTrace state, 195 * including enabling state, probes, ECBs, consumer state, helper state, 196 * etc. Importantly, dtrace_lock is _not_ required when in probe context; 197 * probe context is lock-free -- synchronization is handled via the 198 * dtrace_sync() cross call mechanism. 199 * 200 * (2) dtrace_provider_lock is required when manipulating provider state, or 201 * when provider state must be held constant. 202 * 203 * (3) dtrace_meta_lock is required when manipulating meta provider state, or 204 * when meta provider state must be held constant. 205 * 206 * The lock ordering between these three locks is dtrace_meta_lock before 207 * dtrace_provider_lock before dtrace_lock. (In particular, there are 208 * several places where dtrace_provider_lock is held by the framework as it 209 * calls into the providers -- which then call back into the framework, 210 * grabbing dtrace_lock.) 211 * 212 * There are two other locks in the mix: mod_lock and cpu_lock. With respect 213 * to dtrace_provider_lock and dtrace_lock, cpu_lock continues its historical 214 * role as a coarse-grained lock; it is acquired before both of these locks. 215 * With respect to dtrace_meta_lock, its behavior is stranger: cpu_lock must 216 * be acquired _between_ dtrace_meta_lock and any other DTrace locks. 217 * mod_lock is similar with respect to dtrace_provider_lock in that it must be 218 * acquired _between_ dtrace_provider_lock and dtrace_lock. 219 */ 220 static kmutex_t dtrace_lock; /* probe state lock */ 221 static kmutex_t dtrace_provider_lock; /* provider state lock */ 222 static kmutex_t dtrace_meta_lock; /* meta-provider state lock */ 223 224 /* 225 * DTrace Provider Variables 226 * 227 * These are the variables relating to DTrace as a provider (that is, the 228 * provider of the BEGIN, END, and ERROR probes). 229 */ 230 static dtrace_pattr_t dtrace_provider_attr = { 231 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON }, 232 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN }, 233 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN }, 234 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON }, 235 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON }, 236 }; 237 238 static void 239 dtrace_nullop(void) 240 {} 241 242 static dtrace_pops_t dtrace_provider_ops = { 243 (void (*)(void *, const dtrace_probedesc_t *))dtrace_nullop, 244 (void (*)(void *, struct modctl *))dtrace_nullop, 245 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop, 246 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop, 247 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop, 248 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop, 249 NULL, 250 NULL, 251 NULL, 252 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop 253 }; 254 255 static dtrace_id_t dtrace_probeid_begin; /* special BEGIN probe */ 256 static dtrace_id_t dtrace_probeid_end; /* special END probe */ 257 dtrace_id_t dtrace_probeid_error; /* special ERROR probe */ 258 259 /* 260 * DTrace Helper Tracing Variables 261 */ 262 uint32_t dtrace_helptrace_next = 0; 263 uint32_t dtrace_helptrace_nlocals; 264 char *dtrace_helptrace_buffer; 265 int dtrace_helptrace_bufsize = 512 * 1024; 266 267 #ifdef DEBUG 268 int dtrace_helptrace_enabled = 1; 269 #else 270 int dtrace_helptrace_enabled = 0; 271 #endif 272 273 /* 274 * DTrace Error Hashing 275 * 276 * On DEBUG kernels, DTrace will track the errors that has seen in a hash 277 * table. This is very useful for checking coverage of tests that are 278 * expected to induce DIF or DOF processing errors, and may be useful for 279 * debugging problems in the DIF code generator or in DOF generation . The 280 * error hash may be examined with the ::dtrace_errhash MDB dcmd. 281 */ 282 #ifdef DEBUG 283 static dtrace_errhash_t dtrace_errhash[DTRACE_ERRHASHSZ]; 284 static const char *dtrace_errlast; 285 static kthread_t *dtrace_errthread; 286 static kmutex_t dtrace_errlock; 287 #endif 288 289 /* 290 * DTrace Macros and Constants 291 * 292 * These are various macros that are useful in various spots in the 293 * implementation, along with a few random constants that have no meaning 294 * outside of the implementation. There is no real structure to this cpp 295 * mishmash -- but is there ever? 296 */ 297 #define DTRACE_HASHSTR(hash, probe) \ 298 dtrace_hash_str(*((char **)((uintptr_t)(probe) + (hash)->dth_stroffs))) 299 300 #define DTRACE_HASHNEXT(hash, probe) \ 301 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_nextoffs) 302 303 #define DTRACE_HASHPREV(hash, probe) \ 304 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_prevoffs) 305 306 #define DTRACE_HASHEQ(hash, lhs, rhs) \ 307 (strcmp(*((char **)((uintptr_t)(lhs) + (hash)->dth_stroffs)), \ 308 *((char **)((uintptr_t)(rhs) + (hash)->dth_stroffs))) == 0) 309 310 #define DTRACE_AGGHASHSIZE_SLEW 17 311 312 /* 313 * The key for a thread-local variable consists of the lower 61 bits of the 314 * t_did, plus the 3 bits of the highest active interrupt above LOCK_LEVEL. 315 * We add DIF_VARIABLE_MAX to t_did to assure that the thread key is never 316 * equal to a variable identifier. This is necessary (but not sufficient) to 317 * assure that global associative arrays never collide with thread-local 318 * variables. To guarantee that they cannot collide, we must also define the 319 * order for keying dynamic variables. That order is: 320 * 321 * [ key0 ] ... [ keyn ] [ variable-key ] [ tls-key ] 322 * 323 * Because the variable-key and the tls-key are in orthogonal spaces, there is 324 * no way for a global variable key signature to match a thread-local key 325 * signature. 326 */ 327 #define DTRACE_TLS_THRKEY(where) { \ 328 uint_t intr = 0; \ 329 uint_t actv = CPU->cpu_intr_actv >> (LOCK_LEVEL + 1); \ 330 for (; actv; actv >>= 1) \ 331 intr++; \ 332 ASSERT(intr < (1 << 3)); \ 333 (where) = ((curthread->t_did + DIF_VARIABLE_MAX) & \ 334 (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \ 335 } 336 337 #define DTRACE_STORE(type, tomax, offset, what) \ 338 *((type *)((uintptr_t)(tomax) + (uintptr_t)offset)) = (type)(what); 339 340 #ifndef __i386 341 #define DTRACE_ALIGNCHECK(addr, size, flags) \ 342 if (addr & (size - 1)) { \ 343 *flags |= CPU_DTRACE_BADALIGN; \ 344 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = addr; \ 345 return (0); \ 346 } 347 #else 348 #define DTRACE_ALIGNCHECK(addr, size, flags) 349 #endif 350 351 #define DTRACE_LOADFUNC(bits) \ 352 /*CSTYLED*/ \ 353 uint##bits##_t \ 354 dtrace_load##bits(uintptr_t addr) \ 355 { \ 356 size_t size = bits / NBBY; \ 357 /*CSTYLED*/ \ 358 uint##bits##_t rval; \ 359 int i; \ 360 volatile uint16_t *flags = (volatile uint16_t *) \ 361 &cpu_core[CPU->cpu_id].cpuc_dtrace_flags; \ 362 \ 363 DTRACE_ALIGNCHECK(addr, size, flags); \ 364 \ 365 for (i = 0; i < dtrace_toxranges; i++) { \ 366 if (addr >= dtrace_toxrange[i].dtt_limit) \ 367 continue; \ 368 \ 369 if (addr + size <= dtrace_toxrange[i].dtt_base) \ 370 continue; \ 371 \ 372 /* \ 373 * This address falls within a toxic region; return 0. \ 374 */ \ 375 *flags |= CPU_DTRACE_BADADDR; \ 376 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = addr; \ 377 return (0); \ 378 } \ 379 \ 380 *flags |= CPU_DTRACE_NOFAULT; \ 381 /*CSTYLED*/ \ 382 rval = *((volatile uint##bits##_t *)addr); \ 383 *flags &= ~CPU_DTRACE_NOFAULT; \ 384 \ 385 return (rval); \ 386 } 387 388 #ifdef _LP64 389 #define dtrace_loadptr dtrace_load64 390 #else 391 #define dtrace_loadptr dtrace_load32 392 #endif 393 394 #define DTRACE_MATCH_NEXT 0 395 #define DTRACE_MATCH_DONE 1 396 #define DTRACE_ANCHORED(probe) ((probe)->dtpr_func[0] != '\0') 397 #define DTRACE_STATE_ALIGN 64 398 399 #define DTRACE_FLAGS2FLT(flags) \ 400 (((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR : \ 401 ((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP : \ 402 ((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO : \ 403 ((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV : \ 404 ((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV : \ 405 ((flags) & CPU_DTRACE_TUPOFLOW) ? DTRACEFLT_TUPOFLOW : \ 406 ((flags) & CPU_DTRACE_BADALIGN) ? DTRACEFLT_BADALIGN : \ 407 ((flags) & CPU_DTRACE_NOSCRATCH) ? DTRACEFLT_NOSCRATCH : \ 408 DTRACEFLT_UNKNOWN) 409 410 #define DTRACEACT_ISSTRING(act) \ 411 ((act)->dta_kind == DTRACEACT_DIFEXPR && \ 412 (act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) 413 414 static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id); 415 static void dtrace_enabling_provide(dtrace_provider_t *); 416 static int dtrace_enabling_match(dtrace_enabling_t *, int *); 417 static void dtrace_enabling_matchall(void); 418 static dtrace_state_t *dtrace_anon_grab(void); 419 static uint64_t dtrace_helper(int, dtrace_mstate_t *, 420 dtrace_state_t *, uint64_t, uint64_t); 421 static dtrace_helpers_t *dtrace_helpers_create(proc_t *); 422 static void dtrace_buffer_drop(dtrace_buffer_t *); 423 static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t, 424 dtrace_state_t *, dtrace_mstate_t *); 425 static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t, 426 dtrace_optval_t); 427 static int dtrace_ecb_create_enable(dtrace_probe_t *, void *); 428 429 /* 430 * DTrace Probe Context Functions 431 * 432 * These functions are called from probe context. Because probe context is 433 * any context in which C may be called, arbitrarily locks may be held, 434 * interrupts may be disabled, we may be in arbitrary dispatched state, etc. 435 * As a result, functions called from probe context may only call other DTrace 436 * support functions -- they may not interact at all with the system at large. 437 * (Note that the ASSERT macro is made probe-context safe by redefining it in 438 * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary 439 * loads are to be performed from probe context, they _must_ be in terms of 440 * the safe dtrace_load*() variants. 441 * 442 * Some functions in this block are not actually called from probe context; 443 * for these functions, there will be a comment above the function reading 444 * "Note: not called from probe context." 445 */ 446 void 447 dtrace_panic(const char *format, ...) 448 { 449 va_list alist; 450 451 va_start(alist, format); 452 dtrace_vpanic(format, alist); 453 va_end(alist); 454 } 455 456 int 457 dtrace_assfail(const char *a, const char *f, int l) 458 { 459 dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l); 460 461 /* 462 * We just need something here that even the most clever compiler 463 * cannot optimize away. 464 */ 465 return (a[(uintptr_t)f]); 466 } 467 468 /* 469 * Atomically increment a specified error counter from probe context. 470 */ 471 static void 472 dtrace_error(uint32_t *counter) 473 { 474 /* 475 * Most counters stored to in probe context are per-CPU counters. 476 * However, there are some error conditions that are sufficiently 477 * arcane that they don't merit per-CPU storage. If these counters 478 * are incremented concurrently on different CPUs, scalability will be 479 * adversely affected -- but we don't expect them to be white-hot in a 480 * correctly constructed enabling... 481 */ 482 uint32_t oval, nval; 483 484 do { 485 oval = *counter; 486 487 if ((nval = oval + 1) == 0) { 488 /* 489 * If the counter would wrap, set it to 1 -- assuring 490 * that the counter is never zero when we have seen 491 * errors. (The counter must be 32-bits because we 492 * aren't guaranteed a 64-bit compare&swap operation.) 493 * To save this code both the infamy of being fingered 494 * by a priggish news story and the indignity of being 495 * the target of a neo-puritan witch trial, we're 496 * carefully avoiding any colorful description of the 497 * likelihood of this condition -- but suffice it to 498 * say that it is only slightly more likely than the 499 * overflow of predicate cache IDs, as discussed in 500 * dtrace_predicate_create(). 501 */ 502 nval = 1; 503 } 504 } while (dtrace_cas32(counter, oval, nval) != oval); 505 } 506 507 /* 508 * Use the DTRACE_LOADFUNC macro to define functions for each of loading a 509 * uint8_t, a uint16_t, a uint32_t and a uint64_t. 510 */ 511 DTRACE_LOADFUNC(8) 512 DTRACE_LOADFUNC(16) 513 DTRACE_LOADFUNC(32) 514 DTRACE_LOADFUNC(64) 515 516 static int 517 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate) 518 { 519 if (dest < mstate->dtms_scratch_base) 520 return (0); 521 522 if (dest + size < dest) 523 return (0); 524 525 if (dest + size > mstate->dtms_scratch_ptr) 526 return (0); 527 528 return (1); 529 } 530 531 static int 532 dtrace_canstore_statvar(uint64_t addr, size_t sz, 533 dtrace_statvar_t **svars, int nsvars) 534 { 535 int i; 536 537 for (i = 0; i < nsvars; i++) { 538 dtrace_statvar_t *svar = svars[i]; 539 540 if (svar == NULL || svar->dtsv_size == 0) 541 continue; 542 543 if (addr - svar->dtsv_data < svar->dtsv_size && 544 addr + sz <= svar->dtsv_data + svar->dtsv_size) 545 return (1); 546 } 547 548 return (0); 549 } 550 551 /* 552 * Check to see if the address is within a memory region to which a store may 553 * be issued. This includes the DTrace scratch areas, and any DTrace variable 554 * region. The caller of dtrace_canstore() is responsible for performing any 555 * alignment checks that are needed before stores are actually executed. 556 */ 557 static int 558 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate, 559 dtrace_vstate_t *vstate) 560 { 561 uintptr_t a; 562 size_t s; 563 564 /* 565 * First, check to see if the address is in scratch space... 566 */ 567 a = mstate->dtms_scratch_base; 568 s = mstate->dtms_scratch_size; 569 570 if (addr - a < s && addr + sz <= a + s) 571 return (1); 572 573 /* 574 * Now check to see if it's a dynamic variable. This check will pick 575 * up both thread-local variables and any global dynamically-allocated 576 * variables. 577 */ 578 a = (uintptr_t)vstate->dtvs_dynvars.dtds_base; 579 s = vstate->dtvs_dynvars.dtds_size; 580 if (addr - a < s && addr + sz <= a + s) 581 return (1); 582 583 /* 584 * Finally, check the static local and global variables. These checks 585 * take the longest, so we perform them last. 586 */ 587 if (dtrace_canstore_statvar(addr, sz, 588 vstate->dtvs_locals, vstate->dtvs_nlocals)) 589 return (1); 590 591 if (dtrace_canstore_statvar(addr, sz, 592 vstate->dtvs_globals, vstate->dtvs_nglobals)) 593 return (1); 594 595 return (0); 596 } 597 598 /* 599 * Compare two strings using safe loads. 600 */ 601 static int 602 dtrace_strncmp(char *s1, char *s2, size_t limit) 603 { 604 uint8_t c1, c2; 605 volatile uint16_t *flags; 606 607 if (s1 == s2 || limit == 0) 608 return (0); 609 610 flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags; 611 612 do { 613 if (s1 == NULL) { 614 c1 = '\0'; 615 } else { 616 c1 = dtrace_load8((uintptr_t)s1++); 617 } 618 619 if (s2 == NULL) { 620 c2 = '\0'; 621 } else { 622 c2 = dtrace_load8((uintptr_t)s2++); 623 } 624 625 if (c1 != c2) 626 return (c1 - c2); 627 } while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT)); 628 629 return (0); 630 } 631 632 /* 633 * Compute strlen(s) for a string using safe memory accesses. The additional 634 * len parameter is used to specify a maximum length to ensure completion. 635 */ 636 static size_t 637 dtrace_strlen(const char *s, size_t lim) 638 { 639 uint_t len; 640 641 for (len = 0; len != lim; len++) { 642 if (dtrace_load8((uintptr_t)s++) == '\0') 643 break; 644 } 645 646 return (len); 647 } 648 649 /* 650 * Check if an address falls within a toxic region. 651 */ 652 static int 653 dtrace_istoxic(uintptr_t kaddr, size_t size) 654 { 655 uintptr_t taddr, tsize; 656 int i; 657 658 for (i = 0; i < dtrace_toxranges; i++) { 659 taddr = dtrace_toxrange[i].dtt_base; 660 tsize = dtrace_toxrange[i].dtt_limit - taddr; 661 662 if (kaddr - taddr < tsize) { 663 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR); 664 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = kaddr; 665 return (1); 666 } 667 668 if (taddr - kaddr < size) { 669 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR); 670 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = taddr; 671 return (1); 672 } 673 } 674 675 return (0); 676 } 677 678 /* 679 * Copy src to dst using safe memory accesses. The src is assumed to be unsafe 680 * memory specified by the DIF program. The dst is assumed to be safe memory 681 * that we can store to directly because it is managed by DTrace. As with 682 * standard bcopy, overlapping copies are handled properly. 683 */ 684 static void 685 dtrace_bcopy(const void *src, void *dst, size_t len) 686 { 687 if (len != 0) { 688 uint8_t *s1 = dst; 689 const uint8_t *s2 = src; 690 691 if (s1 <= s2) { 692 do { 693 *s1++ = dtrace_load8((uintptr_t)s2++); 694 } while (--len != 0); 695 } else { 696 s2 += len; 697 s1 += len; 698 699 do { 700 *--s1 = dtrace_load8((uintptr_t)--s2); 701 } while (--len != 0); 702 } 703 } 704 } 705 706 /* 707 * Copy src to dst using safe memory accesses, up to either the specified 708 * length, or the point that a nul byte is encountered. The src is assumed to 709 * be unsafe memory specified by the DIF program. The dst is assumed to be 710 * safe memory that we can store to directly because it is managed by DTrace. 711 * Unlike dtrace_bcopy(), overlapping regions are not handled. 712 */ 713 static void 714 dtrace_strcpy(const void *src, void *dst, size_t len) 715 { 716 if (len != 0) { 717 uint8_t *s1 = dst, c; 718 const uint8_t *s2 = src; 719 720 do { 721 *s1++ = c = dtrace_load8((uintptr_t)s2++); 722 } while (--len != 0 && c != '\0'); 723 } 724 } 725 726 /* 727 * Copy src to dst, deriving the size and type from the specified (BYREF) 728 * variable type. The src is assumed to be unsafe memory specified by the DIF 729 * program. The dst is assumed to be DTrace variable memory that is of the 730 * specified type; we assume that we can store to directly. 731 */ 732 static void 733 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type) 734 { 735 ASSERT(type->dtdt_flags & DIF_TF_BYREF); 736 737 if (type->dtdt_kind == DIF_TYPE_STRING) { 738 dtrace_strcpy(src, dst, type->dtdt_size); 739 } else { 740 dtrace_bcopy(src, dst, type->dtdt_size); 741 } 742 } 743 744 /* 745 * Compare s1 to s2 using safe memory accesses. The s1 data is assumed to be 746 * unsafe memory specified by the DIF program. The s2 data is assumed to be 747 * safe memory that we can access directly because it is managed by DTrace. 748 */ 749 static int 750 dtrace_bcmp(const void *s1, const void *s2, size_t len) 751 { 752 volatile uint16_t *flags; 753 754 flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags; 755 756 if (s1 == s2) 757 return (0); 758 759 if (s1 == NULL || s2 == NULL) 760 return (1); 761 762 if (s1 != s2 && len != 0) { 763 const uint8_t *ps1 = s1; 764 const uint8_t *ps2 = s2; 765 766 do { 767 if (dtrace_load8((uintptr_t)ps1++) != *ps2++) 768 return (1); 769 } while (--len != 0 && !(*flags & CPU_DTRACE_FAULT)); 770 } 771 return (0); 772 } 773 774 /* 775 * Zero the specified region using a simple byte-by-byte loop. Note that this 776 * is for safe DTrace-managed memory only. 777 */ 778 static void 779 dtrace_bzero(void *dst, size_t len) 780 { 781 uchar_t *cp; 782 783 for (cp = dst; len != 0; len--) 784 *cp++ = 0; 785 } 786 787 /* 788 * This privilege checks should be used by actions and subroutines to 789 * verify the credentials of the process that enabled the invoking ECB. 790 */ 791 static int 792 dtrace_priv_proc_common(dtrace_state_t *state) 793 { 794 uid_t uid = state->dts_cred.dcr_uid; 795 gid_t gid = state->dts_cred.dcr_gid; 796 cred_t *cr; 797 proc_t *proc; 798 799 if ((cr = CRED()) != NULL && 800 uid == cr->cr_uid && 801 uid == cr->cr_ruid && 802 uid == cr->cr_suid && 803 gid == cr->cr_gid && 804 gid == cr->cr_rgid && 805 gid == cr->cr_sgid && 806 (proc = ttoproc(curthread)) != NULL && 807 !(proc->p_flag & SNOCD)) 808 return (1); 809 810 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV; 811 812 return (0); 813 } 814 815 static int 816 dtrace_priv_proc_destructive(dtrace_state_t *state) 817 { 818 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_DESTRUCTIVE) 819 return (1); 820 821 return (dtrace_priv_proc_common(state)); 822 } 823 824 static int 825 dtrace_priv_proc_control(dtrace_state_t *state) 826 { 827 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL) 828 return (1); 829 830 return (dtrace_priv_proc_common(state)); 831 } 832 833 static int 834 dtrace_priv_proc(dtrace_state_t *state) 835 { 836 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC) 837 return (1); 838 839 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV; 840 841 return (0); 842 } 843 844 static int 845 dtrace_priv_kernel(dtrace_state_t *state) 846 { 847 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL) 848 return (1); 849 850 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV; 851 852 return (0); 853 } 854 855 static int 856 dtrace_priv_kernel_destructive(dtrace_state_t *state) 857 { 858 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE) 859 return (1); 860 861 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV; 862 863 return (0); 864 } 865 866 /* 867 * Note: not called from probe context. This function is called 868 * asynchronously (and at a regular interval) from outside of probe context to 869 * clean the dirty dynamic variable lists on all CPUs. Dynamic variable 870 * cleaning is explained in detail in <sys/dtrace_impl.h>. 871 */ 872 void 873 dtrace_dynvar_clean(dtrace_dstate_t *dstate) 874 { 875 dtrace_dynvar_t *dirty; 876 dtrace_dstate_percpu_t *dcpu; 877 int i, work = 0; 878 879 for (i = 0; i < NCPU; i++) { 880 dcpu = &dstate->dtds_percpu[i]; 881 882 ASSERT(dcpu->dtdsc_rinsing == NULL); 883 884 /* 885 * If the dirty list is NULL, there is no dirty work to do. 886 */ 887 if (dcpu->dtdsc_dirty == NULL) 888 continue; 889 890 /* 891 * If the clean list is non-NULL, then we're not going to do 892 * any work for this CPU -- it means that there has not been 893 * a dtrace_dynvar() allocation on this CPU (or from this CPU) 894 * since the last time we cleaned house. 895 */ 896 if (dcpu->dtdsc_clean != NULL) 897 continue; 898 899 work = 1; 900 901 /* 902 * Atomically move the dirty list aside. 903 */ 904 do { 905 dirty = dcpu->dtdsc_dirty; 906 907 /* 908 * Before we zap the dirty list, set the rinsing list. 909 * (This allows for a potential assertion in 910 * dtrace_dynvar(): if a free dynamic variable appears 911 * on a hash chain, either the dirty list or the 912 * rinsing list for some CPU must be non-NULL.) 913 */ 914 dcpu->dtdsc_rinsing = dirty; 915 dtrace_membar_producer(); 916 } while (dtrace_casptr(&dcpu->dtdsc_dirty, 917 dirty, NULL) != dirty); 918 } 919 920 if (!work) { 921 /* 922 * We have no work to do; we can simply return. 923 */ 924 return; 925 } 926 927 dtrace_sync(); 928 929 for (i = 0; i < NCPU; i++) { 930 dcpu = &dstate->dtds_percpu[i]; 931 932 if (dcpu->dtdsc_rinsing == NULL) 933 continue; 934 935 /* 936 * We are now guaranteed that no hash chain contains a pointer 937 * into this dirty list; we can make it clean. 938 */ 939 ASSERT(dcpu->dtdsc_clean == NULL); 940 dcpu->dtdsc_clean = dcpu->dtdsc_rinsing; 941 dcpu->dtdsc_rinsing = NULL; 942 } 943 944 /* 945 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make 946 * sure that all CPUs have seen all of the dtdsc_clean pointers. 947 * This prevents a race whereby a CPU incorrectly decides that 948 * the state should be something other than DTRACE_DSTATE_CLEAN 949 * after dtrace_dynvar_clean() has completed. 950 */ 951 dtrace_sync(); 952 953 dstate->dtds_state = DTRACE_DSTATE_CLEAN; 954 } 955 956 /* 957 * Depending on the value of the op parameter, this function looks-up, 958 * allocates or deallocates an arbitrarily-keyed dynamic variable. If an 959 * allocation is requested, this function will return a pointer to a 960 * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no 961 * variable can be allocated. If NULL is returned, the appropriate counter 962 * will be incremented. 963 */ 964 dtrace_dynvar_t * 965 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys, 966 dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op) 967 { 968 uint64_t hashval = 1; 969 dtrace_dynhash_t *hash = dstate->dtds_hash; 970 dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL; 971 processorid_t me = CPU->cpu_id, cpu = me; 972 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me]; 973 size_t bucket, ksize; 974 size_t chunksize = dstate->dtds_chunksize; 975 uintptr_t kdata, lock, nstate; 976 uint_t i; 977 978 ASSERT(nkeys != 0); 979 980 /* 981 * Hash the key. As with aggregations, we use Jenkins' "One-at-a-time" 982 * algorithm. For the by-value portions, we perform the algorithm in 983 * 16-bit chunks (as opposed to 8-bit chunks). This speeds things up a 984 * bit, and seems to have only a minute effect on distribution. For 985 * the by-reference data, we perform "One-at-a-time" iterating (safely) 986 * over each referenced byte. It's painful to do this, but it's much 987 * better than pathological hash distribution. The efficacy of the 988 * hashing algorithm (and a comparison with other algorithms) may be 989 * found by running the ::dtrace_dynstat MDB dcmd. 990 */ 991 for (i = 0; i < nkeys; i++) { 992 if (key[i].dttk_size == 0) { 993 uint64_t val = key[i].dttk_value; 994 995 hashval += (val >> 48) & 0xffff; 996 hashval += (hashval << 10); 997 hashval ^= (hashval >> 6); 998 999 hashval += (val >> 32) & 0xffff; 1000 hashval += (hashval << 10); 1001 hashval ^= (hashval >> 6); 1002 1003 hashval += (val >> 16) & 0xffff; 1004 hashval += (hashval << 10); 1005 hashval ^= (hashval >> 6); 1006 1007 hashval += val & 0xffff; 1008 hashval += (hashval << 10); 1009 hashval ^= (hashval >> 6); 1010 } else { 1011 /* 1012 * This is incredibly painful, but it beats the hell 1013 * out of the alternative. 1014 */ 1015 uint64_t j, size = key[i].dttk_size; 1016 uintptr_t base = (uintptr_t)key[i].dttk_value; 1017 1018 for (j = 0; j < size; j++) { 1019 hashval += dtrace_load8(base + j); 1020 hashval += (hashval << 10); 1021 hashval ^= (hashval >> 6); 1022 } 1023 } 1024 } 1025 1026 hashval += (hashval << 3); 1027 hashval ^= (hashval >> 11); 1028 hashval += (hashval << 15); 1029 1030 /* 1031 * There is a remote chance (ideally, 1 in 2^32) that our hashval 1032 * comes out to be 0. We rely on a zero hashval denoting a free 1033 * element; if this actually happens, we set the hashval to 1. 1034 */ 1035 if (hashval == 0) 1036 hashval = 1; 1037 1038 /* 1039 * Yes, it's painful to do a divide here. If the cycle count becomes 1040 * important here, tricks can be pulled to reduce it. (However, it's 1041 * critical that hash collisions be kept to an absolute minimum; 1042 * they're much more painful than a divide.) It's better to have a 1043 * solution that generates few collisions and still keeps things 1044 * relatively simple. 1045 */ 1046 bucket = hashval % dstate->dtds_hashsize; 1047 1048 if (op == DTRACE_DYNVAR_DEALLOC) { 1049 volatile uintptr_t *lockp = &hash[bucket].dtdh_lock; 1050 1051 for (;;) { 1052 while ((lock = *lockp) & 1) 1053 continue; 1054 1055 if (dtrace_casptr((void *)lockp, 1056 (void *)lock, (void *)(lock + 1)) == (void *)lock) 1057 break; 1058 } 1059 1060 dtrace_membar_producer(); 1061 } 1062 1063 top: 1064 prev = NULL; 1065 lock = hash[bucket].dtdh_lock; 1066 1067 dtrace_membar_consumer(); 1068 1069 start = hash[bucket].dtdh_chain; 1070 ASSERT(start == NULL || start->dtdv_hashval != 0 || 1071 op != DTRACE_DYNVAR_DEALLOC); 1072 1073 for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) { 1074 dtrace_tuple_t *dtuple = &dvar->dtdv_tuple; 1075 dtrace_key_t *dkey = &dtuple->dtt_key[0]; 1076 1077 if (dvar->dtdv_hashval != hashval) { 1078 if (dvar->dtdv_hashval == 0) { 1079 /* 1080 * We've gone off the rails. Somewhere 1081 * along the line, one of the members of this 1082 * hash chain was deleted. We could assert 1083 * that either the dirty list or the rinsing 1084 * list is non-NULL. (The dtrace_sync() in 1085 * dtrace_dynvar_clean() would validate this 1086 * assertion.) 1087 */ 1088 ASSERT(op != DTRACE_DYNVAR_DEALLOC); 1089 goto top; 1090 } 1091 1092 goto next; 1093 } 1094 1095 if (dtuple->dtt_nkeys != nkeys) 1096 goto next; 1097 1098 for (i = 0; i < nkeys; i++, dkey++) { 1099 if (dkey->dttk_size != key[i].dttk_size) 1100 goto next; /* size or type mismatch */ 1101 1102 if (dkey->dttk_size != 0) { 1103 if (dtrace_bcmp( 1104 (void *)(uintptr_t)key[i].dttk_value, 1105 (void *)(uintptr_t)dkey->dttk_value, 1106 dkey->dttk_size)) 1107 goto next; 1108 } else { 1109 if (dkey->dttk_value != key[i].dttk_value) 1110 goto next; 1111 } 1112 } 1113 1114 if (op != DTRACE_DYNVAR_DEALLOC) 1115 return (dvar); 1116 1117 ASSERT(dvar->dtdv_next == NULL || 1118 dvar->dtdv_next->dtdv_hashval != 0); 1119 1120 if (prev != NULL) { 1121 ASSERT(hash[bucket].dtdh_chain != dvar); 1122 ASSERT(start != dvar); 1123 ASSERT(prev->dtdv_next == dvar); 1124 prev->dtdv_next = dvar->dtdv_next; 1125 } else { 1126 if (dtrace_casptr(&hash[bucket].dtdh_chain, 1127 start, dvar->dtdv_next) != start) { 1128 /* 1129 * We have failed to atomically swing the 1130 * hash table head pointer, presumably because 1131 * of a conflicting allocation on another CPU. 1132 * We need to reread the hash chain and try 1133 * again. 1134 */ 1135 goto top; 1136 } 1137 } 1138 1139 dtrace_membar_producer(); 1140 1141 /* 1142 * Now clear the hash value to indicate that it's free. 1143 */ 1144 ASSERT(hash[bucket].dtdh_chain != dvar); 1145 dvar->dtdv_hashval = 0; 1146 1147 dtrace_membar_producer(); 1148 1149 /* 1150 * Set the next pointer to point at the dirty list, and 1151 * atomically swing the dirty pointer to the newly freed dvar. 1152 */ 1153 do { 1154 next = dcpu->dtdsc_dirty; 1155 dvar->dtdv_next = next; 1156 } while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next); 1157 1158 /* 1159 * Finally, unlock this hash bucket. 1160 */ 1161 ASSERT(hash[bucket].dtdh_lock == lock); 1162 ASSERT(lock & 1); 1163 hash[bucket].dtdh_lock++; 1164 1165 return (NULL); 1166 next: 1167 prev = dvar; 1168 continue; 1169 } 1170 1171 if (op != DTRACE_DYNVAR_ALLOC) { 1172 /* 1173 * If we are not to allocate a new variable, we want to 1174 * return NULL now. Before we return, check that the value 1175 * of the lock word hasn't changed. If it has, we may have 1176 * seen an inconsistent snapshot. 1177 */ 1178 if (op == DTRACE_DYNVAR_NOALLOC) { 1179 if (hash[bucket].dtdh_lock != lock) 1180 goto top; 1181 } else { 1182 ASSERT(op == DTRACE_DYNVAR_DEALLOC); 1183 ASSERT(hash[bucket].dtdh_lock == lock); 1184 ASSERT(lock & 1); 1185 hash[bucket].dtdh_lock++; 1186 } 1187 1188 return (NULL); 1189 } 1190 1191 /* 1192 * We need to allocate a new dynamic variable. The size we need is the 1193 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the 1194 * size of any auxiliary key data (rounded up to 8-byte alignment) plus 1195 * the size of any referred-to data (dsize). We then round the final 1196 * size up to the chunksize for allocation. 1197 */ 1198 for (ksize = 0, i = 0; i < nkeys; i++) 1199 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t)); 1200 1201 /* 1202 * This should be pretty much impossible, but could happen if, say, 1203 * strange DIF specified the tuple. Ideally, this should be an 1204 * assertion and not an error condition -- but that requires that the 1205 * chunksize calculation in dtrace_difo_chunksize() be absolutely 1206 * bullet-proof. (That is, it must not be able to be fooled by 1207 * malicious DIF.) Given the lack of backwards branches in DIF, 1208 * solving this would presumably not amount to solving the Halting 1209 * Problem -- but it still seems awfully hard. 1210 */ 1211 if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) + 1212 ksize + dsize > chunksize) { 1213 dcpu->dtdsc_drops++; 1214 return (NULL); 1215 } 1216 1217 nstate = DTRACE_DSTATE_EMPTY; 1218 1219 do { 1220 retry: 1221 free = dcpu->dtdsc_free; 1222 1223 if (free == NULL) { 1224 dtrace_dynvar_t *clean = dcpu->dtdsc_clean; 1225 void *rval; 1226 1227 if (clean == NULL) { 1228 /* 1229 * We're out of dynamic variable space on 1230 * this CPU. Unless we have tried all CPUs, 1231 * we'll try to allocate from a different 1232 * CPU. 1233 */ 1234 switch (dstate->dtds_state) { 1235 case DTRACE_DSTATE_CLEAN: { 1236 void *sp = &dstate->dtds_state; 1237 1238 if (++cpu >= NCPU) 1239 cpu = 0; 1240 1241 if (dcpu->dtdsc_dirty != NULL && 1242 nstate == DTRACE_DSTATE_EMPTY) 1243 nstate = DTRACE_DSTATE_DIRTY; 1244 1245 if (dcpu->dtdsc_rinsing != NULL) 1246 nstate = DTRACE_DSTATE_RINSING; 1247 1248 dcpu = &dstate->dtds_percpu[cpu]; 1249 1250 if (cpu != me) 1251 goto retry; 1252 1253 (void) dtrace_cas32(sp, 1254 DTRACE_DSTATE_CLEAN, nstate); 1255 1256 /* 1257 * To increment the correct bean 1258 * counter, take another lap. 1259 */ 1260 goto retry; 1261 } 1262 1263 case DTRACE_DSTATE_DIRTY: 1264 dcpu->dtdsc_dirty_drops++; 1265 break; 1266 1267 case DTRACE_DSTATE_RINSING: 1268 dcpu->dtdsc_rinsing_drops++; 1269 break; 1270 1271 case DTRACE_DSTATE_EMPTY: 1272 dcpu->dtdsc_drops++; 1273 break; 1274 } 1275 1276 DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP); 1277 return (NULL); 1278 } 1279 1280 /* 1281 * The clean list appears to be non-empty. We want to 1282 * move the clean list to the free list; we start by 1283 * moving the clean pointer aside. 1284 */ 1285 if (dtrace_casptr(&dcpu->dtdsc_clean, 1286 clean, NULL) != clean) { 1287 /* 1288 * We are in one of two situations: 1289 * 1290 * (a) The clean list was switched to the 1291 * free list by another CPU. 1292 * 1293 * (b) The clean list was added to by the 1294 * cleansing cyclic. 1295 * 1296 * In either of these situations, we can 1297 * just reattempt the free list allocation. 1298 */ 1299 goto retry; 1300 } 1301 1302 ASSERT(clean->dtdv_hashval == 0); 1303 1304 /* 1305 * Now we'll move the clean list to the free list. 1306 * It's impossible for this to fail: the only way 1307 * the free list can be updated is through this 1308 * code path, and only one CPU can own the clean list. 1309 * Thus, it would only be possible for this to fail if 1310 * this code were racing with dtrace_dynvar_clean(). 1311 * (That is, if dtrace_dynvar_clean() updated the clean 1312 * list, and we ended up racing to update the free 1313 * list.) This race is prevented by the dtrace_sync() 1314 * in dtrace_dynvar_clean() -- which flushes the 1315 * owners of the clean lists out before resetting 1316 * the clean lists. 1317 */ 1318 rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean); 1319 ASSERT(rval == NULL); 1320 goto retry; 1321 } 1322 1323 dvar = free; 1324 new_free = dvar->dtdv_next; 1325 } while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free); 1326 1327 /* 1328 * We have now allocated a new chunk. We copy the tuple keys into the 1329 * tuple array and copy any referenced key data into the data space 1330 * following the tuple array. As we do this, we relocate dttk_value 1331 * in the final tuple to point to the key data address in the chunk. 1332 */ 1333 kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys]; 1334 dvar->dtdv_data = (void *)(kdata + ksize); 1335 dvar->dtdv_tuple.dtt_nkeys = nkeys; 1336 1337 for (i = 0; i < nkeys; i++) { 1338 dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i]; 1339 size_t kesize = key[i].dttk_size; 1340 1341 if (kesize != 0) { 1342 dtrace_bcopy( 1343 (const void *)(uintptr_t)key[i].dttk_value, 1344 (void *)kdata, kesize); 1345 dkey->dttk_value = kdata; 1346 kdata += P2ROUNDUP(kesize, sizeof (uint64_t)); 1347 } else { 1348 dkey->dttk_value = key[i].dttk_value; 1349 } 1350 1351 dkey->dttk_size = kesize; 1352 } 1353 1354 ASSERT(dvar->dtdv_hashval == 0); 1355 dvar->dtdv_hashval = hashval; 1356 dvar->dtdv_next = start; 1357 1358 if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start) 1359 return (dvar); 1360 1361 /* 1362 * The cas has failed. Either another CPU is adding an element to 1363 * this hash chain, or another CPU is deleting an element from this 1364 * hash chain. The simplest way to deal with both of these cases 1365 * (though not necessarily the most efficient) is to free our 1366 * allocated block and tail-call ourselves. Note that the free is 1367 * to the dirty list and _not_ to the free list. This is to prevent 1368 * races with allocators, above. 1369 */ 1370 dvar->dtdv_hashval = 0; 1371 1372 dtrace_membar_producer(); 1373 1374 do { 1375 free = dcpu->dtdsc_dirty; 1376 dvar->dtdv_next = free; 1377 } while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free); 1378 1379 return (dtrace_dynvar(dstate, nkeys, key, dsize, op)); 1380 } 1381 1382 /*ARGSUSED*/ 1383 static void 1384 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg) 1385 { 1386 if (nval < *oval) 1387 *oval = nval; 1388 } 1389 1390 /*ARGSUSED*/ 1391 static void 1392 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg) 1393 { 1394 if (nval > *oval) 1395 *oval = nval; 1396 } 1397 1398 static void 1399 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr) 1400 { 1401 int i, zero = DTRACE_QUANTIZE_ZEROBUCKET; 1402 int64_t val = (int64_t)nval; 1403 1404 if (val < 0) { 1405 for (i = 0; i < zero; i++) { 1406 if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) { 1407 quanta[i] += incr; 1408 return; 1409 } 1410 } 1411 } else { 1412 for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) { 1413 if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) { 1414 quanta[i - 1] += incr; 1415 return; 1416 } 1417 } 1418 1419 quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr; 1420 return; 1421 } 1422 1423 ASSERT(0); 1424 } 1425 1426 static void 1427 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr) 1428 { 1429 uint64_t arg = *lquanta++; 1430 int32_t base = DTRACE_LQUANTIZE_BASE(arg); 1431 uint16_t step = DTRACE_LQUANTIZE_STEP(arg); 1432 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg); 1433 int32_t val = (int32_t)nval, level; 1434 1435 ASSERT(step != 0); 1436 ASSERT(levels != 0); 1437 1438 if (val < base) { 1439 /* 1440 * This is an underflow. 1441 */ 1442 lquanta[0] += incr; 1443 return; 1444 } 1445 1446 level = (val - base) / step; 1447 1448 if (level < levels) { 1449 lquanta[level + 1] += incr; 1450 return; 1451 } 1452 1453 /* 1454 * This is an overflow. 1455 */ 1456 lquanta[levels + 1] += incr; 1457 } 1458 1459 /*ARGSUSED*/ 1460 static void 1461 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg) 1462 { 1463 data[0]++; 1464 data[1] += nval; 1465 } 1466 1467 /*ARGSUSED*/ 1468 static void 1469 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg) 1470 { 1471 *oval = *oval + 1; 1472 } 1473 1474 /*ARGSUSED*/ 1475 static void 1476 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg) 1477 { 1478 *oval += nval; 1479 } 1480 1481 /* 1482 * Aggregate given the tuple in the principal data buffer, and the aggregating 1483 * action denoted by the specified dtrace_aggregation_t. The aggregation 1484 * buffer is specified as the buf parameter. This routine does not return 1485 * failure; if there is no space in the aggregation buffer, the data will be 1486 * dropped, and a corresponding counter incremented. 1487 */ 1488 static void 1489 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf, 1490 intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg) 1491 { 1492 dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec; 1493 uint32_t i, ndx, size, fsize; 1494 uint32_t align = sizeof (uint64_t) - 1; 1495 dtrace_aggbuffer_t *agb; 1496 dtrace_aggkey_t *key; 1497 uint32_t hashval = 0, limit, isstr; 1498 caddr_t tomax, data, kdata; 1499 dtrace_actkind_t action; 1500 dtrace_action_t *act; 1501 uintptr_t offs; 1502 1503 if (buf == NULL) 1504 return; 1505 1506 if (!agg->dtag_hasarg) { 1507 /* 1508 * Currently, only quantize() and lquantize() take additional 1509 * arguments, and they have the same semantics: an increment 1510 * value that defaults to 1 when not present. If additional 1511 * aggregating actions take arguments, the setting of the 1512 * default argument value will presumably have to become more 1513 * sophisticated... 1514 */ 1515 arg = 1; 1516 } 1517 1518 action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION; 1519 size = rec->dtrd_offset - agg->dtag_base; 1520 fsize = size + rec->dtrd_size; 1521 1522 ASSERT(dbuf->dtb_tomax != NULL); 1523 data = dbuf->dtb_tomax + offset + agg->dtag_base; 1524 1525 if ((tomax = buf->dtb_tomax) == NULL) { 1526 dtrace_buffer_drop(buf); 1527 return; 1528 } 1529 1530 /* 1531 * The metastructure is always at the bottom of the buffer. 1532 */ 1533 agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size - 1534 sizeof (dtrace_aggbuffer_t)); 1535 1536 if (buf->dtb_offset == 0) { 1537 /* 1538 * We just kludge up approximately 1/8th of the size to be 1539 * buckets. If this guess ends up being routinely 1540 * off-the-mark, we may need to dynamically readjust this 1541 * based on past performance. 1542 */ 1543 uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t); 1544 1545 if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) < 1546 (uintptr_t)tomax || hashsize == 0) { 1547 /* 1548 * We've been given a ludicrously small buffer; 1549 * increment our drop count and leave. 1550 */ 1551 dtrace_buffer_drop(buf); 1552 return; 1553 } 1554 1555 /* 1556 * And now, a pathetic attempt to try to get a an odd (or 1557 * perchance, a prime) hash size for better hash distribution. 1558 */ 1559 if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3)) 1560 hashsize -= DTRACE_AGGHASHSIZE_SLEW; 1561 1562 agb->dtagb_hashsize = hashsize; 1563 agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb - 1564 agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *)); 1565 agb->dtagb_free = (uintptr_t)agb->dtagb_hash; 1566 1567 for (i = 0; i < agb->dtagb_hashsize; i++) 1568 agb->dtagb_hash[i] = NULL; 1569 } 1570 1571 ASSERT(agg->dtag_first != NULL); 1572 ASSERT(agg->dtag_first->dta_intuple); 1573 1574 /* 1575 * Calculate the hash value based on the key. Note that we _don't_ 1576 * include the aggid in the hashing (but we will store it as part of 1577 * the key). The hashing algorithm is Bob Jenkins' "One-at-a-time" 1578 * algorithm: a simple, quick algorithm that has no known funnels, and 1579 * gets good distribution in practice. The efficacy of the hashing 1580 * algorithm (and a comparison with other algorithms) may be found by 1581 * running the ::dtrace_aggstat MDB dcmd. 1582 */ 1583 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) { 1584 i = act->dta_rec.dtrd_offset - agg->dtag_base; 1585 limit = i + act->dta_rec.dtrd_size; 1586 ASSERT(limit <= size); 1587 isstr = DTRACEACT_ISSTRING(act); 1588 1589 for (; i < limit; i++) { 1590 hashval += data[i]; 1591 hashval += (hashval << 10); 1592 hashval ^= (hashval >> 6); 1593 1594 if (isstr && data[i] == '\0') 1595 break; 1596 } 1597 } 1598 1599 hashval += (hashval << 3); 1600 hashval ^= (hashval >> 11); 1601 hashval += (hashval << 15); 1602 1603 /* 1604 * Yes, the divide here is expensive -- but it's generally the least 1605 * of the performance issues given the amount of data that we iterate 1606 * over to compute hash values, compare data, etc. 1607 */ 1608 ndx = hashval % agb->dtagb_hashsize; 1609 1610 for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) { 1611 ASSERT((caddr_t)key >= tomax); 1612 ASSERT((caddr_t)key < tomax + buf->dtb_size); 1613 1614 if (hashval != key->dtak_hashval || key->dtak_size != size) 1615 continue; 1616 1617 kdata = key->dtak_data; 1618 ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size); 1619 1620 for (act = agg->dtag_first; act->dta_intuple; 1621 act = act->dta_next) { 1622 i = act->dta_rec.dtrd_offset - agg->dtag_base; 1623 limit = i + act->dta_rec.dtrd_size; 1624 ASSERT(limit <= size); 1625 isstr = DTRACEACT_ISSTRING(act); 1626 1627 for (; i < limit; i++) { 1628 if (kdata[i] != data[i]) 1629 goto next; 1630 1631 if (isstr && data[i] == '\0') 1632 break; 1633 } 1634 } 1635 1636 if (action != key->dtak_action) { 1637 /* 1638 * We are aggregating on the same value in the same 1639 * aggregation with two different aggregating actions. 1640 * (This should have been picked up in the compiler, 1641 * so we may be dealing with errant or devious DIF.) 1642 * This is an error condition; we indicate as much, 1643 * and return. 1644 */ 1645 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP); 1646 return; 1647 } 1648 1649 /* 1650 * This is a hit: we need to apply the aggregator to 1651 * the value at this key. 1652 */ 1653 agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg); 1654 return; 1655 next: 1656 continue; 1657 } 1658 1659 /* 1660 * We didn't find it. We need to allocate some zero-filled space, 1661 * link it into the hash table appropriately, and apply the aggregator 1662 * to the (zero-filled) value. 1663 */ 1664 offs = buf->dtb_offset; 1665 while (offs & (align - 1)) 1666 offs += sizeof (uint32_t); 1667 1668 /* 1669 * If we don't have enough room to both allocate a new key _and_ 1670 * its associated data, increment the drop count and return. 1671 */ 1672 if ((uintptr_t)tomax + offs + fsize > 1673 agb->dtagb_free - sizeof (dtrace_aggkey_t)) { 1674 dtrace_buffer_drop(buf); 1675 return; 1676 } 1677 1678 /*CONSTCOND*/ 1679 ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1))); 1680 key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t)); 1681 agb->dtagb_free -= sizeof (dtrace_aggkey_t); 1682 1683 key->dtak_data = kdata = tomax + offs; 1684 buf->dtb_offset = offs + fsize; 1685 1686 /* 1687 * Now copy the data across. 1688 */ 1689 *((dtrace_aggid_t *)kdata) = agg->dtag_id; 1690 1691 for (i = sizeof (dtrace_aggid_t); i < size; i++) 1692 kdata[i] = data[i]; 1693 1694 /* 1695 * Because strings are not zeroed out by default, we need to iterate 1696 * looking for actions that store strings, and we need to explicitly 1697 * pad these strings out with zeroes. 1698 */ 1699 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) { 1700 int nul; 1701 1702 if (!DTRACEACT_ISSTRING(act)) 1703 continue; 1704 1705 i = act->dta_rec.dtrd_offset - agg->dtag_base; 1706 limit = i + act->dta_rec.dtrd_size; 1707 ASSERT(limit <= size); 1708 1709 for (nul = 0; i < limit; i++) { 1710 if (nul) { 1711 kdata[i] = '\0'; 1712 continue; 1713 } 1714 1715 if (data[i] != '\0') 1716 continue; 1717 1718 nul = 1; 1719 } 1720 } 1721 1722 for (i = size; i < fsize; i++) 1723 kdata[i] = 0; 1724 1725 key->dtak_hashval = hashval; 1726 key->dtak_size = size; 1727 key->dtak_action = action; 1728 key->dtak_next = agb->dtagb_hash[ndx]; 1729 agb->dtagb_hash[ndx] = key; 1730 1731 /* 1732 * Finally, apply the aggregator. 1733 */ 1734 *((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial; 1735 agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg); 1736 } 1737 1738 /* 1739 * Given consumer state, this routine finds a speculation in the INACTIVE 1740 * state and transitions it into the ACTIVE state. If there is no speculation 1741 * in the INACTIVE state, 0 is returned. In this case, no error counter is 1742 * incremented -- it is up to the caller to take appropriate action. 1743 */ 1744 static int 1745 dtrace_speculation(dtrace_state_t *state) 1746 { 1747 int i = 0; 1748 dtrace_speculation_state_t current; 1749 uint32_t *stat = &state->dts_speculations_unavail, count; 1750 1751 while (i < state->dts_nspeculations) { 1752 dtrace_speculation_t *spec = &state->dts_speculations[i]; 1753 1754 current = spec->dtsp_state; 1755 1756 if (current != DTRACESPEC_INACTIVE) { 1757 if (current == DTRACESPEC_COMMITTINGMANY || 1758 current == DTRACESPEC_COMMITTING || 1759 current == DTRACESPEC_DISCARDING) 1760 stat = &state->dts_speculations_busy; 1761 i++; 1762 continue; 1763 } 1764 1765 if (dtrace_cas32((uint32_t *)&spec->dtsp_state, 1766 current, DTRACESPEC_ACTIVE) == current) 1767 return (i + 1); 1768 } 1769 1770 /* 1771 * We couldn't find a speculation. If we found as much as a single 1772 * busy speculation buffer, we'll attribute this failure as "busy" 1773 * instead of "unavail". 1774 */ 1775 do { 1776 count = *stat; 1777 } while (dtrace_cas32(stat, count, count + 1) != count); 1778 1779 return (0); 1780 } 1781 1782 /* 1783 * This routine commits an active speculation. If the specified speculation 1784 * is not in a valid state to perform a commit(), this routine will silently do 1785 * nothing. The state of the specified speculation is transitioned according 1786 * to the state transition diagram outlined in <sys/dtrace_impl.h> 1787 */ 1788 static void 1789 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu, 1790 dtrace_specid_t which) 1791 { 1792 dtrace_speculation_t *spec; 1793 dtrace_buffer_t *src, *dest; 1794 uintptr_t daddr, saddr, dlimit; 1795 dtrace_speculation_state_t current, new; 1796 intptr_t offs; 1797 1798 if (which == 0) 1799 return; 1800 1801 if (which > state->dts_nspeculations) { 1802 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP; 1803 return; 1804 } 1805 1806 spec = &state->dts_speculations[which - 1]; 1807 src = &spec->dtsp_buffer[cpu]; 1808 dest = &state->dts_buffer[cpu]; 1809 1810 do { 1811 current = spec->dtsp_state; 1812 1813 if (current == DTRACESPEC_COMMITTINGMANY) 1814 break; 1815 1816 switch (current) { 1817 case DTRACESPEC_INACTIVE: 1818 case DTRACESPEC_DISCARDING: 1819 return; 1820 1821 case DTRACESPEC_COMMITTING: 1822 /* 1823 * This is only possible if we are (a) commit()'ing 1824 * without having done a prior speculate() on this CPU 1825 * and (b) racing with another commit() on a different 1826 * CPU. There's nothing to do -- we just assert that 1827 * our offset is 0. 1828 */ 1829 ASSERT(src->dtb_offset == 0); 1830 return; 1831 1832 case DTRACESPEC_ACTIVE: 1833 new = DTRACESPEC_COMMITTING; 1834 break; 1835 1836 case DTRACESPEC_ACTIVEONE: 1837 /* 1838 * This speculation is active on one CPU. If our 1839 * buffer offset is non-zero, we know that the one CPU 1840 * must be us. Otherwise, we are committing on a 1841 * different CPU from the speculate(), and we must 1842 * rely on being asynchronously cleaned. 1843 */ 1844 if (src->dtb_offset != 0) { 1845 new = DTRACESPEC_COMMITTING; 1846 break; 1847 } 1848 /*FALLTHROUGH*/ 1849 1850 case DTRACESPEC_ACTIVEMANY: 1851 new = DTRACESPEC_COMMITTINGMANY; 1852 break; 1853 1854 default: 1855 ASSERT(0); 1856 } 1857 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state, 1858 current, new) != current); 1859 1860 /* 1861 * We have set the state to indicate that we are committing this 1862 * speculation. Now reserve the necessary space in the destination 1863 * buffer. 1864 */ 1865 if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset, 1866 sizeof (uint64_t), state, NULL)) < 0) { 1867 dtrace_buffer_drop(dest); 1868 goto out; 1869 } 1870 1871 /* 1872 * We have the space; copy the buffer across. (Note that this is a 1873 * highly subobtimal bcopy(); in the unlikely event that this becomes 1874 * a serious performance issue, a high-performance DTrace-specific 1875 * bcopy() should obviously be invented.) 1876 */ 1877 daddr = (uintptr_t)dest->dtb_tomax + offs; 1878 dlimit = daddr + src->dtb_offset; 1879 saddr = (uintptr_t)src->dtb_tomax; 1880 1881 /* 1882 * First, the aligned portion. 1883 */ 1884 while (dlimit - daddr >= sizeof (uint64_t)) { 1885 *((uint64_t *)daddr) = *((uint64_t *)saddr); 1886 1887 daddr += sizeof (uint64_t); 1888 saddr += sizeof (uint64_t); 1889 } 1890 1891 /* 1892 * Now any left-over bit... 1893 */ 1894 while (dlimit - daddr) 1895 *((uint8_t *)daddr++) = *((uint8_t *)saddr++); 1896 1897 /* 1898 * Finally, commit the reserved space in the destination buffer. 1899 */ 1900 dest->dtb_offset = offs + src->dtb_offset; 1901 1902 out: 1903 /* 1904 * If we're lucky enough to be the only active CPU on this speculation 1905 * buffer, we can just set the state back to DTRACESPEC_INACTIVE. 1906 */ 1907 if (current == DTRACESPEC_ACTIVE || 1908 (current == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) { 1909 uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state, 1910 DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE); 1911 1912 ASSERT(rval == DTRACESPEC_COMMITTING); 1913 } 1914 1915 src->dtb_offset = 0; 1916 src->dtb_xamot_drops += src->dtb_drops; 1917 src->dtb_drops = 0; 1918 } 1919 1920 /* 1921 * This routine discards an active speculation. If the specified speculation 1922 * is not in a valid state to perform a discard(), this routine will silently 1923 * do nothing. The state of the specified speculation is transitioned 1924 * according to the state transition diagram outlined in <sys/dtrace_impl.h> 1925 */ 1926 static void 1927 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu, 1928 dtrace_specid_t which) 1929 { 1930 dtrace_speculation_t *spec; 1931 dtrace_speculation_state_t current, new; 1932 dtrace_buffer_t *buf; 1933 1934 if (which == 0) 1935 return; 1936 1937 if (which > state->dts_nspeculations) { 1938 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP; 1939 return; 1940 } 1941 1942 spec = &state->dts_speculations[which - 1]; 1943 buf = &spec->dtsp_buffer[cpu]; 1944 1945 do { 1946 current = spec->dtsp_state; 1947 1948 switch (current) { 1949 case DTRACESPEC_INACTIVE: 1950 case DTRACESPEC_COMMITTINGMANY: 1951 case DTRACESPEC_COMMITTING: 1952 case DTRACESPEC_DISCARDING: 1953 return; 1954 1955 case DTRACESPEC_ACTIVE: 1956 case DTRACESPEC_ACTIVEMANY: 1957 new = DTRACESPEC_DISCARDING; 1958 break; 1959 1960 case DTRACESPEC_ACTIVEONE: 1961 if (buf->dtb_offset != 0) { 1962 new = DTRACESPEC_INACTIVE; 1963 } else { 1964 new = DTRACESPEC_DISCARDING; 1965 } 1966 break; 1967 1968 default: 1969 ASSERT(0); 1970 } 1971 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state, 1972 current, new) != current); 1973 1974 buf->dtb_offset = 0; 1975 buf->dtb_drops = 0; 1976 } 1977 1978 /* 1979 * Note: not called from probe context. This function is called 1980 * asynchronously from cross call context to clean any speculations that are 1981 * in the COMMITTINGMANY or DISCARDING states. These speculations may not be 1982 * transitioned back to the INACTIVE state until all CPUs have cleaned the 1983 * speculation. 1984 */ 1985 static void 1986 dtrace_speculation_clean_here(dtrace_state_t *state) 1987 { 1988 dtrace_icookie_t cookie; 1989 processorid_t cpu = CPU->cpu_id; 1990 dtrace_buffer_t *dest = &state->dts_buffer[cpu]; 1991 dtrace_specid_t i; 1992 1993 cookie = dtrace_interrupt_disable(); 1994 1995 if (dest->dtb_tomax == NULL) { 1996 dtrace_interrupt_enable(cookie); 1997 return; 1998 } 1999 2000 for (i = 0; i < state->dts_nspeculations; i++) { 2001 dtrace_speculation_t *spec = &state->dts_speculations[i]; 2002 dtrace_buffer_t *src = &spec->dtsp_buffer[cpu]; 2003 2004 if (src->dtb_tomax == NULL) 2005 continue; 2006 2007 if (spec->dtsp_state == DTRACESPEC_DISCARDING) { 2008 src->dtb_offset = 0; 2009 continue; 2010 } 2011 2012 if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY) 2013 continue; 2014 2015 if (src->dtb_offset == 0) 2016 continue; 2017 2018 dtrace_speculation_commit(state, cpu, i + 1); 2019 } 2020 2021 dtrace_interrupt_enable(cookie); 2022 } 2023 2024 /* 2025 * Note: not called from probe context. This function is called 2026 * asynchronously (and at a regular interval) to clean any speculations that 2027 * are in the COMMITTINGMANY or DISCARDING states. If it discovers that there 2028 * is work to be done, it cross calls all CPUs to perform that work; 2029 * COMMITMANY and DISCARDING speculations may not be transitioned back to the 2030 * INACTIVE state until they have been cleaned by all CPUs. 2031 */ 2032 static void 2033 dtrace_speculation_clean(dtrace_state_t *state) 2034 { 2035 int work = 0, rv; 2036 dtrace_specid_t i; 2037 2038 for (i = 0; i < state->dts_nspeculations; i++) { 2039 dtrace_speculation_t *spec = &state->dts_speculations[i]; 2040 2041 ASSERT(!spec->dtsp_cleaning); 2042 2043 if (spec->dtsp_state != DTRACESPEC_DISCARDING && 2044 spec->dtsp_state != DTRACESPEC_COMMITTINGMANY) 2045 continue; 2046 2047 work++; 2048 spec->dtsp_cleaning = 1; 2049 } 2050 2051 if (!work) 2052 return; 2053 2054 dtrace_xcall(DTRACE_CPUALL, 2055 (dtrace_xcall_t)dtrace_speculation_clean_here, state); 2056 2057 /* 2058 * We now know that all CPUs have committed or discarded their 2059 * speculation buffers, as appropriate. We can now set the state 2060 * to inactive. 2061 */ 2062 for (i = 0; i < state->dts_nspeculations; i++) { 2063 dtrace_speculation_t *spec = &state->dts_speculations[i]; 2064 dtrace_speculation_state_t current, new; 2065 2066 if (!spec->dtsp_cleaning) 2067 continue; 2068 2069 current = spec->dtsp_state; 2070 ASSERT(current == DTRACESPEC_DISCARDING || 2071 current == DTRACESPEC_COMMITTINGMANY); 2072 2073 new = DTRACESPEC_INACTIVE; 2074 2075 rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, current, new); 2076 ASSERT(rv == current); 2077 spec->dtsp_cleaning = 0; 2078 } 2079 } 2080 2081 /* 2082 * Called as part of a speculate() to get the speculative buffer associated 2083 * with a given speculation. Returns NULL if the specified speculation is not 2084 * in an ACTIVE state. If the speculation is in the ACTIVEONE state -- and 2085 * the active CPU is not the specified CPU -- the speculation will be 2086 * atomically transitioned into the ACTIVEMANY state. 2087 */ 2088 static dtrace_buffer_t * 2089 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid, 2090 dtrace_specid_t which) 2091 { 2092 dtrace_speculation_t *spec; 2093 dtrace_speculation_state_t current, new; 2094 dtrace_buffer_t *buf; 2095 2096 if (which == 0) 2097 return (NULL); 2098 2099 if (which > state->dts_nspeculations) { 2100 cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP; 2101 return (NULL); 2102 } 2103 2104 spec = &state->dts_speculations[which - 1]; 2105 buf = &spec->dtsp_buffer[cpuid]; 2106 2107 do { 2108 current = spec->dtsp_state; 2109 2110 switch (current) { 2111 case DTRACESPEC_INACTIVE: 2112 case DTRACESPEC_COMMITTINGMANY: 2113 case DTRACESPEC_DISCARDING: 2114 return (NULL); 2115 2116 case DTRACESPEC_COMMITTING: 2117 ASSERT(buf->dtb_offset == 0); 2118 return (NULL); 2119 2120 case DTRACESPEC_ACTIVEONE: 2121 /* 2122 * This speculation is currently active on one CPU. 2123 * Check the offset in the buffer; if it's non-zero, 2124 * that CPU must be us (and we leave the state alone). 2125 * If it's zero, assume that we're starting on a new 2126 * CPU -- and change the state to indicate that the 2127 * speculation is active on more than one CPU. 2128 */ 2129 if (buf->dtb_offset != 0) 2130 return (buf); 2131 2132 new = DTRACESPEC_ACTIVEMANY; 2133 break; 2134 2135 case DTRACESPEC_ACTIVEMANY: 2136 return (buf); 2137 2138 case DTRACESPEC_ACTIVE: 2139 new = DTRACESPEC_ACTIVEONE; 2140 break; 2141 2142 default: 2143 ASSERT(0); 2144 } 2145 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state, 2146 current, new) != current); 2147 2148 ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY); 2149 return (buf); 2150 } 2151 2152 /* 2153 * This function implements the DIF emulator's variable lookups. The emulator 2154 * passes a reserved variable identifier and optional built-in array index. 2155 */ 2156 static uint64_t 2157 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v, 2158 uint64_t ndx) 2159 { 2160 /* 2161 * If we're accessing one of the uncached arguments, we'll turn this 2162 * into a reference in the args array. 2163 */ 2164 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) { 2165 ndx = v - DIF_VAR_ARG0; 2166 v = DIF_VAR_ARGS; 2167 } 2168 2169 switch (v) { 2170 case DIF_VAR_ARGS: 2171 ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS); 2172 if (ndx >= sizeof (mstate->dtms_arg) / 2173 sizeof (mstate->dtms_arg[0])) { 2174 int aframes = mstate->dtms_probe->dtpr_aframes + 2; 2175 dtrace_provider_t *pv; 2176 uint64_t val; 2177 2178 pv = mstate->dtms_probe->dtpr_provider; 2179 if (pv->dtpv_pops.dtps_getargval != NULL) 2180 val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg, 2181 mstate->dtms_probe->dtpr_id, 2182 mstate->dtms_probe->dtpr_arg, ndx, aframes); 2183 else 2184 val = dtrace_getarg(ndx, aframes); 2185 2186 /* 2187 * This is regrettably required to keep the compiler 2188 * from tail-optimizing the call to dtrace_getarg(). 2189 * The condition always evaluates to true, but the 2190 * compiler has no way of figuring that out a priori. 2191 * (None of this would be necessary if the compiler 2192 * could be relied upon to _always_ tail-optimize 2193 * the call to dtrace_getarg() -- but it can't.) 2194 */ 2195 if (mstate->dtms_probe != NULL) 2196 return (val); 2197 2198 ASSERT(0); 2199 } 2200 2201 return (mstate->dtms_arg[ndx]); 2202 2203 case DIF_VAR_UREGS: { 2204 klwp_t *lwp; 2205 2206 if (!dtrace_priv_proc(state)) 2207 return (0); 2208 2209 if ((lwp = curthread->t_lwp) == NULL) { 2210 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR); 2211 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = NULL; 2212 return (0); 2213 } 2214 2215 return (dtrace_getreg(lwp->lwp_regs, ndx)); 2216 } 2217 2218 case DIF_VAR_CURTHREAD: 2219 if (!dtrace_priv_kernel(state)) 2220 return (0); 2221 return ((uint64_t)(uintptr_t)curthread); 2222 2223 case DIF_VAR_TIMESTAMP: 2224 if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) { 2225 mstate->dtms_timestamp = dtrace_gethrtime(); 2226 mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP; 2227 } 2228 return (mstate->dtms_timestamp); 2229 2230 case DIF_VAR_VTIMESTAMP: 2231 ASSERT(dtrace_vtime_references != 0); 2232 return (curthread->t_dtrace_vtime); 2233 2234 case DIF_VAR_WALLTIMESTAMP: 2235 if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) { 2236 mstate->dtms_walltimestamp = dtrace_gethrestime(); 2237 mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP; 2238 } 2239 return (mstate->dtms_walltimestamp); 2240 2241 case DIF_VAR_IPL: 2242 if (!dtrace_priv_kernel(state)) 2243 return (0); 2244 if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) { 2245 mstate->dtms_ipl = dtrace_getipl(); 2246 mstate->dtms_present |= DTRACE_MSTATE_IPL; 2247 } 2248 return (mstate->dtms_ipl); 2249 2250 case DIF_VAR_EPID: 2251 ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID); 2252 return (mstate->dtms_epid); 2253 2254 case DIF_VAR_ID: 2255 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE); 2256 return (mstate->dtms_probe->dtpr_id); 2257 2258 case DIF_VAR_STACKDEPTH: 2259 if (!dtrace_priv_kernel(state)) 2260 return (0); 2261 if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) { 2262 int aframes = mstate->dtms_probe->dtpr_aframes + 2; 2263 2264 mstate->dtms_stackdepth = dtrace_getstackdepth(aframes); 2265 mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH; 2266 } 2267 return (mstate->dtms_stackdepth); 2268 2269 case DIF_VAR_USTACKDEPTH: 2270 if (!dtrace_priv_proc(state)) 2271 return (0); 2272 if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) { 2273 /* 2274 * See comment in DIF_VAR_PID. 2275 */ 2276 if (DTRACE_ANCHORED(mstate->dtms_probe) && 2277 CPU_ON_INTR(CPU)) { 2278 mstate->dtms_ustackdepth = 0; 2279 } else { 2280 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT); 2281 mstate->dtms_ustackdepth = 2282 dtrace_getustackdepth(); 2283 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT); 2284 } 2285 mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH; 2286 } 2287 return (mstate->dtms_ustackdepth); 2288 2289 case DIF_VAR_CALLER: 2290 if (!dtrace_priv_kernel(state)) 2291 return (0); 2292 if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) { 2293 int aframes = mstate->dtms_probe->dtpr_aframes + 2; 2294 2295 if (!DTRACE_ANCHORED(mstate->dtms_probe)) { 2296 /* 2297 * If this is an unanchored probe, we are 2298 * required to go through the slow path: 2299 * dtrace_caller() only guarantees correct 2300 * results for anchored probes. 2301 */ 2302 pc_t caller[2]; 2303 2304 dtrace_getpcstack(caller, 2, aframes, 2305 (uint32_t *)(uintptr_t)mstate->dtms_arg[0]); 2306 mstate->dtms_caller = caller[1]; 2307 } else if ((mstate->dtms_caller = 2308 dtrace_caller(aframes)) == -1) { 2309 /* 2310 * We have failed to do this the quick way; 2311 * we must resort to the slower approach of 2312 * calling dtrace_getpcstack(). 2313 */ 2314 pc_t caller; 2315 2316 dtrace_getpcstack(&caller, 1, aframes, NULL); 2317 mstate->dtms_caller = caller; 2318 } 2319 2320 mstate->dtms_present |= DTRACE_MSTATE_CALLER; 2321 } 2322 return (mstate->dtms_caller); 2323 2324 case DIF_VAR_UCALLER: 2325 if (!dtrace_priv_proc(state)) 2326 return (0); 2327 2328 if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) { 2329 uint64_t ustack[3]; 2330 2331 /* 2332 * dtrace_getupcstack() fills in the first uint64_t 2333 * with the current PID. The second uint64_t will 2334 * be the program counter at user-level. The third 2335 * uint64_t will contain the caller, which is what 2336 * we're after. 2337 */ 2338 ustack[2] = NULL; 2339 dtrace_getupcstack(ustack, 3); 2340 mstate->dtms_ucaller = ustack[2]; 2341 mstate->dtms_present |= DTRACE_MSTATE_UCALLER; 2342 } 2343 2344 return (mstate->dtms_ucaller); 2345 2346 case DIF_VAR_PROBEPROV: 2347 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE); 2348 return ((uint64_t)(uintptr_t) 2349 mstate->dtms_probe->dtpr_provider->dtpv_name); 2350 2351 case DIF_VAR_PROBEMOD: 2352 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE); 2353 return ((uint64_t)(uintptr_t) 2354 mstate->dtms_probe->dtpr_mod); 2355 2356 case DIF_VAR_PROBEFUNC: 2357 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE); 2358 return ((uint64_t)(uintptr_t) 2359 mstate->dtms_probe->dtpr_func); 2360 2361 case DIF_VAR_PROBENAME: 2362 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE); 2363 return ((uint64_t)(uintptr_t) 2364 mstate->dtms_probe->dtpr_name); 2365 2366 case DIF_VAR_PID: 2367 if (!dtrace_priv_proc(state)) 2368 return (0); 2369 2370 /* 2371 * Note that we are assuming that an unanchored probe is 2372 * always due to a high-level interrupt. (And we're assuming 2373 * that there is only a single high level interrupt.) 2374 */ 2375 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU)) 2376 return (pid0.pid_id); 2377 2378 /* 2379 * It is always safe to dereference one's own t_procp pointer: 2380 * it always points to a valid, allocated proc structure. 2381 * Further, it is always safe to dereference the p_pidp member 2382 * of one's own proc structure. (These are truisms becuase 2383 * threads and processes don't clean up their own state -- 2384 * they leave that task to whomever reaps them.) 2385 */ 2386 return ((uint64_t)curthread->t_procp->p_pidp->pid_id); 2387 2388 case DIF_VAR_TID: 2389 /* 2390 * See comment in DIF_VAR_PID. 2391 */ 2392 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU)) 2393 return (0); 2394 2395 return ((uint64_t)curthread->t_tid); 2396 2397 case DIF_VAR_EXECNAME: 2398 if (!dtrace_priv_proc(state)) 2399 return (0); 2400 2401 /* 2402 * See comment in DIF_VAR_PID. 2403 */ 2404 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU)) 2405 return ((uint64_t)(uintptr_t)p0.p_user.u_comm); 2406 2407 /* 2408 * It is always safe to dereference one's own t_procp pointer: 2409 * it always points to a valid, allocated proc structure. 2410 * (This is true because threads don't clean up their own 2411 * state -- they leave that task to whomever reaps them.) 2412 */ 2413 return ((uint64_t)(uintptr_t) 2414 curthread->t_procp->p_user.u_comm); 2415 2416 case DIF_VAR_ZONENAME: 2417 if (!dtrace_priv_proc(state)) 2418 return (0); 2419 2420 /* 2421 * See comment in DIF_VAR_PID. 2422 */ 2423 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU)) 2424 return ((uint64_t)(uintptr_t)p0.p_zone->zone_name); 2425 2426 /* 2427 * It is always safe to dereference one's own t_procp pointer: 2428 * it always points to a valid, allocated proc structure. 2429 * (This is true because threads don't clean up their own 2430 * state -- they leave that task to whomever reaps them.) 2431 */ 2432 return ((uint64_t)(uintptr_t) 2433 curthread->t_procp->p_zone->zone_name); 2434 2435 default: 2436 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP); 2437 return (0); 2438 } 2439 } 2440 2441 /* 2442 * Emulate the execution of DTrace ID subroutines invoked by the call opcode. 2443 * Notice that we don't bother validating the proper number of arguments or 2444 * their types in the tuple stack. This isn't needed because all argument 2445 * interpretation is safe because of our load safety -- the worst that can 2446 * happen is that a bogus program can obtain bogus results. 2447 */ 2448 static void 2449 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs, 2450 dtrace_key_t *tupregs, int nargs, 2451 dtrace_mstate_t *mstate, dtrace_state_t *state) 2452 { 2453 volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags; 2454 volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval; 2455 2456 union { 2457 mutex_impl_t mi; 2458 uint64_t mx; 2459 } m; 2460 2461 union { 2462 krwlock_t ri; 2463 uintptr_t rw; 2464 } r; 2465 2466 switch (subr) { 2467 case DIF_SUBR_RAND: 2468 regs[rd] = (dtrace_gethrtime() * 2416 + 374441) % 1771875; 2469 break; 2470 2471 case DIF_SUBR_MUTEX_OWNED: 2472 m.mx = dtrace_load64(tupregs[0].dttk_value); 2473 if (MUTEX_TYPE_ADAPTIVE(&m.mi)) 2474 regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER; 2475 else 2476 regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock); 2477 break; 2478 2479 case DIF_SUBR_MUTEX_OWNER: 2480 m.mx = dtrace_load64(tupregs[0].dttk_value); 2481 if (MUTEX_TYPE_ADAPTIVE(&m.mi) && 2482 MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER) 2483 regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi); 2484 else 2485 regs[rd] = 0; 2486 break; 2487 2488 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE: 2489 m.mx = dtrace_load64(tupregs[0].dttk_value); 2490 regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi); 2491 break; 2492 2493 case DIF_SUBR_MUTEX_TYPE_SPIN: 2494 m.mx = dtrace_load64(tupregs[0].dttk_value); 2495 regs[rd] = MUTEX_TYPE_SPIN(&m.mi); 2496 break; 2497 2498 case DIF_SUBR_RW_READ_HELD: { 2499 uintptr_t tmp; 2500 2501 r.rw = dtrace_loadptr(tupregs[0].dttk_value); 2502 regs[rd] = _RW_READ_HELD(&r.ri, tmp); 2503 break; 2504 } 2505 2506 case DIF_SUBR_RW_WRITE_HELD: 2507 r.rw = dtrace_loadptr(tupregs[0].dttk_value); 2508 regs[rd] = _RW_WRITE_HELD(&r.ri); 2509 break; 2510 2511 case DIF_SUBR_RW_ISWRITER: 2512 r.rw = dtrace_loadptr(tupregs[0].dttk_value); 2513 regs[rd] = _RW_ISWRITER(&r.ri); 2514 break; 2515 2516 case DIF_SUBR_BCOPY: { 2517 /* 2518 * We need to be sure that the destination is in the scratch 2519 * region -- no other region is allowed. 2520 */ 2521 uintptr_t src = tupregs[0].dttk_value; 2522 uintptr_t dest = tupregs[1].dttk_value; 2523 size_t size = tupregs[2].dttk_value; 2524 2525 if (!dtrace_inscratch(dest, size, mstate)) { 2526 *flags |= CPU_DTRACE_BADADDR; 2527 *illval = regs[rd]; 2528 break; 2529 } 2530 2531 dtrace_bcopy((void *)src, (void *)dest, size); 2532 break; 2533 } 2534 2535 case DIF_SUBR_ALLOCA: 2536 case DIF_SUBR_COPYIN: { 2537 uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8); 2538 uint64_t size = 2539 tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value; 2540 size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size; 2541 2542 /* 2543 * This action doesn't require any credential checks since 2544 * probes will not activate in user contexts to which the 2545 * enabling user does not have permissions. 2546 */ 2547 if (mstate->dtms_scratch_ptr + scratch_size > 2548 mstate->dtms_scratch_base + mstate->dtms_scratch_size) { 2549 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 2550 regs[rd] = NULL; 2551 break; 2552 } 2553 2554 if (subr == DIF_SUBR_COPYIN) { 2555 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT); 2556 dtrace_copyin(tupregs[0].dttk_value, dest, size); 2557 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT); 2558 } 2559 2560 mstate->dtms_scratch_ptr += scratch_size; 2561 regs[rd] = dest; 2562 break; 2563 } 2564 2565 case DIF_SUBR_COPYINTO: { 2566 uint64_t size = tupregs[1].dttk_value; 2567 uintptr_t dest = tupregs[2].dttk_value; 2568 2569 /* 2570 * This action doesn't require any credential checks since 2571 * probes will not activate in user contexts to which the 2572 * enabling user does not have permissions. 2573 */ 2574 if (!dtrace_inscratch(dest, size, mstate)) { 2575 *flags |= CPU_DTRACE_BADADDR; 2576 *illval = regs[rd]; 2577 break; 2578 } 2579 2580 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT); 2581 dtrace_copyin(tupregs[0].dttk_value, dest, size); 2582 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT); 2583 break; 2584 } 2585 2586 case DIF_SUBR_COPYINSTR: { 2587 uintptr_t dest = mstate->dtms_scratch_ptr; 2588 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE]; 2589 2590 if (nargs > 1 && tupregs[1].dttk_value < size) 2591 size = tupregs[1].dttk_value + 1; 2592 2593 /* 2594 * This action doesn't require any credential checks since 2595 * probes will not activate in user contexts to which the 2596 * enabling user does not have permissions. 2597 */ 2598 if (mstate->dtms_scratch_ptr + size > 2599 mstate->dtms_scratch_base + mstate->dtms_scratch_size) { 2600 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 2601 regs[rd] = NULL; 2602 break; 2603 } 2604 2605 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT); 2606 dtrace_copyinstr(tupregs[0].dttk_value, dest, size); 2607 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT); 2608 2609 ((char *)dest)[size - 1] = '\0'; 2610 mstate->dtms_scratch_ptr += size; 2611 regs[rd] = dest; 2612 break; 2613 } 2614 2615 case DIF_SUBR_MSGSIZE: 2616 case DIF_SUBR_MSGDSIZE: { 2617 uintptr_t baddr = tupregs[0].dttk_value, daddr; 2618 uintptr_t wptr, rptr; 2619 size_t count = 0; 2620 int cont = 0; 2621 2622 while (baddr != NULL && !(*flags & CPU_DTRACE_FAULT)) { 2623 wptr = dtrace_loadptr(baddr + 2624 offsetof(mblk_t, b_wptr)); 2625 2626 rptr = dtrace_loadptr(baddr + 2627 offsetof(mblk_t, b_rptr)); 2628 2629 if (wptr < rptr) { 2630 *flags |= CPU_DTRACE_BADADDR; 2631 *illval = tupregs[0].dttk_value; 2632 break; 2633 } 2634 2635 daddr = dtrace_loadptr(baddr + 2636 offsetof(mblk_t, b_datap)); 2637 2638 baddr = dtrace_loadptr(baddr + 2639 offsetof(mblk_t, b_cont)); 2640 2641 /* 2642 * We want to prevent against denial-of-service here, 2643 * so we're only going to search the list for 2644 * dtrace_msgdsize_max mblks. 2645 */ 2646 if (cont++ > dtrace_msgdsize_max) { 2647 *flags |= CPU_DTRACE_ILLOP; 2648 break; 2649 } 2650 2651 if (subr == DIF_SUBR_MSGDSIZE) { 2652 if (dtrace_load8(daddr + 2653 offsetof(dblk_t, db_type)) != M_DATA) 2654 continue; 2655 } 2656 2657 count += wptr - rptr; 2658 } 2659 2660 if (!(*flags & CPU_DTRACE_FAULT)) 2661 regs[rd] = count; 2662 2663 break; 2664 } 2665 2666 case DIF_SUBR_PROGENYOF: { 2667 pid_t pid = tupregs[0].dttk_value; 2668 proc_t *p; 2669 int rval = 0; 2670 2671 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT); 2672 2673 for (p = curthread->t_procp; p != NULL; p = p->p_parent) { 2674 if (p->p_pidp->pid_id == pid) { 2675 rval = 1; 2676 break; 2677 } 2678 } 2679 2680 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT); 2681 2682 regs[rd] = rval; 2683 break; 2684 } 2685 2686 case DIF_SUBR_SPECULATION: 2687 regs[rd] = dtrace_speculation(state); 2688 break; 2689 2690 case DIF_SUBR_COPYOUT: { 2691 uintptr_t kaddr = tupregs[0].dttk_value; 2692 uintptr_t uaddr = tupregs[1].dttk_value; 2693 uint64_t size = tupregs[2].dttk_value; 2694 2695 if (!dtrace_destructive_disallow && 2696 dtrace_priv_proc_control(state) && 2697 !dtrace_istoxic(kaddr, size)) { 2698 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT); 2699 dtrace_copyout(kaddr, uaddr, size); 2700 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT); 2701 } 2702 break; 2703 } 2704 2705 case DIF_SUBR_COPYOUTSTR: { 2706 uintptr_t kaddr = tupregs[0].dttk_value; 2707 uintptr_t uaddr = tupregs[1].dttk_value; 2708 uint64_t size = tupregs[2].dttk_value; 2709 2710 if (!dtrace_destructive_disallow && 2711 dtrace_priv_proc_control(state) && 2712 !dtrace_istoxic(kaddr, size)) { 2713 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT); 2714 dtrace_copyoutstr(kaddr, uaddr, size); 2715 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT); 2716 } 2717 break; 2718 } 2719 2720 case DIF_SUBR_STRLEN: 2721 regs[rd] = dtrace_strlen((char *)(uintptr_t) 2722 tupregs[0].dttk_value, 2723 state->dts_options[DTRACEOPT_STRSIZE]); 2724 break; 2725 2726 case DIF_SUBR_STRCHR: 2727 case DIF_SUBR_STRRCHR: { 2728 /* 2729 * We're going to iterate over the string looking for the 2730 * specified character. We will iterate until we have reached 2731 * the string length or we have found the character. If this 2732 * is DIF_SUBR_STRRCHR, we will look for the last occurrence 2733 * of the specified character instead of the first. 2734 */ 2735 uintptr_t addr = tupregs[0].dttk_value; 2736 uintptr_t limit = addr + state->dts_options[DTRACEOPT_STRSIZE]; 2737 char c, target = (char)tupregs[1].dttk_value; 2738 2739 for (regs[rd] = NULL; addr < limit; addr++) { 2740 if ((c = dtrace_load8(addr)) == target) { 2741 regs[rd] = addr; 2742 2743 if (subr == DIF_SUBR_STRCHR) 2744 break; 2745 } 2746 2747 if (c == '\0') 2748 break; 2749 } 2750 2751 break; 2752 } 2753 2754 case DIF_SUBR_STRSTR: 2755 case DIF_SUBR_INDEX: 2756 case DIF_SUBR_RINDEX: { 2757 /* 2758 * We're going to iterate over the string looking for the 2759 * specified string. We will iterate until we have reached 2760 * the string length or we have found the string. (Yes, this 2761 * is done in the most naive way possible -- but considering 2762 * that the string we're searching for is likely to be 2763 * relatively short, the complexity of Rabin-Karp or similar 2764 * hardly seems merited.) 2765 */ 2766 char *addr = (char *)(uintptr_t)tupregs[0].dttk_value; 2767 char *substr = (char *)(uintptr_t)tupregs[1].dttk_value; 2768 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE]; 2769 size_t len = dtrace_strlen(addr, size); 2770 size_t sublen = dtrace_strlen(substr, size); 2771 char *limit = addr + len, *orig = addr; 2772 int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1; 2773 int inc = 1; 2774 2775 regs[rd] = notfound; 2776 2777 /* 2778 * strstr() and index()/rindex() have similar semantics if 2779 * both strings are the empty string: strstr() returns a 2780 * pointer to the (empty) string, and index() and rindex() 2781 * both return index 0 (regardless of any position argument). 2782 */ 2783 if (sublen == 0 && len == 0) { 2784 if (subr == DIF_SUBR_STRSTR) 2785 regs[rd] = (uintptr_t)addr; 2786 else 2787 regs[rd] = 0; 2788 break; 2789 } 2790 2791 if (subr != DIF_SUBR_STRSTR) { 2792 if (subr == DIF_SUBR_RINDEX) { 2793 limit = orig - 1; 2794 addr += len; 2795 inc = -1; 2796 } 2797 2798 /* 2799 * Both index() and rindex() take an optional position 2800 * argument that denotes the starting position. 2801 */ 2802 if (nargs == 3) { 2803 int64_t pos = (int64_t)tupregs[2].dttk_value; 2804 2805 /* 2806 * If the position argument to index() is 2807 * negative, Perl implicitly clamps it at 2808 * zero. This semantic is a little surprising 2809 * given the special meaning of negative 2810 * positions to similar Perl functions like 2811 * substr(), but it appears to reflect a 2812 * notion that index() can start from a 2813 * negative index and increment its way up to 2814 * the string. Given this notion, Perl's 2815 * rindex() is at least self-consistent in 2816 * that it implicitly clamps positions greater 2817 * than the string length to be the string 2818 * length. Where Perl completely loses 2819 * coherence, however, is when the specified 2820 * substring is the empty string (""). In 2821 * this case, even if the position is 2822 * negative, rindex() returns 0 -- and even if 2823 * the position is greater than the length, 2824 * index() returns the string length. These 2825 * semantics violate the notion that index() 2826 * should never return a value less than the 2827 * specified position and that rindex() should 2828 * never return a value greater than the 2829 * specified position. (One assumes that 2830 * these semantics are artifacts of Perl's 2831 * implementation and not the results of 2832 * deliberate design -- it beggars belief that 2833 * even Larry Wall could desire such oddness.) 2834 * While in the abstract one would wish for 2835 * consistent position semantics across 2836 * substr(), index() and rindex() -- or at the 2837 * very least self-consistent position 2838 * semantics for index() and rindex() -- we 2839 * instead opt to keep with the extant Perl 2840 * semantics, in all their broken glory. (Do 2841 * we have more desire to maintain Perl's 2842 * semantics than Perl does? Probably.) 2843 */ 2844 if (subr == DIF_SUBR_RINDEX) { 2845 if (pos < 0) { 2846 if (sublen == 0) 2847 regs[rd] = 0; 2848 break; 2849 } 2850 2851 if (pos > len) 2852 pos = len; 2853 } else { 2854 if (pos < 0) 2855 pos = 0; 2856 2857 if (pos >= len) { 2858 if (sublen == 0) 2859 regs[rd] = len; 2860 break; 2861 } 2862 } 2863 2864 addr = orig + pos; 2865 } 2866 } 2867 2868 for (regs[rd] = notfound; addr != limit; addr += inc) { 2869 if (dtrace_strncmp(addr, substr, sublen) == 0) { 2870 if (subr != DIF_SUBR_STRSTR) { 2871 /* 2872 * As D index() and rindex() are 2873 * modeled on Perl (and not on awk), 2874 * we return a zero-based (and not a 2875 * one-based) index. (For you Perl 2876 * weenies: no, we're not going to add 2877 * $[ -- and shouldn't you be at a con 2878 * or something?) 2879 */ 2880 regs[rd] = (uintptr_t)(addr - orig); 2881 break; 2882 } 2883 2884 ASSERT(subr == DIF_SUBR_STRSTR); 2885 regs[rd] = (uintptr_t)addr; 2886 break; 2887 } 2888 } 2889 2890 break; 2891 } 2892 2893 case DIF_SUBR_STRTOK: { 2894 uintptr_t addr = tupregs[0].dttk_value; 2895 uintptr_t tokaddr = tupregs[1].dttk_value; 2896 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE]; 2897 uintptr_t limit, toklimit = tokaddr + size; 2898 uint8_t c, tokmap[32]; /* 256 / 8 */ 2899 char *dest = (char *)mstate->dtms_scratch_ptr; 2900 int i; 2901 2902 if (mstate->dtms_scratch_ptr + size > 2903 mstate->dtms_scratch_base + mstate->dtms_scratch_size) { 2904 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 2905 regs[rd] = NULL; 2906 break; 2907 } 2908 2909 if (addr == NULL) { 2910 /* 2911 * If the address specified is NULL, we use our saved 2912 * strtok pointer from the mstate. Note that this 2913 * means that the saved strtok pointer is _only_ 2914 * valid within multiple enablings of the same probe -- 2915 * it behaves like an implicit clause-local variable. 2916 */ 2917 addr = mstate->dtms_strtok; 2918 } 2919 2920 /* 2921 * First, zero the token map, and then process the token 2922 * string -- setting a bit in the map for every character 2923 * found in the token string. 2924 */ 2925 for (i = 0; i < sizeof (tokmap); i++) 2926 tokmap[i] = 0; 2927 2928 for (; tokaddr < toklimit; tokaddr++) { 2929 if ((c = dtrace_load8(tokaddr)) == '\0') 2930 break; 2931 2932 ASSERT((c >> 3) < sizeof (tokmap)); 2933 tokmap[c >> 3] |= (1 << (c & 0x7)); 2934 } 2935 2936 for (limit = addr + size; addr < limit; addr++) { 2937 /* 2938 * We're looking for a character that is _not_ contained 2939 * in the token string. 2940 */ 2941 if ((c = dtrace_load8(addr)) == '\0') 2942 break; 2943 2944 if (!(tokmap[c >> 3] & (1 << (c & 0x7)))) 2945 break; 2946 } 2947 2948 if (c == '\0') { 2949 /* 2950 * We reached the end of the string without finding 2951 * any character that was not in the token string. 2952 * We return NULL in this case, and we set the saved 2953 * address to NULL as well. 2954 */ 2955 regs[rd] = NULL; 2956 mstate->dtms_strtok = NULL; 2957 break; 2958 } 2959 2960 /* 2961 * From here on, we're copying into the destination string. 2962 */ 2963 for (i = 0; addr < limit && i < size - 1; addr++) { 2964 if ((c = dtrace_load8(addr)) == '\0') 2965 break; 2966 2967 if (tokmap[c >> 3] & (1 << (c & 0x7))) 2968 break; 2969 2970 ASSERT(i < size); 2971 dest[i++] = c; 2972 } 2973 2974 ASSERT(i < size); 2975 dest[i] = '\0'; 2976 regs[rd] = (uintptr_t)dest; 2977 mstate->dtms_scratch_ptr += size; 2978 mstate->dtms_strtok = addr; 2979 break; 2980 } 2981 2982 case DIF_SUBR_SUBSTR: { 2983 uintptr_t s = tupregs[0].dttk_value; 2984 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE]; 2985 char *d = (char *)mstate->dtms_scratch_ptr; 2986 int64_t index = (int64_t)tupregs[1].dttk_value; 2987 int64_t remaining = (int64_t)tupregs[2].dttk_value; 2988 size_t len = dtrace_strlen((char *)s, size); 2989 int64_t i = 0; 2990 2991 if (nargs <= 2) 2992 remaining = (int64_t)size; 2993 2994 if (mstate->dtms_scratch_ptr + size > 2995 mstate->dtms_scratch_base + mstate->dtms_scratch_size) { 2996 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 2997 regs[rd] = NULL; 2998 break; 2999 } 3000 3001 if (index < 0) { 3002 index += len; 3003 3004 if (index < 0 && index + remaining > 0) { 3005 remaining += index; 3006 index = 0; 3007 } 3008 } 3009 3010 if (index >= len || index < 0) 3011 index = len; 3012 3013 for (d[0] = '\0'; remaining > 0; remaining--) { 3014 if ((d[i++] = dtrace_load8(s++ + index)) == '\0') 3015 break; 3016 3017 if (i == size) { 3018 d[i - 1] = '\0'; 3019 break; 3020 } 3021 } 3022 3023 mstate->dtms_scratch_ptr += size; 3024 regs[rd] = (uintptr_t)d; 3025 break; 3026 } 3027 3028 case DIF_SUBR_GETMAJOR: 3029 #ifdef _LP64 3030 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64; 3031 #else 3032 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ; 3033 #endif 3034 break; 3035 3036 case DIF_SUBR_GETMINOR: 3037 #ifdef _LP64 3038 regs[rd] = tupregs[0].dttk_value & MAXMIN64; 3039 #else 3040 regs[rd] = tupregs[0].dttk_value & MAXMIN; 3041 #endif 3042 break; 3043 3044 case DIF_SUBR_DDI_PATHNAME: { 3045 /* 3046 * This one is a galactic mess. We are going to roughly 3047 * emulate ddi_pathname(), but it's made more complicated 3048 * by the fact that we (a) want to include the minor name and 3049 * (b) must proceed iteratively instead of recursively. 3050 */ 3051 uintptr_t dest = mstate->dtms_scratch_ptr; 3052 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE]; 3053 char *start = (char *)dest, *end = start + size - 1; 3054 uintptr_t daddr = tupregs[0].dttk_value; 3055 int64_t minor = (int64_t)tupregs[1].dttk_value; 3056 char *s; 3057 int i, len, depth = 0; 3058 3059 if (size == 0 || mstate->dtms_scratch_ptr + size > 3060 mstate->dtms_scratch_base + mstate->dtms_scratch_size) { 3061 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 3062 regs[rd] = NULL; 3063 break; 3064 } 3065 3066 *end = '\0'; 3067 3068 /* 3069 * We want to have a name for the minor. In order to do this, 3070 * we need to walk the minor list from the devinfo. We want 3071 * to be sure that we don't infinitely walk a circular list, 3072 * so we check for circularity by sending a scout pointer 3073 * ahead two elements for every element that we iterate over; 3074 * if the list is circular, these will ultimately point to the 3075 * same element. You may recognize this little trick as the 3076 * answer to a stupid interview question -- one that always 3077 * seems to be asked by those who had to have it laboriously 3078 * explained to them, and who can't even concisely describe 3079 * the conditions under which one would be forced to resort to 3080 * this technique. Needless to say, those conditions are 3081 * found here -- and probably only here. Is this is the only 3082 * use of this infamous trick in shipping, production code? 3083 * If it isn't, it probably should be... 3084 */ 3085 if (minor != -1) { 3086 uintptr_t maddr = dtrace_loadptr(daddr + 3087 offsetof(struct dev_info, devi_minor)); 3088 3089 uintptr_t next = offsetof(struct ddi_minor_data, next); 3090 uintptr_t name = offsetof(struct ddi_minor_data, 3091 d_minor) + offsetof(struct ddi_minor, name); 3092 uintptr_t dev = offsetof(struct ddi_minor_data, 3093 d_minor) + offsetof(struct ddi_minor, dev); 3094 uintptr_t scout; 3095 3096 if (maddr != NULL) 3097 scout = dtrace_loadptr(maddr + next); 3098 3099 while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) { 3100 uint64_t m; 3101 #ifdef _LP64 3102 m = dtrace_load64(maddr + dev) & MAXMIN64; 3103 #else 3104 m = dtrace_load32(maddr + dev) & MAXMIN; 3105 #endif 3106 if (m != minor) { 3107 maddr = dtrace_loadptr(maddr + next); 3108 3109 if (scout == NULL) 3110 continue; 3111 3112 scout = dtrace_loadptr(scout + next); 3113 3114 if (scout == NULL) 3115 continue; 3116 3117 scout = dtrace_loadptr(scout + next); 3118 3119 if (scout == NULL) 3120 continue; 3121 3122 if (scout == maddr) { 3123 *flags |= CPU_DTRACE_ILLOP; 3124 break; 3125 } 3126 3127 continue; 3128 } 3129 3130 /* 3131 * We have the minor data. Now we need to 3132 * copy the minor's name into the end of the 3133 * pathname. 3134 */ 3135 s = (char *)dtrace_loadptr(maddr + name); 3136 len = dtrace_strlen(s, size); 3137 3138 if (*flags & CPU_DTRACE_FAULT) 3139 break; 3140 3141 if (len != 0) { 3142 if ((end -= (len + 1)) < start) 3143 break; 3144 3145 *end = ':'; 3146 } 3147 3148 for (i = 1; i <= len; i++) 3149 end[i] = dtrace_load8((uintptr_t)s++); 3150 break; 3151 } 3152 } 3153 3154 while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) { 3155 ddi_node_state_t devi_state; 3156 3157 devi_state = dtrace_load32(daddr + 3158 offsetof(struct dev_info, devi_node_state)); 3159 3160 if (*flags & CPU_DTRACE_FAULT) 3161 break; 3162 3163 if (devi_state >= DS_INITIALIZED) { 3164 s = (char *)dtrace_loadptr(daddr + 3165 offsetof(struct dev_info, devi_addr)); 3166 len = dtrace_strlen(s, size); 3167 3168 if (*flags & CPU_DTRACE_FAULT) 3169 break; 3170 3171 if (len != 0) { 3172 if ((end -= (len + 1)) < start) 3173 break; 3174 3175 *end = '@'; 3176 } 3177 3178 for (i = 1; i <= len; i++) 3179 end[i] = dtrace_load8((uintptr_t)s++); 3180 } 3181 3182 /* 3183 * Now for the node name... 3184 */ 3185 s = (char *)dtrace_loadptr(daddr + 3186 offsetof(struct dev_info, devi_node_name)); 3187 3188 daddr = dtrace_loadptr(daddr + 3189 offsetof(struct dev_info, devi_parent)); 3190 3191 /* 3192 * If our parent is NULL (that is, if we're the root 3193 * node), we're going to use the special path 3194 * "devices". 3195 */ 3196 if (daddr == NULL) 3197 s = "devices"; 3198 3199 len = dtrace_strlen(s, size); 3200 if (*flags & CPU_DTRACE_FAULT) 3201 break; 3202 3203 if ((end -= (len + 1)) < start) 3204 break; 3205 3206 for (i = 1; i <= len; i++) 3207 end[i] = dtrace_load8((uintptr_t)s++); 3208 *end = '/'; 3209 3210 if (depth++ > dtrace_devdepth_max) { 3211 *flags |= CPU_DTRACE_ILLOP; 3212 break; 3213 } 3214 } 3215 3216 if (end < start) 3217 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 3218 3219 if (daddr == NULL) { 3220 regs[rd] = (uintptr_t)end; 3221 mstate->dtms_scratch_ptr += size; 3222 } 3223 3224 break; 3225 } 3226 3227 case DIF_SUBR_STRJOIN: { 3228 char *d = (char *)mstate->dtms_scratch_ptr; 3229 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE]; 3230 uintptr_t s1 = tupregs[0].dttk_value; 3231 uintptr_t s2 = tupregs[1].dttk_value; 3232 int i = 0; 3233 3234 if (mstate->dtms_scratch_ptr + size > 3235 mstate->dtms_scratch_base + mstate->dtms_scratch_size) { 3236 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 3237 regs[rd] = NULL; 3238 break; 3239 } 3240 3241 for (;;) { 3242 if (i >= size) { 3243 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 3244 regs[rd] = NULL; 3245 break; 3246 } 3247 3248 if ((d[i++] = dtrace_load8(s1++)) == '\0') { 3249 i--; 3250 break; 3251 } 3252 } 3253 3254 for (;;) { 3255 if (i >= size) { 3256 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 3257 regs[rd] = NULL; 3258 break; 3259 } 3260 3261 if ((d[i++] = dtrace_load8(s2++)) == '\0') 3262 break; 3263 } 3264 3265 if (i < size) { 3266 mstate->dtms_scratch_ptr += i; 3267 regs[rd] = (uintptr_t)d; 3268 } 3269 3270 break; 3271 } 3272 3273 case DIF_SUBR_LLTOSTR: { 3274 int64_t i = (int64_t)tupregs[0].dttk_value; 3275 int64_t val = i < 0 ? i * -1 : i; 3276 uint64_t size = 22; /* enough room for 2^64 in decimal */ 3277 char *end = (char *)mstate->dtms_scratch_ptr + size - 1; 3278 3279 if (mstate->dtms_scratch_ptr + size > 3280 mstate->dtms_scratch_base + mstate->dtms_scratch_size) { 3281 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 3282 regs[rd] = NULL; 3283 break; 3284 } 3285 3286 for (*end-- = '\0'; val; val /= 10) 3287 *end-- = '0' + (val % 10); 3288 3289 if (i == 0) 3290 *end-- = '0'; 3291 3292 if (i < 0) 3293 *end-- = '-'; 3294 3295 regs[rd] = (uintptr_t)end + 1; 3296 mstate->dtms_scratch_ptr += size; 3297 break; 3298 } 3299 3300 case DIF_SUBR_DIRNAME: 3301 case DIF_SUBR_BASENAME: { 3302 char *dest = (char *)mstate->dtms_scratch_ptr; 3303 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE]; 3304 uintptr_t src = tupregs[0].dttk_value; 3305 int i, j, len = dtrace_strlen((char *)src, size); 3306 int lastbase = -1, firstbase = -1, lastdir = -1; 3307 int start, end; 3308 3309 if (mstate->dtms_scratch_ptr + size > 3310 mstate->dtms_scratch_base + mstate->dtms_scratch_size) { 3311 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 3312 regs[rd] = NULL; 3313 break; 3314 } 3315 3316 /* 3317 * The basename and dirname for a zero-length string is 3318 * defined to be "." 3319 */ 3320 if (len == 0) { 3321 len = 1; 3322 src = (uintptr_t)"."; 3323 } 3324 3325 /* 3326 * Start from the back of the string, moving back toward the 3327 * front until we see a character that isn't a slash. That 3328 * character is the last character in the basename. 3329 */ 3330 for (i = len - 1; i >= 0; i--) { 3331 if (dtrace_load8(src + i) != '/') 3332 break; 3333 } 3334 3335 if (i >= 0) 3336 lastbase = i; 3337 3338 /* 3339 * Starting from the last character in the basename, move 3340 * towards the front until we find a slash. The character 3341 * that we processed immediately before that is the first 3342 * character in the basename. 3343 */ 3344 for (; i >= 0; i--) { 3345 if (dtrace_load8(src + i) == '/') 3346 break; 3347 } 3348 3349 if (i >= 0) 3350 firstbase = i + 1; 3351 3352 /* 3353 * Now keep going until we find a non-slash character. That 3354 * character is the last character in the dirname. 3355 */ 3356 for (; i >= 0; i--) { 3357 if (dtrace_load8(src + i) != '/') 3358 break; 3359 } 3360 3361 if (i >= 0) 3362 lastdir = i; 3363 3364 ASSERT(!(lastbase == -1 && firstbase != -1)); 3365 ASSERT(!(firstbase == -1 && lastdir != -1)); 3366 3367 if (lastbase == -1) { 3368 /* 3369 * We didn't find a non-slash character. We know that 3370 * the length is non-zero, so the whole string must be 3371 * slashes. In either the dirname or the basename 3372 * case, we return '/'. 3373 */ 3374 ASSERT(firstbase == -1); 3375 firstbase = lastbase = lastdir = 0; 3376 } 3377 3378 if (firstbase == -1) { 3379 /* 3380 * The entire string consists only of a basename 3381 * component. If we're looking for dirname, we need 3382 * to change our string to be just "."; if we're 3383 * looking for a basename, we'll just set the first 3384 * character of the basename to be 0. 3385 */ 3386 if (subr == DIF_SUBR_DIRNAME) { 3387 ASSERT(lastdir == -1); 3388 src = (uintptr_t)"."; 3389 lastdir = 0; 3390 } else { 3391 firstbase = 0; 3392 } 3393 } 3394 3395 if (subr == DIF_SUBR_DIRNAME) { 3396 if (lastdir == -1) { 3397 /* 3398 * We know that we have a slash in the name -- 3399 * or lastdir would be set to 0, above. And 3400 * because lastdir is -1, we know that this 3401 * slash must be the first character. (That 3402 * is, the full string must be of the form 3403 * "/basename".) In this case, the last 3404 * character of the directory name is 0. 3405 */ 3406 lastdir = 0; 3407 } 3408 3409 start = 0; 3410 end = lastdir; 3411 } else { 3412 ASSERT(subr == DIF_SUBR_BASENAME); 3413 ASSERT(firstbase != -1 && lastbase != -1); 3414 start = firstbase; 3415 end = lastbase; 3416 } 3417 3418 for (i = start, j = 0; i <= end && j < size - 1; i++, j++) 3419 dest[j] = dtrace_load8(src + i); 3420 3421 dest[j] = '\0'; 3422 regs[rd] = (uintptr_t)dest; 3423 mstate->dtms_scratch_ptr += size; 3424 break; 3425 } 3426 3427 case DIF_SUBR_CLEANPATH: { 3428 char *dest = (char *)mstate->dtms_scratch_ptr, c; 3429 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE]; 3430 uintptr_t src = tupregs[0].dttk_value; 3431 int i = 0, j = 0; 3432 3433 if (mstate->dtms_scratch_ptr + size > 3434 mstate->dtms_scratch_base + mstate->dtms_scratch_size) { 3435 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 3436 regs[rd] = NULL; 3437 break; 3438 } 3439 3440 /* 3441 * Move forward, loading each character. 3442 */ 3443 do { 3444 c = dtrace_load8(src + i++); 3445 next: 3446 if (j + 5 >= size) /* 5 = strlen("/..c\0") */ 3447 break; 3448 3449 if (c != '/') { 3450 dest[j++] = c; 3451 continue; 3452 } 3453 3454 c = dtrace_load8(src + i++); 3455 3456 if (c == '/') { 3457 /* 3458 * We have two slashes -- we can just advance 3459 * to the next character. 3460 */ 3461 goto next; 3462 } 3463 3464 if (c != '.') { 3465 /* 3466 * This is not "." and it's not ".." -- we can 3467 * just store the "/" and this character and 3468 * drive on. 3469 */ 3470 dest[j++] = '/'; 3471 dest[j++] = c; 3472 continue; 3473 } 3474 3475 c = dtrace_load8(src + i++); 3476 3477 if (c == '/') { 3478 /* 3479 * This is a "/./" component. We're not going 3480 * to store anything in the destination buffer; 3481 * we're just going to go to the next component. 3482 */ 3483 goto next; 3484 } 3485 3486 if (c != '.') { 3487 /* 3488 * This is not ".." -- we can just store the 3489 * "/." and this character and continue 3490 * processing. 3491 */ 3492 dest[j++] = '/'; 3493 dest[j++] = '.'; 3494 dest[j++] = c; 3495 continue; 3496 } 3497 3498 c = dtrace_load8(src + i++); 3499 3500 if (c != '/' && c != '\0') { 3501 /* 3502 * This is not ".." -- it's "..[mumble]". 3503 * We'll store the "/.." and this character 3504 * and continue processing. 3505 */ 3506 dest[j++] = '/'; 3507 dest[j++] = '.'; 3508 dest[j++] = '.'; 3509 dest[j++] = c; 3510 continue; 3511 } 3512 3513 /* 3514 * This is "/../" or "/..\0". We need to back up 3515 * our destination pointer until we find a "/". 3516 */ 3517 i--; 3518 while (j != 0 && dest[--j] != '/') 3519 continue; 3520 3521 if (c == '\0') 3522 dest[++j] = '/'; 3523 } while (c != '\0'); 3524 3525 dest[j] = '\0'; 3526 regs[rd] = (uintptr_t)dest; 3527 mstate->dtms_scratch_ptr += size; 3528 break; 3529 } 3530 } 3531 } 3532 3533 /* 3534 * Emulate the execution of DTrace IR instructions specified by the given 3535 * DIF object. This function is deliberately void of assertions as all of 3536 * the necessary checks are handled by a call to dtrace_difo_validate(). 3537 */ 3538 static uint64_t 3539 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate, 3540 dtrace_vstate_t *vstate, dtrace_state_t *state) 3541 { 3542 const dif_instr_t *text = difo->dtdo_buf; 3543 const uint_t textlen = difo->dtdo_len; 3544 const char *strtab = difo->dtdo_strtab; 3545 const uint64_t *inttab = difo->dtdo_inttab; 3546 3547 uint64_t rval = 0; 3548 dtrace_statvar_t *svar; 3549 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars; 3550 dtrace_difv_t *v; 3551 volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags; 3552 volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval; 3553 3554 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */ 3555 uint64_t regs[DIF_DIR_NREGS]; 3556 uint64_t *tmp; 3557 3558 uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0; 3559 int64_t cc_r; 3560 uint_t pc = 0, id, opc; 3561 uint8_t ttop = 0; 3562 dif_instr_t instr; 3563 uint_t r1, r2, rd; 3564 3565 regs[DIF_REG_R0] = 0; /* %r0 is fixed at zero */ 3566 3567 while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) { 3568 opc = pc; 3569 3570 instr = text[pc++]; 3571 r1 = DIF_INSTR_R1(instr); 3572 r2 = DIF_INSTR_R2(instr); 3573 rd = DIF_INSTR_RD(instr); 3574 3575 switch (DIF_INSTR_OP(instr)) { 3576 case DIF_OP_OR: 3577 regs[rd] = regs[r1] | regs[r2]; 3578 break; 3579 case DIF_OP_XOR: 3580 regs[rd] = regs[r1] ^ regs[r2]; 3581 break; 3582 case DIF_OP_AND: 3583 regs[rd] = regs[r1] & regs[r2]; 3584 break; 3585 case DIF_OP_SLL: 3586 regs[rd] = regs[r1] << regs[r2]; 3587 break; 3588 case DIF_OP_SRL: 3589 regs[rd] = regs[r1] >> regs[r2]; 3590 break; 3591 case DIF_OP_SUB: 3592 regs[rd] = regs[r1] - regs[r2]; 3593 break; 3594 case DIF_OP_ADD: 3595 regs[rd] = regs[r1] + regs[r2]; 3596 break; 3597 case DIF_OP_MUL: 3598 regs[rd] = regs[r1] * regs[r2]; 3599 break; 3600 case DIF_OP_SDIV: 3601 if (regs[r2] == 0) { 3602 regs[rd] = 0; 3603 *flags |= CPU_DTRACE_DIVZERO; 3604 } else { 3605 regs[rd] = (int64_t)regs[r1] / 3606 (int64_t)regs[r2]; 3607 } 3608 break; 3609 3610 case DIF_OP_UDIV: 3611 if (regs[r2] == 0) { 3612 regs[rd] = 0; 3613 *flags |= CPU_DTRACE_DIVZERO; 3614 } else { 3615 regs[rd] = regs[r1] / regs[r2]; 3616 } 3617 break; 3618 3619 case DIF_OP_SREM: 3620 if (regs[r2] == 0) { 3621 regs[rd] = 0; 3622 *flags |= CPU_DTRACE_DIVZERO; 3623 } else { 3624 regs[rd] = (int64_t)regs[r1] % 3625 (int64_t)regs[r2]; 3626 } 3627 break; 3628 3629 case DIF_OP_UREM: 3630 if (regs[r2] == 0) { 3631 regs[rd] = 0; 3632 *flags |= CPU_DTRACE_DIVZERO; 3633 } else { 3634 regs[rd] = regs[r1] % regs[r2]; 3635 } 3636 break; 3637 3638 case DIF_OP_NOT: 3639 regs[rd] = ~regs[r1]; 3640 break; 3641 case DIF_OP_MOV: 3642 regs[rd] = regs[r1]; 3643 break; 3644 case DIF_OP_CMP: 3645 cc_r = regs[r1] - regs[r2]; 3646 cc_n = cc_r < 0; 3647 cc_z = cc_r == 0; 3648 cc_v = 0; 3649 cc_c = regs[r1] < regs[r2]; 3650 break; 3651 case DIF_OP_TST: 3652 cc_n = cc_v = cc_c = 0; 3653 cc_z = regs[r1] == 0; 3654 break; 3655 case DIF_OP_BA: 3656 pc = DIF_INSTR_LABEL(instr); 3657 break; 3658 case DIF_OP_BE: 3659 if (cc_z) 3660 pc = DIF_INSTR_LABEL(instr); 3661 break; 3662 case DIF_OP_BNE: 3663 if (cc_z == 0) 3664 pc = DIF_INSTR_LABEL(instr); 3665 break; 3666 case DIF_OP_BG: 3667 if ((cc_z | (cc_n ^ cc_v)) == 0) 3668 pc = DIF_INSTR_LABEL(instr); 3669 break; 3670 case DIF_OP_BGU: 3671 if ((cc_c | cc_z) == 0) 3672 pc = DIF_INSTR_LABEL(instr); 3673 break; 3674 case DIF_OP_BGE: 3675 if ((cc_n ^ cc_v) == 0) 3676 pc = DIF_INSTR_LABEL(instr); 3677 break; 3678 case DIF_OP_BGEU: 3679 if (cc_c == 0) 3680 pc = DIF_INSTR_LABEL(instr); 3681 break; 3682 case DIF_OP_BL: 3683 if (cc_n ^ cc_v) 3684 pc = DIF_INSTR_LABEL(instr); 3685 break; 3686 case DIF_OP_BLU: 3687 if (cc_c) 3688 pc = DIF_INSTR_LABEL(instr); 3689 break; 3690 case DIF_OP_BLE: 3691 if (cc_z | (cc_n ^ cc_v)) 3692 pc = DIF_INSTR_LABEL(instr); 3693 break; 3694 case DIF_OP_BLEU: 3695 if (cc_c | cc_z) 3696 pc = DIF_INSTR_LABEL(instr); 3697 break; 3698 case DIF_OP_RLDSB: 3699 if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) { 3700 *flags |= CPU_DTRACE_KPRIV; 3701 *illval = regs[r1]; 3702 break; 3703 } 3704 /*FALLTHROUGH*/ 3705 case DIF_OP_LDSB: 3706 regs[rd] = (int8_t)dtrace_load8(regs[r1]); 3707 break; 3708 case DIF_OP_RLDSH: 3709 if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) { 3710 *flags |= CPU_DTRACE_KPRIV; 3711 *illval = regs[r1]; 3712 break; 3713 } 3714 /*FALLTHROUGH*/ 3715 case DIF_OP_LDSH: 3716 regs[rd] = (int16_t)dtrace_load16(regs[r1]); 3717 break; 3718 case DIF_OP_RLDSW: 3719 if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) { 3720 *flags |= CPU_DTRACE_KPRIV; 3721 *illval = regs[r1]; 3722 break; 3723 } 3724 /*FALLTHROUGH*/ 3725 case DIF_OP_LDSW: 3726 regs[rd] = (int32_t)dtrace_load32(regs[r1]); 3727 break; 3728 case DIF_OP_RLDUB: 3729 if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) { 3730 *flags |= CPU_DTRACE_KPRIV; 3731 *illval = regs[r1]; 3732 break; 3733 } 3734 /*FALLTHROUGH*/ 3735 case DIF_OP_LDUB: 3736 regs[rd] = dtrace_load8(regs[r1]); 3737 break; 3738 case DIF_OP_RLDUH: 3739 if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) { 3740 *flags |= CPU_DTRACE_KPRIV; 3741 *illval = regs[r1]; 3742 break; 3743 } 3744 /*FALLTHROUGH*/ 3745 case DIF_OP_LDUH: 3746 regs[rd] = dtrace_load16(regs[r1]); 3747 break; 3748 case DIF_OP_RLDUW: 3749 if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) { 3750 *flags |= CPU_DTRACE_KPRIV; 3751 *illval = regs[r1]; 3752 break; 3753 } 3754 /*FALLTHROUGH*/ 3755 case DIF_OP_LDUW: 3756 regs[rd] = dtrace_load32(regs[r1]); 3757 break; 3758 case DIF_OP_RLDX: 3759 if (!dtrace_canstore(regs[r1], 8, mstate, vstate)) { 3760 *flags |= CPU_DTRACE_KPRIV; 3761 *illval = regs[r1]; 3762 break; 3763 } 3764 /*FALLTHROUGH*/ 3765 case DIF_OP_LDX: 3766 regs[rd] = dtrace_load64(regs[r1]); 3767 break; 3768 case DIF_OP_ULDSB: 3769 regs[rd] = (int8_t) 3770 dtrace_fuword8((void *)(uintptr_t)regs[r1]); 3771 break; 3772 case DIF_OP_ULDSH: 3773 regs[rd] = (int16_t) 3774 dtrace_fuword16((void *)(uintptr_t)regs[r1]); 3775 break; 3776 case DIF_OP_ULDSW: 3777 regs[rd] = (int32_t) 3778 dtrace_fuword32((void *)(uintptr_t)regs[r1]); 3779 break; 3780 case DIF_OP_ULDUB: 3781 regs[rd] = 3782 dtrace_fuword8((void *)(uintptr_t)regs[r1]); 3783 break; 3784 case DIF_OP_ULDUH: 3785 regs[rd] = 3786 dtrace_fuword16((void *)(uintptr_t)regs[r1]); 3787 break; 3788 case DIF_OP_ULDUW: 3789 regs[rd] = 3790 dtrace_fuword32((void *)(uintptr_t)regs[r1]); 3791 break; 3792 case DIF_OP_ULDX: 3793 regs[rd] = 3794 dtrace_fuword64((void *)(uintptr_t)regs[r1]); 3795 break; 3796 case DIF_OP_RET: 3797 rval = regs[rd]; 3798 break; 3799 case DIF_OP_NOP: 3800 break; 3801 case DIF_OP_SETX: 3802 regs[rd] = inttab[DIF_INSTR_INTEGER(instr)]; 3803 break; 3804 case DIF_OP_SETS: 3805 regs[rd] = (uint64_t)(uintptr_t) 3806 (strtab + DIF_INSTR_STRING(instr)); 3807 break; 3808 case DIF_OP_SCMP: 3809 cc_r = dtrace_strncmp((char *)(uintptr_t)regs[r1], 3810 (char *)(uintptr_t)regs[r2], 3811 state->dts_options[DTRACEOPT_STRSIZE]); 3812 3813 cc_n = cc_r < 0; 3814 cc_z = cc_r == 0; 3815 cc_v = cc_c = 0; 3816 break; 3817 case DIF_OP_LDGA: 3818 regs[rd] = dtrace_dif_variable(mstate, state, 3819 r1, regs[r2]); 3820 break; 3821 case DIF_OP_LDGS: 3822 id = DIF_INSTR_VAR(instr); 3823 3824 if (id >= DIF_VAR_OTHER_UBASE) { 3825 uintptr_t a; 3826 3827 id -= DIF_VAR_OTHER_UBASE; 3828 svar = vstate->dtvs_globals[id]; 3829 ASSERT(svar != NULL); 3830 v = &svar->dtsv_var; 3831 3832 if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) { 3833 regs[rd] = svar->dtsv_data; 3834 break; 3835 } 3836 3837 a = (uintptr_t)svar->dtsv_data; 3838 3839 if (*(uint8_t *)a == UINT8_MAX) { 3840 /* 3841 * If the 0th byte is set to UINT8_MAX 3842 * then this is to be treated as a 3843 * reference to a NULL variable. 3844 */ 3845 regs[rd] = NULL; 3846 } else { 3847 regs[rd] = a + sizeof (uint64_t); 3848 } 3849 3850 break; 3851 } 3852 3853 regs[rd] = dtrace_dif_variable(mstate, state, id, 0); 3854 break; 3855 3856 case DIF_OP_STGS: 3857 id = DIF_INSTR_VAR(instr); 3858 3859 ASSERT(id >= DIF_VAR_OTHER_UBASE); 3860 id -= DIF_VAR_OTHER_UBASE; 3861 3862 svar = vstate->dtvs_globals[id]; 3863 ASSERT(svar != NULL); 3864 v = &svar->dtsv_var; 3865 3866 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) { 3867 uintptr_t a = (uintptr_t)svar->dtsv_data; 3868 3869 ASSERT(a != NULL); 3870 ASSERT(svar->dtsv_size != 0); 3871 3872 if (regs[rd] == NULL) { 3873 *(uint8_t *)a = UINT8_MAX; 3874 break; 3875 } else { 3876 *(uint8_t *)a = 0; 3877 a += sizeof (uint64_t); 3878 } 3879 3880 dtrace_vcopy((void *)(uintptr_t)regs[rd], 3881 (void *)a, &v->dtdv_type); 3882 break; 3883 } 3884 3885 svar->dtsv_data = regs[rd]; 3886 break; 3887 3888 case DIF_OP_LDTA: 3889 /* 3890 * There are no DTrace built-in thread-local arrays at 3891 * present. This opcode is saved for future work. 3892 */ 3893 *flags |= CPU_DTRACE_ILLOP; 3894 regs[rd] = 0; 3895 break; 3896 3897 case DIF_OP_LDLS: 3898 id = DIF_INSTR_VAR(instr); 3899 3900 if (id < DIF_VAR_OTHER_UBASE) { 3901 /* 3902 * For now, this has no meaning. 3903 */ 3904 regs[rd] = 0; 3905 break; 3906 } 3907 3908 id -= DIF_VAR_OTHER_UBASE; 3909 3910 ASSERT(id < vstate->dtvs_nlocals); 3911 ASSERT(vstate->dtvs_locals != NULL); 3912 3913 svar = vstate->dtvs_locals[id]; 3914 ASSERT(svar != NULL); 3915 v = &svar->dtsv_var; 3916 3917 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) { 3918 uintptr_t a = (uintptr_t)svar->dtsv_data; 3919 size_t sz = v->dtdv_type.dtdt_size; 3920 3921 sz += sizeof (uint64_t); 3922 ASSERT(svar->dtsv_size == NCPU * sz); 3923 a += CPU->cpu_id * sz; 3924 3925 if (*(uint8_t *)a == UINT8_MAX) { 3926 /* 3927 * If the 0th byte is set to UINT8_MAX 3928 * then this is to be treated as a 3929 * reference to a NULL variable. 3930 */ 3931 regs[rd] = NULL; 3932 } else { 3933 regs[rd] = a + sizeof (uint64_t); 3934 } 3935 3936 break; 3937 } 3938 3939 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t)); 3940 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data; 3941 regs[rd] = tmp[CPU->cpu_id]; 3942 break; 3943 3944 case DIF_OP_STLS: 3945 id = DIF_INSTR_VAR(instr); 3946 3947 ASSERT(id >= DIF_VAR_OTHER_UBASE); 3948 id -= DIF_VAR_OTHER_UBASE; 3949 ASSERT(id < vstate->dtvs_nlocals); 3950 3951 ASSERT(vstate->dtvs_locals != NULL); 3952 svar = vstate->dtvs_locals[id]; 3953 ASSERT(svar != NULL); 3954 v = &svar->dtsv_var; 3955 3956 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) { 3957 uintptr_t a = (uintptr_t)svar->dtsv_data; 3958 size_t sz = v->dtdv_type.dtdt_size; 3959 3960 sz += sizeof (uint64_t); 3961 ASSERT(svar->dtsv_size == NCPU * sz); 3962 a += CPU->cpu_id * sz; 3963 3964 if (regs[rd] == NULL) { 3965 *(uint8_t *)a = UINT8_MAX; 3966 break; 3967 } else { 3968 *(uint8_t *)a = 0; 3969 a += sizeof (uint64_t); 3970 } 3971 3972 dtrace_vcopy((void *)(uintptr_t)regs[rd], 3973 (void *)a, &v->dtdv_type); 3974 break; 3975 } 3976 3977 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t)); 3978 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data; 3979 tmp[CPU->cpu_id] = regs[rd]; 3980 break; 3981 3982 case DIF_OP_LDTS: { 3983 dtrace_dynvar_t *dvar; 3984 dtrace_key_t *key; 3985 3986 id = DIF_INSTR_VAR(instr); 3987 ASSERT(id >= DIF_VAR_OTHER_UBASE); 3988 id -= DIF_VAR_OTHER_UBASE; 3989 v = &vstate->dtvs_tlocals[id]; 3990 3991 key = &tupregs[DIF_DTR_NREGS]; 3992 key[0].dttk_value = (uint64_t)id; 3993 key[0].dttk_size = 0; 3994 DTRACE_TLS_THRKEY(key[1].dttk_value); 3995 key[1].dttk_size = 0; 3996 3997 dvar = dtrace_dynvar(dstate, 2, key, 3998 sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC); 3999 4000 if (dvar == NULL) { 4001 regs[rd] = 0; 4002 break; 4003 } 4004 4005 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) { 4006 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data; 4007 } else { 4008 regs[rd] = *((uint64_t *)dvar->dtdv_data); 4009 } 4010 4011 break; 4012 } 4013 4014 case DIF_OP_STTS: { 4015 dtrace_dynvar_t *dvar; 4016 dtrace_key_t *key; 4017 4018 id = DIF_INSTR_VAR(instr); 4019 ASSERT(id >= DIF_VAR_OTHER_UBASE); 4020 id -= DIF_VAR_OTHER_UBASE; 4021 4022 key = &tupregs[DIF_DTR_NREGS]; 4023 key[0].dttk_value = (uint64_t)id; 4024 key[0].dttk_size = 0; 4025 DTRACE_TLS_THRKEY(key[1].dttk_value); 4026 key[1].dttk_size = 0; 4027 v = &vstate->dtvs_tlocals[id]; 4028 4029 dvar = dtrace_dynvar(dstate, 2, key, 4030 v->dtdv_type.dtdt_size > sizeof (uint64_t) ? 4031 v->dtdv_type.dtdt_size : sizeof (uint64_t), 4032 regs[rd] ? DTRACE_DYNVAR_ALLOC : 4033 DTRACE_DYNVAR_DEALLOC); 4034 4035 /* 4036 * Given that we're storing to thread-local data, 4037 * we need to flush our predicate cache. 4038 */ 4039 curthread->t_predcache = NULL; 4040 4041 if (dvar == NULL) 4042 break; 4043 4044 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) { 4045 dtrace_vcopy((void *)(uintptr_t)regs[rd], 4046 dvar->dtdv_data, &v->dtdv_type); 4047 } else { 4048 *((uint64_t *)dvar->dtdv_data) = regs[rd]; 4049 } 4050 4051 break; 4052 } 4053 4054 case DIF_OP_SRA: 4055 regs[rd] = (int64_t)regs[r1] >> regs[r2]; 4056 break; 4057 4058 case DIF_OP_CALL: 4059 dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd, 4060 regs, tupregs, ttop, mstate, state); 4061 break; 4062 4063 case DIF_OP_PUSHTR: 4064 if (ttop == DIF_DTR_NREGS) { 4065 *flags |= CPU_DTRACE_TUPOFLOW; 4066 break; 4067 } 4068 4069 if (r1 == DIF_TYPE_STRING) { 4070 /* 4071 * If this is a string type and the size is 0, 4072 * we'll use the system-wide default string 4073 * size. Note that we are _not_ looking at 4074 * the value of the DTRACEOPT_STRSIZE option; 4075 * had this been set, we would expect to have 4076 * a non-zero size value in the "pushtr". 4077 */ 4078 tupregs[ttop].dttk_size = 4079 dtrace_strlen((char *)(uintptr_t)regs[rd], 4080 regs[r2] ? regs[r2] : 4081 dtrace_strsize_default) + 1; 4082 } else { 4083 tupregs[ttop].dttk_size = regs[r2]; 4084 } 4085 4086 tupregs[ttop++].dttk_value = regs[rd]; 4087 break; 4088 4089 case DIF_OP_PUSHTV: 4090 if (ttop == DIF_DTR_NREGS) { 4091 *flags |= CPU_DTRACE_TUPOFLOW; 4092 break; 4093 } 4094 4095 tupregs[ttop].dttk_value = regs[rd]; 4096 tupregs[ttop++].dttk_size = 0; 4097 break; 4098 4099 case DIF_OP_POPTS: 4100 if (ttop != 0) 4101 ttop--; 4102 break; 4103 4104 case DIF_OP_FLUSHTS: 4105 ttop = 0; 4106 break; 4107 4108 case DIF_OP_LDGAA: 4109 case DIF_OP_LDTAA: { 4110 dtrace_dynvar_t *dvar; 4111 dtrace_key_t *key = tupregs; 4112 uint_t nkeys = ttop; 4113 4114 id = DIF_INSTR_VAR(instr); 4115 ASSERT(id >= DIF_VAR_OTHER_UBASE); 4116 id -= DIF_VAR_OTHER_UBASE; 4117 4118 key[nkeys].dttk_value = (uint64_t)id; 4119 key[nkeys++].dttk_size = 0; 4120 4121 if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) { 4122 DTRACE_TLS_THRKEY(key[nkeys].dttk_value); 4123 key[nkeys++].dttk_size = 0; 4124 v = &vstate->dtvs_tlocals[id]; 4125 } else { 4126 v = &vstate->dtvs_globals[id]->dtsv_var; 4127 } 4128 4129 dvar = dtrace_dynvar(dstate, nkeys, key, 4130 v->dtdv_type.dtdt_size > sizeof (uint64_t) ? 4131 v->dtdv_type.dtdt_size : sizeof (uint64_t), 4132 DTRACE_DYNVAR_NOALLOC); 4133 4134 if (dvar == NULL) { 4135 regs[rd] = 0; 4136 break; 4137 } 4138 4139 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) { 4140 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data; 4141 } else { 4142 regs[rd] = *((uint64_t *)dvar->dtdv_data); 4143 } 4144 4145 break; 4146 } 4147 4148 case DIF_OP_STGAA: 4149 case DIF_OP_STTAA: { 4150 dtrace_dynvar_t *dvar; 4151 dtrace_key_t *key = tupregs; 4152 uint_t nkeys = ttop; 4153 4154 id = DIF_INSTR_VAR(instr); 4155 ASSERT(id >= DIF_VAR_OTHER_UBASE); 4156 id -= DIF_VAR_OTHER_UBASE; 4157 4158 key[nkeys].dttk_value = (uint64_t)id; 4159 key[nkeys++].dttk_size = 0; 4160 4161 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) { 4162 DTRACE_TLS_THRKEY(key[nkeys].dttk_value); 4163 key[nkeys++].dttk_size = 0; 4164 v = &vstate->dtvs_tlocals[id]; 4165 } else { 4166 v = &vstate->dtvs_globals[id]->dtsv_var; 4167 } 4168 4169 dvar = dtrace_dynvar(dstate, nkeys, key, 4170 v->dtdv_type.dtdt_size > sizeof (uint64_t) ? 4171 v->dtdv_type.dtdt_size : sizeof (uint64_t), 4172 regs[rd] ? DTRACE_DYNVAR_ALLOC : 4173 DTRACE_DYNVAR_DEALLOC); 4174 4175 if (dvar == NULL) 4176 break; 4177 4178 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) { 4179 dtrace_vcopy((void *)(uintptr_t)regs[rd], 4180 dvar->dtdv_data, &v->dtdv_type); 4181 } else { 4182 *((uint64_t *)dvar->dtdv_data) = regs[rd]; 4183 } 4184 4185 break; 4186 } 4187 4188 case DIF_OP_ALLOCS: { 4189 uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8); 4190 size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1]; 4191 4192 if (mstate->dtms_scratch_ptr + size > 4193 mstate->dtms_scratch_base + 4194 mstate->dtms_scratch_size) { 4195 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 4196 regs[rd] = NULL; 4197 } else { 4198 dtrace_bzero((void *) 4199 mstate->dtms_scratch_ptr, size); 4200 mstate->dtms_scratch_ptr += size; 4201 regs[rd] = ptr; 4202 } 4203 break; 4204 } 4205 4206 case DIF_OP_COPYS: 4207 if (!dtrace_canstore(regs[rd], regs[r2], 4208 mstate, vstate)) { 4209 *flags |= CPU_DTRACE_BADADDR; 4210 *illval = regs[rd]; 4211 break; 4212 } 4213 4214 dtrace_bcopy((void *)(uintptr_t)regs[r1], 4215 (void *)(uintptr_t)regs[rd], (size_t)regs[r2]); 4216 break; 4217 4218 case DIF_OP_STB: 4219 if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) { 4220 *flags |= CPU_DTRACE_BADADDR; 4221 *illval = regs[rd]; 4222 break; 4223 } 4224 *((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1]; 4225 break; 4226 4227 case DIF_OP_STH: 4228 if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) { 4229 *flags |= CPU_DTRACE_BADADDR; 4230 *illval = regs[rd]; 4231 break; 4232 } 4233 if (regs[rd] & 1) { 4234 *flags |= CPU_DTRACE_BADALIGN; 4235 *illval = regs[rd]; 4236 break; 4237 } 4238 *((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1]; 4239 break; 4240 4241 case DIF_OP_STW: 4242 if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) { 4243 *flags |= CPU_DTRACE_BADADDR; 4244 *illval = regs[rd]; 4245 break; 4246 } 4247 if (regs[rd] & 3) { 4248 *flags |= CPU_DTRACE_BADALIGN; 4249 *illval = regs[rd]; 4250 break; 4251 } 4252 *((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1]; 4253 break; 4254 4255 case DIF_OP_STX: 4256 if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) { 4257 *flags |= CPU_DTRACE_BADADDR; 4258 *illval = regs[rd]; 4259 break; 4260 } 4261 if (regs[rd] & 7) { 4262 *flags |= CPU_DTRACE_BADALIGN; 4263 *illval = regs[rd]; 4264 break; 4265 } 4266 *((uint64_t *)(uintptr_t)regs[rd]) = regs[r1]; 4267 break; 4268 } 4269 } 4270 4271 if (!(*flags & CPU_DTRACE_FAULT)) 4272 return (rval); 4273 4274 mstate->dtms_fltoffs = opc * sizeof (dif_instr_t); 4275 mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS; 4276 4277 return (0); 4278 } 4279 4280 static void 4281 dtrace_action_breakpoint(dtrace_ecb_t *ecb) 4282 { 4283 dtrace_probe_t *probe = ecb->dte_probe; 4284 dtrace_provider_t *prov = probe->dtpr_provider; 4285 char c[DTRACE_FULLNAMELEN + 80], *str; 4286 char *msg = "dtrace: breakpoint action at probe "; 4287 char *ecbmsg = " (ecb "; 4288 uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4)); 4289 uintptr_t val = (uintptr_t)ecb; 4290 int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0; 4291 4292 if (dtrace_destructive_disallow) 4293 return; 4294 4295 /* 4296 * It's impossible to be taking action on the NULL probe. 4297 */ 4298 ASSERT(probe != NULL); 4299 4300 /* 4301 * This is a poor man's (destitute man's?) sprintf(): we want to 4302 * print the provider name, module name, function name and name of 4303 * the probe, along with the hex address of the ECB with the breakpoint 4304 * action -- all of which we must place in the character buffer by 4305 * hand. 4306 */ 4307 while (*msg != '\0') 4308 c[i++] = *msg++; 4309 4310 for (str = prov->dtpv_name; *str != '\0'; str++) 4311 c[i++] = *str; 4312 c[i++] = ':'; 4313 4314 for (str = probe->dtpr_mod; *str != '\0'; str++) 4315 c[i++] = *str; 4316 c[i++] = ':'; 4317 4318 for (str = probe->dtpr_func; *str != '\0'; str++) 4319 c[i++] = *str; 4320 c[i++] = ':'; 4321 4322 for (str = probe->dtpr_name; *str != '\0'; str++) 4323 c[i++] = *str; 4324 4325 while (*ecbmsg != '\0') 4326 c[i++] = *ecbmsg++; 4327 4328 while (shift >= 0) { 4329 mask = (uintptr_t)0xf << shift; 4330 4331 if (val >= ((uintptr_t)1 << shift)) 4332 c[i++] = "0123456789abcdef"[(val & mask) >> shift]; 4333 shift -= 4; 4334 } 4335 4336 c[i++] = ')'; 4337 c[i] = '\0'; 4338 4339 debug_enter(c); 4340 } 4341 4342 static void 4343 dtrace_action_panic(dtrace_ecb_t *ecb) 4344 { 4345 dtrace_probe_t *probe = ecb->dte_probe; 4346 4347 /* 4348 * It's impossible to be taking action on the NULL probe. 4349 */ 4350 ASSERT(probe != NULL); 4351 4352 if (dtrace_destructive_disallow) 4353 return; 4354 4355 if (dtrace_panicked != NULL) 4356 return; 4357 4358 if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL) 4359 return; 4360 4361 /* 4362 * We won the right to panic. (We want to be sure that only one 4363 * thread calls panic() from dtrace_probe(), and that panic() is 4364 * called exactly once.) 4365 */ 4366 dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)", 4367 probe->dtpr_provider->dtpv_name, probe->dtpr_mod, 4368 probe->dtpr_func, probe->dtpr_name, (void *)ecb); 4369 } 4370 4371 static void 4372 dtrace_action_raise(uint64_t sig) 4373 { 4374 if (dtrace_destructive_disallow) 4375 return; 4376 4377 if (sig >= NSIG) { 4378 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP); 4379 return; 4380 } 4381 4382 /* 4383 * raise() has a queue depth of 1 -- we ignore all subsequent 4384 * invocations of the raise() action. 4385 */ 4386 if (curthread->t_dtrace_sig == 0) 4387 curthread->t_dtrace_sig = (uint8_t)sig; 4388 4389 curthread->t_sig_check = 1; 4390 aston(curthread); 4391 } 4392 4393 static void 4394 dtrace_action_stop(void) 4395 { 4396 if (dtrace_destructive_disallow) 4397 return; 4398 4399 if (!curthread->t_dtrace_stop) { 4400 curthread->t_dtrace_stop = 1; 4401 curthread->t_sig_check = 1; 4402 aston(curthread); 4403 } 4404 } 4405 4406 static void 4407 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val) 4408 { 4409 hrtime_t now; 4410 volatile uint16_t *flags; 4411 cpu_t *cpu = CPU; 4412 4413 if (dtrace_destructive_disallow) 4414 return; 4415 4416 flags = (volatile uint16_t *)&cpu_core[cpu->cpu_id].cpuc_dtrace_flags; 4417 4418 now = dtrace_gethrtime(); 4419 4420 if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) { 4421 /* 4422 * We need to advance the mark to the current time. 4423 */ 4424 cpu->cpu_dtrace_chillmark = now; 4425 cpu->cpu_dtrace_chilled = 0; 4426 } 4427 4428 /* 4429 * Now check to see if the requested chill time would take us over 4430 * the maximum amount of time allowed in the chill interval. (Or 4431 * worse, if the calculation itself induces overflow.) 4432 */ 4433 if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max || 4434 cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) { 4435 *flags |= CPU_DTRACE_ILLOP; 4436 return; 4437 } 4438 4439 while (dtrace_gethrtime() - now < val) 4440 continue; 4441 4442 /* 4443 * Normally, we assure that the value of the variable "timestamp" does 4444 * not change within an ECB. The presence of chill() represents an 4445 * exception to this rule, however. 4446 */ 4447 mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP; 4448 cpu->cpu_dtrace_chilled += val; 4449 } 4450 4451 static void 4452 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state, 4453 uint64_t *buf, uint64_t arg) 4454 { 4455 int nframes = DTRACE_USTACK_NFRAMES(arg); 4456 int strsize = DTRACE_USTACK_STRSIZE(arg); 4457 uint64_t *pcs = &buf[1], *fps; 4458 char *str = (char *)&pcs[nframes]; 4459 int size, offs = 0, i, j; 4460 uintptr_t old = mstate->dtms_scratch_ptr, saved; 4461 uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags; 4462 char *sym; 4463 4464 /* 4465 * Should be taking a faster path if string space has not been 4466 * allocated. 4467 */ 4468 ASSERT(strsize != 0); 4469 4470 /* 4471 * We will first allocate some temporary space for the frame pointers. 4472 */ 4473 fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8); 4474 size = (uintptr_t)fps - mstate->dtms_scratch_ptr + 4475 (nframes * sizeof (uint64_t)); 4476 4477 if (mstate->dtms_scratch_ptr + size > 4478 mstate->dtms_scratch_base + mstate->dtms_scratch_size) { 4479 /* 4480 * Not enough room for our frame pointers -- need to indicate 4481 * that we ran out of scratch space. 4482 */ 4483 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 4484 return; 4485 } 4486 4487 mstate->dtms_scratch_ptr += size; 4488 saved = mstate->dtms_scratch_ptr; 4489 4490 /* 4491 * Now get a stack with both program counters and frame pointers. 4492 */ 4493 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT); 4494 dtrace_getufpstack(buf, fps, nframes + 1); 4495 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT); 4496 4497 /* 4498 * If that faulted, we're cooked. 4499 */ 4500 if (*flags & CPU_DTRACE_FAULT) 4501 goto out; 4502 4503 /* 4504 * Now we want to walk up the stack, calling the USTACK helper. For 4505 * each iteration, we restore the scratch pointer. 4506 */ 4507 for (i = 0; i < nframes; i++) { 4508 mstate->dtms_scratch_ptr = saved; 4509 4510 if (offs >= strsize) 4511 break; 4512 4513 sym = (char *)(uintptr_t)dtrace_helper( 4514 DTRACE_HELPER_ACTION_USTACK, 4515 mstate, state, pcs[i], fps[i]); 4516 4517 /* 4518 * If we faulted while running the helper, we're going to 4519 * clear the fault and null out the corresponding string. 4520 */ 4521 if (*flags & CPU_DTRACE_FAULT) { 4522 *flags &= ~CPU_DTRACE_FAULT; 4523 str[offs++] = '\0'; 4524 continue; 4525 } 4526 4527 if (sym == NULL) { 4528 str[offs++] = '\0'; 4529 continue; 4530 } 4531 4532 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT); 4533 4534 /* 4535 * Now copy in the string that the helper returned to us. 4536 */ 4537 for (j = 0; offs + j < strsize; j++) { 4538 if ((str[offs + j] = sym[j]) == '\0') 4539 break; 4540 } 4541 4542 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT); 4543 4544 offs += j + 1; 4545 } 4546 4547 if (offs >= strsize) { 4548 /* 4549 * If we didn't have room for all of the strings, we don't 4550 * abort processing -- this needn't be a fatal error -- but we 4551 * still want to increment a counter (dts_stkstroverflows) to 4552 * allow this condition to be warned about. (If this is from 4553 * a jstack() action, it is easily tuned via jstackstrsize.) 4554 */ 4555 dtrace_error(&state->dts_stkstroverflows); 4556 } 4557 4558 while (offs < strsize) 4559 str[offs++] = '\0'; 4560 4561 out: 4562 mstate->dtms_scratch_ptr = old; 4563 } 4564 4565 /* 4566 * If you're looking for the epicenter of DTrace, you just found it. This 4567 * is the function called by the provider to fire a probe -- from which all 4568 * subsequent probe-context DTrace activity emanates. 4569 */ 4570 void 4571 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1, 4572 uintptr_t arg2, uintptr_t arg3, uintptr_t arg4) 4573 { 4574 processorid_t cpuid; 4575 dtrace_icookie_t cookie; 4576 dtrace_probe_t *probe; 4577 dtrace_mstate_t mstate; 4578 dtrace_ecb_t *ecb; 4579 dtrace_action_t *act; 4580 intptr_t offs; 4581 size_t size; 4582 int vtime, onintr; 4583 volatile uint16_t *flags; 4584 hrtime_t now; 4585 4586 /* 4587 * Kick out immediately if this CPU is still being born (in which case 4588 * curthread will be set to -1) 4589 */ 4590 if ((uintptr_t)curthread & 1) 4591 return; 4592 4593 cookie = dtrace_interrupt_disable(); 4594 probe = dtrace_probes[id - 1]; 4595 cpuid = CPU->cpu_id; 4596 onintr = CPU_ON_INTR(CPU); 4597 4598 if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE && 4599 probe->dtpr_predcache == curthread->t_predcache) { 4600 /* 4601 * We have hit in the predicate cache; we know that 4602 * this predicate would evaluate to be false. 4603 */ 4604 dtrace_interrupt_enable(cookie); 4605 return; 4606 } 4607 4608 if (panic_quiesce) { 4609 /* 4610 * We don't trace anything if we're panicking. 4611 */ 4612 dtrace_interrupt_enable(cookie); 4613 return; 4614 } 4615 4616 now = dtrace_gethrtime(); 4617 vtime = dtrace_vtime_references != 0; 4618 4619 if (vtime && curthread->t_dtrace_start) 4620 curthread->t_dtrace_vtime += now - curthread->t_dtrace_start; 4621 4622 mstate.dtms_probe = probe; 4623 mstate.dtms_arg[0] = arg0; 4624 mstate.dtms_arg[1] = arg1; 4625 mstate.dtms_arg[2] = arg2; 4626 mstate.dtms_arg[3] = arg3; 4627 mstate.dtms_arg[4] = arg4; 4628 4629 flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags; 4630 4631 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) { 4632 dtrace_predicate_t *pred = ecb->dte_predicate; 4633 dtrace_state_t *state = ecb->dte_state; 4634 dtrace_buffer_t *buf = &state->dts_buffer[cpuid]; 4635 dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid]; 4636 dtrace_vstate_t *vstate = &state->dts_vstate; 4637 dtrace_provider_t *prov = probe->dtpr_provider; 4638 int committed = 0; 4639 caddr_t tomax; 4640 4641 /* 4642 * A little subtlety with the following (seemingly innocuous) 4643 * declaration of the automatic 'val': by looking at the 4644 * code, you might think that it could be declared in the 4645 * action processing loop, below. (That is, it's only used in 4646 * the action processing loop.) However, it must be declared 4647 * out of that scope because in the case of DIF expression 4648 * arguments to aggregating actions, one iteration of the 4649 * action loop will use the last iteration's value. 4650 */ 4651 #ifdef lint 4652 uint64_t val = 0; 4653 #else 4654 uint64_t val; 4655 #endif 4656 4657 mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE; 4658 *flags &= ~CPU_DTRACE_ERROR; 4659 4660 if (prov == dtrace_provider) { 4661 /* 4662 * If dtrace itself is the provider of this probe, 4663 * we're only going to continue processing the ECB if 4664 * arg0 (the dtrace_state_t) is equal to the ECB's 4665 * creating state. (This prevents disjoint consumers 4666 * from seeing one another's metaprobes.) 4667 */ 4668 if (arg0 != (uint64_t)(uintptr_t)state) 4669 continue; 4670 } 4671 4672 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) { 4673 /* 4674 * We're not currently active. If our provider isn't 4675 * the dtrace pseudo provider, we're not interested. 4676 */ 4677 if (prov != dtrace_provider) 4678 continue; 4679 4680 /* 4681 * Now we must further check if we are in the BEGIN 4682 * probe. If we are, we will only continue processing 4683 * if we're still in WARMUP -- if one BEGIN enabling 4684 * has invoked the exit() action, we don't want to 4685 * evaluate subsequent BEGIN enablings. 4686 */ 4687 if (probe->dtpr_id == dtrace_probeid_begin && 4688 state->dts_activity != DTRACE_ACTIVITY_WARMUP) { 4689 ASSERT(state->dts_activity == 4690 DTRACE_ACTIVITY_DRAINING); 4691 continue; 4692 } 4693 } 4694 4695 if (ecb->dte_cond) { 4696 /* 4697 * If the dte_cond bits indicate that this 4698 * consumer is only allowed to see user-mode firings 4699 * of this probe, call the provider's dtps_usermode() 4700 * entry point to check that the probe was fired 4701 * while in a user context. Skip this ECB if that's 4702 * not the case. 4703 */ 4704 if ((ecb->dte_cond & DTRACE_COND_USERMODE) && 4705 prov->dtpv_pops.dtps_usermode(prov->dtpv_arg, 4706 probe->dtpr_id, probe->dtpr_arg) == 0) 4707 continue; 4708 4709 /* 4710 * This is more subtle than it looks. We have to be 4711 * absolutely certain that CRED() isn't going to 4712 * change out from under us so it's only legit to 4713 * examine that structure if we're in constrained 4714 * situations. Currently, the only times we'll this 4715 * check is if a non-super-user has enabled the 4716 * profile or syscall providers -- providers that 4717 * allow visibility of all processes. For the 4718 * profile case, the check above will ensure that 4719 * we're examining a user context. 4720 */ 4721 if (ecb->dte_cond & DTRACE_COND_OWNER) { 4722 uid_t uid = ecb->dte_state->dts_cred.dcr_uid; 4723 gid_t gid = ecb->dte_state->dts_cred.dcr_gid; 4724 cred_t *cr; 4725 proc_t *proc; 4726 4727 if ((cr = CRED()) == NULL || 4728 uid != cr->cr_uid || 4729 uid != cr->cr_ruid || 4730 uid != cr->cr_suid || 4731 gid != cr->cr_gid || 4732 gid != cr->cr_rgid || 4733 gid != cr->cr_sgid || 4734 (proc = ttoproc(curthread)) == NULL || 4735 (proc->p_flag & SNOCD)) 4736 continue; 4737 4738 } 4739 } 4740 4741 if (now - state->dts_alive > dtrace_deadman_timeout) { 4742 /* 4743 * We seem to be dead. Unless we (a) have kernel 4744 * destructive permissions (b) have expicitly enabled 4745 * destructive actions and (c) destructive actions have 4746 * not been disabled, we're going to transition into 4747 * the KILLED state, from which no further processing 4748 * on this state will be performed. 4749 */ 4750 if (!dtrace_priv_kernel_destructive(state) || 4751 !state->dts_cred.dcr_destructive || 4752 dtrace_destructive_disallow) { 4753 void *activity = &state->dts_activity; 4754 dtrace_activity_t current; 4755 4756 do { 4757 current = state->dts_activity; 4758 } while (dtrace_cas32(activity, current, 4759 DTRACE_ACTIVITY_KILLED) != current); 4760 4761 continue; 4762 } 4763 } 4764 4765 if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed, 4766 ecb->dte_alignment, state, &mstate)) < 0) 4767 continue; 4768 4769 tomax = buf->dtb_tomax; 4770 ASSERT(tomax != NULL); 4771 4772 if (ecb->dte_size != 0) 4773 DTRACE_STORE(uint32_t, tomax, offs, ecb->dte_epid); 4774 4775 mstate.dtms_epid = ecb->dte_epid; 4776 mstate.dtms_present |= DTRACE_MSTATE_EPID; 4777 4778 if (pred != NULL) { 4779 dtrace_difo_t *dp = pred->dtp_difo; 4780 int rval; 4781 4782 rval = dtrace_dif_emulate(dp, &mstate, vstate, state); 4783 4784 if (!(*flags & CPU_DTRACE_ERROR) && !rval) { 4785 dtrace_cacheid_t cid = probe->dtpr_predcache; 4786 4787 if (cid != DTRACE_CACHEIDNONE && !onintr) { 4788 /* 4789 * Update the predicate cache... 4790 */ 4791 ASSERT(cid == pred->dtp_cacheid); 4792 curthread->t_predcache = cid; 4793 } 4794 4795 continue; 4796 } 4797 } 4798 4799 for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) && 4800 act != NULL; act = act->dta_next) { 4801 size_t valoffs; 4802 dtrace_difo_t *dp; 4803 dtrace_recdesc_t *rec = &act->dta_rec; 4804 4805 size = rec->dtrd_size; 4806 valoffs = offs + rec->dtrd_offset; 4807 4808 if (DTRACEACT_ISAGG(act->dta_kind)) { 4809 uint64_t v = 0xbad; 4810 dtrace_aggregation_t *agg; 4811 4812 agg = (dtrace_aggregation_t *)act; 4813 4814 if ((dp = act->dta_difo) != NULL) 4815 v = dtrace_dif_emulate(dp, 4816 &mstate, vstate, state); 4817 4818 if (*flags & CPU_DTRACE_ERROR) 4819 continue; 4820 4821 /* 4822 * Note that we always pass the expression 4823 * value from the previous iteration of the 4824 * action loop. This value will only be used 4825 * if there is an expression argument to the 4826 * aggregating action, denoted by the 4827 * dtag_hasarg field. 4828 */ 4829 dtrace_aggregate(agg, buf, 4830 offs, aggbuf, v, val); 4831 continue; 4832 } 4833 4834 switch (act->dta_kind) { 4835 case DTRACEACT_STOP: 4836 if (dtrace_priv_proc_destructive(state)) 4837 dtrace_action_stop(); 4838 continue; 4839 4840 case DTRACEACT_BREAKPOINT: 4841 if (dtrace_priv_kernel_destructive(state)) 4842 dtrace_action_breakpoint(ecb); 4843 continue; 4844 4845 case DTRACEACT_PANIC: 4846 if (dtrace_priv_kernel_destructive(state)) 4847 dtrace_action_panic(ecb); 4848 continue; 4849 4850 case DTRACEACT_STACK: 4851 if (!dtrace_priv_kernel(state)) 4852 continue; 4853 4854 dtrace_getpcstack((pc_t *)(tomax + valoffs), 4855 size / sizeof (pc_t), probe->dtpr_aframes, 4856 DTRACE_ANCHORED(probe) ? NULL : 4857 (uint32_t *)arg0); 4858 4859 continue; 4860 4861 case DTRACEACT_JSTACK: 4862 case DTRACEACT_USTACK: 4863 if (!dtrace_priv_proc(state)) 4864 continue; 4865 4866 /* 4867 * See comment in DIF_VAR_PID. 4868 */ 4869 if (DTRACE_ANCHORED(mstate.dtms_probe) && 4870 CPU_ON_INTR(CPU)) { 4871 int depth = DTRACE_USTACK_NFRAMES( 4872 rec->dtrd_arg) + 1; 4873 4874 dtrace_bzero((void *)(tomax + valoffs), 4875 DTRACE_USTACK_STRSIZE(rec->dtrd_arg) 4876 + depth * sizeof (uint64_t)); 4877 4878 continue; 4879 } 4880 4881 if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 && 4882 curproc->p_dtrace_helpers != NULL) { 4883 /* 4884 * This is the slow path -- we have 4885 * allocated string space, and we're 4886 * getting the stack of a process that 4887 * has helpers. Call into a separate 4888 * routine to perform this processing. 4889 */ 4890 dtrace_action_ustack(&mstate, state, 4891 (uint64_t *)(tomax + valoffs), 4892 rec->dtrd_arg); 4893 continue; 4894 } 4895 4896 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT); 4897 dtrace_getupcstack((uint64_t *) 4898 (tomax + valoffs), 4899 DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1); 4900 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT); 4901 continue; 4902 4903 default: 4904 break; 4905 } 4906 4907 dp = act->dta_difo; 4908 ASSERT(dp != NULL); 4909 4910 val = dtrace_dif_emulate(dp, &mstate, vstate, state); 4911 4912 if (*flags & CPU_DTRACE_ERROR) 4913 continue; 4914 4915 switch (act->dta_kind) { 4916 case DTRACEACT_SPECULATE: 4917 ASSERT(buf == &state->dts_buffer[cpuid]); 4918 buf = dtrace_speculation_buffer(state, 4919 cpuid, val); 4920 4921 if (buf == NULL) { 4922 *flags |= CPU_DTRACE_DROP; 4923 continue; 4924 } 4925 4926 offs = dtrace_buffer_reserve(buf, 4927 ecb->dte_needed, ecb->dte_alignment, 4928 state, NULL); 4929 4930 if (offs < 0) { 4931 *flags |= CPU_DTRACE_DROP; 4932 continue; 4933 } 4934 4935 tomax = buf->dtb_tomax; 4936 ASSERT(tomax != NULL); 4937 4938 if (ecb->dte_size != 0) 4939 DTRACE_STORE(uint32_t, tomax, offs, 4940 ecb->dte_epid); 4941 continue; 4942 4943 case DTRACEACT_CHILL: 4944 if (dtrace_priv_kernel_destructive(state)) 4945 dtrace_action_chill(&mstate, val); 4946 continue; 4947 4948 case DTRACEACT_RAISE: 4949 if (dtrace_priv_proc_destructive(state)) 4950 dtrace_action_raise(val); 4951 continue; 4952 4953 case DTRACEACT_COMMIT: 4954 ASSERT(!committed); 4955 4956 /* 4957 * We need to commit our buffer state. 4958 */ 4959 if (ecb->dte_size) 4960 buf->dtb_offset = offs + ecb->dte_size; 4961 buf = &state->dts_buffer[cpuid]; 4962 dtrace_speculation_commit(state, cpuid, val); 4963 committed = 1; 4964 continue; 4965 4966 case DTRACEACT_DISCARD: 4967 dtrace_speculation_discard(state, cpuid, val); 4968 continue; 4969 4970 case DTRACEACT_DIFEXPR: 4971 case DTRACEACT_LIBACT: 4972 case DTRACEACT_PRINTF: 4973 case DTRACEACT_PRINTA: 4974 case DTRACEACT_SYSTEM: 4975 case DTRACEACT_FREOPEN: 4976 break; 4977 4978 case DTRACEACT_SYM: 4979 case DTRACEACT_MOD: 4980 if (!dtrace_priv_kernel(state)) 4981 continue; 4982 break; 4983 4984 case DTRACEACT_USYM: 4985 case DTRACEACT_UMOD: 4986 case DTRACEACT_UADDR: { 4987 struct pid *pid = curthread->t_procp->p_pidp; 4988 4989 if (!dtrace_priv_proc(state)) 4990 continue; 4991 4992 DTRACE_STORE(uint64_t, tomax, 4993 valoffs, (uint64_t)pid->pid_id); 4994 DTRACE_STORE(uint64_t, tomax, 4995 valoffs + sizeof (uint64_t), val); 4996 4997 continue; 4998 } 4999 5000 case DTRACEACT_EXIT: { 5001 /* 5002 * For the exit action, we are going to attempt 5003 * to atomically set our activity to be 5004 * draining. If this fails (either because 5005 * another CPU has beat us to the exit action, 5006 * or because our current activity is something 5007 * other than ACTIVE or WARMUP), we will 5008 * continue. This assures that the exit action 5009 * can be successfully recorded at most once 5010 * when we're in the ACTIVE state. If we're 5011 * encountering the exit() action while in 5012 * COOLDOWN, however, we want to honor the new 5013 * status code. (We know that we're the only 5014 * thread in COOLDOWN, so there is no race.) 5015 */ 5016 void *activity = &state->dts_activity; 5017 dtrace_activity_t current = state->dts_activity; 5018 5019 if (current == DTRACE_ACTIVITY_COOLDOWN) 5020 break; 5021 5022 if (current != DTRACE_ACTIVITY_WARMUP) 5023 current = DTRACE_ACTIVITY_ACTIVE; 5024 5025 if (dtrace_cas32(activity, current, 5026 DTRACE_ACTIVITY_DRAINING) != current) { 5027 *flags |= CPU_DTRACE_DROP; 5028 continue; 5029 } 5030 5031 break; 5032 } 5033 5034 default: 5035 ASSERT(0); 5036 } 5037 5038 if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF) { 5039 uintptr_t end = valoffs + size; 5040 5041 /* 5042 * If this is a string, we're going to only 5043 * load until we find the zero byte -- after 5044 * which we'll store zero bytes. 5045 */ 5046 if (dp->dtdo_rtype.dtdt_kind == 5047 DIF_TYPE_STRING) { 5048 char c = '\0' + 1; 5049 int intuple = act->dta_intuple; 5050 size_t s; 5051 5052 for (s = 0; s < size; s++) { 5053 if (c != '\0') 5054 c = dtrace_load8(val++); 5055 5056 DTRACE_STORE(uint8_t, tomax, 5057 valoffs++, c); 5058 5059 if (c == '\0' && intuple) 5060 break; 5061 } 5062 5063 continue; 5064 } 5065 5066 while (valoffs < end) { 5067 DTRACE_STORE(uint8_t, tomax, valoffs++, 5068 dtrace_load8(val++)); 5069 } 5070 5071 continue; 5072 } 5073 5074 switch (size) { 5075 case 0: 5076 break; 5077 5078 case sizeof (uint8_t): 5079 DTRACE_STORE(uint8_t, tomax, valoffs, val); 5080 break; 5081 case sizeof (uint16_t): 5082 DTRACE_STORE(uint16_t, tomax, valoffs, val); 5083 break; 5084 case sizeof (uint32_t): 5085 DTRACE_STORE(uint32_t, tomax, valoffs, val); 5086 break; 5087 case sizeof (uint64_t): 5088 DTRACE_STORE(uint64_t, tomax, valoffs, val); 5089 break; 5090 default: 5091 /* 5092 * Any other size should have been returned by 5093 * reference, not by value. 5094 */ 5095 ASSERT(0); 5096 break; 5097 } 5098 } 5099 5100 if (*flags & CPU_DTRACE_DROP) 5101 continue; 5102 5103 if (*flags & CPU_DTRACE_FAULT) { 5104 int ndx; 5105 dtrace_action_t *err; 5106 5107 buf->dtb_errors++; 5108 5109 if (probe->dtpr_id == dtrace_probeid_error) { 5110 /* 5111 * There's nothing we can do -- we had an 5112 * error on the error probe. We bump an 5113 * error counter to at least indicate that 5114 * this condition happened. 5115 */ 5116 dtrace_error(&state->dts_dblerrors); 5117 continue; 5118 } 5119 5120 if (vtime) { 5121 /* 5122 * Before recursing on dtrace_probe(), we 5123 * need to explicitly clear out our start 5124 * time to prevent it from being accumulated 5125 * into t_dtrace_vtime. 5126 */ 5127 curthread->t_dtrace_start = 0; 5128 } 5129 5130 /* 5131 * Iterate over the actions to figure out which action 5132 * we were processing when we experienced the error. 5133 * Note that act points _past_ the faulting action; if 5134 * act is ecb->dte_action, the fault was in the 5135 * predicate, if it's ecb->dte_action->dta_next it's 5136 * in action #1, and so on. 5137 */ 5138 for (err = ecb->dte_action, ndx = 0; 5139 err != act; err = err->dta_next, ndx++) 5140 continue; 5141 5142 dtrace_probe_error(state, ecb->dte_epid, ndx, 5143 (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ? 5144 mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags), 5145 cpu_core[cpuid].cpuc_dtrace_illval); 5146 5147 continue; 5148 } 5149 5150 if (!committed) 5151 buf->dtb_offset = offs + ecb->dte_size; 5152 } 5153 5154 if (vtime) 5155 curthread->t_dtrace_start = dtrace_gethrtime(); 5156 5157 dtrace_interrupt_enable(cookie); 5158 } 5159 5160 /* 5161 * DTrace Probe Hashing Functions 5162 * 5163 * The functions in this section (and indeed, the functions in remaining 5164 * sections) are not _called_ from probe context. (Any exceptions to this are 5165 * marked with a "Note:".) Rather, they are called from elsewhere in the 5166 * DTrace framework to look-up probes in, add probes to and remove probes from 5167 * the DTrace probe hashes. (Each probe is hashed by each element of the 5168 * probe tuple -- allowing for fast lookups, regardless of what was 5169 * specified.) 5170 */ 5171 static uint_t 5172 dtrace_hash_str(char *p) 5173 { 5174 unsigned int g; 5175 uint_t hval = 0; 5176 5177 while (*p) { 5178 hval = (hval << 4) + *p++; 5179 if ((g = (hval & 0xf0000000)) != 0) 5180 hval ^= g >> 24; 5181 hval &= ~g; 5182 } 5183 return (hval); 5184 } 5185 5186 static dtrace_hash_t * 5187 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs) 5188 { 5189 dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP); 5190 5191 hash->dth_stroffs = stroffs; 5192 hash->dth_nextoffs = nextoffs; 5193 hash->dth_prevoffs = prevoffs; 5194 5195 hash->dth_size = 1; 5196 hash->dth_mask = hash->dth_size - 1; 5197 5198 hash->dth_tab = kmem_zalloc(hash->dth_size * 5199 sizeof (dtrace_hashbucket_t *), KM_SLEEP); 5200 5201 return (hash); 5202 } 5203 5204 static void 5205 dtrace_hash_destroy(dtrace_hash_t *hash) 5206 { 5207 #ifdef DEBUG 5208 int i; 5209 5210 for (i = 0; i < hash->dth_size; i++) 5211 ASSERT(hash->dth_tab[i] == NULL); 5212 #endif 5213 5214 kmem_free(hash->dth_tab, 5215 hash->dth_size * sizeof (dtrace_hashbucket_t *)); 5216 kmem_free(hash, sizeof (dtrace_hash_t)); 5217 } 5218 5219 static void 5220 dtrace_hash_resize(dtrace_hash_t *hash) 5221 { 5222 int size = hash->dth_size, i, ndx; 5223 int new_size = hash->dth_size << 1; 5224 int new_mask = new_size - 1; 5225 dtrace_hashbucket_t **new_tab, *bucket, *next; 5226 5227 ASSERT((new_size & new_mask) == 0); 5228 5229 new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP); 5230 5231 for (i = 0; i < size; i++) { 5232 for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) { 5233 dtrace_probe_t *probe = bucket->dthb_chain; 5234 5235 ASSERT(probe != NULL); 5236 ndx = DTRACE_HASHSTR(hash, probe) & new_mask; 5237 5238 next = bucket->dthb_next; 5239 bucket->dthb_next = new_tab[ndx]; 5240 new_tab[ndx] = bucket; 5241 } 5242 } 5243 5244 kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *)); 5245 hash->dth_tab = new_tab; 5246 hash->dth_size = new_size; 5247 hash->dth_mask = new_mask; 5248 } 5249 5250 static void 5251 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new) 5252 { 5253 int hashval = DTRACE_HASHSTR(hash, new); 5254 int ndx = hashval & hash->dth_mask; 5255 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx]; 5256 dtrace_probe_t **nextp, **prevp; 5257 5258 for (; bucket != NULL; bucket = bucket->dthb_next) { 5259 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new)) 5260 goto add; 5261 } 5262 5263 if ((hash->dth_nbuckets >> 1) > hash->dth_size) { 5264 dtrace_hash_resize(hash); 5265 dtrace_hash_add(hash, new); 5266 return; 5267 } 5268 5269 bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP); 5270 bucket->dthb_next = hash->dth_tab[ndx]; 5271 hash->dth_tab[ndx] = bucket; 5272 hash->dth_nbuckets++; 5273 5274 add: 5275 nextp = DTRACE_HASHNEXT(hash, new); 5276 ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL); 5277 *nextp = bucket->dthb_chain; 5278 5279 if (bucket->dthb_chain != NULL) { 5280 prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain); 5281 ASSERT(*prevp == NULL); 5282 *prevp = new; 5283 } 5284 5285 bucket->dthb_chain = new; 5286 bucket->dthb_len++; 5287 } 5288 5289 static dtrace_probe_t * 5290 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template) 5291 { 5292 int hashval = DTRACE_HASHSTR(hash, template); 5293 int ndx = hashval & hash->dth_mask; 5294 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx]; 5295 5296 for (; bucket != NULL; bucket = bucket->dthb_next) { 5297 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template)) 5298 return (bucket->dthb_chain); 5299 } 5300 5301 return (NULL); 5302 } 5303 5304 static int 5305 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template) 5306 { 5307 int hashval = DTRACE_HASHSTR(hash, template); 5308 int ndx = hashval & hash->dth_mask; 5309 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx]; 5310 5311 for (; bucket != NULL; bucket = bucket->dthb_next) { 5312 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template)) 5313 return (bucket->dthb_len); 5314 } 5315 5316 return (NULL); 5317 } 5318 5319 static void 5320 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe) 5321 { 5322 int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask; 5323 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx]; 5324 5325 dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe); 5326 dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe); 5327 5328 /* 5329 * Find the bucket that we're removing this probe from. 5330 */ 5331 for (; bucket != NULL; bucket = bucket->dthb_next) { 5332 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe)) 5333 break; 5334 } 5335 5336 ASSERT(bucket != NULL); 5337 5338 if (*prevp == NULL) { 5339 if (*nextp == NULL) { 5340 /* 5341 * The removed probe was the only probe on this 5342 * bucket; we need to remove the bucket. 5343 */ 5344 dtrace_hashbucket_t *b = hash->dth_tab[ndx]; 5345 5346 ASSERT(bucket->dthb_chain == probe); 5347 ASSERT(b != NULL); 5348 5349 if (b == bucket) { 5350 hash->dth_tab[ndx] = bucket->dthb_next; 5351 } else { 5352 while (b->dthb_next != bucket) 5353 b = b->dthb_next; 5354 b->dthb_next = bucket->dthb_next; 5355 } 5356 5357 ASSERT(hash->dth_nbuckets > 0); 5358 hash->dth_nbuckets--; 5359 kmem_free(bucket, sizeof (dtrace_hashbucket_t)); 5360 return; 5361 } 5362 5363 bucket->dthb_chain = *nextp; 5364 } else { 5365 *(DTRACE_HASHNEXT(hash, *prevp)) = *nextp; 5366 } 5367 5368 if (*nextp != NULL) 5369 *(DTRACE_HASHPREV(hash, *nextp)) = *prevp; 5370 } 5371 5372 /* 5373 * DTrace Utility Functions 5374 * 5375 * These are random utility functions that are _not_ called from probe context. 5376 */ 5377 static int 5378 dtrace_badattr(const dtrace_attribute_t *a) 5379 { 5380 return (a->dtat_name > DTRACE_STABILITY_MAX || 5381 a->dtat_data > DTRACE_STABILITY_MAX || 5382 a->dtat_class > DTRACE_CLASS_MAX); 5383 } 5384 5385 /* 5386 * Return a duplicate copy of a string. If the specified string is NULL, 5387 * this function returns a zero-length string. 5388 */ 5389 static char * 5390 dtrace_strdup(const char *str) 5391 { 5392 char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP); 5393 5394 if (str != NULL) 5395 (void) strcpy(new, str); 5396 5397 return (new); 5398 } 5399 5400 #define DTRACE_ISALPHA(c) \ 5401 (((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z')) 5402 5403 static int 5404 dtrace_badname(const char *s) 5405 { 5406 char c; 5407 5408 if (s == NULL || (c = *s++) == '\0') 5409 return (0); 5410 5411 if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.') 5412 return (1); 5413 5414 while ((c = *s++) != '\0') { 5415 if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') && 5416 c != '-' && c != '_' && c != '.' && c != '`') 5417 return (1); 5418 } 5419 5420 return (0); 5421 } 5422 5423 static void 5424 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp) 5425 { 5426 uint32_t priv; 5427 5428 *uidp = crgetuid(cr); 5429 if (PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) { 5430 priv = DTRACE_PRIV_ALL; 5431 } else { 5432 priv = 0; 5433 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) 5434 priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER; 5435 else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) 5436 priv |= DTRACE_PRIV_USER; 5437 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) 5438 priv |= DTRACE_PRIV_PROC; 5439 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) 5440 priv |= DTRACE_PRIV_OWNER; 5441 } 5442 5443 *privp = priv; 5444 } 5445 5446 #ifdef DTRACE_ERRDEBUG 5447 static void 5448 dtrace_errdebug(const char *str) 5449 { 5450 int hval = dtrace_hash_str((char *)str) % DTRACE_ERRHASHSZ; 5451 int occupied = 0; 5452 5453 mutex_enter(&dtrace_errlock); 5454 dtrace_errlast = str; 5455 dtrace_errthread = curthread; 5456 5457 while (occupied++ < DTRACE_ERRHASHSZ) { 5458 if (dtrace_errhash[hval].dter_msg == str) { 5459 dtrace_errhash[hval].dter_count++; 5460 goto out; 5461 } 5462 5463 if (dtrace_errhash[hval].dter_msg != NULL) { 5464 hval = (hval + 1) % DTRACE_ERRHASHSZ; 5465 continue; 5466 } 5467 5468 dtrace_errhash[hval].dter_msg = str; 5469 dtrace_errhash[hval].dter_count = 1; 5470 goto out; 5471 } 5472 5473 panic("dtrace: undersized error hash"); 5474 out: 5475 mutex_exit(&dtrace_errlock); 5476 } 5477 #endif 5478 5479 /* 5480 * DTrace Matching Functions 5481 * 5482 * These functions are used to match groups of probes, given some elements of 5483 * a probe tuple, or some globbed expressions for elements of a probe tuple. 5484 */ 5485 static int 5486 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid) 5487 { 5488 if (priv != DTRACE_PRIV_ALL) { 5489 uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags; 5490 uint32_t match = priv & ppriv; 5491 5492 /* 5493 * No PRIV_DTRACE_* privileges... 5494 */ 5495 if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER | 5496 DTRACE_PRIV_KERNEL)) == 0) 5497 return (0); 5498 5499 /* 5500 * No matching bits, but there were bits to match... 5501 */ 5502 if (match == 0 && ppriv != 0) 5503 return (0); 5504 5505 /* 5506 * Need to have permissions to the process, but don't... 5507 */ 5508 if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 && 5509 uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) 5510 return (0); 5511 } 5512 5513 return (1); 5514 } 5515 5516 /* 5517 * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which 5518 * consists of input pattern strings and an ops-vector to evaluate them. 5519 * This function returns >0 for match, 0 for no match, and <0 for error. 5520 */ 5521 static int 5522 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp, 5523 uint32_t priv, uid_t uid) 5524 { 5525 dtrace_provider_t *pvp = prp->dtpr_provider; 5526 int rv; 5527 5528 if (pvp->dtpv_defunct) 5529 return (0); 5530 5531 if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0) 5532 return (rv); 5533 5534 if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0) 5535 return (rv); 5536 5537 if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0) 5538 return (rv); 5539 5540 if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0) 5541 return (rv); 5542 5543 if (dtrace_match_priv(prp, priv, uid) == 0) 5544 return (0); 5545 5546 return (rv); 5547 } 5548 5549 /* 5550 * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN) 5551 * interface for matching a glob pattern 'p' to an input string 's'. Unlike 5552 * libc's version, the kernel version only applies to 8-bit ASCII strings. 5553 * In addition, all of the recursion cases except for '*' matching have been 5554 * unwound. For '*', we still implement recursive evaluation, but a depth 5555 * counter is maintained and matching is aborted if we recurse too deep. 5556 * The function returns 0 if no match, >0 if match, and <0 if recursion error. 5557 */ 5558 static int 5559 dtrace_match_glob(const char *s, const char *p, int depth) 5560 { 5561 const char *olds; 5562 char s1, c; 5563 int gs; 5564 5565 if (depth > DTRACE_PROBEKEY_MAXDEPTH) 5566 return (-1); 5567 5568 if (s == NULL) 5569 s = ""; /* treat NULL as empty string */ 5570 5571 top: 5572 olds = s; 5573 s1 = *s++; 5574 5575 if (p == NULL) 5576 return (0); 5577 5578 if ((c = *p++) == '\0') 5579 return (s1 == '\0'); 5580 5581 switch (c) { 5582 case '[': { 5583 int ok = 0, notflag = 0; 5584 char lc = '\0'; 5585 5586 if (s1 == '\0') 5587 return (0); 5588 5589 if (*p == '!') { 5590 notflag = 1; 5591 p++; 5592 } 5593 5594 if ((c = *p++) == '\0') 5595 return (0); 5596 5597 do { 5598 if (c == '-' && lc != '\0' && *p != ']') { 5599 if ((c = *p++) == '\0') 5600 return (0); 5601 if (c == '\\' && (c = *p++) == '\0') 5602 return (0); 5603 5604 if (notflag) { 5605 if (s1 < lc || s1 > c) 5606 ok++; 5607 else 5608 return (0); 5609 } else if (lc <= s1 && s1 <= c) 5610 ok++; 5611 5612 } else if (c == '\\' && (c = *p++) == '\0') 5613 return (0); 5614 5615 lc = c; /* save left-hand 'c' for next iteration */ 5616 5617 if (notflag) { 5618 if (s1 != c) 5619 ok++; 5620 else 5621 return (0); 5622 } else if (s1 == c) 5623 ok++; 5624 5625 if ((c = *p++) == '\0') 5626 return (0); 5627 5628 } while (c != ']'); 5629 5630 if (ok) 5631 goto top; 5632 5633 return (0); 5634 } 5635 5636 case '\\': 5637 if ((c = *p++) == '\0') 5638 return (0); 5639 /*FALLTHRU*/ 5640 5641 default: 5642 if (c != s1) 5643 return (0); 5644 /*FALLTHRU*/ 5645 5646 case '?': 5647 if (s1 != '\0') 5648 goto top; 5649 return (0); 5650 5651 case '*': 5652 while (*p == '*') 5653 p++; /* consecutive *'s are identical to a single one */ 5654 5655 if (*p == '\0') 5656 return (1); 5657 5658 for (s = olds; *s != '\0'; s++) { 5659 if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0) 5660 return (gs); 5661 } 5662 5663 return (0); 5664 } 5665 } 5666 5667 /*ARGSUSED*/ 5668 static int 5669 dtrace_match_string(const char *s, const char *p, int depth) 5670 { 5671 return (s != NULL && strcmp(s, p) == 0); 5672 } 5673 5674 /*ARGSUSED*/ 5675 static int 5676 dtrace_match_nul(const char *s, const char *p, int depth) 5677 { 5678 return (1); /* always match the empty pattern */ 5679 } 5680 5681 /*ARGSUSED*/ 5682 static int 5683 dtrace_match_nonzero(const char *s, const char *p, int depth) 5684 { 5685 return (s != NULL && s[0] != '\0'); 5686 } 5687 5688 static int 5689 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid, 5690 int (*matched)(dtrace_probe_t *, void *), void *arg) 5691 { 5692 dtrace_probe_t template, *probe; 5693 dtrace_hash_t *hash = NULL; 5694 int len, best = INT_MAX, nmatched = 0; 5695 dtrace_id_t i; 5696 5697 ASSERT(MUTEX_HELD(&dtrace_lock)); 5698 5699 /* 5700 * If the probe ID is specified in the key, just lookup by ID and 5701 * invoke the match callback once if a matching probe is found. 5702 */ 5703 if (pkp->dtpk_id != DTRACE_IDNONE) { 5704 if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL && 5705 dtrace_match_probe(probe, pkp, priv, uid) > 0) { 5706 (void) (*matched)(probe, arg); 5707 nmatched++; 5708 } 5709 return (nmatched); 5710 } 5711 5712 template.dtpr_mod = (char *)pkp->dtpk_mod; 5713 template.dtpr_func = (char *)pkp->dtpk_func; 5714 template.dtpr_name = (char *)pkp->dtpk_name; 5715 5716 /* 5717 * We want to find the most distinct of the module name, function 5718 * name, and name. So for each one that is not a glob pattern or 5719 * empty string, we perform a lookup in the corresponding hash and 5720 * use the hash table with the fewest collisions to do our search. 5721 */ 5722 if (pkp->dtpk_mmatch == &dtrace_match_string && 5723 (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) { 5724 best = len; 5725 hash = dtrace_bymod; 5726 } 5727 5728 if (pkp->dtpk_fmatch == &dtrace_match_string && 5729 (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) { 5730 best = len; 5731 hash = dtrace_byfunc; 5732 } 5733 5734 if (pkp->dtpk_nmatch == &dtrace_match_string && 5735 (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) { 5736 best = len; 5737 hash = dtrace_byname; 5738 } 5739 5740 /* 5741 * If we did not select a hash table, iterate over every probe and 5742 * invoke our callback for each one that matches our input probe key. 5743 */ 5744 if (hash == NULL) { 5745 for (i = 0; i < dtrace_nprobes; i++) { 5746 if ((probe = dtrace_probes[i]) == NULL || 5747 dtrace_match_probe(probe, pkp, priv, uid) <= 0) 5748 continue; 5749 5750 nmatched++; 5751 5752 if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT) 5753 break; 5754 } 5755 5756 return (nmatched); 5757 } 5758 5759 /* 5760 * If we selected a hash table, iterate over each probe of the same key 5761 * name and invoke the callback for every probe that matches the other 5762 * attributes of our input probe key. 5763 */ 5764 for (probe = dtrace_hash_lookup(hash, &template); probe != NULL; 5765 probe = *(DTRACE_HASHNEXT(hash, probe))) { 5766 5767 if (dtrace_match_probe(probe, pkp, priv, uid) <= 0) 5768 continue; 5769 5770 nmatched++; 5771 5772 if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT) 5773 break; 5774 } 5775 5776 return (nmatched); 5777 } 5778 5779 /* 5780 * Return the function pointer dtrace_probecmp() should use to compare the 5781 * specified pattern with a string. For NULL or empty patterns, we select 5782 * dtrace_match_nul(). For glob pattern strings, we use dtrace_match_glob(). 5783 * For non-empty non-glob strings, we use dtrace_match_string(). 5784 */ 5785 static dtrace_probekey_f * 5786 dtrace_probekey_func(const char *p) 5787 { 5788 char c; 5789 5790 if (p == NULL || *p == '\0') 5791 return (&dtrace_match_nul); 5792 5793 while ((c = *p++) != '\0') { 5794 if (c == '[' || c == '?' || c == '*' || c == '\\') 5795 return (&dtrace_match_glob); 5796 } 5797 5798 return (&dtrace_match_string); 5799 } 5800 5801 /* 5802 * Build a probe comparison key for use with dtrace_match_probe() from the 5803 * given probe description. By convention, a null key only matches anchored 5804 * probes: if each field is the empty string, reset dtpk_fmatch to 5805 * dtrace_match_nonzero(). 5806 */ 5807 static void 5808 dtrace_probekey(const dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp) 5809 { 5810 pkp->dtpk_prov = pdp->dtpd_provider; 5811 pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider); 5812 5813 pkp->dtpk_mod = pdp->dtpd_mod; 5814 pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod); 5815 5816 pkp->dtpk_func = pdp->dtpd_func; 5817 pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func); 5818 5819 pkp->dtpk_name = pdp->dtpd_name; 5820 pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name); 5821 5822 pkp->dtpk_id = pdp->dtpd_id; 5823 5824 if (pkp->dtpk_id == DTRACE_IDNONE && 5825 pkp->dtpk_pmatch == &dtrace_match_nul && 5826 pkp->dtpk_mmatch == &dtrace_match_nul && 5827 pkp->dtpk_fmatch == &dtrace_match_nul && 5828 pkp->dtpk_nmatch == &dtrace_match_nul) 5829 pkp->dtpk_fmatch = &dtrace_match_nonzero; 5830 } 5831 5832 /* 5833 * DTrace Provider-to-Framework API Functions 5834 * 5835 * These functions implement much of the Provider-to-Framework API, as 5836 * described in <sys/dtrace.h>. The parts of the API not in this section are 5837 * the functions in the API for probe management (found below), and 5838 * dtrace_probe() itself (found above). 5839 */ 5840 5841 /* 5842 * Register the calling provider with the DTrace framework. This should 5843 * generally be called by DTrace providers in their attach(9E) entry point. 5844 */ 5845 int 5846 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv, 5847 uid_t uid, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp) 5848 { 5849 dtrace_provider_t *provider; 5850 5851 if (name == NULL || pap == NULL || pops == NULL || idp == NULL) { 5852 cmn_err(CE_WARN, "failed to register provider '%s': invalid " 5853 "arguments", name ? name : "<NULL>"); 5854 return (EINVAL); 5855 } 5856 5857 if (name[0] == '\0' || dtrace_badname(name)) { 5858 cmn_err(CE_WARN, "failed to register provider '%s': invalid " 5859 "provider name", name); 5860 return (EINVAL); 5861 } 5862 5863 if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) || 5864 pops->dtps_enable == NULL || pops->dtps_disable == NULL || 5865 pops->dtps_destroy == NULL || 5866 ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) { 5867 cmn_err(CE_WARN, "failed to register provider '%s': invalid " 5868 "provider ops", name); 5869 return (EINVAL); 5870 } 5871 5872 if (dtrace_badattr(&pap->dtpa_provider) || 5873 dtrace_badattr(&pap->dtpa_mod) || 5874 dtrace_badattr(&pap->dtpa_func) || 5875 dtrace_badattr(&pap->dtpa_name) || 5876 dtrace_badattr(&pap->dtpa_args)) { 5877 cmn_err(CE_WARN, "failed to register provider '%s': invalid " 5878 "provider attributes", name); 5879 return (EINVAL); 5880 } 5881 5882 if (priv & ~DTRACE_PRIV_ALL) { 5883 cmn_err(CE_WARN, "failed to register provider '%s': invalid " 5884 "privilege attributes", name); 5885 return (EINVAL); 5886 } 5887 5888 if ((priv & DTRACE_PRIV_KERNEL) && 5889 (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) && 5890 pops->dtps_usermode == NULL) { 5891 cmn_err(CE_WARN, "failed to register provider '%s': need " 5892 "dtps_usermode() op for given privilege attributes", name); 5893 return (EINVAL); 5894 } 5895 5896 provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP); 5897 provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP); 5898 (void) strcpy(provider->dtpv_name, name); 5899 5900 provider->dtpv_attr = *pap; 5901 provider->dtpv_priv.dtpp_flags = priv; 5902 provider->dtpv_priv.dtpp_uid = uid; 5903 provider->dtpv_pops = *pops; 5904 5905 if (pops->dtps_provide == NULL) { 5906 ASSERT(pops->dtps_provide_module != NULL); 5907 provider->dtpv_pops.dtps_provide = 5908 (void (*)(void *, const dtrace_probedesc_t *))dtrace_nullop; 5909 } 5910 5911 if (pops->dtps_provide_module == NULL) { 5912 ASSERT(pops->dtps_provide != NULL); 5913 provider->dtpv_pops.dtps_provide_module = 5914 (void (*)(void *, struct modctl *))dtrace_nullop; 5915 } 5916 5917 if (pops->dtps_suspend == NULL) { 5918 ASSERT(pops->dtps_resume == NULL); 5919 provider->dtpv_pops.dtps_suspend = 5920 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop; 5921 provider->dtpv_pops.dtps_resume = 5922 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop; 5923 } 5924 5925 provider->dtpv_arg = arg; 5926 *idp = (dtrace_provider_id_t)provider; 5927 5928 if (pops == &dtrace_provider_ops) { 5929 ASSERT(MUTEX_HELD(&dtrace_provider_lock)); 5930 ASSERT(MUTEX_HELD(&dtrace_lock)); 5931 ASSERT(dtrace_anon.dta_enabling == NULL); 5932 5933 /* 5934 * We make sure that the DTrace provider is at the head of 5935 * the provider chain. 5936 */ 5937 provider->dtpv_next = dtrace_provider; 5938 dtrace_provider = provider; 5939 return (0); 5940 } 5941 5942 mutex_enter(&dtrace_provider_lock); 5943 mutex_enter(&dtrace_lock); 5944 5945 /* 5946 * If there is at least one provider registered, we'll add this 5947 * provider after the first provider. 5948 */ 5949 if (dtrace_provider != NULL) { 5950 provider->dtpv_next = dtrace_provider->dtpv_next; 5951 dtrace_provider->dtpv_next = provider; 5952 } else { 5953 dtrace_provider = provider; 5954 } 5955 5956 if (dtrace_retained != NULL) { 5957 dtrace_enabling_provide(provider); 5958 5959 /* 5960 * Now we need to call dtrace_enabling_matchall() -- which 5961 * will acquire cpu_lock and dtrace_lock. We therefore need 5962 * to drop all of our locks before calling into it... 5963 */ 5964 mutex_exit(&dtrace_lock); 5965 mutex_exit(&dtrace_provider_lock); 5966 dtrace_enabling_matchall(); 5967 5968 return (0); 5969 } 5970 5971 mutex_exit(&dtrace_lock); 5972 mutex_exit(&dtrace_provider_lock); 5973 5974 return (0); 5975 } 5976 5977 /* 5978 * Unregister the specified provider from the DTrace framework. This should 5979 * generally be called by DTrace providers in their detach(9E) entry point. 5980 */ 5981 int 5982 dtrace_unregister(dtrace_provider_id_t id) 5983 { 5984 dtrace_provider_t *old = (dtrace_provider_t *)id; 5985 dtrace_provider_t *prev = NULL; 5986 int i, self = 0; 5987 dtrace_probe_t *probe, *first = NULL; 5988 5989 if (old->dtpv_pops.dtps_enable == 5990 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop) { 5991 /* 5992 * If DTrace itself is the provider, we're called with locks 5993 * already held. 5994 */ 5995 ASSERT(old == dtrace_provider); 5996 ASSERT(dtrace_devi != NULL); 5997 ASSERT(MUTEX_HELD(&dtrace_provider_lock)); 5998 ASSERT(MUTEX_HELD(&dtrace_lock)); 5999 self = 1; 6000 6001 if (dtrace_provider->dtpv_next != NULL) { 6002 /* 6003 * There's another provider here; return failure. 6004 */ 6005 return (EBUSY); 6006 } 6007 } else { 6008 mutex_enter(&dtrace_provider_lock); 6009 mutex_enter(&mod_lock); 6010 mutex_enter(&dtrace_lock); 6011 } 6012 6013 /* 6014 * If anyone has /dev/dtrace open, or if there are anonymous enabled 6015 * probes, we refuse to let providers slither away, unless this 6016 * provider has already been explicitly invalidated. 6017 */ 6018 if (!old->dtpv_defunct && 6019 (dtrace_opens || (dtrace_anon.dta_state != NULL && 6020 dtrace_anon.dta_state->dts_necbs > 0))) { 6021 if (!self) { 6022 mutex_exit(&dtrace_lock); 6023 mutex_exit(&mod_lock); 6024 mutex_exit(&dtrace_provider_lock); 6025 } 6026 return (EBUSY); 6027 } 6028 6029 /* 6030 * Attempt to destroy the probes associated with this provider. 6031 */ 6032 for (i = 0; i < dtrace_nprobes; i++) { 6033 if ((probe = dtrace_probes[i]) == NULL) 6034 continue; 6035 6036 if (probe->dtpr_provider != old) 6037 continue; 6038 6039 if (probe->dtpr_ecb == NULL) 6040 continue; 6041 6042 /* 6043 * We have at least one ECB; we can't remove this provider. 6044 */ 6045 if (!self) { 6046 mutex_exit(&dtrace_lock); 6047 mutex_exit(&mod_lock); 6048 mutex_exit(&dtrace_provider_lock); 6049 } 6050 return (EBUSY); 6051 } 6052 6053 /* 6054 * All of the probes for this provider are disabled; we can safely 6055 * remove all of them from their hash chains and from the probe array. 6056 */ 6057 for (i = 0; i < dtrace_nprobes; i++) { 6058 if ((probe = dtrace_probes[i]) == NULL) 6059 continue; 6060 6061 if (probe->dtpr_provider != old) 6062 continue; 6063 6064 dtrace_probes[i] = NULL; 6065 6066 dtrace_hash_remove(dtrace_bymod, probe); 6067 dtrace_hash_remove(dtrace_byfunc, probe); 6068 dtrace_hash_remove(dtrace_byname, probe); 6069 6070 if (first == NULL) { 6071 first = probe; 6072 probe->dtpr_nextmod = NULL; 6073 } else { 6074 probe->dtpr_nextmod = first; 6075 first = probe; 6076 } 6077 } 6078 6079 /* 6080 * The provider's probes have been removed from the hash chains and 6081 * from the probe array. Now issue a dtrace_sync() to be sure that 6082 * everyone has cleared out from any probe array processing. 6083 */ 6084 dtrace_sync(); 6085 6086 for (probe = first; probe != NULL; probe = first) { 6087 first = probe->dtpr_nextmod; 6088 6089 old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id, 6090 probe->dtpr_arg); 6091 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1); 6092 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1); 6093 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1); 6094 vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1); 6095 kmem_free(probe, sizeof (dtrace_probe_t)); 6096 } 6097 6098 if ((prev = dtrace_provider) == old) { 6099 ASSERT(self || dtrace_devi == NULL); 6100 ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL); 6101 dtrace_provider = old->dtpv_next; 6102 } else { 6103 while (prev != NULL && prev->dtpv_next != old) 6104 prev = prev->dtpv_next; 6105 6106 if (prev == NULL) { 6107 panic("attempt to unregister non-existent " 6108 "dtrace provider %p\n", (void *)id); 6109 } 6110 6111 prev->dtpv_next = old->dtpv_next; 6112 } 6113 6114 if (!self) { 6115 mutex_exit(&dtrace_lock); 6116 mutex_exit(&mod_lock); 6117 mutex_exit(&dtrace_provider_lock); 6118 } 6119 6120 kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1); 6121 kmem_free(old, sizeof (dtrace_provider_t)); 6122 6123 return (0); 6124 } 6125 6126 /* 6127 * Invalidate the specified provider. All subsequent probe lookups for the 6128 * specified provider will fail, but its probes will not be removed. 6129 */ 6130 void 6131 dtrace_invalidate(dtrace_provider_id_t id) 6132 { 6133 dtrace_provider_t *pvp = (dtrace_provider_t *)id; 6134 6135 ASSERT(pvp->dtpv_pops.dtps_enable != 6136 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop); 6137 6138 mutex_enter(&dtrace_provider_lock); 6139 mutex_enter(&dtrace_lock); 6140 6141 pvp->dtpv_defunct = 1; 6142 6143 mutex_exit(&dtrace_lock); 6144 mutex_exit(&dtrace_provider_lock); 6145 } 6146 6147 /* 6148 * Indicate whether or not DTrace has attached. 6149 */ 6150 int 6151 dtrace_attached(void) 6152 { 6153 /* 6154 * dtrace_provider will be non-NULL iff the DTrace driver has 6155 * attached. (It's non-NULL because DTrace is always itself a 6156 * provider.) 6157 */ 6158 return (dtrace_provider != NULL); 6159 } 6160 6161 /* 6162 * Remove all the unenabled probes for the given provider. This function is 6163 * not unlike dtrace_unregister(), except that it doesn't remove the provider 6164 * -- just as many of its associated probes as it can. 6165 */ 6166 int 6167 dtrace_condense(dtrace_provider_id_t id) 6168 { 6169 dtrace_provider_t *prov = (dtrace_provider_t *)id; 6170 int i; 6171 dtrace_probe_t *probe; 6172 6173 /* 6174 * Make sure this isn't the dtrace provider itself. 6175 */ 6176 ASSERT(prov->dtpv_pops.dtps_enable != 6177 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop); 6178 6179 mutex_enter(&dtrace_provider_lock); 6180 mutex_enter(&dtrace_lock); 6181 6182 /* 6183 * Attempt to destroy the probes associated with this provider. 6184 */ 6185 for (i = 0; i < dtrace_nprobes; i++) { 6186 if ((probe = dtrace_probes[i]) == NULL) 6187 continue; 6188 6189 if (probe->dtpr_provider != prov) 6190 continue; 6191 6192 if (probe->dtpr_ecb != NULL) 6193 continue; 6194 6195 dtrace_probes[i] = NULL; 6196 6197 dtrace_hash_remove(dtrace_bymod, probe); 6198 dtrace_hash_remove(dtrace_byfunc, probe); 6199 dtrace_hash_remove(dtrace_byname, probe); 6200 6201 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1, 6202 probe->dtpr_arg); 6203 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1); 6204 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1); 6205 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1); 6206 kmem_free(probe, sizeof (dtrace_probe_t)); 6207 vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1); 6208 } 6209 6210 mutex_exit(&dtrace_lock); 6211 mutex_exit(&dtrace_provider_lock); 6212 6213 return (0); 6214 } 6215 6216 /* 6217 * DTrace Probe Management Functions 6218 * 6219 * The functions in this section perform the DTrace probe management, 6220 * including functions to create probes, look-up probes, and call into the 6221 * providers to request that probes be provided. Some of these functions are 6222 * in the Provider-to-Framework API; these functions can be identified by the 6223 * fact that they are not declared "static". 6224 */ 6225 6226 /* 6227 * Create a probe with the specified module name, function name, and name. 6228 */ 6229 dtrace_id_t 6230 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod, 6231 const char *func, const char *name, int aframes, void *arg) 6232 { 6233 dtrace_probe_t *probe, **probes; 6234 dtrace_provider_t *provider = (dtrace_provider_t *)prov; 6235 dtrace_id_t id; 6236 6237 if (provider == dtrace_provider) { 6238 ASSERT(MUTEX_HELD(&dtrace_lock)); 6239 } else { 6240 mutex_enter(&dtrace_lock); 6241 } 6242 6243 id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1, 6244 VM_BESTFIT | VM_SLEEP); 6245 probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP); 6246 6247 probe->dtpr_id = id; 6248 probe->dtpr_gen = dtrace_probegen++; 6249 probe->dtpr_mod = dtrace_strdup(mod); 6250 probe->dtpr_func = dtrace_strdup(func); 6251 probe->dtpr_name = dtrace_strdup(name); 6252 probe->dtpr_arg = arg; 6253 probe->dtpr_aframes = aframes; 6254 probe->dtpr_provider = provider; 6255 6256 dtrace_hash_add(dtrace_bymod, probe); 6257 dtrace_hash_add(dtrace_byfunc, probe); 6258 dtrace_hash_add(dtrace_byname, probe); 6259 6260 if (id - 1 >= dtrace_nprobes) { 6261 size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *); 6262 size_t nsize = osize << 1; 6263 6264 if (nsize == 0) { 6265 ASSERT(osize == 0); 6266 ASSERT(dtrace_probes == NULL); 6267 nsize = sizeof (dtrace_probe_t *); 6268 } 6269 6270 probes = kmem_zalloc(nsize, KM_SLEEP); 6271 6272 if (dtrace_probes == NULL) { 6273 ASSERT(osize == 0); 6274 dtrace_probes = probes; 6275 dtrace_nprobes = 1; 6276 } else { 6277 dtrace_probe_t **oprobes = dtrace_probes; 6278 6279 bcopy(oprobes, probes, osize); 6280 dtrace_membar_producer(); 6281 dtrace_probes = probes; 6282 6283 dtrace_sync(); 6284 6285 /* 6286 * All CPUs are now seeing the new probes array; we can 6287 * safely free the old array. 6288 */ 6289 kmem_free(oprobes, osize); 6290 dtrace_nprobes <<= 1; 6291 } 6292 6293 ASSERT(id - 1 < dtrace_nprobes); 6294 } 6295 6296 ASSERT(dtrace_probes[id - 1] == NULL); 6297 dtrace_probes[id - 1] = probe; 6298 6299 if (provider != dtrace_provider) 6300 mutex_exit(&dtrace_lock); 6301 6302 return (id); 6303 } 6304 6305 static dtrace_probe_t * 6306 dtrace_probe_lookup_id(dtrace_id_t id) 6307 { 6308 ASSERT(MUTEX_HELD(&dtrace_lock)); 6309 6310 if (id == 0 || id > dtrace_nprobes) 6311 return (NULL); 6312 6313 return (dtrace_probes[id - 1]); 6314 } 6315 6316 static int 6317 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg) 6318 { 6319 *((dtrace_id_t *)arg) = probe->dtpr_id; 6320 6321 return (DTRACE_MATCH_DONE); 6322 } 6323 6324 /* 6325 * Look up a probe based on provider and one or more of module name, function 6326 * name and probe name. 6327 */ 6328 dtrace_id_t 6329 dtrace_probe_lookup(dtrace_provider_id_t prid, const char *mod, 6330 const char *func, const char *name) 6331 { 6332 dtrace_probekey_t pkey; 6333 dtrace_id_t id; 6334 int match; 6335 6336 pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name; 6337 pkey.dtpk_pmatch = &dtrace_match_string; 6338 pkey.dtpk_mod = mod; 6339 pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul; 6340 pkey.dtpk_func = func; 6341 pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul; 6342 pkey.dtpk_name = name; 6343 pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul; 6344 pkey.dtpk_id = DTRACE_IDNONE; 6345 6346 mutex_enter(&dtrace_lock); 6347 match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 6348 dtrace_probe_lookup_match, &id); 6349 mutex_exit(&dtrace_lock); 6350 6351 ASSERT(match == 1 || match == 0); 6352 return (match ? id : 0); 6353 } 6354 6355 /* 6356 * Returns the probe argument associated with the specified probe. 6357 */ 6358 void * 6359 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid) 6360 { 6361 dtrace_probe_t *probe; 6362 void *rval = NULL; 6363 6364 mutex_enter(&dtrace_lock); 6365 6366 if ((probe = dtrace_probe_lookup_id(pid)) != NULL && 6367 probe->dtpr_provider == (dtrace_provider_t *)id) 6368 rval = probe->dtpr_arg; 6369 6370 mutex_exit(&dtrace_lock); 6371 6372 return (rval); 6373 } 6374 6375 /* 6376 * Copy a probe into a probe description. 6377 */ 6378 static void 6379 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp) 6380 { 6381 bzero(pdp, sizeof (dtrace_probedesc_t)); 6382 pdp->dtpd_id = prp->dtpr_id; 6383 6384 (void) strncpy(pdp->dtpd_provider, 6385 prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1); 6386 6387 (void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1); 6388 (void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1); 6389 (void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1); 6390 } 6391 6392 /* 6393 * Called to indicate that a probe -- or probes -- should be provided by a 6394 * specfied provider. If the specified description is NULL, the provider will 6395 * be told to provide all of its probes. (This is done whenever a new 6396 * consumer comes along, or whenever a retained enabling is to be matched.) If 6397 * the specified description is non-NULL, the provider is given the 6398 * opportunity to dynamically provide the specified probe, allowing providers 6399 * to support the creation of probes on-the-fly. (So-called _autocreated_ 6400 * probes.) If the provider is NULL, the operations will be applied to all 6401 * providers; if the provider is non-NULL the operations will only be applied 6402 * to the specified provider. The dtrace_provider_lock must be held, and the 6403 * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation 6404 * will need to grab the dtrace_lock when it reenters the framework through 6405 * dtrace_probe_lookup(), dtrace_probe_create(), etc. 6406 */ 6407 static void 6408 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv) 6409 { 6410 struct modctl *ctl; 6411 int all = 0; 6412 6413 ASSERT(MUTEX_HELD(&dtrace_provider_lock)); 6414 6415 if (prv == NULL) { 6416 all = 1; 6417 prv = dtrace_provider; 6418 } 6419 6420 do { 6421 /* 6422 * First, call the blanket provide operation. 6423 */ 6424 prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc); 6425 6426 /* 6427 * Now call the per-module provide operation. We will grab 6428 * mod_lock to prevent the list from being modified. Note 6429 * that this also prevents the mod_busy bits from changing. 6430 * (mod_busy can only be changed with mod_lock held.) 6431 */ 6432 mutex_enter(&mod_lock); 6433 6434 ctl = &modules; 6435 do { 6436 if (ctl->mod_busy || ctl->mod_mp == NULL) 6437 continue; 6438 6439 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl); 6440 6441 } while ((ctl = ctl->mod_next) != &modules); 6442 6443 mutex_exit(&mod_lock); 6444 } while (all && (prv = prv->dtpv_next) != NULL); 6445 } 6446 6447 /* 6448 * Iterate over each probe, and call the Framework-to-Provider API function 6449 * denoted by offs. 6450 */ 6451 static void 6452 dtrace_probe_foreach(uintptr_t offs) 6453 { 6454 dtrace_provider_t *prov; 6455 void (*func)(void *, dtrace_id_t, void *); 6456 dtrace_probe_t *probe; 6457 dtrace_icookie_t cookie; 6458 int i; 6459 6460 /* 6461 * We disable interrupts to walk through the probe array. This is 6462 * safe -- the dtrace_sync() in dtrace_unregister() assures that we 6463 * won't see stale data. 6464 */ 6465 cookie = dtrace_interrupt_disable(); 6466 6467 for (i = 0; i < dtrace_nprobes; i++) { 6468 if ((probe = dtrace_probes[i]) == NULL) 6469 continue; 6470 6471 if (probe->dtpr_ecb == NULL) { 6472 /* 6473 * This probe isn't enabled -- don't call the function. 6474 */ 6475 continue; 6476 } 6477 6478 prov = probe->dtpr_provider; 6479 func = *((void(**)(void *, dtrace_id_t, void *)) 6480 ((uintptr_t)&prov->dtpv_pops + offs)); 6481 6482 func(prov->dtpv_arg, i + 1, probe->dtpr_arg); 6483 } 6484 6485 dtrace_interrupt_enable(cookie); 6486 } 6487 6488 static int 6489 dtrace_probe_enable(const dtrace_probedesc_t *desc, dtrace_enabling_t *enab) 6490 { 6491 dtrace_probekey_t pkey; 6492 uint32_t priv; 6493 uid_t uid; 6494 6495 ASSERT(MUTEX_HELD(&dtrace_lock)); 6496 dtrace_ecb_create_cache = NULL; 6497 6498 if (desc == NULL) { 6499 /* 6500 * If we're passed a NULL description, we're being asked to 6501 * create an ECB with a NULL probe. 6502 */ 6503 (void) dtrace_ecb_create_enable(NULL, enab); 6504 return (0); 6505 } 6506 6507 dtrace_probekey(desc, &pkey); 6508 dtrace_cred2priv(CRED(), &priv, &uid); 6509 6510 return (dtrace_match(&pkey, priv, uid, dtrace_ecb_create_enable, enab)); 6511 } 6512 6513 /* 6514 * DTrace Helper Provider Functions 6515 */ 6516 static void 6517 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr) 6518 { 6519 attr->dtat_name = DOF_ATTR_NAME(dofattr); 6520 attr->dtat_data = DOF_ATTR_DATA(dofattr); 6521 attr->dtat_class = DOF_ATTR_CLASS(dofattr); 6522 } 6523 6524 static void 6525 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov, 6526 const dof_provider_t *dofprov, char *strtab) 6527 { 6528 hprov->dthpv_provname = strtab + dofprov->dofpv_name; 6529 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider, 6530 dofprov->dofpv_provattr); 6531 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod, 6532 dofprov->dofpv_modattr); 6533 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func, 6534 dofprov->dofpv_funcattr); 6535 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name, 6536 dofprov->dofpv_nameattr); 6537 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args, 6538 dofprov->dofpv_argsattr); 6539 } 6540 6541 static void 6542 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid) 6543 { 6544 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof; 6545 dof_hdr_t *dof = (dof_hdr_t *)daddr; 6546 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec; 6547 dof_provider_t *provider; 6548 dof_probe_t *probe; 6549 uint32_t *off; 6550 uint8_t *arg; 6551 char *strtab; 6552 uint_t i, nprobes; 6553 dtrace_helper_provdesc_t dhpv; 6554 dtrace_helper_probedesc_t dhpb; 6555 dtrace_meta_t *meta = dtrace_meta_pid; 6556 dtrace_mops_t *mops = &meta->dtm_mops; 6557 void *parg; 6558 6559 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset); 6560 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff + 6561 provider->dofpv_strtab * dof->dofh_secsize); 6562 prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff + 6563 provider->dofpv_probes * dof->dofh_secsize); 6564 arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff + 6565 provider->dofpv_prargs * dof->dofh_secsize); 6566 off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff + 6567 provider->dofpv_proffs * dof->dofh_secsize); 6568 6569 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset); 6570 off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset); 6571 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset); 6572 6573 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize; 6574 6575 /* 6576 * Create the provider. 6577 */ 6578 dtrace_dofprov2hprov(&dhpv, provider, strtab); 6579 6580 if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL) 6581 return; 6582 6583 meta->dtm_count++; 6584 6585 /* 6586 * Create the probes. 6587 */ 6588 for (i = 0; i < nprobes; i++) { 6589 probe = (dof_probe_t *)(uintptr_t)(daddr + 6590 prb_sec->dofs_offset + i * prb_sec->dofs_entsize); 6591 6592 dhpb.dthpb_mod = dhp->dofhp_mod; 6593 dhpb.dthpb_func = strtab + probe->dofpr_func; 6594 dhpb.dthpb_name = strtab + probe->dofpr_name; 6595 dhpb.dthpb_base = probe->dofpr_addr; 6596 dhpb.dthpb_offs = off + probe->dofpr_offidx; 6597 dhpb.dthpb_noffs = probe->dofpr_noffs; 6598 dhpb.dthpb_args = arg + probe->dofpr_argidx; 6599 dhpb.dthpb_nargc = probe->dofpr_nargc; 6600 dhpb.dthpb_xargc = probe->dofpr_xargc; 6601 dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv; 6602 dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv; 6603 6604 mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb); 6605 } 6606 } 6607 6608 static void 6609 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid) 6610 { 6611 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof; 6612 dof_hdr_t *dof = (dof_hdr_t *)daddr; 6613 int i; 6614 6615 ASSERT(MUTEX_HELD(&dtrace_meta_lock)); 6616 6617 for (i = 0; i < dof->dofh_secnum; i++) { 6618 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr + 6619 dof->dofh_secoff + i * dof->dofh_secsize); 6620 6621 if (sec->dofs_type != DOF_SECT_PROVIDER) 6622 continue; 6623 6624 dtrace_helper_provide_one(dhp, sec, pid); 6625 } 6626 6627 /* 6628 * We may have just created probes, so we must now rematch against 6629 * any retained enablings. Note that this call will acquire both 6630 * cpu_lock and dtrace_lock; the fact that we are holding 6631 * dtrace_meta_lock now is what defines the ordering with respect to 6632 * these three locks. 6633 */ 6634 dtrace_enabling_matchall(); 6635 } 6636 6637 static void 6638 dtrace_helper_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid) 6639 { 6640 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof; 6641 dof_hdr_t *dof = (dof_hdr_t *)daddr; 6642 dof_sec_t *str_sec; 6643 dof_provider_t *provider; 6644 char *strtab; 6645 dtrace_helper_provdesc_t dhpv; 6646 dtrace_meta_t *meta = dtrace_meta_pid; 6647 dtrace_mops_t *mops = &meta->dtm_mops; 6648 6649 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset); 6650 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff + 6651 provider->dofpv_strtab * dof->dofh_secsize); 6652 6653 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset); 6654 6655 /* 6656 * Create the provider. 6657 */ 6658 dtrace_dofprov2hprov(&dhpv, provider, strtab); 6659 6660 mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid); 6661 6662 meta->dtm_count--; 6663 } 6664 6665 static void 6666 dtrace_helper_remove(dof_helper_t *dhp, pid_t pid) 6667 { 6668 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof; 6669 dof_hdr_t *dof = (dof_hdr_t *)daddr; 6670 int i; 6671 6672 ASSERT(MUTEX_HELD(&dtrace_meta_lock)); 6673 6674 for (i = 0; i < dof->dofh_secnum; i++) { 6675 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr + 6676 dof->dofh_secoff + i * dof->dofh_secsize); 6677 6678 if (sec->dofs_type != DOF_SECT_PROVIDER) 6679 continue; 6680 6681 dtrace_helper_remove_one(dhp, sec, pid); 6682 } 6683 } 6684 6685 /* 6686 * DTrace Meta Provider-to-Framework API Functions 6687 * 6688 * These functions implement the Meta Provider-to-Framework API, as described 6689 * in <sys/dtrace.h>. 6690 */ 6691 int 6692 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg, 6693 dtrace_meta_provider_id_t *idp) 6694 { 6695 dtrace_meta_t *meta; 6696 dtrace_helpers_t *help, *next; 6697 int i; 6698 6699 *idp = DTRACE_METAPROVNONE; 6700 6701 /* 6702 * We strictly don't need the name, but we hold onto it for 6703 * debuggability. All hail error queues! 6704 */ 6705 if (name == NULL) { 6706 cmn_err(CE_WARN, "failed to register meta-provider: " 6707 "invalid name"); 6708 return (EINVAL); 6709 } 6710 6711 if (mops == NULL || 6712 mops->dtms_create_probe == NULL || 6713 mops->dtms_provide_pid == NULL || 6714 mops->dtms_remove_pid == NULL) { 6715 cmn_err(CE_WARN, "failed to register meta-register %s: " 6716 "invalid ops", name); 6717 return (EINVAL); 6718 } 6719 6720 meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP); 6721 meta->dtm_mops = *mops; 6722 meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP); 6723 (void) strcpy(meta->dtm_name, name); 6724 meta->dtm_arg = arg; 6725 6726 mutex_enter(&dtrace_meta_lock); 6727 mutex_enter(&dtrace_lock); 6728 6729 if (dtrace_meta_pid != NULL) { 6730 mutex_exit(&dtrace_lock); 6731 mutex_exit(&dtrace_meta_lock); 6732 cmn_err(CE_WARN, "failed to register meta-register %s: " 6733 "user-land meta-provider exists", name); 6734 kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1); 6735 kmem_free(meta, sizeof (dtrace_meta_t)); 6736 return (EINVAL); 6737 } 6738 6739 dtrace_meta_pid = meta; 6740 *idp = (dtrace_meta_provider_id_t)meta; 6741 6742 /* 6743 * If there are providers and probes ready to go, pass them 6744 * off to the new meta provider now. 6745 */ 6746 6747 help = dtrace_deferred_pid; 6748 dtrace_deferred_pid = NULL; 6749 6750 mutex_exit(&dtrace_lock); 6751 6752 while (help != NULL) { 6753 for (i = 0; i < help->dthps_nprovs; i++) { 6754 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov, 6755 help->dthps_pid); 6756 } 6757 6758 next = help->dthps_next; 6759 help->dthps_next = NULL; 6760 help->dthps_prev = NULL; 6761 help->dthps_deferred = 0; 6762 help = next; 6763 } 6764 6765 mutex_exit(&dtrace_meta_lock); 6766 6767 return (0); 6768 } 6769 6770 int 6771 dtrace_meta_unregister(dtrace_meta_provider_id_t id) 6772 { 6773 dtrace_meta_t **pp, *old = (dtrace_meta_t *)id; 6774 6775 mutex_enter(&dtrace_meta_lock); 6776 mutex_enter(&dtrace_lock); 6777 6778 if (old == dtrace_meta_pid) { 6779 pp = &dtrace_meta_pid; 6780 } else { 6781 panic("attempt to unregister non-existent " 6782 "dtrace meta-provider %p\n", (void *)old); 6783 } 6784 6785 if (old->dtm_count != 0) { 6786 mutex_exit(&dtrace_lock); 6787 mutex_exit(&dtrace_meta_lock); 6788 return (EBUSY); 6789 } 6790 6791 *pp = NULL; 6792 6793 mutex_exit(&dtrace_lock); 6794 mutex_exit(&dtrace_meta_lock); 6795 6796 kmem_free(old->dtm_name, strlen(old->dtm_name) + 1); 6797 kmem_free(old, sizeof (dtrace_meta_t)); 6798 6799 return (0); 6800 } 6801 6802 6803 /* 6804 * DTrace DIF Object Functions 6805 */ 6806 static int 6807 dtrace_difo_err(uint_t pc, const char *format, ...) 6808 { 6809 if (dtrace_err_verbose) { 6810 va_list alist; 6811 6812 (void) uprintf("dtrace DIF object error: [%u]: ", pc); 6813 va_start(alist, format); 6814 (void) vuprintf(format, alist); 6815 va_end(alist); 6816 } 6817 6818 #ifdef DTRACE_ERRDEBUG 6819 dtrace_errdebug(format); 6820 #endif 6821 return (1); 6822 } 6823 6824 /* 6825 * Validate a DTrace DIF object by checking the IR instructions. The following 6826 * rules are currently enforced by dtrace_difo_validate(): 6827 * 6828 * 1. Each instruction must have a valid opcode 6829 * 2. Each register, string, variable, or subroutine reference must be valid 6830 * 3. No instruction can modify register %r0 (must be zero) 6831 * 4. All instruction reserved bits must be set to zero 6832 * 5. The last instruction must be a "ret" instruction 6833 * 6. All branch targets must reference a valid instruction _after_ the branch 6834 */ 6835 static int 6836 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs, 6837 cred_t *cr) 6838 { 6839 int err = 0, i; 6840 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err; 6841 int kcheck; 6842 uint_t pc; 6843 6844 kcheck = cr == NULL || 6845 PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE) == 0; 6846 6847 dp->dtdo_destructive = 0; 6848 6849 for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) { 6850 dif_instr_t instr = dp->dtdo_buf[pc]; 6851 6852 uint_t r1 = DIF_INSTR_R1(instr); 6853 uint_t r2 = DIF_INSTR_R2(instr); 6854 uint_t rd = DIF_INSTR_RD(instr); 6855 uint_t rs = DIF_INSTR_RS(instr); 6856 uint_t label = DIF_INSTR_LABEL(instr); 6857 uint_t v = DIF_INSTR_VAR(instr); 6858 uint_t subr = DIF_INSTR_SUBR(instr); 6859 uint_t type = DIF_INSTR_TYPE(instr); 6860 uint_t op = DIF_INSTR_OP(instr); 6861 6862 switch (op) { 6863 case DIF_OP_OR: 6864 case DIF_OP_XOR: 6865 case DIF_OP_AND: 6866 case DIF_OP_SLL: 6867 case DIF_OP_SRL: 6868 case DIF_OP_SRA: 6869 case DIF_OP_SUB: 6870 case DIF_OP_ADD: 6871 case DIF_OP_MUL: 6872 case DIF_OP_SDIV: 6873 case DIF_OP_UDIV: 6874 case DIF_OP_SREM: 6875 case DIF_OP_UREM: 6876 case DIF_OP_COPYS: 6877 if (r1 >= nregs) 6878 err += efunc(pc, "invalid register %u\n", r1); 6879 if (r2 >= nregs) 6880 err += efunc(pc, "invalid register %u\n", r2); 6881 if (rd >= nregs) 6882 err += efunc(pc, "invalid register %u\n", rd); 6883 if (rd == 0) 6884 err += efunc(pc, "cannot write to %r0\n"); 6885 break; 6886 case DIF_OP_NOT: 6887 case DIF_OP_MOV: 6888 case DIF_OP_ALLOCS: 6889 if (r1 >= nregs) 6890 err += efunc(pc, "invalid register %u\n", r1); 6891 if (r2 != 0) 6892 err += efunc(pc, "non-zero reserved bits\n"); 6893 if (rd >= nregs) 6894 err += efunc(pc, "invalid register %u\n", rd); 6895 if (rd == 0) 6896 err += efunc(pc, "cannot write to %r0\n"); 6897 break; 6898 case DIF_OP_LDSB: 6899 case DIF_OP_LDSH: 6900 case DIF_OP_LDSW: 6901 case DIF_OP_LDUB: 6902 case DIF_OP_LDUH: 6903 case DIF_OP_LDUW: 6904 case DIF_OP_LDX: 6905 if (r1 >= nregs) 6906 err += efunc(pc, "invalid register %u\n", r1); 6907 if (r2 != 0) 6908 err += efunc(pc, "non-zero reserved bits\n"); 6909 if (rd >= nregs) 6910 err += efunc(pc, "invalid register %u\n", rd); 6911 if (rd == 0) 6912 err += efunc(pc, "cannot write to %r0\n"); 6913 if (kcheck) 6914 dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op + 6915 DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd); 6916 break; 6917 case DIF_OP_RLDSB: 6918 case DIF_OP_RLDSH: 6919 case DIF_OP_RLDSW: 6920 case DIF_OP_RLDUB: 6921 case DIF_OP_RLDUH: 6922 case DIF_OP_RLDUW: 6923 case DIF_OP_RLDX: 6924 if (r1 >= nregs) 6925 err += efunc(pc, "invalid register %u\n", r1); 6926 if (r2 != 0) 6927 err += efunc(pc, "non-zero reserved bits\n"); 6928 if (rd >= nregs) 6929 err += efunc(pc, "invalid register %u\n", rd); 6930 if (rd == 0) 6931 err += efunc(pc, "cannot write to %r0\n"); 6932 break; 6933 case DIF_OP_ULDSB: 6934 case DIF_OP_ULDSH: 6935 case DIF_OP_ULDSW: 6936 case DIF_OP_ULDUB: 6937 case DIF_OP_ULDUH: 6938 case DIF_OP_ULDUW: 6939 case DIF_OP_ULDX: 6940 if (r1 >= nregs) 6941 err += efunc(pc, "invalid register %u\n", r1); 6942 if (r2 != 0) 6943 err += efunc(pc, "non-zero reserved bits\n"); 6944 if (rd >= nregs) 6945 err += efunc(pc, "invalid register %u\n", rd); 6946 if (rd == 0) 6947 err += efunc(pc, "cannot write to %r0\n"); 6948 break; 6949 case DIF_OP_STB: 6950 case DIF_OP_STH: 6951 case DIF_OP_STW: 6952 case DIF_OP_STX: 6953 if (r1 >= nregs) 6954 err += efunc(pc, "invalid register %u\n", r1); 6955 if (r2 != 0) 6956 err += efunc(pc, "non-zero reserved bits\n"); 6957 if (rd >= nregs) 6958 err += efunc(pc, "invalid register %u\n", rd); 6959 if (rd == 0) 6960 err += efunc(pc, "cannot write to 0 address\n"); 6961 break; 6962 case DIF_OP_CMP: 6963 case DIF_OP_SCMP: 6964 if (r1 >= nregs) 6965 err += efunc(pc, "invalid register %u\n", r1); 6966 if (r2 >= nregs) 6967 err += efunc(pc, "invalid register %u\n", r2); 6968 if (rd != 0) 6969 err += efunc(pc, "non-zero reserved bits\n"); 6970 break; 6971 case DIF_OP_TST: 6972 if (r1 >= nregs) 6973 err += efunc(pc, "invalid register %u\n", r1); 6974 if (r2 != 0 || rd != 0) 6975 err += efunc(pc, "non-zero reserved bits\n"); 6976 break; 6977 case DIF_OP_BA: 6978 case DIF_OP_BE: 6979 case DIF_OP_BNE: 6980 case DIF_OP_BG: 6981 case DIF_OP_BGU: 6982 case DIF_OP_BGE: 6983 case DIF_OP_BGEU: 6984 case DIF_OP_BL: 6985 case DIF_OP_BLU: 6986 case DIF_OP_BLE: 6987 case DIF_OP_BLEU: 6988 if (label >= dp->dtdo_len) { 6989 err += efunc(pc, "invalid branch target %u\n", 6990 label); 6991 } 6992 if (label <= pc) { 6993 err += efunc(pc, "backward branch to %u\n", 6994 label); 6995 } 6996 break; 6997 case DIF_OP_RET: 6998 if (r1 != 0 || r2 != 0) 6999 err += efunc(pc, "non-zero reserved bits\n"); 7000 if (rd >= nregs) 7001 err += efunc(pc, "invalid register %u\n", rd); 7002 break; 7003 case DIF_OP_NOP: 7004 case DIF_OP_POPTS: 7005 case DIF_OP_FLUSHTS: 7006 if (r1 != 0 || r2 != 0 || rd != 0) 7007 err += efunc(pc, "non-zero reserved bits\n"); 7008 break; 7009 case DIF_OP_SETX: 7010 if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) { 7011 err += efunc(pc, "invalid integer ref %u\n", 7012 DIF_INSTR_INTEGER(instr)); 7013 } 7014 if (rd >= nregs) 7015 err += efunc(pc, "invalid register %u\n", rd); 7016 if (rd == 0) 7017 err += efunc(pc, "cannot write to %r0\n"); 7018 break; 7019 case DIF_OP_SETS: 7020 if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) { 7021 err += efunc(pc, "invalid string ref %u\n", 7022 DIF_INSTR_STRING(instr)); 7023 } 7024 if (rd >= nregs) 7025 err += efunc(pc, "invalid register %u\n", rd); 7026 if (rd == 0) 7027 err += efunc(pc, "cannot write to %r0\n"); 7028 break; 7029 case DIF_OP_LDGA: 7030 case DIF_OP_LDTA: 7031 if (r1 > DIF_VAR_ARRAY_MAX) 7032 err += efunc(pc, "invalid array %u\n", r1); 7033 if (r2 >= nregs) 7034 err += efunc(pc, "invalid register %u\n", r2); 7035 if (rd >= nregs) 7036 err += efunc(pc, "invalid register %u\n", rd); 7037 if (rd == 0) 7038 err += efunc(pc, "cannot write to %r0\n"); 7039 break; 7040 case DIF_OP_LDGS: 7041 case DIF_OP_LDTS: 7042 case DIF_OP_LDLS: 7043 case DIF_OP_LDGAA: 7044 case DIF_OP_LDTAA: 7045 if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX) 7046 err += efunc(pc, "invalid variable %u\n", v); 7047 if (rd >= nregs) 7048 err += efunc(pc, "invalid register %u\n", rd); 7049 if (rd == 0) 7050 err += efunc(pc, "cannot write to %r0\n"); 7051 break; 7052 case DIF_OP_STGS: 7053 case DIF_OP_STTS: 7054 case DIF_OP_STLS: 7055 case DIF_OP_STGAA: 7056 case DIF_OP_STTAA: 7057 if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX) 7058 err += efunc(pc, "invalid variable %u\n", v); 7059 if (rs >= nregs) 7060 err += efunc(pc, "invalid register %u\n", rd); 7061 break; 7062 case DIF_OP_CALL: 7063 if (subr > DIF_SUBR_MAX) 7064 err += efunc(pc, "invalid subr %u\n", subr); 7065 if (rd >= nregs) 7066 err += efunc(pc, "invalid register %u\n", rd); 7067 if (rd == 0) 7068 err += efunc(pc, "cannot write to %r0\n"); 7069 7070 if (subr == DIF_SUBR_COPYOUT || 7071 subr == DIF_SUBR_COPYOUTSTR) { 7072 dp->dtdo_destructive = 1; 7073 } 7074 break; 7075 case DIF_OP_PUSHTR: 7076 if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF) 7077 err += efunc(pc, "invalid ref type %u\n", type); 7078 if (r2 >= nregs) 7079 err += efunc(pc, "invalid register %u\n", r2); 7080 if (rs >= nregs) 7081 err += efunc(pc, "invalid register %u\n", rs); 7082 break; 7083 case DIF_OP_PUSHTV: 7084 if (type != DIF_TYPE_CTF) 7085 err += efunc(pc, "invalid val type %u\n", type); 7086 if (r2 >= nregs) 7087 err += efunc(pc, "invalid register %u\n", r2); 7088 if (rs >= nregs) 7089 err += efunc(pc, "invalid register %u\n", rs); 7090 break; 7091 default: 7092 err += efunc(pc, "invalid opcode %u\n", 7093 DIF_INSTR_OP(instr)); 7094 } 7095 } 7096 7097 if (dp->dtdo_len != 0 && 7098 DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) { 7099 err += efunc(dp->dtdo_len - 1, 7100 "expected 'ret' as last DIF instruction\n"); 7101 } 7102 7103 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) { 7104 /* 7105 * If we're not returning by reference, the size must be either 7106 * 0 or the size of one of the base types. 7107 */ 7108 switch (dp->dtdo_rtype.dtdt_size) { 7109 case 0: 7110 case sizeof (uint8_t): 7111 case sizeof (uint16_t): 7112 case sizeof (uint32_t): 7113 case sizeof (uint64_t): 7114 break; 7115 7116 default: 7117 err += efunc(dp->dtdo_len - 1, "bad return size"); 7118 } 7119 } 7120 7121 for (i = 0; i < dp->dtdo_varlen && err == 0; i++) { 7122 dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL; 7123 dtrace_diftype_t *vt, *et; 7124 uint_t id, ndx; 7125 7126 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL && 7127 v->dtdv_scope != DIFV_SCOPE_THREAD && 7128 v->dtdv_scope != DIFV_SCOPE_LOCAL) { 7129 err += efunc(i, "unrecognized variable scope %d\n", 7130 v->dtdv_scope); 7131 break; 7132 } 7133 7134 if (v->dtdv_kind != DIFV_KIND_ARRAY && 7135 v->dtdv_kind != DIFV_KIND_SCALAR) { 7136 err += efunc(i, "unrecognized variable type %d\n", 7137 v->dtdv_kind); 7138 break; 7139 } 7140 7141 if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) { 7142 err += efunc(i, "%d exceeds variable id limit\n", id); 7143 break; 7144 } 7145 7146 if (id < DIF_VAR_OTHER_UBASE) 7147 continue; 7148 7149 /* 7150 * For user-defined variables, we need to check that this 7151 * definition is identical to any previous definition that we 7152 * encountered. 7153 */ 7154 ndx = id - DIF_VAR_OTHER_UBASE; 7155 7156 switch (v->dtdv_scope) { 7157 case DIFV_SCOPE_GLOBAL: 7158 if (ndx < vstate->dtvs_nglobals) { 7159 dtrace_statvar_t *svar; 7160 7161 if ((svar = vstate->dtvs_globals[ndx]) != NULL) 7162 existing = &svar->dtsv_var; 7163 } 7164 7165 break; 7166 7167 case DIFV_SCOPE_THREAD: 7168 if (ndx < vstate->dtvs_ntlocals) 7169 existing = &vstate->dtvs_tlocals[ndx]; 7170 break; 7171 7172 case DIFV_SCOPE_LOCAL: 7173 if (ndx < vstate->dtvs_nlocals) { 7174 dtrace_statvar_t *svar; 7175 7176 if ((svar = vstate->dtvs_locals[ndx]) != NULL) 7177 existing = &svar->dtsv_var; 7178 } 7179 7180 break; 7181 } 7182 7183 vt = &v->dtdv_type; 7184 7185 if (vt->dtdt_flags & DIF_TF_BYREF) { 7186 if (vt->dtdt_size == 0) { 7187 err += efunc(i, "zero-sized variable\n"); 7188 break; 7189 } 7190 7191 if (v->dtdv_scope == DIFV_SCOPE_GLOBAL && 7192 vt->dtdt_size > dtrace_global_maxsize) { 7193 err += efunc(i, "oversized by-ref global\n"); 7194 break; 7195 } 7196 } 7197 7198 if (existing == NULL || existing->dtdv_id == 0) 7199 continue; 7200 7201 ASSERT(existing->dtdv_id == v->dtdv_id); 7202 ASSERT(existing->dtdv_scope == v->dtdv_scope); 7203 7204 if (existing->dtdv_kind != v->dtdv_kind) 7205 err += efunc(i, "%d changed variable kind\n", id); 7206 7207 et = &existing->dtdv_type; 7208 7209 if (vt->dtdt_flags != et->dtdt_flags) { 7210 err += efunc(i, "%d changed variable type flags\n", id); 7211 break; 7212 } 7213 7214 if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) { 7215 err += efunc(i, "%d changed variable type size\n", id); 7216 break; 7217 } 7218 } 7219 7220 return (err); 7221 } 7222 7223 /* 7224 * Validate a DTrace DIF object that it is to be used as a helper. Helpers 7225 * are much more constrained than normal DIFOs. Specifically, they may 7226 * not: 7227 * 7228 * 1. Make calls to subroutines other than copyin(), copyinstr() or 7229 * miscellaneous string routines 7230 * 2. Access DTrace variables other than the args[] array, and the 7231 * curthread, pid, tid and execname variables. 7232 * 3. Have thread-local variables. 7233 * 4. Have dynamic variables. 7234 */ 7235 static int 7236 dtrace_difo_validate_helper(dtrace_difo_t *dp) 7237 { 7238 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err; 7239 int err = 0; 7240 uint_t pc; 7241 7242 for (pc = 0; pc < dp->dtdo_len; pc++) { 7243 dif_instr_t instr = dp->dtdo_buf[pc]; 7244 7245 uint_t v = DIF_INSTR_VAR(instr); 7246 uint_t subr = DIF_INSTR_SUBR(instr); 7247 uint_t op = DIF_INSTR_OP(instr); 7248 7249 switch (op) { 7250 case DIF_OP_OR: 7251 case DIF_OP_XOR: 7252 case DIF_OP_AND: 7253 case DIF_OP_SLL: 7254 case DIF_OP_SRL: 7255 case DIF_OP_SRA: 7256 case DIF_OP_SUB: 7257 case DIF_OP_ADD: 7258 case DIF_OP_MUL: 7259 case DIF_OP_SDIV: 7260 case DIF_OP_UDIV: 7261 case DIF_OP_SREM: 7262 case DIF_OP_UREM: 7263 case DIF_OP_COPYS: 7264 case DIF_OP_NOT: 7265 case DIF_OP_MOV: 7266 case DIF_OP_RLDSB: 7267 case DIF_OP_RLDSH: 7268 case DIF_OP_RLDSW: 7269 case DIF_OP_RLDUB: 7270 case DIF_OP_RLDUH: 7271 case DIF_OP_RLDUW: 7272 case DIF_OP_RLDX: 7273 case DIF_OP_ULDSB: 7274 case DIF_OP_ULDSH: 7275 case DIF_OP_ULDSW: 7276 case DIF_OP_ULDUB: 7277 case DIF_OP_ULDUH: 7278 case DIF_OP_ULDUW: 7279 case DIF_OP_ULDX: 7280 case DIF_OP_STB: 7281 case DIF_OP_STH: 7282 case DIF_OP_STW: 7283 case DIF_OP_STX: 7284 case DIF_OP_ALLOCS: 7285 case DIF_OP_CMP: 7286 case DIF_OP_SCMP: 7287 case DIF_OP_TST: 7288 case DIF_OP_BA: 7289 case DIF_OP_BE: 7290 case DIF_OP_BNE: 7291 case DIF_OP_BG: 7292 case DIF_OP_BGU: 7293 case DIF_OP_BGE: 7294 case DIF_OP_BGEU: 7295 case DIF_OP_BL: 7296 case DIF_OP_BLU: 7297 case DIF_OP_BLE: 7298 case DIF_OP_BLEU: 7299 case DIF_OP_RET: 7300 case DIF_OP_NOP: 7301 case DIF_OP_POPTS: 7302 case DIF_OP_FLUSHTS: 7303 case DIF_OP_SETX: 7304 case DIF_OP_SETS: 7305 case DIF_OP_LDGA: 7306 case DIF_OP_LDLS: 7307 case DIF_OP_STGS: 7308 case DIF_OP_STLS: 7309 case DIF_OP_PUSHTR: 7310 case DIF_OP_PUSHTV: 7311 break; 7312 7313 case DIF_OP_LDGS: 7314 if (v >= DIF_VAR_OTHER_UBASE) 7315 break; 7316 7317 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) 7318 break; 7319 7320 if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID || 7321 v == DIF_VAR_TID || v == DIF_VAR_EXECNAME || 7322 v == DIF_VAR_ZONENAME) 7323 break; 7324 7325 err += efunc(pc, "illegal variable %u\n", v); 7326 break; 7327 7328 case DIF_OP_LDTA: 7329 case DIF_OP_LDTS: 7330 case DIF_OP_LDGAA: 7331 case DIF_OP_LDTAA: 7332 err += efunc(pc, "illegal dynamic variable load\n"); 7333 break; 7334 7335 case DIF_OP_STTS: 7336 case DIF_OP_STGAA: 7337 case DIF_OP_STTAA: 7338 err += efunc(pc, "illegal dynamic variable store\n"); 7339 break; 7340 7341 case DIF_OP_CALL: 7342 if (subr == DIF_SUBR_ALLOCA || 7343 subr == DIF_SUBR_BCOPY || 7344 subr == DIF_SUBR_COPYIN || 7345 subr == DIF_SUBR_COPYINTO || 7346 subr == DIF_SUBR_COPYINSTR || 7347 subr == DIF_SUBR_INDEX || 7348 subr == DIF_SUBR_LLTOSTR || 7349 subr == DIF_SUBR_RINDEX || 7350 subr == DIF_SUBR_STRCHR || 7351 subr == DIF_SUBR_STRJOIN || 7352 subr == DIF_SUBR_STRRCHR || 7353 subr == DIF_SUBR_STRSTR) 7354 break; 7355 7356 err += efunc(pc, "invalid subr %u\n", subr); 7357 break; 7358 7359 default: 7360 err += efunc(pc, "invalid opcode %u\n", 7361 DIF_INSTR_OP(instr)); 7362 } 7363 } 7364 7365 return (err); 7366 } 7367 7368 /* 7369 * Returns 1 if the expression in the DIF object can be cached on a per-thread 7370 * basis; 0 if not. 7371 */ 7372 static int 7373 dtrace_difo_cacheable(dtrace_difo_t *dp) 7374 { 7375 int i; 7376 7377 if (dp == NULL) 7378 return (0); 7379 7380 for (i = 0; i < dp->dtdo_varlen; i++) { 7381 dtrace_difv_t *v = &dp->dtdo_vartab[i]; 7382 7383 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL) 7384 continue; 7385 7386 switch (v->dtdv_id) { 7387 case DIF_VAR_CURTHREAD: 7388 case DIF_VAR_PID: 7389 case DIF_VAR_TID: 7390 case DIF_VAR_EXECNAME: 7391 case DIF_VAR_ZONENAME: 7392 break; 7393 7394 default: 7395 return (0); 7396 } 7397 } 7398 7399 /* 7400 * This DIF object may be cacheable. Now we need to look for any 7401 * array loading instructions, any memory loading instructions, or 7402 * any stores to thread-local variables. 7403 */ 7404 for (i = 0; i < dp->dtdo_len; i++) { 7405 uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]); 7406 7407 if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) || 7408 (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) || 7409 (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) || 7410 op == DIF_OP_LDGA || op == DIF_OP_STTS) 7411 return (0); 7412 } 7413 7414 return (1); 7415 } 7416 7417 static void 7418 dtrace_difo_hold(dtrace_difo_t *dp) 7419 { 7420 int i; 7421 7422 ASSERT(MUTEX_HELD(&dtrace_lock)); 7423 7424 dp->dtdo_refcnt++; 7425 ASSERT(dp->dtdo_refcnt != 0); 7426 7427 /* 7428 * We need to check this DIF object for references to the variable 7429 * DIF_VAR_VTIMESTAMP. 7430 */ 7431 for (i = 0; i < dp->dtdo_varlen; i++) { 7432 dtrace_difv_t *v = &dp->dtdo_vartab[i]; 7433 7434 if (v->dtdv_id != DIF_VAR_VTIMESTAMP) 7435 continue; 7436 7437 if (dtrace_vtime_references++ == 0) 7438 dtrace_vtime_enable(); 7439 } 7440 } 7441 7442 /* 7443 * This routine calculates the dynamic variable chunksize for a given DIF 7444 * object. The calculation is not fool-proof, and can probably be tricked by 7445 * malicious DIF -- but it works for all compiler-generated DIF. Because this 7446 * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail 7447 * if a dynamic variable size exceeds the chunksize. 7448 */ 7449 static void 7450 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate) 7451 { 7452 uint64_t sval; 7453 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */ 7454 const dif_instr_t *text = dp->dtdo_buf; 7455 uint_t pc, srd = 0; 7456 uint_t ttop = 0; 7457 size_t size, ksize; 7458 uint_t id, i; 7459 7460 for (pc = 0; pc < dp->dtdo_len; pc++) { 7461 dif_instr_t instr = text[pc]; 7462 uint_t op = DIF_INSTR_OP(instr); 7463 uint_t rd = DIF_INSTR_RD(instr); 7464 uint_t r1 = DIF_INSTR_R1(instr); 7465 uint_t nkeys = 0; 7466 uchar_t scope; 7467 7468 dtrace_key_t *key = tupregs; 7469 7470 switch (op) { 7471 case DIF_OP_SETX: 7472 sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)]; 7473 srd = rd; 7474 continue; 7475 7476 case DIF_OP_STTS: 7477 key = &tupregs[DIF_DTR_NREGS]; 7478 key[0].dttk_size = 0; 7479 key[1].dttk_size = 0; 7480 nkeys = 2; 7481 scope = DIFV_SCOPE_THREAD; 7482 break; 7483 7484 case DIF_OP_STGAA: 7485 case DIF_OP_STTAA: 7486 nkeys = ttop; 7487 7488 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) 7489 key[nkeys++].dttk_size = 0; 7490 7491 key[nkeys++].dttk_size = 0; 7492 7493 if (op == DIF_OP_STTAA) { 7494 scope = DIFV_SCOPE_THREAD; 7495 } else { 7496 scope = DIFV_SCOPE_GLOBAL; 7497 } 7498 7499 break; 7500 7501 case DIF_OP_PUSHTR: 7502 if (ttop == DIF_DTR_NREGS) 7503 return; 7504 7505 if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) { 7506 /* 7507 * If the register for the size of the "pushtr" 7508 * is %r0 (or the value is 0) and the type is 7509 * a string, we'll use the system-wide default 7510 * string size. 7511 */ 7512 tupregs[ttop++].dttk_size = 7513 dtrace_strsize_default; 7514 } else { 7515 if (srd == 0) 7516 return; 7517 7518 tupregs[ttop++].dttk_size = sval; 7519 } 7520 7521 break; 7522 7523 case DIF_OP_PUSHTV: 7524 if (ttop == DIF_DTR_NREGS) 7525 return; 7526 7527 tupregs[ttop++].dttk_size = 0; 7528 break; 7529 7530 case DIF_OP_FLUSHTS: 7531 ttop = 0; 7532 break; 7533 7534 case DIF_OP_POPTS: 7535 if (ttop != 0) 7536 ttop--; 7537 break; 7538 } 7539 7540 sval = 0; 7541 srd = 0; 7542 7543 if (nkeys == 0) 7544 continue; 7545 7546 /* 7547 * We have a dynamic variable allocation; calculate its size. 7548 */ 7549 for (ksize = 0, i = 0; i < nkeys; i++) 7550 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t)); 7551 7552 size = sizeof (dtrace_dynvar_t); 7553 size += sizeof (dtrace_key_t) * (nkeys - 1); 7554 size += ksize; 7555 7556 /* 7557 * Now we need to determine the size of the stored data. 7558 */ 7559 id = DIF_INSTR_VAR(instr); 7560 7561 for (i = 0; i < dp->dtdo_varlen; i++) { 7562 dtrace_difv_t *v = &dp->dtdo_vartab[i]; 7563 7564 if (v->dtdv_id == id && v->dtdv_scope == scope) { 7565 size += v->dtdv_type.dtdt_size; 7566 break; 7567 } 7568 } 7569 7570 if (i == dp->dtdo_varlen) 7571 return; 7572 7573 /* 7574 * We have the size. If this is larger than the chunk size 7575 * for our dynamic variable state, reset the chunk size. 7576 */ 7577 size = P2ROUNDUP(size, sizeof (uint64_t)); 7578 7579 if (size > vstate->dtvs_dynvars.dtds_chunksize) 7580 vstate->dtvs_dynvars.dtds_chunksize = size; 7581 } 7582 } 7583 7584 static void 7585 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate) 7586 { 7587 int i, oldsvars, osz, nsz, otlocals, ntlocals; 7588 uint_t id; 7589 7590 ASSERT(MUTEX_HELD(&dtrace_lock)); 7591 ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0); 7592 7593 for (i = 0; i < dp->dtdo_varlen; i++) { 7594 dtrace_difv_t *v = &dp->dtdo_vartab[i]; 7595 dtrace_statvar_t *svar, ***svarp; 7596 size_t dsize = 0; 7597 uint8_t scope = v->dtdv_scope; 7598 int *np; 7599 7600 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE) 7601 continue; 7602 7603 id -= DIF_VAR_OTHER_UBASE; 7604 7605 switch (scope) { 7606 case DIFV_SCOPE_THREAD: 7607 while (id >= (otlocals = vstate->dtvs_ntlocals)) { 7608 dtrace_difv_t *tlocals; 7609 7610 if ((ntlocals = (otlocals << 1)) == 0) 7611 ntlocals = 1; 7612 7613 osz = otlocals * sizeof (dtrace_difv_t); 7614 nsz = ntlocals * sizeof (dtrace_difv_t); 7615 7616 tlocals = kmem_zalloc(nsz, KM_SLEEP); 7617 7618 if (osz != 0) { 7619 bcopy(vstate->dtvs_tlocals, 7620 tlocals, osz); 7621 kmem_free(vstate->dtvs_tlocals, osz); 7622 } 7623 7624 vstate->dtvs_tlocals = tlocals; 7625 vstate->dtvs_ntlocals = ntlocals; 7626 } 7627 7628 vstate->dtvs_tlocals[id] = *v; 7629 continue; 7630 7631 case DIFV_SCOPE_LOCAL: 7632 np = &vstate->dtvs_nlocals; 7633 svarp = &vstate->dtvs_locals; 7634 7635 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) 7636 dsize = NCPU * (v->dtdv_type.dtdt_size + 7637 sizeof (uint64_t)); 7638 else 7639 dsize = NCPU * sizeof (uint64_t); 7640 7641 break; 7642 7643 case DIFV_SCOPE_GLOBAL: 7644 np = &vstate->dtvs_nglobals; 7645 svarp = &vstate->dtvs_globals; 7646 7647 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) 7648 dsize = v->dtdv_type.dtdt_size + 7649 sizeof (uint64_t); 7650 7651 break; 7652 7653 default: 7654 ASSERT(0); 7655 } 7656 7657 while (id >= (oldsvars = *np)) { 7658 dtrace_statvar_t **statics; 7659 int newsvars, oldsize, newsize; 7660 7661 if ((newsvars = (oldsvars << 1)) == 0) 7662 newsvars = 1; 7663 7664 oldsize = oldsvars * sizeof (dtrace_statvar_t *); 7665 newsize = newsvars * sizeof (dtrace_statvar_t *); 7666 7667 statics = kmem_zalloc(newsize, KM_SLEEP); 7668 7669 if (oldsize != 0) { 7670 bcopy(*svarp, statics, oldsize); 7671 kmem_free(*svarp, oldsize); 7672 } 7673 7674 *svarp = statics; 7675 *np = newsvars; 7676 } 7677 7678 if ((svar = (*svarp)[id]) == NULL) { 7679 svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP); 7680 svar->dtsv_var = *v; 7681 7682 if ((svar->dtsv_size = dsize) != 0) { 7683 svar->dtsv_data = (uint64_t)(uintptr_t) 7684 kmem_zalloc(dsize, KM_SLEEP); 7685 } 7686 7687 (*svarp)[id] = svar; 7688 } 7689 7690 svar->dtsv_refcnt++; 7691 } 7692 7693 dtrace_difo_chunksize(dp, vstate); 7694 dtrace_difo_hold(dp); 7695 } 7696 7697 static dtrace_difo_t * 7698 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate) 7699 { 7700 dtrace_difo_t *new; 7701 size_t sz; 7702 7703 ASSERT(dp->dtdo_buf != NULL); 7704 ASSERT(dp->dtdo_refcnt != 0); 7705 7706 new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP); 7707 7708 ASSERT(dp->dtdo_buf != NULL); 7709 sz = dp->dtdo_len * sizeof (dif_instr_t); 7710 new->dtdo_buf = kmem_alloc(sz, KM_SLEEP); 7711 bcopy(dp->dtdo_buf, new->dtdo_buf, sz); 7712 new->dtdo_len = dp->dtdo_len; 7713 7714 if (dp->dtdo_strtab != NULL) { 7715 ASSERT(dp->dtdo_strlen != 0); 7716 new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP); 7717 bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen); 7718 new->dtdo_strlen = dp->dtdo_strlen; 7719 } 7720 7721 if (dp->dtdo_inttab != NULL) { 7722 ASSERT(dp->dtdo_intlen != 0); 7723 sz = dp->dtdo_intlen * sizeof (uint64_t); 7724 new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP); 7725 bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz); 7726 new->dtdo_intlen = dp->dtdo_intlen; 7727 } 7728 7729 if (dp->dtdo_vartab != NULL) { 7730 ASSERT(dp->dtdo_varlen != 0); 7731 sz = dp->dtdo_varlen * sizeof (dtrace_difv_t); 7732 new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP); 7733 bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz); 7734 new->dtdo_varlen = dp->dtdo_varlen; 7735 } 7736 7737 dtrace_difo_init(new, vstate); 7738 return (new); 7739 } 7740 7741 static void 7742 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate) 7743 { 7744 int i; 7745 7746 ASSERT(dp->dtdo_refcnt == 0); 7747 7748 for (i = 0; i < dp->dtdo_varlen; i++) { 7749 dtrace_difv_t *v = &dp->dtdo_vartab[i]; 7750 dtrace_statvar_t *svar, **svarp; 7751 uint_t id; 7752 uint8_t scope = v->dtdv_scope; 7753 int *np; 7754 7755 switch (scope) { 7756 case DIFV_SCOPE_THREAD: 7757 continue; 7758 7759 case DIFV_SCOPE_LOCAL: 7760 np = &vstate->dtvs_nlocals; 7761 svarp = vstate->dtvs_locals; 7762 break; 7763 7764 case DIFV_SCOPE_GLOBAL: 7765 np = &vstate->dtvs_nglobals; 7766 svarp = vstate->dtvs_globals; 7767 break; 7768 7769 default: 7770 ASSERT(0); 7771 } 7772 7773 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE) 7774 continue; 7775 7776 id -= DIF_VAR_OTHER_UBASE; 7777 ASSERT(id < *np); 7778 7779 svar = svarp[id]; 7780 ASSERT(svar != NULL); 7781 ASSERT(svar->dtsv_refcnt > 0); 7782 7783 if (--svar->dtsv_refcnt > 0) 7784 continue; 7785 7786 if (svar->dtsv_size != 0) { 7787 ASSERT(svar->dtsv_data != NULL); 7788 kmem_free((void *)(uintptr_t)svar->dtsv_data, 7789 svar->dtsv_size); 7790 } 7791 7792 kmem_free(svar, sizeof (dtrace_statvar_t)); 7793 svarp[id] = NULL; 7794 } 7795 7796 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t)); 7797 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t)); 7798 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen); 7799 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t)); 7800 7801 kmem_free(dp, sizeof (dtrace_difo_t)); 7802 } 7803 7804 static void 7805 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate) 7806 { 7807 int i; 7808 7809 ASSERT(MUTEX_HELD(&dtrace_lock)); 7810 ASSERT(dp->dtdo_refcnt != 0); 7811 7812 for (i = 0; i < dp->dtdo_varlen; i++) { 7813 dtrace_difv_t *v = &dp->dtdo_vartab[i]; 7814 7815 if (v->dtdv_id != DIF_VAR_VTIMESTAMP) 7816 continue; 7817 7818 ASSERT(dtrace_vtime_references > 0); 7819 if (--dtrace_vtime_references == 0) 7820 dtrace_vtime_disable(); 7821 } 7822 7823 if (--dp->dtdo_refcnt == 0) 7824 dtrace_difo_destroy(dp, vstate); 7825 } 7826 7827 /* 7828 * DTrace Format Functions 7829 */ 7830 static uint16_t 7831 dtrace_format_add(dtrace_state_t *state, char *str) 7832 { 7833 char *fmt, **new; 7834 uint16_t ndx, len = strlen(str) + 1; 7835 7836 fmt = kmem_zalloc(len, KM_SLEEP); 7837 bcopy(str, fmt, len); 7838 7839 for (ndx = 0; ndx < state->dts_nformats; ndx++) { 7840 if (state->dts_formats[ndx] == NULL) { 7841 state->dts_formats[ndx] = fmt; 7842 return (ndx + 1); 7843 } 7844 } 7845 7846 if (state->dts_nformats == USHRT_MAX) { 7847 /* 7848 * This is only likely if a denial-of-service attack is being 7849 * attempted. As such, it's okay to fail silently here. 7850 */ 7851 kmem_free(fmt, len); 7852 return (0); 7853 } 7854 7855 /* 7856 * For simplicity, we always resize the formats array to be exactly the 7857 * number of formats. 7858 */ 7859 ndx = state->dts_nformats++; 7860 new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP); 7861 7862 if (state->dts_formats != NULL) { 7863 ASSERT(ndx != 0); 7864 bcopy(state->dts_formats, new, ndx * sizeof (char *)); 7865 kmem_free(state->dts_formats, ndx * sizeof (char *)); 7866 } 7867 7868 state->dts_formats = new; 7869 state->dts_formats[ndx] = fmt; 7870 7871 return (ndx + 1); 7872 } 7873 7874 static void 7875 dtrace_format_remove(dtrace_state_t *state, uint16_t format) 7876 { 7877 char *fmt; 7878 7879 ASSERT(state->dts_formats != NULL); 7880 ASSERT(format <= state->dts_nformats); 7881 ASSERT(state->dts_formats[format - 1] != NULL); 7882 7883 fmt = state->dts_formats[format - 1]; 7884 kmem_free(fmt, strlen(fmt) + 1); 7885 state->dts_formats[format - 1] = NULL; 7886 } 7887 7888 static void 7889 dtrace_format_destroy(dtrace_state_t *state) 7890 { 7891 int i; 7892 7893 if (state->dts_nformats == 0) { 7894 ASSERT(state->dts_formats == NULL); 7895 return; 7896 } 7897 7898 ASSERT(state->dts_formats != NULL); 7899 7900 for (i = 0; i < state->dts_nformats; i++) { 7901 char *fmt = state->dts_formats[i]; 7902 7903 if (fmt == NULL) 7904 continue; 7905 7906 kmem_free(fmt, strlen(fmt) + 1); 7907 } 7908 7909 kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *)); 7910 state->dts_nformats = 0; 7911 state->dts_formats = NULL; 7912 } 7913 7914 /* 7915 * DTrace Predicate Functions 7916 */ 7917 static dtrace_predicate_t * 7918 dtrace_predicate_create(dtrace_difo_t *dp) 7919 { 7920 dtrace_predicate_t *pred; 7921 7922 ASSERT(MUTEX_HELD(&dtrace_lock)); 7923 ASSERT(dp->dtdo_refcnt != 0); 7924 7925 pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP); 7926 pred->dtp_difo = dp; 7927 pred->dtp_refcnt = 1; 7928 7929 if (!dtrace_difo_cacheable(dp)) 7930 return (pred); 7931 7932 if (dtrace_predcache_id == DTRACE_CACHEIDNONE) { 7933 /* 7934 * This is only theoretically possible -- we have had 2^32 7935 * cacheable predicates on this machine. We cannot allow any 7936 * more predicates to become cacheable: as unlikely as it is, 7937 * there may be a thread caching a (now stale) predicate cache 7938 * ID. (N.B.: the temptation is being successfully resisted to 7939 * have this cmn_err() "Holy shit -- we executed this code!") 7940 */ 7941 return (pred); 7942 } 7943 7944 pred->dtp_cacheid = dtrace_predcache_id++; 7945 7946 return (pred); 7947 } 7948 7949 static void 7950 dtrace_predicate_hold(dtrace_predicate_t *pred) 7951 { 7952 ASSERT(MUTEX_HELD(&dtrace_lock)); 7953 ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0); 7954 ASSERT(pred->dtp_refcnt > 0); 7955 7956 pred->dtp_refcnt++; 7957 } 7958 7959 static void 7960 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate) 7961 { 7962 dtrace_difo_t *dp = pred->dtp_difo; 7963 7964 ASSERT(MUTEX_HELD(&dtrace_lock)); 7965 ASSERT(dp != NULL && dp->dtdo_refcnt != 0); 7966 ASSERT(pred->dtp_refcnt > 0); 7967 7968 if (--pred->dtp_refcnt == 0) { 7969 dtrace_difo_release(pred->dtp_difo, vstate); 7970 kmem_free(pred, sizeof (dtrace_predicate_t)); 7971 } 7972 } 7973 7974 /* 7975 * DTrace Action Description Functions 7976 */ 7977 static dtrace_actdesc_t * 7978 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple, 7979 uint64_t uarg, uint64_t arg) 7980 { 7981 dtrace_actdesc_t *act; 7982 7983 ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL && 7984 arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA)); 7985 7986 act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP); 7987 act->dtad_kind = kind; 7988 act->dtad_ntuple = ntuple; 7989 act->dtad_uarg = uarg; 7990 act->dtad_arg = arg; 7991 act->dtad_refcnt = 1; 7992 7993 return (act); 7994 } 7995 7996 static void 7997 dtrace_actdesc_hold(dtrace_actdesc_t *act) 7998 { 7999 ASSERT(act->dtad_refcnt >= 1); 8000 act->dtad_refcnt++; 8001 } 8002 8003 static void 8004 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate) 8005 { 8006 dtrace_actkind_t kind = act->dtad_kind; 8007 dtrace_difo_t *dp; 8008 8009 ASSERT(act->dtad_refcnt >= 1); 8010 8011 if (--act->dtad_refcnt != 0) 8012 return; 8013 8014 if ((dp = act->dtad_difo) != NULL) 8015 dtrace_difo_release(dp, vstate); 8016 8017 if (DTRACEACT_ISPRINTFLIKE(kind)) { 8018 char *str = (char *)(uintptr_t)act->dtad_arg; 8019 8020 ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) || 8021 (str == NULL && act->dtad_kind == DTRACEACT_PRINTA)); 8022 8023 if (str != NULL) 8024 kmem_free(str, strlen(str) + 1); 8025 } 8026 8027 kmem_free(act, sizeof (dtrace_actdesc_t)); 8028 } 8029 8030 /* 8031 * DTrace ECB Functions 8032 */ 8033 static dtrace_ecb_t * 8034 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe) 8035 { 8036 dtrace_ecb_t *ecb; 8037 dtrace_epid_t epid; 8038 8039 ASSERT(MUTEX_HELD(&dtrace_lock)); 8040 8041 ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP); 8042 ecb->dte_predicate = NULL; 8043 ecb->dte_probe = probe; 8044 8045 /* 8046 * The default size is the size of the default action: recording 8047 * the epid. 8048 */ 8049 ecb->dte_size = ecb->dte_needed = sizeof (dtrace_epid_t); 8050 ecb->dte_alignment = sizeof (dtrace_epid_t); 8051 8052 epid = state->dts_epid++; 8053 8054 if (epid - 1 >= state->dts_necbs) { 8055 dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs; 8056 int necbs = state->dts_necbs << 1; 8057 8058 ASSERT(epid == state->dts_necbs + 1); 8059 8060 if (necbs == 0) { 8061 ASSERT(oecbs == NULL); 8062 necbs = 1; 8063 } 8064 8065 ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP); 8066 8067 if (oecbs != NULL) 8068 bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs)); 8069 8070 dtrace_membar_producer(); 8071 state->dts_ecbs = ecbs; 8072 8073 if (oecbs != NULL) { 8074 /* 8075 * If this state is active, we must dtrace_sync() 8076 * before we can free the old dts_ecbs array: we're 8077 * coming in hot, and there may be active ring 8078 * buffer processing (which indexes into the dts_ecbs 8079 * array) on another CPU. 8080 */ 8081 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) 8082 dtrace_sync(); 8083 8084 kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs)); 8085 } 8086 8087 dtrace_membar_producer(); 8088 state->dts_necbs = necbs; 8089 } 8090 8091 ecb->dte_state = state; 8092 8093 ASSERT(state->dts_ecbs[epid - 1] == NULL); 8094 dtrace_membar_producer(); 8095 state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb; 8096 8097 return (ecb); 8098 } 8099 8100 static void 8101 dtrace_ecb_enable(dtrace_ecb_t *ecb) 8102 { 8103 dtrace_probe_t *probe = ecb->dte_probe; 8104 8105 ASSERT(MUTEX_HELD(&cpu_lock)); 8106 ASSERT(MUTEX_HELD(&dtrace_lock)); 8107 ASSERT(ecb->dte_next == NULL); 8108 8109 if (probe == NULL) { 8110 /* 8111 * This is the NULL probe -- there's nothing to do. 8112 */ 8113 return; 8114 } 8115 8116 if (probe->dtpr_ecb == NULL) { 8117 dtrace_provider_t *prov = probe->dtpr_provider; 8118 8119 /* 8120 * We're the first ECB on this probe. 8121 */ 8122 probe->dtpr_ecb = probe->dtpr_ecb_last = ecb; 8123 8124 if (ecb->dte_predicate != NULL) 8125 probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid; 8126 8127 prov->dtpv_pops.dtps_enable(prov->dtpv_arg, 8128 probe->dtpr_id, probe->dtpr_arg); 8129 } else { 8130 /* 8131 * This probe is already active. Swing the last pointer to 8132 * point to the new ECB, and issue a dtrace_sync() to assure 8133 * that all CPUs have seen the change. 8134 */ 8135 ASSERT(probe->dtpr_ecb_last != NULL); 8136 probe->dtpr_ecb_last->dte_next = ecb; 8137 probe->dtpr_ecb_last = ecb; 8138 probe->dtpr_predcache = 0; 8139 8140 dtrace_sync(); 8141 } 8142 } 8143 8144 static void 8145 dtrace_ecb_resize(dtrace_ecb_t *ecb) 8146 { 8147 uint32_t maxalign = sizeof (dtrace_epid_t); 8148 uint32_t align = sizeof (uint8_t), offs, diff; 8149 dtrace_action_t *act; 8150 int wastuple = 0; 8151 uint32_t aggbase = UINT32_MAX; 8152 dtrace_state_t *state = ecb->dte_state; 8153 8154 /* 8155 * If we record anything, we always record the epid. (And we always 8156 * record it first.) 8157 */ 8158 offs = sizeof (dtrace_epid_t); 8159 ecb->dte_size = ecb->dte_needed = sizeof (dtrace_epid_t); 8160 8161 for (act = ecb->dte_action; act != NULL; act = act->dta_next) { 8162 dtrace_recdesc_t *rec = &act->dta_rec; 8163 8164 if ((align = rec->dtrd_alignment) > maxalign) 8165 maxalign = align; 8166 8167 if (!wastuple && act->dta_intuple) { 8168 /* 8169 * This is the first record in a tuple. Align the 8170 * offset to be at offset 4 in an 8-byte aligned 8171 * block. 8172 */ 8173 diff = offs + sizeof (dtrace_aggid_t); 8174 8175 if (diff = (diff & (sizeof (uint64_t) - 1))) 8176 offs += sizeof (uint64_t) - diff; 8177 8178 aggbase = offs - sizeof (dtrace_aggid_t); 8179 ASSERT(!(aggbase & (sizeof (uint64_t) - 1))); 8180 } 8181 8182 /*LINTED*/ 8183 if (rec->dtrd_size != 0 && (diff = (offs & (align - 1)))) { 8184 /* 8185 * The current offset is not properly aligned; align it. 8186 */ 8187 offs += align - diff; 8188 } 8189 8190 rec->dtrd_offset = offs; 8191 8192 if (offs + rec->dtrd_size > ecb->dte_needed) { 8193 ecb->dte_needed = offs + rec->dtrd_size; 8194 8195 if (ecb->dte_needed > state->dts_needed) 8196 state->dts_needed = ecb->dte_needed; 8197 } 8198 8199 if (DTRACEACT_ISAGG(act->dta_kind)) { 8200 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act; 8201 dtrace_action_t *first = agg->dtag_first, *prev; 8202 8203 ASSERT(rec->dtrd_size != 0 && first != NULL); 8204 ASSERT(wastuple); 8205 ASSERT(aggbase != UINT32_MAX); 8206 8207 agg->dtag_base = aggbase; 8208 8209 while ((prev = first->dta_prev) != NULL && 8210 DTRACEACT_ISAGG(prev->dta_kind)) { 8211 agg = (dtrace_aggregation_t *)prev; 8212 first = agg->dtag_first; 8213 } 8214 8215 if (prev != NULL) { 8216 offs = prev->dta_rec.dtrd_offset + 8217 prev->dta_rec.dtrd_size; 8218 } else { 8219 offs = sizeof (dtrace_epid_t); 8220 } 8221 wastuple = 0; 8222 } else { 8223 if (!act->dta_intuple) 8224 ecb->dte_size = offs + rec->dtrd_size; 8225 8226 offs += rec->dtrd_size; 8227 } 8228 8229 wastuple = act->dta_intuple; 8230 } 8231 8232 if ((act = ecb->dte_action) != NULL && 8233 !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) && 8234 ecb->dte_size == sizeof (dtrace_epid_t)) { 8235 /* 8236 * If the size is still sizeof (dtrace_epid_t), then all 8237 * actions store no data; set the size to 0. 8238 */ 8239 ecb->dte_alignment = maxalign; 8240 ecb->dte_size = 0; 8241 8242 /* 8243 * If the needed space is still sizeof (dtrace_epid_t), then 8244 * all actions need no additional space; set the needed 8245 * size to 0. 8246 */ 8247 if (ecb->dte_needed == sizeof (dtrace_epid_t)) 8248 ecb->dte_needed = 0; 8249 8250 return; 8251 } 8252 8253 /* 8254 * Set our alignment, and make sure that the dte_size and dte_needed 8255 * are aligned to the size of an EPID. 8256 */ 8257 ecb->dte_alignment = maxalign; 8258 ecb->dte_size = (ecb->dte_size + (sizeof (dtrace_epid_t) - 1)) & 8259 ~(sizeof (dtrace_epid_t) - 1); 8260 ecb->dte_needed = (ecb->dte_needed + (sizeof (dtrace_epid_t) - 1)) & 8261 ~(sizeof (dtrace_epid_t) - 1); 8262 ASSERT(ecb->dte_size <= ecb->dte_needed); 8263 } 8264 8265 static dtrace_action_t * 8266 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc) 8267 { 8268 dtrace_aggregation_t *agg; 8269 size_t size = sizeof (uint64_t); 8270 int ntuple = desc->dtad_ntuple; 8271 dtrace_action_t *act; 8272 dtrace_recdesc_t *frec; 8273 dtrace_aggid_t aggid; 8274 dtrace_state_t *state = ecb->dte_state; 8275 8276 agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP); 8277 agg->dtag_ecb = ecb; 8278 8279 ASSERT(DTRACEACT_ISAGG(desc->dtad_kind)); 8280 8281 switch (desc->dtad_kind) { 8282 case DTRACEAGG_MIN: 8283 agg->dtag_initial = UINT64_MAX; 8284 agg->dtag_aggregate = dtrace_aggregate_min; 8285 break; 8286 8287 case DTRACEAGG_MAX: 8288 agg->dtag_aggregate = dtrace_aggregate_max; 8289 break; 8290 8291 case DTRACEAGG_COUNT: 8292 agg->dtag_aggregate = dtrace_aggregate_count; 8293 break; 8294 8295 case DTRACEAGG_QUANTIZE: 8296 agg->dtag_aggregate = dtrace_aggregate_quantize; 8297 size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) * 8298 sizeof (uint64_t); 8299 break; 8300 8301 case DTRACEAGG_LQUANTIZE: { 8302 uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg); 8303 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg); 8304 8305 agg->dtag_initial = desc->dtad_arg; 8306 agg->dtag_aggregate = dtrace_aggregate_lquantize; 8307 8308 if (step == 0 || levels == 0) 8309 goto err; 8310 8311 size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t); 8312 break; 8313 } 8314 8315 case DTRACEAGG_AVG: 8316 agg->dtag_aggregate = dtrace_aggregate_avg; 8317 size = sizeof (uint64_t) * 2; 8318 break; 8319 8320 case DTRACEAGG_SUM: 8321 agg->dtag_aggregate = dtrace_aggregate_sum; 8322 break; 8323 8324 default: 8325 goto err; 8326 } 8327 8328 agg->dtag_action.dta_rec.dtrd_size = size; 8329 8330 if (ntuple == 0) 8331 goto err; 8332 8333 /* 8334 * We must make sure that we have enough actions for the n-tuple. 8335 */ 8336 for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) { 8337 if (DTRACEACT_ISAGG(act->dta_kind)) 8338 break; 8339 8340 if (--ntuple == 0) { 8341 /* 8342 * This is the action with which our n-tuple begins. 8343 */ 8344 agg->dtag_first = act; 8345 goto success; 8346 } 8347 } 8348 8349 /* 8350 * This n-tuple is short by ntuple elements. Return failure. 8351 */ 8352 ASSERT(ntuple != 0); 8353 err: 8354 kmem_free(agg, sizeof (dtrace_aggregation_t)); 8355 return (NULL); 8356 8357 success: 8358 /* 8359 * If the last action in the tuple has a size of zero, it's actually 8360 * an expression argument for the aggregating action. 8361 */ 8362 ASSERT(ecb->dte_action_last != NULL); 8363 act = ecb->dte_action_last; 8364 8365 if (act->dta_kind == DTRACEACT_DIFEXPR) { 8366 ASSERT(act->dta_difo != NULL); 8367 8368 if (act->dta_difo->dtdo_rtype.dtdt_size == 0) 8369 agg->dtag_hasarg = 1; 8370 } 8371 8372 /* 8373 * We need to allocate an id for this aggregation. 8374 */ 8375 aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1, 8376 VM_BESTFIT | VM_SLEEP); 8377 8378 if (aggid - 1 >= state->dts_naggregations) { 8379 dtrace_aggregation_t **oaggs = state->dts_aggregations; 8380 dtrace_aggregation_t **aggs; 8381 int naggs = state->dts_naggregations << 1; 8382 int onaggs = state->dts_naggregations; 8383 8384 ASSERT(aggid == state->dts_naggregations + 1); 8385 8386 if (naggs == 0) { 8387 ASSERT(oaggs == NULL); 8388 naggs = 1; 8389 } 8390 8391 aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP); 8392 8393 if (oaggs != NULL) { 8394 bcopy(oaggs, aggs, onaggs * sizeof (*aggs)); 8395 kmem_free(oaggs, onaggs * sizeof (*aggs)); 8396 } 8397 8398 state->dts_aggregations = aggs; 8399 state->dts_naggregations = naggs; 8400 } 8401 8402 ASSERT(state->dts_aggregations[aggid - 1] == NULL); 8403 state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg; 8404 8405 frec = &agg->dtag_first->dta_rec; 8406 if (frec->dtrd_alignment < sizeof (dtrace_aggid_t)) 8407 frec->dtrd_alignment = sizeof (dtrace_aggid_t); 8408 8409 for (act = agg->dtag_first; act != NULL; act = act->dta_next) { 8410 ASSERT(!act->dta_intuple); 8411 act->dta_intuple = 1; 8412 } 8413 8414 return (&agg->dtag_action); 8415 } 8416 8417 static void 8418 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act) 8419 { 8420 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act; 8421 dtrace_state_t *state = ecb->dte_state; 8422 dtrace_aggid_t aggid = agg->dtag_id; 8423 8424 ASSERT(DTRACEACT_ISAGG(act->dta_kind)); 8425 vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1); 8426 8427 ASSERT(state->dts_aggregations[aggid - 1] == agg); 8428 state->dts_aggregations[aggid - 1] = NULL; 8429 8430 kmem_free(agg, sizeof (dtrace_aggregation_t)); 8431 } 8432 8433 static int 8434 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc) 8435 { 8436 dtrace_action_t *action, *last; 8437 dtrace_difo_t *dp = desc->dtad_difo; 8438 uint32_t size = 0, align = sizeof (uint8_t), mask; 8439 uint16_t format = 0; 8440 dtrace_recdesc_t *rec; 8441 dtrace_state_t *state = ecb->dte_state; 8442 dtrace_optval_t *opt = state->dts_options, nframes, strsize; 8443 uint64_t arg = desc->dtad_arg; 8444 8445 ASSERT(MUTEX_HELD(&dtrace_lock)); 8446 ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1); 8447 8448 if (DTRACEACT_ISAGG(desc->dtad_kind)) { 8449 /* 8450 * If this is an aggregating action, there must be neither 8451 * a speculate nor a commit on the action chain. 8452 */ 8453 dtrace_action_t *act; 8454 8455 for (act = ecb->dte_action; act != NULL; act = act->dta_next) { 8456 if (act->dta_kind == DTRACEACT_COMMIT) 8457 return (EINVAL); 8458 8459 if (act->dta_kind == DTRACEACT_SPECULATE) 8460 return (EINVAL); 8461 } 8462 8463 action = dtrace_ecb_aggregation_create(ecb, desc); 8464 8465 if (action == NULL) 8466 return (EINVAL); 8467 } else { 8468 if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) || 8469 (desc->dtad_kind == DTRACEACT_DIFEXPR && 8470 dp != NULL && dp->dtdo_destructive)) { 8471 state->dts_destructive = 1; 8472 } 8473 8474 switch (desc->dtad_kind) { 8475 case DTRACEACT_PRINTF: 8476 case DTRACEACT_PRINTA: 8477 case DTRACEACT_SYSTEM: 8478 case DTRACEACT_FREOPEN: 8479 /* 8480 * We know that our arg is a string -- turn it into a 8481 * format. 8482 */ 8483 if (arg == NULL) { 8484 ASSERT(desc->dtad_kind == DTRACEACT_PRINTA); 8485 format = 0; 8486 } else { 8487 ASSERT(arg != NULL); 8488 ASSERT(arg > KERNELBASE); 8489 format = dtrace_format_add(state, 8490 (char *)(uintptr_t)arg); 8491 } 8492 8493 /*FALLTHROUGH*/ 8494 case DTRACEACT_LIBACT: 8495 case DTRACEACT_DIFEXPR: 8496 if (dp == NULL) 8497 return (EINVAL); 8498 8499 if ((size = dp->dtdo_rtype.dtdt_size) != 0) 8500 break; 8501 8502 if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) { 8503 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) 8504 return (EINVAL); 8505 8506 size = opt[DTRACEOPT_STRSIZE]; 8507 } 8508 8509 break; 8510 8511 case DTRACEACT_STACK: 8512 if ((nframes = arg) == 0) { 8513 nframes = opt[DTRACEOPT_STACKFRAMES]; 8514 ASSERT(nframes > 0); 8515 arg = nframes; 8516 } 8517 8518 size = nframes * sizeof (pc_t); 8519 break; 8520 8521 case DTRACEACT_JSTACK: 8522 if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0) 8523 strsize = opt[DTRACEOPT_JSTACKSTRSIZE]; 8524 8525 if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) 8526 nframes = opt[DTRACEOPT_JSTACKFRAMES]; 8527 8528 arg = DTRACE_USTACK_ARG(nframes, strsize); 8529 8530 /*FALLTHROUGH*/ 8531 case DTRACEACT_USTACK: 8532 if (desc->dtad_kind != DTRACEACT_JSTACK && 8533 (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) { 8534 strsize = DTRACE_USTACK_STRSIZE(arg); 8535 nframes = opt[DTRACEOPT_USTACKFRAMES]; 8536 ASSERT(nframes > 0); 8537 arg = DTRACE_USTACK_ARG(nframes, strsize); 8538 } 8539 8540 /* 8541 * Save a slot for the pid. 8542 */ 8543 size = (nframes + 1) * sizeof (uint64_t); 8544 size += DTRACE_USTACK_STRSIZE(arg); 8545 size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t))); 8546 8547 break; 8548 8549 case DTRACEACT_SYM: 8550 case DTRACEACT_MOD: 8551 if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) != 8552 sizeof (uint64_t)) || 8553 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) 8554 return (EINVAL); 8555 break; 8556 8557 case DTRACEACT_USYM: 8558 case DTRACEACT_UMOD: 8559 case DTRACEACT_UADDR: 8560 if (dp == NULL || 8561 (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) || 8562 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) 8563 return (EINVAL); 8564 8565 /* 8566 * We have a slot for the pid, plus a slot for the 8567 * argument. To keep things simple (aligned with 8568 * bitness-neutral sizing), we store each as a 64-bit 8569 * quantity. 8570 */ 8571 size = 2 * sizeof (uint64_t); 8572 break; 8573 8574 case DTRACEACT_STOP: 8575 case DTRACEACT_BREAKPOINT: 8576 case DTRACEACT_PANIC: 8577 break; 8578 8579 case DTRACEACT_CHILL: 8580 case DTRACEACT_DISCARD: 8581 case DTRACEACT_RAISE: 8582 if (dp == NULL) 8583 return (EINVAL); 8584 break; 8585 8586 case DTRACEACT_EXIT: 8587 if (dp == NULL || 8588 (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) || 8589 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) 8590 return (EINVAL); 8591 break; 8592 8593 case DTRACEACT_SPECULATE: 8594 if (ecb->dte_size > sizeof (dtrace_epid_t)) 8595 return (EINVAL); 8596 8597 if (dp == NULL) 8598 return (EINVAL); 8599 8600 state->dts_speculates = 1; 8601 break; 8602 8603 case DTRACEACT_COMMIT: { 8604 dtrace_action_t *act = ecb->dte_action; 8605 8606 for (; act != NULL; act = act->dta_next) { 8607 if (act->dta_kind == DTRACEACT_COMMIT) 8608 return (EINVAL); 8609 } 8610 8611 if (dp == NULL) 8612 return (EINVAL); 8613 break; 8614 } 8615 8616 default: 8617 return (EINVAL); 8618 } 8619 8620 if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) { 8621 /* 8622 * If this is a data-storing action or a speculate, 8623 * we must be sure that there isn't a commit on the 8624 * action chain. 8625 */ 8626 dtrace_action_t *act = ecb->dte_action; 8627 8628 for (; act != NULL; act = act->dta_next) { 8629 if (act->dta_kind == DTRACEACT_COMMIT) 8630 return (EINVAL); 8631 } 8632 } 8633 8634 action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP); 8635 action->dta_rec.dtrd_size = size; 8636 } 8637 8638 action->dta_refcnt = 1; 8639 rec = &action->dta_rec; 8640 size = rec->dtrd_size; 8641 8642 for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) { 8643 if (!(size & mask)) { 8644 align = mask + 1; 8645 break; 8646 } 8647 } 8648 8649 action->dta_kind = desc->dtad_kind; 8650 8651 if ((action->dta_difo = dp) != NULL) 8652 dtrace_difo_hold(dp); 8653 8654 rec->dtrd_action = action->dta_kind; 8655 rec->dtrd_arg = arg; 8656 rec->dtrd_uarg = desc->dtad_uarg; 8657 rec->dtrd_alignment = (uint16_t)align; 8658 rec->dtrd_format = format; 8659 8660 if ((last = ecb->dte_action_last) != NULL) { 8661 ASSERT(ecb->dte_action != NULL); 8662 action->dta_prev = last; 8663 last->dta_next = action; 8664 } else { 8665 ASSERT(ecb->dte_action == NULL); 8666 ecb->dte_action = action; 8667 } 8668 8669 ecb->dte_action_last = action; 8670 8671 return (0); 8672 } 8673 8674 static void 8675 dtrace_ecb_action_remove(dtrace_ecb_t *ecb) 8676 { 8677 dtrace_action_t *act = ecb->dte_action, *next; 8678 dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate; 8679 dtrace_difo_t *dp; 8680 uint16_t format; 8681 8682 if (act != NULL && act->dta_refcnt > 1) { 8683 ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1); 8684 act->dta_refcnt--; 8685 } else { 8686 for (; act != NULL; act = next) { 8687 next = act->dta_next; 8688 ASSERT(next != NULL || act == ecb->dte_action_last); 8689 ASSERT(act->dta_refcnt == 1); 8690 8691 if ((format = act->dta_rec.dtrd_format) != 0) 8692 dtrace_format_remove(ecb->dte_state, format); 8693 8694 if ((dp = act->dta_difo) != NULL) 8695 dtrace_difo_release(dp, vstate); 8696 8697 if (DTRACEACT_ISAGG(act->dta_kind)) { 8698 dtrace_ecb_aggregation_destroy(ecb, act); 8699 } else { 8700 kmem_free(act, sizeof (dtrace_action_t)); 8701 } 8702 } 8703 } 8704 8705 ecb->dte_action = NULL; 8706 ecb->dte_action_last = NULL; 8707 ecb->dte_size = sizeof (dtrace_epid_t); 8708 } 8709 8710 static void 8711 dtrace_ecb_disable(dtrace_ecb_t *ecb) 8712 { 8713 /* 8714 * We disable the ECB by removing it from its probe. 8715 */ 8716 dtrace_ecb_t *pecb, *prev = NULL; 8717 dtrace_probe_t *probe = ecb->dte_probe; 8718 8719 ASSERT(MUTEX_HELD(&dtrace_lock)); 8720 8721 if (probe == NULL) { 8722 /* 8723 * This is the NULL probe; there is nothing to disable. 8724 */ 8725 return; 8726 } 8727 8728 for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) { 8729 if (pecb == ecb) 8730 break; 8731 prev = pecb; 8732 } 8733 8734 ASSERT(pecb != NULL); 8735 8736 if (prev == NULL) { 8737 probe->dtpr_ecb = ecb->dte_next; 8738 } else { 8739 prev->dte_next = ecb->dte_next; 8740 } 8741 8742 if (ecb == probe->dtpr_ecb_last) { 8743 ASSERT(ecb->dte_next == NULL); 8744 probe->dtpr_ecb_last = prev; 8745 } 8746 8747 /* 8748 * The ECB has been disconnected from the probe; now sync to assure 8749 * that all CPUs have seen the change before returning. 8750 */ 8751 dtrace_sync(); 8752 8753 if (probe->dtpr_ecb == NULL) { 8754 /* 8755 * That was the last ECB on the probe; clear the predicate 8756 * cache ID for the probe, disable it and sync one more time 8757 * to assure that we'll never hit it again. 8758 */ 8759 dtrace_provider_t *prov = probe->dtpr_provider; 8760 8761 ASSERT(ecb->dte_next == NULL); 8762 ASSERT(probe->dtpr_ecb_last == NULL); 8763 probe->dtpr_predcache = DTRACE_CACHEIDNONE; 8764 prov->dtpv_pops.dtps_disable(prov->dtpv_arg, 8765 probe->dtpr_id, probe->dtpr_arg); 8766 dtrace_sync(); 8767 } else { 8768 /* 8769 * There is at least one ECB remaining on the probe. If there 8770 * is _exactly_ one, set the probe's predicate cache ID to be 8771 * the predicate cache ID of the remaining ECB. 8772 */ 8773 ASSERT(probe->dtpr_ecb_last != NULL); 8774 ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE); 8775 8776 if (probe->dtpr_ecb == probe->dtpr_ecb_last) { 8777 dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate; 8778 8779 ASSERT(probe->dtpr_ecb->dte_next == NULL); 8780 8781 if (p != NULL) 8782 probe->dtpr_predcache = p->dtp_cacheid; 8783 } 8784 8785 ecb->dte_next = NULL; 8786 } 8787 } 8788 8789 static void 8790 dtrace_ecb_destroy(dtrace_ecb_t *ecb) 8791 { 8792 dtrace_state_t *state = ecb->dte_state; 8793 dtrace_vstate_t *vstate = &state->dts_vstate; 8794 dtrace_predicate_t *pred; 8795 dtrace_epid_t epid = ecb->dte_epid; 8796 8797 ASSERT(MUTEX_HELD(&dtrace_lock)); 8798 ASSERT(ecb->dte_next == NULL); 8799 ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb); 8800 8801 if ((pred = ecb->dte_predicate) != NULL) 8802 dtrace_predicate_release(pred, vstate); 8803 8804 dtrace_ecb_action_remove(ecb); 8805 8806 ASSERT(state->dts_ecbs[epid - 1] == ecb); 8807 state->dts_ecbs[epid - 1] = NULL; 8808 8809 kmem_free(ecb, sizeof (dtrace_ecb_t)); 8810 } 8811 8812 static dtrace_ecb_t * 8813 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe, 8814 dtrace_enabling_t *enab) 8815 { 8816 dtrace_ecb_t *ecb; 8817 dtrace_predicate_t *pred; 8818 dtrace_actdesc_t *act; 8819 dtrace_provider_t *prov; 8820 dtrace_ecbdesc_t *desc = enab->dten_current; 8821 8822 ASSERT(MUTEX_HELD(&dtrace_lock)); 8823 ASSERT(state != NULL); 8824 8825 ecb = dtrace_ecb_add(state, probe); 8826 ecb->dte_uarg = desc->dted_uarg; 8827 8828 if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) { 8829 dtrace_predicate_hold(pred); 8830 ecb->dte_predicate = pred; 8831 } 8832 8833 if (probe != NULL) { 8834 /* 8835 * If the provider shows more leg than the consumer is old 8836 * enough to see, we need to enable the appropriate implicit 8837 * predicate bits to prevent the ecb from activating at 8838 * revealing times. 8839 */ 8840 prov = probe->dtpr_provider; 8841 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) && 8842 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER)) 8843 ecb->dte_cond |= DTRACE_COND_OWNER; 8844 8845 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) && 8846 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL)) 8847 ecb->dte_cond |= DTRACE_COND_USERMODE; 8848 } 8849 8850 if (dtrace_ecb_create_cache != NULL) { 8851 /* 8852 * If we have a cached ecb, we'll use its action list instead 8853 * of creating our own (saving both time and space). 8854 */ 8855 dtrace_ecb_t *cached = dtrace_ecb_create_cache; 8856 dtrace_action_t *act = cached->dte_action; 8857 8858 if (act != NULL) { 8859 ASSERT(act->dta_refcnt > 0); 8860 act->dta_refcnt++; 8861 ecb->dte_action = act; 8862 ecb->dte_action_last = cached->dte_action_last; 8863 ecb->dte_needed = cached->dte_needed; 8864 ecb->dte_size = cached->dte_size; 8865 ecb->dte_alignment = cached->dte_alignment; 8866 } 8867 8868 return (ecb); 8869 } 8870 8871 for (act = desc->dted_action; act != NULL; act = act->dtad_next) { 8872 if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) { 8873 dtrace_ecb_destroy(ecb); 8874 return (NULL); 8875 } 8876 } 8877 8878 dtrace_ecb_resize(ecb); 8879 8880 return (dtrace_ecb_create_cache = ecb); 8881 } 8882 8883 static int 8884 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg) 8885 { 8886 dtrace_ecb_t *ecb; 8887 dtrace_enabling_t *enab = arg; 8888 dtrace_state_t *state = enab->dten_vstate->dtvs_state; 8889 8890 ASSERT(state != NULL); 8891 8892 if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) { 8893 /* 8894 * This probe was created in a generation for which this 8895 * enabling has previously created ECBs; we don't want to 8896 * enable it again, so just kick out. 8897 */ 8898 return (DTRACE_MATCH_NEXT); 8899 } 8900 8901 if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL) 8902 return (DTRACE_MATCH_DONE); 8903 8904 dtrace_ecb_enable(ecb); 8905 return (DTRACE_MATCH_NEXT); 8906 } 8907 8908 static dtrace_ecb_t * 8909 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id) 8910 { 8911 dtrace_ecb_t *ecb; 8912 8913 ASSERT(MUTEX_HELD(&dtrace_lock)); 8914 8915 if (id == 0 || id > state->dts_necbs) 8916 return (NULL); 8917 8918 ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL); 8919 ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id); 8920 8921 return (state->dts_ecbs[id - 1]); 8922 } 8923 8924 static dtrace_aggregation_t * 8925 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id) 8926 { 8927 dtrace_aggregation_t *agg; 8928 8929 ASSERT(MUTEX_HELD(&dtrace_lock)); 8930 8931 if (id == 0 || id > state->dts_naggregations) 8932 return (NULL); 8933 8934 ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL); 8935 ASSERT((agg = state->dts_aggregations[id - 1]) == NULL || 8936 agg->dtag_id == id); 8937 8938 return (state->dts_aggregations[id - 1]); 8939 } 8940 8941 /* 8942 * DTrace Buffer Functions 8943 * 8944 * The following functions manipulate DTrace buffers. Most of these functions 8945 * are called in the context of establishing or processing consumer state; 8946 * exceptions are explicitly noted. 8947 */ 8948 8949 /* 8950 * Note: called from cross call context. This function switches the two 8951 * buffers on a given CPU. The atomicity of this operation is assured by 8952 * disabling interrupts while the actual switch takes place; the disabling of 8953 * interrupts serializes the execution with any execution of dtrace_probe() on 8954 * the same CPU. 8955 */ 8956 static void 8957 dtrace_buffer_switch(dtrace_buffer_t *buf) 8958 { 8959 caddr_t tomax = buf->dtb_tomax; 8960 caddr_t xamot = buf->dtb_xamot; 8961 dtrace_icookie_t cookie; 8962 8963 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH)); 8964 ASSERT(!(buf->dtb_flags & DTRACEBUF_RING)); 8965 8966 cookie = dtrace_interrupt_disable(); 8967 buf->dtb_tomax = xamot; 8968 buf->dtb_xamot = tomax; 8969 buf->dtb_xamot_drops = buf->dtb_drops; 8970 buf->dtb_xamot_offset = buf->dtb_offset; 8971 buf->dtb_xamot_errors = buf->dtb_errors; 8972 buf->dtb_xamot_flags = buf->dtb_flags; 8973 buf->dtb_offset = 0; 8974 buf->dtb_drops = 0; 8975 buf->dtb_errors = 0; 8976 buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED); 8977 dtrace_interrupt_enable(cookie); 8978 } 8979 8980 /* 8981 * Note: called from cross call context. This function activates a buffer 8982 * on a CPU. As with dtrace_buffer_switch(), the atomicity of the operation 8983 * is guaranteed by the disabling of interrupts. 8984 */ 8985 static void 8986 dtrace_buffer_activate(dtrace_state_t *state) 8987 { 8988 dtrace_buffer_t *buf; 8989 dtrace_icookie_t cookie = dtrace_interrupt_disable(); 8990 8991 buf = &state->dts_buffer[CPU->cpu_id]; 8992 8993 if (buf->dtb_tomax != NULL) { 8994 /* 8995 * We might like to assert that the buffer is marked inactive, 8996 * but this isn't necessarily true: the buffer for the CPU 8997 * that processes the BEGIN probe has its buffer activated 8998 * manually. In this case, we take the (harmless) action 8999 * re-clearing the bit INACTIVE bit. 9000 */ 9001 buf->dtb_flags &= ~DTRACEBUF_INACTIVE; 9002 } 9003 9004 dtrace_interrupt_enable(cookie); 9005 } 9006 9007 static int 9008 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags, 9009 processorid_t cpu) 9010 { 9011 cpu_t *cp; 9012 dtrace_buffer_t *buf; 9013 9014 ASSERT(MUTEX_HELD(&cpu_lock)); 9015 ASSERT(MUTEX_HELD(&dtrace_lock)); 9016 9017 if (crgetuid(CRED()) != 0 && size > dtrace_nonroot_maxsize) 9018 return (EFBIG); 9019 9020 cp = cpu_list; 9021 9022 do { 9023 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id) 9024 continue; 9025 9026 buf = &bufs[cp->cpu_id]; 9027 9028 /* 9029 * If there is already a buffer allocated for this CPU, it 9030 * is only possible that this is a DR event. In this case, 9031 * the buffer size must match our specified size. 9032 */ 9033 if (buf->dtb_tomax != NULL) { 9034 ASSERT(buf->dtb_size == size); 9035 continue; 9036 } 9037 9038 ASSERT(buf->dtb_xamot == NULL); 9039 9040 if ((buf->dtb_tomax = kmem_zalloc(size, KM_NOSLEEP)) == NULL) 9041 goto err; 9042 9043 buf->dtb_size = size; 9044 buf->dtb_flags = flags; 9045 buf->dtb_offset = 0; 9046 buf->dtb_drops = 0; 9047 9048 if (flags & DTRACEBUF_NOSWITCH) 9049 continue; 9050 9051 if ((buf->dtb_xamot = kmem_zalloc(size, KM_NOSLEEP)) == NULL) 9052 goto err; 9053 } while ((cp = cp->cpu_next) != cpu_list); 9054 9055 return (0); 9056 9057 err: 9058 cp = cpu_list; 9059 9060 do { 9061 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id) 9062 continue; 9063 9064 buf = &bufs[cp->cpu_id]; 9065 9066 if (buf->dtb_xamot != NULL) { 9067 ASSERT(buf->dtb_tomax != NULL); 9068 ASSERT(buf->dtb_size == size); 9069 kmem_free(buf->dtb_xamot, size); 9070 } 9071 9072 if (buf->dtb_tomax != NULL) { 9073 ASSERT(buf->dtb_size == size); 9074 kmem_free(buf->dtb_tomax, size); 9075 } 9076 9077 buf->dtb_tomax = NULL; 9078 buf->dtb_xamot = NULL; 9079 buf->dtb_size = 0; 9080 } while ((cp = cp->cpu_next) != cpu_list); 9081 9082 return (ENOMEM); 9083 } 9084 9085 /* 9086 * Note: called from probe context. This function just increments the drop 9087 * count on a buffer. It has been made a function to allow for the 9088 * possibility of understanding the source of mysterious drop counts. (A 9089 * problem for which one may be particularly disappointed that DTrace cannot 9090 * be used to understand DTrace.) 9091 */ 9092 static void 9093 dtrace_buffer_drop(dtrace_buffer_t *buf) 9094 { 9095 buf->dtb_drops++; 9096 } 9097 9098 /* 9099 * Note: called from probe context. This function is called to reserve space 9100 * in a buffer. If mstate is non-NULL, sets the scratch base and size in the 9101 * mstate. Returns the new offset in the buffer, or a negative value if an 9102 * error has occurred. 9103 */ 9104 static intptr_t 9105 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align, 9106 dtrace_state_t *state, dtrace_mstate_t *mstate) 9107 { 9108 intptr_t offs = buf->dtb_offset, soffs; 9109 intptr_t woffs; 9110 caddr_t tomax; 9111 size_t total; 9112 9113 if (buf->dtb_flags & DTRACEBUF_INACTIVE) 9114 return (-1); 9115 9116 if ((tomax = buf->dtb_tomax) == NULL) { 9117 dtrace_buffer_drop(buf); 9118 return (-1); 9119 } 9120 9121 if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) { 9122 while (offs & (align - 1)) { 9123 /* 9124 * Assert that our alignment is off by a number which 9125 * is itself sizeof (uint32_t) aligned. 9126 */ 9127 ASSERT(!((align - (offs & (align - 1))) & 9128 (sizeof (uint32_t) - 1))); 9129 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE); 9130 offs += sizeof (uint32_t); 9131 } 9132 9133 if ((soffs = offs + needed) > buf->dtb_size) { 9134 dtrace_buffer_drop(buf); 9135 return (-1); 9136 } 9137 9138 if (mstate == NULL) 9139 return (offs); 9140 9141 mstate->dtms_scratch_base = (uintptr_t)tomax + soffs; 9142 mstate->dtms_scratch_size = buf->dtb_size - soffs; 9143 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base; 9144 9145 return (offs); 9146 } 9147 9148 if (buf->dtb_flags & DTRACEBUF_FILL) { 9149 if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN && 9150 (buf->dtb_flags & DTRACEBUF_FULL)) 9151 return (-1); 9152 goto out; 9153 } 9154 9155 total = needed + (offs & (align - 1)); 9156 9157 /* 9158 * For a ring buffer, life is quite a bit more complicated. Before 9159 * we can store any padding, we need to adjust our wrapping offset. 9160 * (If we've never before wrapped or we're not about to, no adjustment 9161 * is required.) 9162 */ 9163 if ((buf->dtb_flags & DTRACEBUF_WRAPPED) || 9164 offs + total > buf->dtb_size) { 9165 woffs = buf->dtb_xamot_offset; 9166 9167 if (offs + total > buf->dtb_size) { 9168 /* 9169 * We can't fit in the end of the buffer. First, a 9170 * sanity check that we can fit in the buffer at all. 9171 */ 9172 if (total > buf->dtb_size) { 9173 dtrace_buffer_drop(buf); 9174 return (-1); 9175 } 9176 9177 /* 9178 * We're going to be storing at the top of the buffer, 9179 * so now we need to deal with the wrapped offset. We 9180 * only reset our wrapped offset to 0 if it is 9181 * currently greater than the current offset. If it 9182 * is less than the current offset, it is because a 9183 * previous allocation induced a wrap -- but the 9184 * allocation didn't subsequently take the space due 9185 * to an error or false predicate evaluation. In this 9186 * case, we'll just leave the wrapped offset alone: if 9187 * the wrapped offset hasn't been advanced far enough 9188 * for this allocation, it will be adjusted in the 9189 * lower loop. 9190 */ 9191 if (buf->dtb_flags & DTRACEBUF_WRAPPED) { 9192 if (woffs >= offs) 9193 woffs = 0; 9194 } else { 9195 woffs = 0; 9196 } 9197 9198 /* 9199 * Now we know that we're going to be storing to the 9200 * top of the buffer and that there is room for us 9201 * there. We need to clear the buffer from the current 9202 * offset to the end (there may be old gunk there). 9203 */ 9204 while (offs < buf->dtb_size) 9205 tomax[offs++] = 0; 9206 9207 /* 9208 * We need to set our offset to zero. And because we 9209 * are wrapping, we need to set the bit indicating as 9210 * much. We can also adjust our needed space back 9211 * down to the space required by the ECB -- we know 9212 * that the top of the buffer is aligned. 9213 */ 9214 offs = 0; 9215 total = needed; 9216 buf->dtb_flags |= DTRACEBUF_WRAPPED; 9217 } else { 9218 /* 9219 * There is room for us in the buffer, so we simply 9220 * need to check the wrapped offset. 9221 */ 9222 if (woffs < offs) { 9223 /* 9224 * The wrapped offset is less than the offset. 9225 * This can happen if we allocated buffer space 9226 * that induced a wrap, but then we didn't 9227 * subsequently take the space due to an error 9228 * or false predicate evaluation. This is 9229 * okay; we know that _this_ allocation isn't 9230 * going to induce a wrap. We still can't 9231 * reset the wrapped offset to be zero, 9232 * however: the space may have been trashed in 9233 * the previous failed probe attempt. But at 9234 * least the wrapped offset doesn't need to 9235 * be adjusted at all... 9236 */ 9237 goto out; 9238 } 9239 } 9240 9241 while (offs + total > woffs) { 9242 dtrace_epid_t epid = *(uint32_t *)(tomax + woffs); 9243 size_t size; 9244 9245 if (epid == DTRACE_EPIDNONE) { 9246 size = sizeof (uint32_t); 9247 } else { 9248 ASSERT(epid <= state->dts_necbs); 9249 ASSERT(state->dts_ecbs[epid - 1] != NULL); 9250 9251 size = state->dts_ecbs[epid - 1]->dte_size; 9252 } 9253 9254 ASSERT(woffs + size <= buf->dtb_size); 9255 ASSERT(size != 0); 9256 9257 if (woffs + size == buf->dtb_size) { 9258 /* 9259 * We've reached the end of the buffer; we want 9260 * to set the wrapped offset to 0 and break 9261 * out. However, if the offs is 0, then we're 9262 * in a strange edge-condition: the amount of 9263 * space that we want to reserve plus the size 9264 * of the record that we're overwriting is 9265 * greater than the size of the buffer. This 9266 * is problematic because if we reserve the 9267 * space but subsequently don't consume it (due 9268 * to a failed predicate or error) the wrapped 9269 * offset will be 0 -- yet the EPID at offset 0 9270 * will not be committed. This situation is 9271 * relatively easy to deal with: if we're in 9272 * this case, the buffer is indistinguishable 9273 * from one that hasn't wrapped; we need only 9274 * finish the job by clearing the wrapped bit, 9275 * explicitly setting the offset to be 0, and 9276 * zero'ing out the old data in the buffer. 9277 */ 9278 if (offs == 0) { 9279 buf->dtb_flags &= ~DTRACEBUF_WRAPPED; 9280 buf->dtb_offset = 0; 9281 woffs = total; 9282 9283 while (woffs < buf->dtb_size) 9284 tomax[woffs++] = 0; 9285 } 9286 9287 woffs = 0; 9288 break; 9289 } 9290 9291 woffs += size; 9292 } 9293 9294 /* 9295 * We have a wrapped offset. It may be that the wrapped offset 9296 * has become zero -- that's okay. 9297 */ 9298 buf->dtb_xamot_offset = woffs; 9299 } 9300 9301 out: 9302 /* 9303 * Now we can plow the buffer with any necessary padding. 9304 */ 9305 while (offs & (align - 1)) { 9306 /* 9307 * Assert that our alignment is off by a number which 9308 * is itself sizeof (uint32_t) aligned. 9309 */ 9310 ASSERT(!((align - (offs & (align - 1))) & 9311 (sizeof (uint32_t) - 1))); 9312 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE); 9313 offs += sizeof (uint32_t); 9314 } 9315 9316 if (buf->dtb_flags & DTRACEBUF_FILL) { 9317 if (offs + needed > buf->dtb_size - state->dts_reserve) { 9318 buf->dtb_flags |= DTRACEBUF_FULL; 9319 return (-1); 9320 } 9321 } 9322 9323 if (mstate == NULL) 9324 return (offs); 9325 9326 /* 9327 * For ring buffers and fill buffers, the scratch space is always 9328 * the inactive buffer. 9329 */ 9330 mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot; 9331 mstate->dtms_scratch_size = buf->dtb_size; 9332 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base; 9333 9334 return (offs); 9335 } 9336 9337 static void 9338 dtrace_buffer_polish(dtrace_buffer_t *buf) 9339 { 9340 ASSERT(buf->dtb_flags & DTRACEBUF_RING); 9341 ASSERT(MUTEX_HELD(&dtrace_lock)); 9342 9343 if (!(buf->dtb_flags & DTRACEBUF_WRAPPED)) 9344 return; 9345 9346 /* 9347 * We need to polish the ring buffer. There are three cases: 9348 * 9349 * - The first (and presumably most common) is that there is no gap 9350 * between the buffer offset and the wrapped offset. In this case, 9351 * there is nothing in the buffer that isn't valid data; we can 9352 * mark the buffer as polished and return. 9353 * 9354 * - The second (less common than the first but still more common 9355 * than the third) is that there is a gap between the buffer offset 9356 * and the wrapped offset, and the wrapped offset is larger than the 9357 * buffer offset. This can happen because of an alignment issue, or 9358 * can happen because of a call to dtrace_buffer_reserve() that 9359 * didn't subsequently consume the buffer space. In this case, 9360 * we need to zero the data from the buffer offset to the wrapped 9361 * offset. 9362 * 9363 * - The third (and least common) is that there is a gap between the 9364 * buffer offset and the wrapped offset, but the wrapped offset is 9365 * _less_ than the buffer offset. This can only happen because a 9366 * call to dtrace_buffer_reserve() induced a wrap, but the space 9367 * was not subsequently consumed. In this case, we need to zero the 9368 * space from the offset to the end of the buffer _and_ from the 9369 * top of the buffer to the wrapped offset. 9370 */ 9371 if (buf->dtb_offset < buf->dtb_xamot_offset) { 9372 bzero(buf->dtb_tomax + buf->dtb_offset, 9373 buf->dtb_xamot_offset - buf->dtb_offset); 9374 } 9375 9376 if (buf->dtb_offset > buf->dtb_xamot_offset) { 9377 bzero(buf->dtb_tomax + buf->dtb_offset, 9378 buf->dtb_size - buf->dtb_offset); 9379 bzero(buf->dtb_tomax, buf->dtb_xamot_offset); 9380 } 9381 } 9382 9383 static void 9384 dtrace_buffer_free(dtrace_buffer_t *bufs) 9385 { 9386 int i; 9387 9388 for (i = 0; i < NCPU; i++) { 9389 dtrace_buffer_t *buf = &bufs[i]; 9390 9391 if (buf->dtb_tomax == NULL) { 9392 ASSERT(buf->dtb_xamot == NULL); 9393 ASSERT(buf->dtb_size == 0); 9394 continue; 9395 } 9396 9397 if (buf->dtb_xamot != NULL) { 9398 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH)); 9399 kmem_free(buf->dtb_xamot, buf->dtb_size); 9400 } 9401 9402 kmem_free(buf->dtb_tomax, buf->dtb_size); 9403 buf->dtb_size = 0; 9404 buf->dtb_tomax = NULL; 9405 buf->dtb_xamot = NULL; 9406 } 9407 } 9408 9409 /* 9410 * DTrace Enabling Functions 9411 */ 9412 static dtrace_enabling_t * 9413 dtrace_enabling_create(dtrace_vstate_t *vstate) 9414 { 9415 dtrace_enabling_t *enab; 9416 9417 enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP); 9418 enab->dten_vstate = vstate; 9419 9420 return (enab); 9421 } 9422 9423 static void 9424 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb) 9425 { 9426 dtrace_ecbdesc_t **ndesc; 9427 size_t osize, nsize; 9428 9429 /* 9430 * We can't add to enablings after we've enabled them, or after we've 9431 * retained them. 9432 */ 9433 ASSERT(enab->dten_probegen == 0); 9434 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL); 9435 9436 if (enab->dten_ndesc < enab->dten_maxdesc) { 9437 enab->dten_desc[enab->dten_ndesc++] = ecb; 9438 return; 9439 } 9440 9441 osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *); 9442 9443 if (enab->dten_maxdesc == 0) { 9444 enab->dten_maxdesc = 1; 9445 } else { 9446 enab->dten_maxdesc <<= 1; 9447 } 9448 9449 ASSERT(enab->dten_ndesc < enab->dten_maxdesc); 9450 9451 nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *); 9452 ndesc = kmem_zalloc(nsize, KM_SLEEP); 9453 bcopy(enab->dten_desc, ndesc, osize); 9454 kmem_free(enab->dten_desc, osize); 9455 9456 enab->dten_desc = ndesc; 9457 enab->dten_desc[enab->dten_ndesc++] = ecb; 9458 } 9459 9460 static void 9461 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb, 9462 dtrace_probedesc_t *pd) 9463 { 9464 dtrace_ecbdesc_t *new; 9465 dtrace_predicate_t *pred; 9466 dtrace_actdesc_t *act; 9467 9468 /* 9469 * We're going to create a new ECB description that matches the 9470 * specified ECB in every way, but has the specified probe description. 9471 */ 9472 new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP); 9473 9474 if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL) 9475 dtrace_predicate_hold(pred); 9476 9477 for (act = ecb->dted_action; act != NULL; act = act->dtad_next) 9478 dtrace_actdesc_hold(act); 9479 9480 new->dted_action = ecb->dted_action; 9481 new->dted_pred = ecb->dted_pred; 9482 new->dted_probe = *pd; 9483 new->dted_uarg = ecb->dted_uarg; 9484 9485 dtrace_enabling_add(enab, new); 9486 } 9487 9488 static void 9489 dtrace_enabling_dump(dtrace_enabling_t *enab) 9490 { 9491 int i; 9492 9493 for (i = 0; i < enab->dten_ndesc; i++) { 9494 dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe; 9495 9496 cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i, 9497 desc->dtpd_provider, desc->dtpd_mod, 9498 desc->dtpd_func, desc->dtpd_name); 9499 } 9500 } 9501 9502 static void 9503 dtrace_enabling_destroy(dtrace_enabling_t *enab) 9504 { 9505 int i; 9506 dtrace_ecbdesc_t *ep; 9507 dtrace_vstate_t *vstate = enab->dten_vstate; 9508 9509 ASSERT(MUTEX_HELD(&dtrace_lock)); 9510 9511 for (i = 0; i < enab->dten_ndesc; i++) { 9512 dtrace_actdesc_t *act, *next; 9513 dtrace_predicate_t *pred; 9514 9515 ep = enab->dten_desc[i]; 9516 9517 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) 9518 dtrace_predicate_release(pred, vstate); 9519 9520 for (act = ep->dted_action; act != NULL; act = next) { 9521 next = act->dtad_next; 9522 dtrace_actdesc_release(act, vstate); 9523 } 9524 9525 kmem_free(ep, sizeof (dtrace_ecbdesc_t)); 9526 } 9527 9528 kmem_free(enab->dten_desc, 9529 enab->dten_maxdesc * sizeof (dtrace_enabling_t *)); 9530 9531 /* 9532 * If this was a retained enabling, decrement the dts_nretained count 9533 * and take it off of the dtrace_retained list. 9534 */ 9535 if (enab->dten_prev != NULL || enab->dten_next != NULL || 9536 dtrace_retained == enab) { 9537 ASSERT(enab->dten_vstate->dtvs_state != NULL); 9538 ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0); 9539 enab->dten_vstate->dtvs_state->dts_nretained--; 9540 } 9541 9542 if (enab->dten_prev == NULL) { 9543 if (dtrace_retained == enab) { 9544 dtrace_retained = enab->dten_next; 9545 9546 if (dtrace_retained != NULL) 9547 dtrace_retained->dten_prev = NULL; 9548 } 9549 } else { 9550 ASSERT(enab != dtrace_retained); 9551 ASSERT(dtrace_retained != NULL); 9552 enab->dten_prev->dten_next = enab->dten_next; 9553 } 9554 9555 if (enab->dten_next != NULL) { 9556 ASSERT(dtrace_retained != NULL); 9557 enab->dten_next->dten_prev = enab->dten_prev; 9558 } 9559 9560 kmem_free(enab, sizeof (dtrace_enabling_t)); 9561 } 9562 9563 static int 9564 dtrace_enabling_retain(dtrace_enabling_t *enab) 9565 { 9566 dtrace_state_t *state; 9567 9568 ASSERT(MUTEX_HELD(&dtrace_lock)); 9569 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL); 9570 ASSERT(enab->dten_vstate != NULL); 9571 9572 state = enab->dten_vstate->dtvs_state; 9573 ASSERT(state != NULL); 9574 9575 /* 9576 * We only allow each state to retain dtrace_retain_max enablings. 9577 */ 9578 if (state->dts_nretained >= dtrace_retain_max) 9579 return (ENOSPC); 9580 9581 state->dts_nretained++; 9582 9583 if (dtrace_retained == NULL) { 9584 dtrace_retained = enab; 9585 return (0); 9586 } 9587 9588 enab->dten_next = dtrace_retained; 9589 dtrace_retained->dten_prev = enab; 9590 dtrace_retained = enab; 9591 9592 return (0); 9593 } 9594 9595 static int 9596 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match, 9597 dtrace_probedesc_t *create) 9598 { 9599 dtrace_enabling_t *new, *enab; 9600 int found = 0, err = ENOENT; 9601 9602 ASSERT(MUTEX_HELD(&dtrace_lock)); 9603 ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN); 9604 ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN); 9605 ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN); 9606 ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN); 9607 9608 new = dtrace_enabling_create(&state->dts_vstate); 9609 9610 /* 9611 * Iterate over all retained enablings, looking for enablings that 9612 * match the specified state. 9613 */ 9614 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) { 9615 int i; 9616 9617 /* 9618 * dtvs_state can only be NULL for helper enablings -- and 9619 * helper enablings can't be retained. 9620 */ 9621 ASSERT(enab->dten_vstate->dtvs_state != NULL); 9622 9623 if (enab->dten_vstate->dtvs_state != state) 9624 continue; 9625 9626 /* 9627 * Now iterate over each probe description; we're looking for 9628 * an exact match to the specified probe description. 9629 */ 9630 for (i = 0; i < enab->dten_ndesc; i++) { 9631 dtrace_ecbdesc_t *ep = enab->dten_desc[i]; 9632 dtrace_probedesc_t *pd = &ep->dted_probe; 9633 9634 if (strcmp(pd->dtpd_provider, match->dtpd_provider)) 9635 continue; 9636 9637 if (strcmp(pd->dtpd_mod, match->dtpd_mod)) 9638 continue; 9639 9640 if (strcmp(pd->dtpd_func, match->dtpd_func)) 9641 continue; 9642 9643 if (strcmp(pd->dtpd_name, match->dtpd_name)) 9644 continue; 9645 9646 /* 9647 * We have a winning probe! Add it to our growing 9648 * enabling. 9649 */ 9650 found = 1; 9651 dtrace_enabling_addlike(new, ep, create); 9652 } 9653 } 9654 9655 if (!found || (err = dtrace_enabling_retain(new)) != 0) { 9656 dtrace_enabling_destroy(new); 9657 return (err); 9658 } 9659 9660 return (0); 9661 } 9662 9663 static void 9664 dtrace_enabling_retract(dtrace_state_t *state) 9665 { 9666 dtrace_enabling_t *enab, *next; 9667 9668 ASSERT(MUTEX_HELD(&dtrace_lock)); 9669 9670 /* 9671 * Iterate over all retained enablings, destroy the enablings retained 9672 * for the specified state. 9673 */ 9674 for (enab = dtrace_retained; enab != NULL; enab = next) { 9675 next = enab->dten_next; 9676 9677 /* 9678 * dtvs_state can only be NULL for helper enablings -- and 9679 * helper enablings can't be retained. 9680 */ 9681 ASSERT(enab->dten_vstate->dtvs_state != NULL); 9682 9683 if (enab->dten_vstate->dtvs_state == state) { 9684 ASSERT(state->dts_nretained > 0); 9685 dtrace_enabling_destroy(enab); 9686 } 9687 } 9688 9689 ASSERT(state->dts_nretained == 0); 9690 } 9691 9692 static int 9693 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched) 9694 { 9695 int i = 0; 9696 int matched = 0; 9697 9698 ASSERT(MUTEX_HELD(&cpu_lock)); 9699 ASSERT(MUTEX_HELD(&dtrace_lock)); 9700 9701 for (i = 0; i < enab->dten_ndesc; i++) { 9702 dtrace_ecbdesc_t *ep = enab->dten_desc[i]; 9703 9704 enab->dten_current = ep; 9705 enab->dten_error = 0; 9706 9707 matched += dtrace_probe_enable(&ep->dted_probe, enab); 9708 9709 if (enab->dten_error != 0) { 9710 /* 9711 * If we get an error half-way through enabling the 9712 * probes, we kick out -- perhaps with some number of 9713 * them enabled. Leaving enabled probes enabled may 9714 * be slightly confusing for user-level, but we expect 9715 * that no one will attempt to actually drive on in 9716 * the face of such errors. If this is an anonymous 9717 * enabling (indicated with a NULL nmatched pointer), 9718 * we cmn_err() a message. We aren't expecting to 9719 * get such an error -- such as it can exist at all, 9720 * it would be a result of corrupted DOF in the driver 9721 * properties. 9722 */ 9723 if (nmatched == NULL) { 9724 cmn_err(CE_WARN, "dtrace_enabling_match() " 9725 "error on %p: %d", (void *)ep, 9726 enab->dten_error); 9727 } 9728 9729 return (enab->dten_error); 9730 } 9731 } 9732 9733 enab->dten_probegen = dtrace_probegen; 9734 if (nmatched != NULL) 9735 *nmatched = matched; 9736 9737 return (0); 9738 } 9739 9740 static void 9741 dtrace_enabling_matchall(void) 9742 { 9743 dtrace_enabling_t *enab; 9744 9745 mutex_enter(&cpu_lock); 9746 mutex_enter(&dtrace_lock); 9747 9748 /* 9749 * Because we can be called after dtrace_detach() has been called, we 9750 * cannot assert that there are retained enablings. We can safely 9751 * load from dtrace_retained, however: the taskq_destroy() at the 9752 * end of dtrace_detach() will block pending our completion. 9753 */ 9754 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) 9755 (void) dtrace_enabling_match(enab, NULL); 9756 9757 mutex_exit(&dtrace_lock); 9758 mutex_exit(&cpu_lock); 9759 } 9760 9761 static int 9762 dtrace_enabling_matchstate(dtrace_state_t *state, int *nmatched) 9763 { 9764 dtrace_enabling_t *enab; 9765 int matched, total = 0, err; 9766 9767 ASSERT(MUTEX_HELD(&cpu_lock)); 9768 ASSERT(MUTEX_HELD(&dtrace_lock)); 9769 9770 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) { 9771 ASSERT(enab->dten_vstate->dtvs_state != NULL); 9772 9773 if (enab->dten_vstate->dtvs_state != state) 9774 continue; 9775 9776 if ((err = dtrace_enabling_match(enab, &matched)) != 0) 9777 return (err); 9778 9779 total += matched; 9780 } 9781 9782 if (nmatched != NULL) 9783 *nmatched = total; 9784 9785 return (0); 9786 } 9787 9788 /* 9789 * If an enabling is to be enabled without having matched probes (that is, if 9790 * dtrace_state_go() is to be called on the underlying dtrace_state_t), the 9791 * enabling must be _primed_ by creating an ECB for every ECB description. 9792 * This must be done to assure that we know the number of speculations, the 9793 * number of aggregations, the minimum buffer size needed, etc. before we 9794 * transition out of DTRACE_ACTIVITY_INACTIVE. To do this without actually 9795 * enabling any probes, we create ECBs for every ECB decription, but with a 9796 * NULL probe -- which is exactly what this function does. 9797 */ 9798 static void 9799 dtrace_enabling_prime(dtrace_state_t *state) 9800 { 9801 dtrace_enabling_t *enab; 9802 int i; 9803 9804 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) { 9805 ASSERT(enab->dten_vstate->dtvs_state != NULL); 9806 9807 if (enab->dten_vstate->dtvs_state != state) 9808 continue; 9809 9810 /* 9811 * We don't want to prime an enabling more than once, lest 9812 * we allow a malicious user to induce resource exhaustion. 9813 * (The ECBs that result from priming an enabling aren't 9814 * leaked -- but they also aren't deallocated until the 9815 * consumer state is destroyed.) 9816 */ 9817 if (enab->dten_primed) 9818 continue; 9819 9820 for (i = 0; i < enab->dten_ndesc; i++) { 9821 enab->dten_current = enab->dten_desc[i]; 9822 (void) dtrace_probe_enable(NULL, enab); 9823 } 9824 9825 enab->dten_primed = 1; 9826 } 9827 } 9828 9829 /* 9830 * Called to indicate that probes should be provided due to retained 9831 * enablings. This is implemented in terms of dtrace_probe_provide(), but it 9832 * must take an initial lap through the enabling calling the dtps_provide() 9833 * entry point explicitly to allow for autocreated probes. 9834 */ 9835 static void 9836 dtrace_enabling_provide(dtrace_provider_t *prv) 9837 { 9838 int i, all = 0; 9839 dtrace_probedesc_t desc; 9840 9841 ASSERT(MUTEX_HELD(&dtrace_lock)); 9842 ASSERT(MUTEX_HELD(&dtrace_provider_lock)); 9843 9844 if (prv == NULL) { 9845 all = 1; 9846 prv = dtrace_provider; 9847 } 9848 9849 do { 9850 dtrace_enabling_t *enab = dtrace_retained; 9851 void *parg = prv->dtpv_arg; 9852 9853 for (; enab != NULL; enab = enab->dten_next) { 9854 for (i = 0; i < enab->dten_ndesc; i++) { 9855 desc = enab->dten_desc[i]->dted_probe; 9856 mutex_exit(&dtrace_lock); 9857 prv->dtpv_pops.dtps_provide(parg, &desc); 9858 mutex_enter(&dtrace_lock); 9859 } 9860 } 9861 } while (all && (prv = prv->dtpv_next) != NULL); 9862 9863 mutex_exit(&dtrace_lock); 9864 dtrace_probe_provide(NULL, all ? NULL : prv); 9865 mutex_enter(&dtrace_lock); 9866 } 9867 9868 /* 9869 * DTrace DOF Functions 9870 */ 9871 /*ARGSUSED*/ 9872 static void 9873 dtrace_dof_error(dof_hdr_t *dof, const char *str) 9874 { 9875 if (dtrace_err_verbose) 9876 cmn_err(CE_WARN, "failed to process DOF: %s", str); 9877 9878 #ifdef DTRACE_ERRDEBUG 9879 dtrace_errdebug(str); 9880 #endif 9881 } 9882 9883 /* 9884 * Create DOF out of a currently enabled state. Right now, we only create 9885 * DOF containing the run-time options -- but this could be expanded to create 9886 * complete DOF representing the enabled state. 9887 */ 9888 static dof_hdr_t * 9889 dtrace_dof_create(dtrace_state_t *state) 9890 { 9891 dof_hdr_t *dof; 9892 dof_sec_t *sec; 9893 dof_optdesc_t *opt; 9894 int i, len = sizeof (dof_hdr_t) + 9895 roundup(sizeof (dof_sec_t), sizeof (uint64_t)) + 9896 sizeof (dof_optdesc_t) * DTRACEOPT_MAX; 9897 9898 ASSERT(MUTEX_HELD(&dtrace_lock)); 9899 9900 dof = kmem_zalloc(len, KM_SLEEP); 9901 dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0; 9902 dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1; 9903 dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2; 9904 dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3; 9905 9906 dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE; 9907 dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE; 9908 dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION_1; 9909 dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION; 9910 dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS; 9911 dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS; 9912 9913 dof->dofh_flags = 0; 9914 dof->dofh_hdrsize = sizeof (dof_hdr_t); 9915 dof->dofh_secsize = sizeof (dof_sec_t); 9916 dof->dofh_secnum = 1; /* only DOF_SECT_OPTDESC */ 9917 dof->dofh_secoff = sizeof (dof_hdr_t); 9918 dof->dofh_loadsz = len; 9919 dof->dofh_filesz = len; 9920 dof->dofh_pad = 0; 9921 9922 /* 9923 * Fill in the option section header... 9924 */ 9925 sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t)); 9926 sec->dofs_type = DOF_SECT_OPTDESC; 9927 sec->dofs_align = sizeof (uint64_t); 9928 sec->dofs_flags = DOF_SECF_LOAD; 9929 sec->dofs_entsize = sizeof (dof_optdesc_t); 9930 9931 opt = (dof_optdesc_t *)((uintptr_t)sec + 9932 roundup(sizeof (dof_sec_t), sizeof (uint64_t))); 9933 9934 sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof; 9935 sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX; 9936 9937 for (i = 0; i < DTRACEOPT_MAX; i++) { 9938 opt[i].dofo_option = i; 9939 opt[i].dofo_strtab = DOF_SECIDX_NONE; 9940 opt[i].dofo_value = state->dts_options[i]; 9941 } 9942 9943 return (dof); 9944 } 9945 9946 static dof_hdr_t * 9947 dtrace_dof_copyin(uintptr_t uarg, int *errp) 9948 { 9949 dof_hdr_t hdr, *dof; 9950 9951 ASSERT(!MUTEX_HELD(&dtrace_lock)); 9952 9953 /* 9954 * First, we're going to copyin() the sizeof (dof_hdr_t). 9955 */ 9956 if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) { 9957 dtrace_dof_error(NULL, "failed to copyin DOF header"); 9958 *errp = EFAULT; 9959 return (NULL); 9960 } 9961 9962 /* 9963 * Now we'll allocate the entire DOF and copy it in -- provided 9964 * that the length isn't outrageous. 9965 */ 9966 if (hdr.dofh_loadsz >= dtrace_dof_maxsize) { 9967 dtrace_dof_error(&hdr, "load size exceeds maximum"); 9968 *errp = E2BIG; 9969 return (NULL); 9970 } 9971 9972 if (hdr.dofh_loadsz < sizeof (hdr)) { 9973 dtrace_dof_error(&hdr, "invalid load size"); 9974 *errp = EINVAL; 9975 return (NULL); 9976 } 9977 9978 dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP); 9979 9980 if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0) { 9981 kmem_free(dof, hdr.dofh_loadsz); 9982 *errp = EFAULT; 9983 return (NULL); 9984 } 9985 9986 return (dof); 9987 } 9988 9989 static dof_hdr_t * 9990 dtrace_dof_property(const char *name) 9991 { 9992 uchar_t *buf; 9993 uint64_t loadsz; 9994 unsigned int len, i; 9995 dof_hdr_t *dof; 9996 9997 /* 9998 * Unfortunately, array of values in .conf files are always (and 9999 * only) interpreted to be integer arrays. We must read our DOF 10000 * as an integer array, and then squeeze it into a byte array. 10001 */ 10002 if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0, 10003 (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS) 10004 return (NULL); 10005 10006 for (i = 0; i < len; i++) 10007 buf[i] = (uchar_t)(((int *)buf)[i]); 10008 10009 if (len < sizeof (dof_hdr_t)) { 10010 ddi_prop_free(buf); 10011 dtrace_dof_error(NULL, "truncated header"); 10012 return (NULL); 10013 } 10014 10015 if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) { 10016 ddi_prop_free(buf); 10017 dtrace_dof_error(NULL, "truncated DOF"); 10018 return (NULL); 10019 } 10020 10021 if (loadsz >= dtrace_dof_maxsize) { 10022 ddi_prop_free(buf); 10023 dtrace_dof_error(NULL, "oversized DOF"); 10024 return (NULL); 10025 } 10026 10027 dof = kmem_alloc(loadsz, KM_SLEEP); 10028 bcopy(buf, dof, loadsz); 10029 ddi_prop_free(buf); 10030 10031 return (dof); 10032 } 10033 10034 static void 10035 dtrace_dof_destroy(dof_hdr_t *dof) 10036 { 10037 kmem_free(dof, dof->dofh_loadsz); 10038 } 10039 10040 /* 10041 * Return the dof_sec_t pointer corresponding to a given section index. If the 10042 * index is not valid, dtrace_dof_error() is called and NULL is returned. If 10043 * a type other than DOF_SECT_NONE is specified, the header is checked against 10044 * this type and NULL is returned if the types do not match. 10045 */ 10046 static dof_sec_t * 10047 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i) 10048 { 10049 dof_sec_t *sec = (dof_sec_t *)(uintptr_t) 10050 ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize); 10051 10052 if (i >= dof->dofh_secnum) { 10053 dtrace_dof_error(dof, "referenced section index is invalid"); 10054 return (NULL); 10055 } 10056 10057 if (!(sec->dofs_flags & DOF_SECF_LOAD)) { 10058 dtrace_dof_error(dof, "referenced section is not loadable"); 10059 return (NULL); 10060 } 10061 10062 if (type != DOF_SECT_NONE && type != sec->dofs_type) { 10063 dtrace_dof_error(dof, "referenced section is the wrong type"); 10064 return (NULL); 10065 } 10066 10067 return (sec); 10068 } 10069 10070 static dtrace_probedesc_t * 10071 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc) 10072 { 10073 dof_probedesc_t *probe; 10074 dof_sec_t *strtab; 10075 uintptr_t daddr = (uintptr_t)dof; 10076 uintptr_t str; 10077 size_t size; 10078 10079 if (sec->dofs_type != DOF_SECT_PROBEDESC) { 10080 dtrace_dof_error(dof, "invalid probe section"); 10081 return (NULL); 10082 } 10083 10084 if (sec->dofs_align != sizeof (dof_secidx_t)) { 10085 dtrace_dof_error(dof, "bad alignment in probe description"); 10086 return (NULL); 10087 } 10088 10089 if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) { 10090 dtrace_dof_error(dof, "truncated probe description"); 10091 return (NULL); 10092 } 10093 10094 probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset); 10095 strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab); 10096 10097 if (strtab == NULL) 10098 return (NULL); 10099 10100 str = daddr + strtab->dofs_offset; 10101 size = strtab->dofs_size; 10102 10103 if (probe->dofp_provider >= strtab->dofs_size) { 10104 dtrace_dof_error(dof, "corrupt probe provider"); 10105 return (NULL); 10106 } 10107 10108 (void) strncpy(desc->dtpd_provider, 10109 (char *)(str + probe->dofp_provider), 10110 MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider)); 10111 10112 if (probe->dofp_mod >= strtab->dofs_size) { 10113 dtrace_dof_error(dof, "corrupt probe module"); 10114 return (NULL); 10115 } 10116 10117 (void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod), 10118 MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod)); 10119 10120 if (probe->dofp_func >= strtab->dofs_size) { 10121 dtrace_dof_error(dof, "corrupt probe function"); 10122 return (NULL); 10123 } 10124 10125 (void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func), 10126 MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func)); 10127 10128 if (probe->dofp_name >= strtab->dofs_size) { 10129 dtrace_dof_error(dof, "corrupt probe name"); 10130 return (NULL); 10131 } 10132 10133 (void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name), 10134 MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name)); 10135 10136 return (desc); 10137 } 10138 10139 static dtrace_difo_t * 10140 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate, 10141 cred_t *cr) 10142 { 10143 dtrace_difo_t *dp; 10144 size_t ttl = 0; 10145 dof_difohdr_t *dofd; 10146 uintptr_t daddr = (uintptr_t)dof; 10147 size_t max = dtrace_difo_maxsize; 10148 int i, l, n; 10149 10150 static const struct { 10151 int section; 10152 int bufoffs; 10153 int lenoffs; 10154 int entsize; 10155 int align; 10156 const char *msg; 10157 } difo[] = { 10158 { DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf), 10159 offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t), 10160 sizeof (dif_instr_t), "multiple DIF sections" }, 10161 10162 { DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab), 10163 offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t), 10164 sizeof (uint64_t), "multiple integer tables" }, 10165 10166 { DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab), 10167 offsetof(dtrace_difo_t, dtdo_strlen), 0, 10168 sizeof (char), "multiple string tables" }, 10169 10170 { DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab), 10171 offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t), 10172 sizeof (uint_t), "multiple variable tables" }, 10173 10174 { DOF_SECT_NONE, 0, 0, 0, NULL } 10175 }; 10176 10177 if (sec->dofs_type != DOF_SECT_DIFOHDR) { 10178 dtrace_dof_error(dof, "invalid DIFO header section"); 10179 return (NULL); 10180 } 10181 10182 if (sec->dofs_align != sizeof (dof_secidx_t)) { 10183 dtrace_dof_error(dof, "bad alignment in DIFO header"); 10184 return (NULL); 10185 } 10186 10187 if (sec->dofs_size < sizeof (dof_difohdr_t) || 10188 sec->dofs_size % sizeof (dof_secidx_t)) { 10189 dtrace_dof_error(dof, "bad size in DIFO header"); 10190 return (NULL); 10191 } 10192 10193 dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset); 10194 n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1; 10195 10196 dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP); 10197 dp->dtdo_rtype = dofd->dofd_rtype; 10198 10199 for (l = 0; l < n; l++) { 10200 dof_sec_t *subsec; 10201 void **bufp; 10202 uint32_t *lenp; 10203 10204 if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE, 10205 dofd->dofd_links[l])) == NULL) 10206 goto err; /* invalid section link */ 10207 10208 if (ttl + subsec->dofs_size > max) { 10209 dtrace_dof_error(dof, "exceeds maximum size"); 10210 goto err; 10211 } 10212 10213 ttl += subsec->dofs_size; 10214 10215 for (i = 0; difo[i].section != DOF_SECT_NONE; i++) { 10216 if (subsec->dofs_type != difo[i].section) 10217 continue; 10218 10219 if (!(subsec->dofs_flags & DOF_SECF_LOAD)) { 10220 dtrace_dof_error(dof, "section not loaded"); 10221 goto err; 10222 } 10223 10224 if (subsec->dofs_align != difo[i].align) { 10225 dtrace_dof_error(dof, "bad alignment"); 10226 goto err; 10227 } 10228 10229 bufp = (void **)((uintptr_t)dp + difo[i].bufoffs); 10230 lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs); 10231 10232 if (*bufp != NULL) { 10233 dtrace_dof_error(dof, difo[i].msg); 10234 goto err; 10235 } 10236 10237 if (difo[i].entsize != subsec->dofs_entsize) { 10238 dtrace_dof_error(dof, "entry size mismatch"); 10239 goto err; 10240 } 10241 10242 if (subsec->dofs_entsize != 0 && 10243 (subsec->dofs_size % subsec->dofs_entsize) != 0) { 10244 dtrace_dof_error(dof, "corrupt entry size"); 10245 goto err; 10246 } 10247 10248 *lenp = subsec->dofs_size; 10249 *bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP); 10250 bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset), 10251 *bufp, subsec->dofs_size); 10252 10253 if (subsec->dofs_entsize != 0) 10254 *lenp /= subsec->dofs_entsize; 10255 10256 break; 10257 } 10258 10259 /* 10260 * If we encounter a loadable DIFO sub-section that is not 10261 * known to us, assume this is a broken program and fail. 10262 */ 10263 if (difo[i].section == DOF_SECT_NONE && 10264 (subsec->dofs_flags & DOF_SECF_LOAD)) { 10265 dtrace_dof_error(dof, "unrecognized DIFO subsection"); 10266 goto err; 10267 } 10268 } 10269 10270 if (dp->dtdo_buf == NULL) { 10271 /* 10272 * We can't have a DIF object without DIF text. 10273 */ 10274 dtrace_dof_error(dof, "missing DIF text"); 10275 goto err; 10276 } 10277 10278 /* 10279 * Before we validate the DIF object, run through the variable table 10280 * looking for the strings -- if any of their size are under, we'll set 10281 * their size to be the system-wide default string size. Note that 10282 * this should _not_ happen if the "strsize" option has been set -- 10283 * in this case, the compiler should have set the size to reflect the 10284 * setting of the option. 10285 */ 10286 for (i = 0; i < dp->dtdo_varlen; i++) { 10287 dtrace_difv_t *v = &dp->dtdo_vartab[i]; 10288 dtrace_diftype_t *t = &v->dtdv_type; 10289 10290 if (v->dtdv_id < DIF_VAR_OTHER_UBASE) 10291 continue; 10292 10293 if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0) 10294 t->dtdt_size = dtrace_strsize_default; 10295 } 10296 10297 if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0) 10298 goto err; 10299 10300 dtrace_difo_init(dp, vstate); 10301 return (dp); 10302 10303 err: 10304 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t)); 10305 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t)); 10306 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen); 10307 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t)); 10308 10309 kmem_free(dp, sizeof (dtrace_difo_t)); 10310 return (NULL); 10311 } 10312 10313 static dtrace_predicate_t * 10314 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate, 10315 cred_t *cr) 10316 { 10317 dtrace_difo_t *dp; 10318 10319 if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL) 10320 return (NULL); 10321 10322 return (dtrace_predicate_create(dp)); 10323 } 10324 10325 static dtrace_actdesc_t * 10326 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate, 10327 cred_t *cr) 10328 { 10329 dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next; 10330 dof_actdesc_t *desc; 10331 dof_sec_t *difosec; 10332 size_t offs; 10333 uintptr_t daddr = (uintptr_t)dof; 10334 uint64_t arg; 10335 dtrace_actkind_t kind; 10336 10337 if (sec->dofs_type != DOF_SECT_ACTDESC) { 10338 dtrace_dof_error(dof, "invalid action section"); 10339 return (NULL); 10340 } 10341 10342 if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) { 10343 dtrace_dof_error(dof, "truncated action description"); 10344 return (NULL); 10345 } 10346 10347 if (sec->dofs_align != sizeof (uint64_t)) { 10348 dtrace_dof_error(dof, "bad alignment in action description"); 10349 return (NULL); 10350 } 10351 10352 if (sec->dofs_size < sec->dofs_entsize) { 10353 dtrace_dof_error(dof, "section entry size exceeds total size"); 10354 return (NULL); 10355 } 10356 10357 if (sec->dofs_entsize != sizeof (dof_actdesc_t)) { 10358 dtrace_dof_error(dof, "bad entry size in action description"); 10359 return (NULL); 10360 } 10361 10362 if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) { 10363 dtrace_dof_error(dof, "actions exceed dtrace_actions_max"); 10364 return (NULL); 10365 } 10366 10367 for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) { 10368 desc = (dof_actdesc_t *)(daddr + 10369 (uintptr_t)sec->dofs_offset + offs); 10370 kind = (dtrace_actkind_t)desc->dofa_kind; 10371 10372 if (DTRACEACT_ISPRINTFLIKE(kind) && 10373 (kind != DTRACEACT_PRINTA || 10374 desc->dofa_strtab != DOF_SECIDX_NONE)) { 10375 dof_sec_t *strtab; 10376 char *str, *fmt; 10377 uint64_t i; 10378 10379 /* 10380 * printf()-like actions must have a format string. 10381 */ 10382 if ((strtab = dtrace_dof_sect(dof, 10383 DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL) 10384 goto err; 10385 10386 str = (char *)((uintptr_t)dof + 10387 (uintptr_t)strtab->dofs_offset); 10388 10389 for (i = desc->dofa_arg; i < strtab->dofs_size; i++) { 10390 if (str[i] == '\0') 10391 break; 10392 } 10393 10394 if (i >= strtab->dofs_size) { 10395 dtrace_dof_error(dof, "bogus format string"); 10396 goto err; 10397 } 10398 10399 if (i == desc->dofa_arg) { 10400 dtrace_dof_error(dof, "empty format string"); 10401 goto err; 10402 } 10403 10404 i -= desc->dofa_arg; 10405 fmt = kmem_alloc(i + 1, KM_SLEEP); 10406 bcopy(&str[desc->dofa_arg], fmt, i + 1); 10407 arg = (uint64_t)(uintptr_t)fmt; 10408 } else { 10409 if (kind == DTRACEACT_PRINTA) { 10410 ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE); 10411 arg = 0; 10412 } else { 10413 arg = desc->dofa_arg; 10414 } 10415 } 10416 10417 act = dtrace_actdesc_create(kind, desc->dofa_ntuple, 10418 desc->dofa_uarg, arg); 10419 10420 if (last != NULL) { 10421 last->dtad_next = act; 10422 } else { 10423 first = act; 10424 } 10425 10426 last = act; 10427 10428 if (desc->dofa_difo == DOF_SECIDX_NONE) 10429 continue; 10430 10431 if ((difosec = dtrace_dof_sect(dof, 10432 DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL) 10433 goto err; 10434 10435 act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr); 10436 10437 if (act->dtad_difo == NULL) 10438 goto err; 10439 } 10440 10441 ASSERT(first != NULL); 10442 return (first); 10443 10444 err: 10445 for (act = first; act != NULL; act = next) { 10446 next = act->dtad_next; 10447 dtrace_actdesc_release(act, vstate); 10448 } 10449 10450 return (NULL); 10451 } 10452 10453 static dtrace_ecbdesc_t * 10454 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate, 10455 cred_t *cr) 10456 { 10457 dtrace_ecbdesc_t *ep; 10458 dof_ecbdesc_t *ecb; 10459 dtrace_probedesc_t *desc; 10460 dtrace_predicate_t *pred = NULL; 10461 10462 if (sec->dofs_size < sizeof (dof_ecbdesc_t)) { 10463 dtrace_dof_error(dof, "truncated ECB description"); 10464 return (NULL); 10465 } 10466 10467 if (sec->dofs_align != sizeof (uint64_t)) { 10468 dtrace_dof_error(dof, "bad alignment in ECB description"); 10469 return (NULL); 10470 } 10471 10472 ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset); 10473 sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes); 10474 10475 if (sec == NULL) 10476 return (NULL); 10477 10478 ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP); 10479 ep->dted_uarg = ecb->dofe_uarg; 10480 desc = &ep->dted_probe; 10481 10482 if (dtrace_dof_probedesc(dof, sec, desc) == NULL) 10483 goto err; 10484 10485 if (ecb->dofe_pred != DOF_SECIDX_NONE) { 10486 if ((sec = dtrace_dof_sect(dof, 10487 DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL) 10488 goto err; 10489 10490 if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL) 10491 goto err; 10492 10493 ep->dted_pred.dtpdd_predicate = pred; 10494 } 10495 10496 if (ecb->dofe_actions != DOF_SECIDX_NONE) { 10497 if ((sec = dtrace_dof_sect(dof, 10498 DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL) 10499 goto err; 10500 10501 ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr); 10502 10503 if (ep->dted_action == NULL) 10504 goto err; 10505 } 10506 10507 return (ep); 10508 10509 err: 10510 if (pred != NULL) 10511 dtrace_predicate_release(pred, vstate); 10512 kmem_free(ep, sizeof (dtrace_ecbdesc_t)); 10513 return (NULL); 10514 } 10515 10516 /* 10517 * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the 10518 * specified DOF. At present, this amounts to simply adding 'ubase' to the 10519 * site of any user SETX relocations to account for load object base address. 10520 * In the future, if we need other relocations, this function can be extended. 10521 */ 10522 static int 10523 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase) 10524 { 10525 uintptr_t daddr = (uintptr_t)dof; 10526 dof_relohdr_t *dofr = 10527 (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset); 10528 dof_sec_t *ss, *rs, *ts; 10529 dof_relodesc_t *r; 10530 uint_t i, n; 10531 10532 if (sec->dofs_size < sizeof (dof_relohdr_t) || 10533 sec->dofs_align != sizeof (dof_secidx_t)) { 10534 dtrace_dof_error(dof, "invalid relocation header"); 10535 return (-1); 10536 } 10537 10538 ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab); 10539 rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec); 10540 ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec); 10541 10542 if (ss == NULL || rs == NULL || ts == NULL) 10543 return (-1); /* dtrace_dof_error() has been called already */ 10544 10545 if (rs->dofs_entsize < sizeof (dof_relodesc_t) || 10546 rs->dofs_align != sizeof (uint64_t)) { 10547 dtrace_dof_error(dof, "invalid relocation section"); 10548 return (-1); 10549 } 10550 10551 r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset); 10552 n = rs->dofs_size / rs->dofs_entsize; 10553 10554 for (i = 0; i < n; i++) { 10555 uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset; 10556 10557 switch (r->dofr_type) { 10558 case DOF_RELO_NONE: 10559 break; 10560 case DOF_RELO_SETX: 10561 if (r->dofr_offset >= ts->dofs_size || r->dofr_offset + 10562 sizeof (uint64_t) > ts->dofs_size) { 10563 dtrace_dof_error(dof, "bad relocation offset"); 10564 return (-1); 10565 } 10566 10567 if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) { 10568 dtrace_dof_error(dof, "misaligned setx relo"); 10569 return (-1); 10570 } 10571 10572 *(uint64_t *)taddr += ubase; 10573 break; 10574 default: 10575 dtrace_dof_error(dof, "invalid relocation type"); 10576 return (-1); 10577 } 10578 10579 r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize); 10580 } 10581 10582 return (0); 10583 } 10584 10585 /* 10586 * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated 10587 * header: it should be at the front of a memory region that is at least 10588 * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in 10589 * size. It need not be validated in any other way. 10590 */ 10591 static int 10592 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr, 10593 dtrace_enabling_t **enabp, uint64_t ubase, int noprobes) 10594 { 10595 uint64_t len = dof->dofh_loadsz, seclen; 10596 uintptr_t daddr = (uintptr_t)dof; 10597 dtrace_ecbdesc_t *ep; 10598 dtrace_enabling_t *enab; 10599 uint_t i; 10600 10601 ASSERT(MUTEX_HELD(&dtrace_lock)); 10602 ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t)); 10603 10604 /* 10605 * Check the DOF header identification bytes. In addition to checking 10606 * valid settings, we also verify that unused bits/bytes are zeroed so 10607 * we can use them later without fear of regressing existing binaries. 10608 */ 10609 if (bcmp(&dof->dofh_ident[DOF_ID_MAG0], 10610 DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) { 10611 dtrace_dof_error(dof, "DOF magic string mismatch"); 10612 return (-1); 10613 } 10614 10615 if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 && 10616 dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) { 10617 dtrace_dof_error(dof, "DOF has invalid data model"); 10618 return (-1); 10619 } 10620 10621 if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) { 10622 dtrace_dof_error(dof, "DOF encoding mismatch"); 10623 return (-1); 10624 } 10625 10626 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) { 10627 dtrace_dof_error(dof, "DOF version mismatch"); 10628 return (-1); 10629 } 10630 10631 if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) { 10632 dtrace_dof_error(dof, "DOF uses unsupported instruction set"); 10633 return (-1); 10634 } 10635 10636 if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) { 10637 dtrace_dof_error(dof, "DOF uses too many integer registers"); 10638 return (-1); 10639 } 10640 10641 if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) { 10642 dtrace_dof_error(dof, "DOF uses too many tuple registers"); 10643 return (-1); 10644 } 10645 10646 for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) { 10647 if (dof->dofh_ident[i] != 0) { 10648 dtrace_dof_error(dof, "DOF has invalid ident byte set"); 10649 return (-1); 10650 } 10651 } 10652 10653 if (dof->dofh_flags & ~DOF_FL_VALID) { 10654 dtrace_dof_error(dof, "DOF has invalid flag bits set"); 10655 return (-1); 10656 } 10657 10658 if (dof->dofh_secsize == 0) { 10659 dtrace_dof_error(dof, "zero section header size"); 10660 return (-1); 10661 } 10662 10663 /* 10664 * Check that the section headers don't exceed the amount of DOF 10665 * data. Note that we cast the section size and number of sections 10666 * to uint64_t's to prevent possible overflow in the multiplication. 10667 */ 10668 seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize; 10669 10670 if (dof->dofh_secoff > len || seclen > len || 10671 dof->dofh_secoff + seclen > len) { 10672 dtrace_dof_error(dof, "truncated section headers"); 10673 return (-1); 10674 } 10675 10676 if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) { 10677 dtrace_dof_error(dof, "misaligned section headers"); 10678 return (-1); 10679 } 10680 10681 if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) { 10682 dtrace_dof_error(dof, "misaligned section size"); 10683 return (-1); 10684 } 10685 10686 /* 10687 * Take an initial pass through the section headers to be sure that 10688 * the headers don't have stray offsets. If the 'noprobes' flag is 10689 * set, do not permit sections relating to providers, probes, or args. 10690 */ 10691 for (i = 0; i < dof->dofh_secnum; i++) { 10692 dof_sec_t *sec = (dof_sec_t *)(daddr + 10693 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize); 10694 10695 if (noprobes) { 10696 switch (sec->dofs_type) { 10697 case DOF_SECT_PROVIDER: 10698 case DOF_SECT_PROBES: 10699 case DOF_SECT_PRARGS: 10700 case DOF_SECT_PROFFS: 10701 dtrace_dof_error(dof, "illegal sections " 10702 "for enabling"); 10703 return (-1); 10704 } 10705 } 10706 10707 if (!(sec->dofs_flags & DOF_SECF_LOAD)) 10708 continue; /* just ignore non-loadable sections */ 10709 10710 if (sec->dofs_align & (sec->dofs_align - 1)) { 10711 dtrace_dof_error(dof, "bad section alignment"); 10712 return (-1); 10713 } 10714 10715 if (sec->dofs_offset & (sec->dofs_align - 1)) { 10716 dtrace_dof_error(dof, "misaligned section"); 10717 return (-1); 10718 } 10719 10720 if (sec->dofs_offset > len || sec->dofs_size > len || 10721 sec->dofs_offset + sec->dofs_size > len) { 10722 dtrace_dof_error(dof, "corrupt section header"); 10723 return (-1); 10724 } 10725 10726 if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr + 10727 sec->dofs_offset + sec->dofs_size - 1) != '\0') { 10728 dtrace_dof_error(dof, "non-terminating string table"); 10729 return (-1); 10730 } 10731 } 10732 10733 /* 10734 * Take a second pass through the sections and locate and perform any 10735 * relocations that are present. We do this after the first pass to 10736 * be sure that all sections have had their headers validated. 10737 */ 10738 for (i = 0; i < dof->dofh_secnum; i++) { 10739 dof_sec_t *sec = (dof_sec_t *)(daddr + 10740 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize); 10741 10742 if (!(sec->dofs_flags & DOF_SECF_LOAD)) 10743 continue; /* skip sections that are not loadable */ 10744 10745 switch (sec->dofs_type) { 10746 case DOF_SECT_URELHDR: 10747 if (dtrace_dof_relocate(dof, sec, ubase) != 0) 10748 return (-1); 10749 break; 10750 } 10751 } 10752 10753 if ((enab = *enabp) == NULL) 10754 enab = *enabp = dtrace_enabling_create(vstate); 10755 10756 for (i = 0; i < dof->dofh_secnum; i++) { 10757 dof_sec_t *sec = (dof_sec_t *)(daddr + 10758 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize); 10759 10760 if (sec->dofs_type != DOF_SECT_ECBDESC) 10761 continue; 10762 10763 if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) { 10764 dtrace_enabling_destroy(enab); 10765 *enabp = NULL; 10766 return (-1); 10767 } 10768 10769 dtrace_enabling_add(enab, ep); 10770 } 10771 10772 return (0); 10773 } 10774 10775 /* 10776 * Process DOF for any options. This routine assumes that the DOF has been 10777 * at least processed by dtrace_dof_slurp(). 10778 */ 10779 static int 10780 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state) 10781 { 10782 int i, rval; 10783 uint32_t entsize; 10784 size_t offs; 10785 dof_optdesc_t *desc; 10786 10787 for (i = 0; i < dof->dofh_secnum; i++) { 10788 dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof + 10789 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize); 10790 10791 if (sec->dofs_type != DOF_SECT_OPTDESC) 10792 continue; 10793 10794 if (sec->dofs_align != sizeof (uint64_t)) { 10795 dtrace_dof_error(dof, "bad alignment in " 10796 "option description"); 10797 return (EINVAL); 10798 } 10799 10800 if ((entsize = sec->dofs_entsize) == 0) { 10801 dtrace_dof_error(dof, "zeroed option entry size"); 10802 return (EINVAL); 10803 } 10804 10805 if (entsize < sizeof (dof_optdesc_t)) { 10806 dtrace_dof_error(dof, "bad option entry size"); 10807 return (EINVAL); 10808 } 10809 10810 for (offs = 0; offs < sec->dofs_size; offs += entsize) { 10811 desc = (dof_optdesc_t *)((uintptr_t)dof + 10812 (uintptr_t)sec->dofs_offset + offs); 10813 10814 if (desc->dofo_strtab != DOF_SECIDX_NONE) { 10815 dtrace_dof_error(dof, "non-zero option string"); 10816 return (EINVAL); 10817 } 10818 10819 if (desc->dofo_value == DTRACEOPT_UNSET) { 10820 dtrace_dof_error(dof, "unset option"); 10821 return (EINVAL); 10822 } 10823 10824 if ((rval = dtrace_state_option(state, 10825 desc->dofo_option, desc->dofo_value)) != 0) { 10826 dtrace_dof_error(dof, "rejected option"); 10827 return (rval); 10828 } 10829 } 10830 } 10831 10832 return (0); 10833 } 10834 10835 /* 10836 * DTrace Consumer State Functions 10837 */ 10838 int 10839 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size) 10840 { 10841 size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize; 10842 void *base; 10843 uintptr_t limit; 10844 dtrace_dynvar_t *dvar, *next, *start; 10845 int i; 10846 10847 ASSERT(MUTEX_HELD(&dtrace_lock)); 10848 ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL); 10849 10850 bzero(dstate, sizeof (dtrace_dstate_t)); 10851 10852 if ((dstate->dtds_chunksize = chunksize) == 0) 10853 dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE; 10854 10855 if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t))) 10856 size = min; 10857 10858 if ((base = kmem_zalloc(size, KM_NOSLEEP)) == NULL) 10859 return (ENOMEM); 10860 10861 dstate->dtds_size = size; 10862 dstate->dtds_base = base; 10863 dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP); 10864 bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t)); 10865 10866 hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)); 10867 10868 if (hashsize != 1 && (hashsize & 1)) 10869 hashsize--; 10870 10871 dstate->dtds_hashsize = hashsize; 10872 dstate->dtds_hash = dstate->dtds_base; 10873 10874 /* 10875 * Determine number of active CPUs. Divide free list evenly among 10876 * active CPUs. 10877 */ 10878 start = (dtrace_dynvar_t *) 10879 ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t)); 10880 limit = (uintptr_t)base + size; 10881 10882 maxper = (limit - (uintptr_t)start) / NCPU; 10883 maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize; 10884 10885 for (i = 0; i < NCPU; i++) { 10886 dstate->dtds_percpu[i].dtdsc_free = dvar = start; 10887 10888 /* 10889 * If we don't even have enough chunks to make it once through 10890 * NCPUs, we're just going to allocate everything to the first 10891 * CPU. And if we're on the last CPU, we're going to allocate 10892 * whatever is left over. In either case, we set the limit to 10893 * be the limit of the dynamic variable space. 10894 */ 10895 if (maxper == 0 || i == NCPU - 1) { 10896 limit = (uintptr_t)base + size; 10897 start = NULL; 10898 } else { 10899 limit = (uintptr_t)start + maxper; 10900 start = (dtrace_dynvar_t *)limit; 10901 } 10902 10903 ASSERT(limit <= (uintptr_t)base + size); 10904 10905 for (;;) { 10906 next = (dtrace_dynvar_t *)((uintptr_t)dvar + 10907 dstate->dtds_chunksize); 10908 10909 if ((uintptr_t)next + dstate->dtds_chunksize >= limit) 10910 break; 10911 10912 dvar->dtdv_next = next; 10913 dvar = next; 10914 } 10915 10916 if (maxper == 0) 10917 break; 10918 } 10919 10920 return (0); 10921 } 10922 10923 void 10924 dtrace_dstate_fini(dtrace_dstate_t *dstate) 10925 { 10926 ASSERT(MUTEX_HELD(&cpu_lock)); 10927 10928 if (dstate->dtds_base == NULL) 10929 return; 10930 10931 kmem_free(dstate->dtds_base, dstate->dtds_size); 10932 kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu); 10933 } 10934 10935 static void 10936 dtrace_vstate_fini(dtrace_vstate_t *vstate) 10937 { 10938 /* 10939 * Logical XOR, where are you? 10940 */ 10941 ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL)); 10942 10943 if (vstate->dtvs_nglobals > 0) { 10944 kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals * 10945 sizeof (dtrace_statvar_t *)); 10946 } 10947 10948 if (vstate->dtvs_ntlocals > 0) { 10949 kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals * 10950 sizeof (dtrace_difv_t)); 10951 } 10952 10953 ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL)); 10954 10955 if (vstate->dtvs_nlocals > 0) { 10956 kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals * 10957 sizeof (dtrace_statvar_t *)); 10958 } 10959 } 10960 10961 static void 10962 dtrace_state_clean(dtrace_state_t *state) 10963 { 10964 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) 10965 return; 10966 10967 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars); 10968 dtrace_speculation_clean(state); 10969 } 10970 10971 static void 10972 dtrace_state_deadman(dtrace_state_t *state) 10973 { 10974 hrtime_t now; 10975 10976 dtrace_sync(); 10977 10978 now = dtrace_gethrtime(); 10979 10980 if (state != dtrace_anon.dta_state && 10981 now - state->dts_laststatus >= dtrace_deadman_user) 10982 return; 10983 10984 /* 10985 * We must be sure that dts_alive never appears to be less than the 10986 * value upon entry to dtrace_state_deadman(), and because we lack a 10987 * dtrace_cas64(), we cannot store to it atomically. We thus instead 10988 * store INT64_MAX to it, followed by a memory barrier, followed by 10989 * the new value. This assures that dts_alive never appears to be 10990 * less than its true value, regardless of the order in which the 10991 * stores to the underlying storage are issued. 10992 */ 10993 state->dts_alive = INT64_MAX; 10994 dtrace_membar_producer(); 10995 state->dts_alive = now; 10996 } 10997 10998 dtrace_state_t * 10999 dtrace_state_create(dev_t *devp, cred_t *cr) 11000 { 11001 minor_t minor; 11002 major_t major; 11003 char c[30]; 11004 dtrace_state_t *state; 11005 dtrace_optval_t *opt; 11006 int bufsize = NCPU * sizeof (dtrace_buffer_t), i; 11007 11008 ASSERT(MUTEX_HELD(&dtrace_lock)); 11009 ASSERT(MUTEX_HELD(&cpu_lock)); 11010 11011 minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1, 11012 VM_BESTFIT | VM_SLEEP); 11013 11014 if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) { 11015 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1); 11016 return (NULL); 11017 } 11018 11019 state = ddi_get_soft_state(dtrace_softstate, minor); 11020 state->dts_epid = DTRACE_EPIDNONE + 1; 11021 11022 (void) snprintf(c, sizeof (c), "dtrace_aggid_%d", minor); 11023 state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1, 11024 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER); 11025 11026 if (devp != NULL) { 11027 major = getemajor(*devp); 11028 } else { 11029 major = ddi_driver_major(dtrace_devi); 11030 } 11031 11032 state->dts_dev = makedevice(major, minor); 11033 11034 if (devp != NULL) 11035 *devp = state->dts_dev; 11036 11037 /* 11038 * We allocate NCPU buffers. On the one hand, this can be quite 11039 * a bit of memory per instance (nearly 36K on a Starcat). On the 11040 * other hand, it saves an additional memory reference in the probe 11041 * path. 11042 */ 11043 state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP); 11044 state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP); 11045 state->dts_cleaner = CYCLIC_NONE; 11046 state->dts_deadman = CYCLIC_NONE; 11047 state->dts_vstate.dtvs_state = state; 11048 11049 for (i = 0; i < DTRACEOPT_MAX; i++) 11050 state->dts_options[i] = DTRACEOPT_UNSET; 11051 11052 /* 11053 * Set the default options. 11054 */ 11055 opt = state->dts_options; 11056 opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH; 11057 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO; 11058 opt[DTRACEOPT_NSPEC] = dtrace_nspec_default; 11059 opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default; 11060 opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL; 11061 opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default; 11062 opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default; 11063 opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default; 11064 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default; 11065 opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default; 11066 opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default; 11067 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default; 11068 opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default; 11069 opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default; 11070 11071 state->dts_activity = DTRACE_ACTIVITY_INACTIVE; 11072 11073 /* 11074 * Set up the credentials for this instantiation. 11075 */ 11076 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) { 11077 state->dts_cred.dcr_visible = DTRACE_CRV_ALL; 11078 state->dts_cred.dcr_action = DTRACE_CRA_ALL; 11079 } else { 11080 state->dts_cred.dcr_uid = crgetuid(cr); 11081 state->dts_cred.dcr_gid = crgetgid(cr); 11082 11083 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) || 11084 PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) { 11085 state->dts_cred.dcr_action |= DTRACE_CRA_PROC; 11086 } 11087 11088 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) && 11089 PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) { 11090 state->dts_cred.dcr_visible |= DTRACE_CRV_ALLPROC; 11091 state->dts_cred.dcr_action |= 11092 DTRACE_CRA_PROC_DESTRUCTIVE; 11093 } 11094 11095 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) { 11096 state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL | 11097 DTRACE_CRV_ALLPROC; 11098 state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL | 11099 DTRACE_CRA_PROC; 11100 11101 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) 11102 state->dts_cred.dcr_action |= 11103 DTRACE_CRA_PROC_DESTRUCTIVE; 11104 } 11105 } 11106 11107 return (state); 11108 } 11109 11110 static int 11111 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which) 11112 { 11113 dtrace_optval_t *opt = state->dts_options, size; 11114 processorid_t cpu; 11115 int flags = 0, rval; 11116 11117 ASSERT(MUTEX_HELD(&dtrace_lock)); 11118 ASSERT(MUTEX_HELD(&cpu_lock)); 11119 ASSERT(which < DTRACEOPT_MAX); 11120 ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE || 11121 (state == dtrace_anon.dta_state && 11122 state->dts_activity == DTRACE_ACTIVITY_ACTIVE)); 11123 11124 if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0) 11125 return (0); 11126 11127 if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET) 11128 cpu = opt[DTRACEOPT_CPU]; 11129 11130 if (which == DTRACEOPT_SPECSIZE) 11131 flags |= DTRACEBUF_NOSWITCH; 11132 11133 if (which == DTRACEOPT_BUFSIZE) { 11134 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING) 11135 flags |= DTRACEBUF_RING; 11136 11137 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL) 11138 flags |= DTRACEBUF_FILL; 11139 11140 flags |= DTRACEBUF_INACTIVE; 11141 } 11142 11143 for (size = opt[which]; size >= sizeof (uint64_t); size >>= 1) { 11144 /* 11145 * The size must be 8-byte aligned. If the size is not 8-byte 11146 * aligned, drop it down by the difference. 11147 */ 11148 if (size & (sizeof (uint64_t) - 1)) 11149 size -= size & (sizeof (uint64_t) - 1); 11150 11151 if (size < state->dts_reserve) { 11152 /* 11153 * Buffers always must be large enough to accommodate 11154 * their prereserved space. We return E2BIG instead 11155 * of ENOMEM in this case to allow for user-level 11156 * software to differentiate the cases. 11157 */ 11158 return (E2BIG); 11159 } 11160 11161 rval = dtrace_buffer_alloc(buf, size, flags, cpu); 11162 11163 if (rval != ENOMEM) { 11164 opt[which] = size; 11165 return (rval); 11166 } 11167 11168 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL) 11169 return (rval); 11170 } 11171 11172 return (ENOMEM); 11173 } 11174 11175 static int 11176 dtrace_state_buffers(dtrace_state_t *state) 11177 { 11178 dtrace_speculation_t *spec = state->dts_speculations; 11179 int rval, i; 11180 11181 if ((rval = dtrace_state_buffer(state, state->dts_buffer, 11182 DTRACEOPT_BUFSIZE)) != 0) 11183 return (rval); 11184 11185 if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer, 11186 DTRACEOPT_AGGSIZE)) != 0) 11187 return (rval); 11188 11189 for (i = 0; i < state->dts_nspeculations; i++) { 11190 if ((rval = dtrace_state_buffer(state, 11191 spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0) 11192 return (rval); 11193 } 11194 11195 return (0); 11196 } 11197 11198 static void 11199 dtrace_state_prereserve(dtrace_state_t *state) 11200 { 11201 dtrace_ecb_t *ecb; 11202 dtrace_probe_t *probe; 11203 11204 state->dts_reserve = 0; 11205 11206 if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL) 11207 return; 11208 11209 /* 11210 * If our buffer policy is a "fill" buffer policy, we need to set the 11211 * prereserved space to be the space required by the END probes. 11212 */ 11213 probe = dtrace_probes[dtrace_probeid_end - 1]; 11214 ASSERT(probe != NULL); 11215 11216 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) { 11217 if (ecb->dte_state != state) 11218 continue; 11219 11220 state->dts_reserve += ecb->dte_needed + ecb->dte_alignment; 11221 } 11222 } 11223 11224 static int 11225 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu) 11226 { 11227 dtrace_optval_t *opt = state->dts_options, sz, nspec; 11228 dtrace_speculation_t *spec; 11229 dtrace_buffer_t *buf; 11230 cyc_handler_t hdlr; 11231 cyc_time_t when; 11232 int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t); 11233 dtrace_icookie_t cookie; 11234 11235 mutex_enter(&cpu_lock); 11236 mutex_enter(&dtrace_lock); 11237 11238 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) { 11239 rval = EBUSY; 11240 goto out; 11241 } 11242 11243 /* 11244 * Before we can perform any checks, we must prime all of the 11245 * retained enablings that correspond to this state. 11246 */ 11247 dtrace_enabling_prime(state); 11248 11249 if (state->dts_destructive && !state->dts_cred.dcr_destructive) { 11250 rval = EACCES; 11251 goto out; 11252 } 11253 11254 dtrace_state_prereserve(state); 11255 11256 /* 11257 * Now we want to do is try to allocate our speculations. 11258 * We do not automatically resize the number of speculations; if 11259 * this fails, we will fail the operation. 11260 */ 11261 nspec = opt[DTRACEOPT_NSPEC]; 11262 ASSERT(nspec != DTRACEOPT_UNSET); 11263 11264 if (nspec > INT_MAX) { 11265 rval = ENOMEM; 11266 goto out; 11267 } 11268 11269 spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t), KM_NOSLEEP); 11270 11271 if (spec == NULL) { 11272 rval = ENOMEM; 11273 goto out; 11274 } 11275 11276 state->dts_speculations = spec; 11277 state->dts_nspeculations = (int)nspec; 11278 11279 for (i = 0; i < nspec; i++) { 11280 if ((buf = kmem_zalloc(bufsize, KM_NOSLEEP)) == NULL) { 11281 rval = ENOMEM; 11282 goto err; 11283 } 11284 11285 spec[i].dtsp_buffer = buf; 11286 } 11287 11288 if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) { 11289 if (dtrace_anon.dta_state == NULL) { 11290 rval = ENOENT; 11291 goto out; 11292 } 11293 11294 if (state->dts_necbs != 0) { 11295 rval = EALREADY; 11296 goto out; 11297 } 11298 11299 state->dts_anon = dtrace_anon_grab(); 11300 ASSERT(state->dts_anon != NULL); 11301 state = state->dts_anon; 11302 11303 /* 11304 * We want "grabanon" to be set in the grabbed state, so we'll 11305 * copy that option value from the grabbing state into the 11306 * grabbed state. 11307 */ 11308 state->dts_options[DTRACEOPT_GRABANON] = 11309 opt[DTRACEOPT_GRABANON]; 11310 11311 *cpu = dtrace_anon.dta_beganon; 11312 11313 /* 11314 * If the anonymous state is active (as it almost certainly 11315 * is if the anonymous enabling ultimately matched anything), 11316 * we don't allow any further option processing -- but we 11317 * don't return failure. 11318 */ 11319 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) 11320 goto out; 11321 } 11322 11323 if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET && 11324 opt[DTRACEOPT_AGGSIZE] != 0) { 11325 if (state->dts_aggregations == NULL) { 11326 /* 11327 * We're not going to create an aggregation buffer 11328 * because we don't have any ECBs that contain 11329 * aggregations -- set this option to 0. 11330 */ 11331 opt[DTRACEOPT_AGGSIZE] = 0; 11332 } else { 11333 /* 11334 * If we have an aggregation buffer, we must also have 11335 * a buffer to use as scratch. 11336 */ 11337 if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET || 11338 opt[DTRACEOPT_BUFSIZE] < state->dts_needed) { 11339 opt[DTRACEOPT_BUFSIZE] = state->dts_needed; 11340 } 11341 } 11342 } 11343 11344 if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET && 11345 opt[DTRACEOPT_SPECSIZE] != 0) { 11346 if (!state->dts_speculates) { 11347 /* 11348 * We're not going to create speculation buffers 11349 * because we don't have any ECBs that actually 11350 * speculate -- set the speculation size to 0. 11351 */ 11352 opt[DTRACEOPT_SPECSIZE] = 0; 11353 } 11354 } 11355 11356 /* 11357 * The bare minimum size for any buffer that we're actually going to 11358 * do anything to is sizeof (uint64_t). 11359 */ 11360 sz = sizeof (uint64_t); 11361 11362 if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) || 11363 (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) || 11364 (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) { 11365 /* 11366 * A buffer size has been explicitly set to 0 (or to a size 11367 * that will be adjusted to 0) and we need the space -- we 11368 * need to return failure. We return ENOSPC to differentiate 11369 * it from failing to allocate a buffer due to failure to meet 11370 * the reserve (for which we return E2BIG). 11371 */ 11372 rval = ENOSPC; 11373 goto out; 11374 } 11375 11376 if ((rval = dtrace_state_buffers(state)) != 0) 11377 goto err; 11378 11379 if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET) 11380 sz = dtrace_dstate_defsize; 11381 11382 do { 11383 rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz); 11384 11385 if (rval == 0) 11386 break; 11387 11388 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL) 11389 goto err; 11390 } while (sz >>= 1); 11391 11392 opt[DTRACEOPT_DYNVARSIZE] = sz; 11393 11394 if (rval != 0) 11395 goto err; 11396 11397 if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max) 11398 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max; 11399 11400 if (opt[DTRACEOPT_CLEANRATE] == 0) 11401 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max; 11402 11403 if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min) 11404 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min; 11405 11406 if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max) 11407 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max; 11408 11409 hdlr.cyh_func = (cyc_func_t)dtrace_state_clean; 11410 hdlr.cyh_arg = state; 11411 hdlr.cyh_level = CY_LOW_LEVEL; 11412 11413 when.cyt_when = 0; 11414 when.cyt_interval = opt[DTRACEOPT_CLEANRATE]; 11415 11416 state->dts_cleaner = cyclic_add(&hdlr, &when); 11417 11418 hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman; 11419 hdlr.cyh_arg = state; 11420 hdlr.cyh_level = CY_LOW_LEVEL; 11421 11422 when.cyt_when = 0; 11423 when.cyt_interval = dtrace_deadman_interval; 11424 11425 state->dts_alive = state->dts_laststatus = dtrace_gethrtime(); 11426 state->dts_deadman = cyclic_add(&hdlr, &when); 11427 11428 state->dts_activity = DTRACE_ACTIVITY_WARMUP; 11429 11430 /* 11431 * Now it's time to actually fire the BEGIN probe. We need to disable 11432 * interrupts here both to record the CPU on which we fired the BEGIN 11433 * probe (the data from this CPU will be processed first at user 11434 * level) and to manually activate the buffer for this CPU. 11435 */ 11436 cookie = dtrace_interrupt_disable(); 11437 *cpu = CPU->cpu_id; 11438 ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE); 11439 state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE; 11440 11441 dtrace_probe(dtrace_probeid_begin, 11442 (uint64_t)(uintptr_t)state, 0, 0, 0, 0); 11443 dtrace_interrupt_enable(cookie); 11444 /* 11445 * We may have had an exit action from a BEGIN probe; only change our 11446 * state to ACTIVE if we're still in WARMUP. 11447 */ 11448 ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP || 11449 state->dts_activity == DTRACE_ACTIVITY_DRAINING); 11450 11451 if (state->dts_activity == DTRACE_ACTIVITY_WARMUP) 11452 state->dts_activity = DTRACE_ACTIVITY_ACTIVE; 11453 11454 /* 11455 * Regardless of whether or not now we're in ACTIVE or DRAINING, we 11456 * want each CPU to transition its principal buffer out of the 11457 * INACTIVE state. Doing this assures that no CPU will suddenly begin 11458 * processing an ECB halfway down a probe's ECB chain; all CPUs will 11459 * atomically transition from processing none of a state's ECBs to 11460 * processing all of them. 11461 */ 11462 dtrace_xcall(DTRACE_CPUALL, 11463 (dtrace_xcall_t)dtrace_buffer_activate, state); 11464 goto out; 11465 11466 err: 11467 dtrace_buffer_free(state->dts_buffer); 11468 dtrace_buffer_free(state->dts_aggbuffer); 11469 11470 if ((nspec = state->dts_nspeculations) == 0) { 11471 ASSERT(state->dts_speculations == NULL); 11472 goto out; 11473 } 11474 11475 spec = state->dts_speculations; 11476 ASSERT(spec != NULL); 11477 11478 for (i = 0; i < state->dts_nspeculations; i++) { 11479 if ((buf = spec[i].dtsp_buffer) == NULL) 11480 break; 11481 11482 dtrace_buffer_free(buf); 11483 kmem_free(buf, bufsize); 11484 } 11485 11486 kmem_free(spec, nspec * sizeof (dtrace_speculation_t)); 11487 state->dts_nspeculations = 0; 11488 state->dts_speculations = NULL; 11489 11490 out: 11491 mutex_exit(&dtrace_lock); 11492 mutex_exit(&cpu_lock); 11493 11494 return (rval); 11495 } 11496 11497 static int 11498 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu) 11499 { 11500 dtrace_icookie_t cookie; 11501 11502 ASSERT(MUTEX_HELD(&dtrace_lock)); 11503 11504 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE && 11505 state->dts_activity != DTRACE_ACTIVITY_DRAINING) 11506 return (EINVAL); 11507 11508 /* 11509 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync 11510 * to be sure that every CPU has seen it. See below for the details 11511 * on why this is done. 11512 */ 11513 state->dts_activity = DTRACE_ACTIVITY_DRAINING; 11514 dtrace_sync(); 11515 11516 /* 11517 * By this point, it is impossible for any CPU to be still processing 11518 * with DTRACE_ACTIVITY_ACTIVE. We can thus set our activity to 11519 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any 11520 * other CPU in dtrace_buffer_reserve(). This allows dtrace_probe() 11521 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN 11522 * iff we're in the END probe. 11523 */ 11524 state->dts_activity = DTRACE_ACTIVITY_COOLDOWN; 11525 dtrace_sync(); 11526 ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN); 11527 11528 /* 11529 * Finally, we can release the reserve and call the END probe. We 11530 * disable interrupts across calling the END probe to allow us to 11531 * return the CPU on which we actually called the END probe. This 11532 * allows user-land to be sure that this CPU's principal buffer is 11533 * processed last. 11534 */ 11535 state->dts_reserve = 0; 11536 11537 cookie = dtrace_interrupt_disable(); 11538 *cpu = CPU->cpu_id; 11539 dtrace_probe(dtrace_probeid_end, 11540 (uint64_t)(uintptr_t)state, 0, 0, 0, 0); 11541 dtrace_interrupt_enable(cookie); 11542 11543 state->dts_activity = DTRACE_ACTIVITY_STOPPED; 11544 dtrace_sync(); 11545 11546 return (0); 11547 } 11548 11549 static int 11550 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option, 11551 dtrace_optval_t val) 11552 { 11553 ASSERT(MUTEX_HELD(&dtrace_lock)); 11554 11555 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) 11556 return (EBUSY); 11557 11558 if (option >= DTRACEOPT_MAX) 11559 return (EINVAL); 11560 11561 if (option != DTRACEOPT_CPU && val < 0) 11562 return (EINVAL); 11563 11564 switch (option) { 11565 case DTRACEOPT_DESTRUCTIVE: 11566 if (dtrace_destructive_disallow) 11567 return (EACCES); 11568 11569 state->dts_cred.dcr_destructive = 1; 11570 break; 11571 11572 case DTRACEOPT_BUFSIZE: 11573 case DTRACEOPT_DYNVARSIZE: 11574 case DTRACEOPT_AGGSIZE: 11575 case DTRACEOPT_SPECSIZE: 11576 case DTRACEOPT_STRSIZE: 11577 if (val < 0) 11578 return (EINVAL); 11579 11580 if (val >= LONG_MAX) { 11581 /* 11582 * If this is an otherwise negative value, set it to 11583 * the highest multiple of 128m less than LONG_MAX. 11584 * Technically, we're adjusting the size without 11585 * regard to the buffer resizing policy, but in fact, 11586 * this has no effect -- if we set the buffer size to 11587 * ~LONG_MAX and the buffer policy is ultimately set to 11588 * be "manual", the buffer allocation is guaranteed to 11589 * fail, if only because the allocation requires two 11590 * buffers. (We set the the size to the highest 11591 * multiple of 128m because it ensures that the size 11592 * will remain a multiple of a megabyte when 11593 * repeatedly halved -- all the way down to 15m.) 11594 */ 11595 val = LONG_MAX - (1 << 27) + 1; 11596 } 11597 } 11598 11599 state->dts_options[option] = val; 11600 11601 return (0); 11602 } 11603 11604 static void 11605 dtrace_state_destroy(dtrace_state_t *state) 11606 { 11607 dtrace_ecb_t *ecb; 11608 dtrace_vstate_t *vstate = &state->dts_vstate; 11609 minor_t minor = getminor(state->dts_dev); 11610 int i, bufsize = NCPU * sizeof (dtrace_buffer_t); 11611 dtrace_speculation_t *spec = state->dts_speculations; 11612 int nspec = state->dts_nspeculations; 11613 uint32_t match; 11614 11615 ASSERT(MUTEX_HELD(&dtrace_lock)); 11616 ASSERT(MUTEX_HELD(&cpu_lock)); 11617 11618 /* 11619 * First, retract any retained enablings for this state. 11620 */ 11621 dtrace_enabling_retract(state); 11622 ASSERT(state->dts_nretained == 0); 11623 11624 /* 11625 * Now we need to disable and destroy any enabled probes. Because any 11626 * DTRACE_PRIV_KERNEL probes may actually be slowing our progress 11627 * (especially if they're all enabled), we take two passes through 11628 * the ECBs: in the first, we disable just DTRACE_PRIV_KERNEL probes, 11629 * and in the second we disable whatever is left over. 11630 */ 11631 for (match = DTRACE_PRIV_KERNEL; ; match = 0) { 11632 for (i = 0; i < state->dts_necbs; i++) { 11633 if ((ecb = state->dts_ecbs[i]) == NULL) 11634 continue; 11635 11636 if (match && ecb->dte_probe != NULL) { 11637 dtrace_probe_t *probe = ecb->dte_probe; 11638 dtrace_provider_t *prov = probe->dtpr_provider; 11639 11640 if (!(prov->dtpv_priv.dtpp_flags & match)) 11641 continue; 11642 } 11643 11644 dtrace_ecb_disable(ecb); 11645 dtrace_ecb_destroy(ecb); 11646 } 11647 11648 if (!match) 11649 break; 11650 } 11651 11652 /* 11653 * Before we free the buffers, perform one more sync to assure that 11654 * every CPU is out of probe context. 11655 */ 11656 dtrace_sync(); 11657 11658 dtrace_buffer_free(state->dts_buffer); 11659 dtrace_buffer_free(state->dts_aggbuffer); 11660 11661 for (i = 0; i < nspec; i++) 11662 dtrace_buffer_free(spec[i].dtsp_buffer); 11663 11664 if (state->dts_cleaner != CYCLIC_NONE) 11665 cyclic_remove(state->dts_cleaner); 11666 11667 if (state->dts_deadman != CYCLIC_NONE) 11668 cyclic_remove(state->dts_deadman); 11669 11670 dtrace_dstate_fini(&vstate->dtvs_dynvars); 11671 dtrace_vstate_fini(vstate); 11672 kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *)); 11673 11674 if (state->dts_aggregations != NULL) { 11675 #ifdef DEBUG 11676 for (i = 0; i < state->dts_naggregations; i++) 11677 ASSERT(state->dts_aggregations[i] == NULL); 11678 #endif 11679 ASSERT(state->dts_naggregations > 0); 11680 kmem_free(state->dts_aggregations, 11681 state->dts_naggregations * sizeof (dtrace_aggregation_t *)); 11682 } 11683 11684 kmem_free(state->dts_buffer, bufsize); 11685 kmem_free(state->dts_aggbuffer, bufsize); 11686 11687 for (i = 0; i < nspec; i++) 11688 kmem_free(spec[i].dtsp_buffer, bufsize); 11689 11690 kmem_free(spec, nspec * sizeof (dtrace_speculation_t)); 11691 11692 dtrace_format_destroy(state); 11693 11694 vmem_destroy(state->dts_aggid_arena); 11695 ddi_soft_state_free(dtrace_softstate, minor); 11696 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1); 11697 } 11698 11699 /* 11700 * DTrace Anonymous Enabling Functions 11701 */ 11702 static dtrace_state_t * 11703 dtrace_anon_grab(void) 11704 { 11705 dtrace_state_t *state; 11706 11707 ASSERT(MUTEX_HELD(&dtrace_lock)); 11708 11709 if ((state = dtrace_anon.dta_state) == NULL) { 11710 ASSERT(dtrace_anon.dta_enabling == NULL); 11711 return (NULL); 11712 } 11713 11714 ASSERT(dtrace_anon.dta_enabling != NULL); 11715 ASSERT(dtrace_retained != NULL); 11716 11717 dtrace_enabling_destroy(dtrace_anon.dta_enabling); 11718 dtrace_anon.dta_enabling = NULL; 11719 dtrace_anon.dta_state = NULL; 11720 11721 return (state); 11722 } 11723 11724 static void 11725 dtrace_anon_property(void) 11726 { 11727 int i, rv; 11728 dtrace_state_t *state; 11729 dof_hdr_t *dof; 11730 char c[32]; /* enough for "dof-data-" + digits */ 11731 11732 ASSERT(MUTEX_HELD(&dtrace_lock)); 11733 ASSERT(MUTEX_HELD(&cpu_lock)); 11734 11735 for (i = 0; ; i++) { 11736 (void) snprintf(c, sizeof (c), "dof-data-%d", i); 11737 11738 dtrace_err_verbose = 1; 11739 11740 if ((dof = dtrace_dof_property(c)) == NULL) { 11741 dtrace_err_verbose = 0; 11742 break; 11743 } 11744 11745 /* 11746 * We want to create anonymous state, so we need to transition 11747 * the kernel debugger to indicate that DTrace is active. If 11748 * this fails (e.g. because the debugger has modified text in 11749 * some way), we won't continue with the processing. 11750 */ 11751 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) { 11752 cmn_err(CE_NOTE, "kernel debugger active; anonymous " 11753 "enabling ignored."); 11754 dtrace_dof_destroy(dof); 11755 break; 11756 } 11757 11758 /* 11759 * If we haven't allocated an anonymous state, we'll do so now. 11760 */ 11761 if ((state = dtrace_anon.dta_state) == NULL) { 11762 state = dtrace_state_create(NULL, NULL); 11763 dtrace_anon.dta_state = state; 11764 11765 if (state == NULL) { 11766 /* 11767 * This basically shouldn't happen: the only 11768 * failure mode from dtrace_state_create() is a 11769 * failure of ddi_soft_state_zalloc() that 11770 * itself should never happen. Still, the 11771 * interface allows for a failure mode, and 11772 * we want to fail as gracefully as possible: 11773 * we'll emit an error message and cease 11774 * processing anonymous state in this case. 11775 */ 11776 cmn_err(CE_WARN, "failed to create " 11777 "anonymous state"); 11778 dtrace_dof_destroy(dof); 11779 break; 11780 } 11781 } 11782 11783 rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(), 11784 &dtrace_anon.dta_enabling, 0, B_TRUE); 11785 11786 if (rv == 0) 11787 rv = dtrace_dof_options(dof, state); 11788 11789 dtrace_err_verbose = 0; 11790 dtrace_dof_destroy(dof); 11791 11792 if (rv != 0) { 11793 /* 11794 * This is malformed DOF; chuck any anonymous state 11795 * that we created. 11796 */ 11797 ASSERT(dtrace_anon.dta_enabling == NULL); 11798 dtrace_state_destroy(state); 11799 dtrace_anon.dta_state = NULL; 11800 break; 11801 } 11802 11803 ASSERT(dtrace_anon.dta_enabling != NULL); 11804 } 11805 11806 if (dtrace_anon.dta_enabling != NULL) { 11807 int rval; 11808 11809 /* 11810 * dtrace_enabling_retain() can only fail because we are 11811 * trying to retain more enablings than are allowed -- but 11812 * we only have one anonymous enabling, and we are guaranteed 11813 * to be allowed at least one retained enabling; we assert 11814 * that dtrace_enabling_retain() returns success. 11815 */ 11816 rval = dtrace_enabling_retain(dtrace_anon.dta_enabling); 11817 ASSERT(rval == 0); 11818 11819 dtrace_enabling_dump(dtrace_anon.dta_enabling); 11820 } 11821 } 11822 11823 /* 11824 * DTrace Helper Functions 11825 */ 11826 static void 11827 dtrace_helper_trace(dtrace_helper_action_t *helper, 11828 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where) 11829 { 11830 uint32_t size, next, nnext, i; 11831 dtrace_helptrace_t *ent; 11832 uint16_t flags = cpu_core[CPU->cpu_id].cpuc_dtrace_flags; 11833 11834 if (!dtrace_helptrace_enabled) 11835 return; 11836 11837 ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals); 11838 11839 /* 11840 * What would a tracing framework be without its own tracing 11841 * framework? (Well, a hell of a lot simpler, for starters...) 11842 */ 11843 size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals * 11844 sizeof (uint64_t) - sizeof (uint64_t); 11845 11846 /* 11847 * Iterate until we can allocate a slot in the trace buffer. 11848 */ 11849 do { 11850 next = dtrace_helptrace_next; 11851 11852 if (next + size < dtrace_helptrace_bufsize) { 11853 nnext = next + size; 11854 } else { 11855 nnext = size; 11856 } 11857 } while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next); 11858 11859 /* 11860 * We have our slot; fill it in. 11861 */ 11862 if (nnext == size) 11863 next = 0; 11864 11865 ent = (dtrace_helptrace_t *)&dtrace_helptrace_buffer[next]; 11866 ent->dtht_helper = helper; 11867 ent->dtht_where = where; 11868 ent->dtht_nlocals = vstate->dtvs_nlocals; 11869 11870 ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ? 11871 mstate->dtms_fltoffs : -1; 11872 ent->dtht_fault = DTRACE_FLAGS2FLT(flags); 11873 ent->dtht_illval = cpu_core[CPU->cpu_id].cpuc_dtrace_illval; 11874 11875 for (i = 0; i < vstate->dtvs_nlocals; i++) { 11876 dtrace_statvar_t *svar; 11877 11878 if ((svar = vstate->dtvs_locals[i]) == NULL) 11879 continue; 11880 11881 ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t)); 11882 ent->dtht_locals[i] = 11883 ((uint64_t *)(uintptr_t)svar->dtsv_data)[CPU->cpu_id]; 11884 } 11885 } 11886 11887 static uint64_t 11888 dtrace_helper(int which, dtrace_mstate_t *mstate, 11889 dtrace_state_t *state, uint64_t arg0, uint64_t arg1) 11890 { 11891 uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags; 11892 uint64_t sarg0 = mstate->dtms_arg[0]; 11893 uint64_t sarg1 = mstate->dtms_arg[1]; 11894 uint64_t rval; 11895 dtrace_helpers_t *helpers = curproc->p_dtrace_helpers; 11896 dtrace_helper_action_t *helper; 11897 dtrace_vstate_t *vstate; 11898 dtrace_difo_t *pred; 11899 int i, trace = dtrace_helptrace_enabled; 11900 11901 ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS); 11902 11903 if (helpers == NULL) 11904 return (0); 11905 11906 if ((helper = helpers->dthps_actions[which]) == NULL) 11907 return (0); 11908 11909 vstate = &helpers->dthps_vstate; 11910 mstate->dtms_arg[0] = arg0; 11911 mstate->dtms_arg[1] = arg1; 11912 11913 /* 11914 * Now iterate over each helper. If its predicate evaluates to 'true', 11915 * we'll call the corresponding actions. Note that the below calls 11916 * to dtrace_dif_emulate() may set faults in machine state. This is 11917 * okay: our caller (the outer dtrace_dif_emulate()) will simply plow 11918 * the stored DIF offset with its own (which is the desired behavior). 11919 * Also, note the calls to dtrace_dif_emulate() may allocate scratch 11920 * from machine state; this is okay, too. 11921 */ 11922 for (; helper != NULL; helper = helper->dthp_next) { 11923 if ((pred = helper->dthp_predicate) != NULL) { 11924 if (trace) 11925 dtrace_helper_trace(helper, mstate, vstate, 0); 11926 11927 if (!dtrace_dif_emulate(pred, mstate, vstate, state)) 11928 goto next; 11929 11930 if (*flags & CPU_DTRACE_FAULT) 11931 goto err; 11932 } 11933 11934 for (i = 0; i < helper->dthp_nactions; i++) { 11935 if (trace) 11936 dtrace_helper_trace(helper, 11937 mstate, vstate, i + 1); 11938 11939 rval = dtrace_dif_emulate(helper->dthp_actions[i], 11940 mstate, vstate, state); 11941 11942 if (*flags & CPU_DTRACE_FAULT) 11943 goto err; 11944 } 11945 11946 next: 11947 if (trace) 11948 dtrace_helper_trace(helper, mstate, vstate, 11949 DTRACE_HELPTRACE_NEXT); 11950 } 11951 11952 if (trace) 11953 dtrace_helper_trace(helper, mstate, vstate, 11954 DTRACE_HELPTRACE_DONE); 11955 11956 /* 11957 * Restore the arg0 that we saved upon entry. 11958 */ 11959 mstate->dtms_arg[0] = sarg0; 11960 mstate->dtms_arg[1] = sarg1; 11961 11962 return (rval); 11963 11964 err: 11965 if (trace) 11966 dtrace_helper_trace(helper, mstate, vstate, 11967 DTRACE_HELPTRACE_ERR); 11968 11969 /* 11970 * Restore the arg0 that we saved upon entry. 11971 */ 11972 mstate->dtms_arg[0] = sarg0; 11973 mstate->dtms_arg[1] = sarg1; 11974 11975 return (NULL); 11976 } 11977 11978 static void 11979 dtrace_helper_destroy(dtrace_helper_action_t *helper, dtrace_vstate_t *vstate) 11980 { 11981 int i; 11982 11983 if (helper->dthp_predicate != NULL) 11984 dtrace_difo_release(helper->dthp_predicate, vstate); 11985 11986 for (i = 0; i < helper->dthp_nactions; i++) { 11987 ASSERT(helper->dthp_actions[i] != NULL); 11988 dtrace_difo_release(helper->dthp_actions[i], vstate); 11989 } 11990 11991 kmem_free(helper->dthp_actions, 11992 helper->dthp_nactions * sizeof (dtrace_difo_t *)); 11993 kmem_free(helper, sizeof (dtrace_helper_action_t)); 11994 } 11995 11996 static int 11997 dtrace_helper_destroygen(int gen) 11998 { 11999 dtrace_helpers_t *help = curproc->p_dtrace_helpers; 12000 dtrace_vstate_t *vstate; 12001 int i; 12002 12003 ASSERT(MUTEX_HELD(&dtrace_lock)); 12004 12005 if (help == NULL || gen > help->dthps_generation) 12006 return (EINVAL); 12007 12008 vstate = &help->dthps_vstate; 12009 12010 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) { 12011 dtrace_helper_action_t *last = NULL, *h, *next; 12012 12013 for (h = help->dthps_actions[i]; h != NULL; h = next) { 12014 next = h->dthp_next; 12015 12016 if (h->dthp_generation == gen) { 12017 if (last != NULL) { 12018 last->dthp_next = next; 12019 } else { 12020 help->dthps_actions[i] = next; 12021 } 12022 12023 dtrace_helper_destroy(h, vstate); 12024 } else { 12025 last = h; 12026 } 12027 } 12028 } 12029 12030 return (0); 12031 } 12032 12033 static int 12034 dtrace_helper_validate(dtrace_helper_action_t *helper) 12035 { 12036 int err = 0, i; 12037 dtrace_difo_t *dp; 12038 12039 if ((dp = helper->dthp_predicate) != NULL) 12040 err += dtrace_difo_validate_helper(dp); 12041 12042 for (i = 0; i < helper->dthp_nactions; i++) 12043 err += dtrace_difo_validate_helper(helper->dthp_actions[i]); 12044 12045 return (err == 0); 12046 } 12047 12048 static int 12049 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep) 12050 { 12051 dtrace_helpers_t *help; 12052 dtrace_helper_action_t *helper, *last; 12053 dtrace_actdesc_t *act; 12054 dtrace_vstate_t *vstate; 12055 dtrace_predicate_t *pred; 12056 int count = 0, nactions = 0, i; 12057 12058 if (which < 0 || which >= DTRACE_NHELPER_ACTIONS) 12059 return (EINVAL); 12060 12061 help = curproc->p_dtrace_helpers; 12062 last = help->dthps_actions[which]; 12063 vstate = &help->dthps_vstate; 12064 12065 for (count = 0; last != NULL; last = last->dthp_next) { 12066 count++; 12067 if (last->dthp_next == NULL) 12068 break; 12069 } 12070 12071 /* 12072 * If we already have dtrace_helper_actions_max helper actions for this 12073 * helper action type, we'll refuse to add a new one. 12074 */ 12075 if (count >= dtrace_helper_actions_max) 12076 return (ENOSPC); 12077 12078 helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP); 12079 helper->dthp_generation = help->dthps_generation; 12080 12081 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) { 12082 ASSERT(pred->dtp_difo != NULL); 12083 dtrace_difo_hold(pred->dtp_difo); 12084 helper->dthp_predicate = pred->dtp_difo; 12085 } 12086 12087 for (act = ep->dted_action; act != NULL; act = act->dtad_next) { 12088 if (act->dtad_kind != DTRACEACT_DIFEXPR) 12089 goto err; 12090 12091 if (act->dtad_difo == NULL) 12092 goto err; 12093 12094 nactions++; 12095 } 12096 12097 helper->dthp_actions = kmem_zalloc(sizeof (dtrace_difo_t *) * 12098 (helper->dthp_nactions = nactions), KM_SLEEP); 12099 12100 for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) { 12101 dtrace_difo_hold(act->dtad_difo); 12102 helper->dthp_actions[i++] = act->dtad_difo; 12103 } 12104 12105 if (!dtrace_helper_validate(helper)) 12106 goto err; 12107 12108 if (last == NULL) { 12109 help->dthps_actions[which] = helper; 12110 } else { 12111 last->dthp_next = helper; 12112 } 12113 12114 if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) { 12115 dtrace_helptrace_nlocals = vstate->dtvs_nlocals; 12116 dtrace_helptrace_next = 0; 12117 } 12118 12119 return (0); 12120 err: 12121 dtrace_helper_destroy(helper, vstate); 12122 return (EINVAL); 12123 } 12124 12125 static void 12126 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help, 12127 dof_helper_t *dofhp) 12128 { 12129 ASSERT(MUTEX_NOT_HELD(&dtrace_lock)); 12130 12131 mutex_enter(&dtrace_meta_lock); 12132 mutex_enter(&dtrace_lock); 12133 12134 if (!dtrace_attached() || dtrace_meta_pid == NULL) { 12135 /* 12136 * If the dtrace module is loaded but not attached, or if 12137 * there aren't isn't a meta provider registered to deal with 12138 * these provider descriptions, we need to postpone creating 12139 * the actual providers until later. 12140 */ 12141 12142 if (help->dthps_next == NULL && help->dthps_prev == NULL && 12143 dtrace_deferred_pid != help) { 12144 help->dthps_deferred = 1; 12145 help->dthps_pid = p->p_pid; 12146 help->dthps_next = dtrace_deferred_pid; 12147 help->dthps_prev = NULL; 12148 if (dtrace_deferred_pid != NULL) 12149 dtrace_deferred_pid->dthps_prev = help; 12150 dtrace_deferred_pid = help; 12151 } 12152 12153 mutex_exit(&dtrace_lock); 12154 12155 } else if (dofhp != NULL) { 12156 /* 12157 * If the dtrace module is loaded and we have a particular 12158 * helper provider description, pass that off to the 12159 * meta provider. 12160 */ 12161 12162 mutex_exit(&dtrace_lock); 12163 12164 dtrace_helper_provide(dofhp, p->p_pid); 12165 12166 } else { 12167 /* 12168 * Otherwise, just pass all the helper provider descriptions 12169 * off to the meta provider. 12170 */ 12171 12172 int i; 12173 mutex_exit(&dtrace_lock); 12174 12175 for (i = 0; i < help->dthps_nprovs; i++) { 12176 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov, 12177 p->p_pid); 12178 } 12179 } 12180 12181 mutex_exit(&dtrace_meta_lock); 12182 } 12183 12184 static int 12185 dtrace_helper_provider_add(dof_helper_t *dofhp) 12186 { 12187 dtrace_helpers_t *help; 12188 dtrace_helper_provider_t *hprov, **tmp_provs; 12189 uint_t tmp_nprovs, i; 12190 12191 help = curproc->p_dtrace_helpers; 12192 ASSERT(help != NULL); 12193 12194 /* 12195 * If we already have dtrace_helper_providers_max helper providers, 12196 * we're refuse to add a new one. 12197 */ 12198 if (help->dthps_nprovs >= dtrace_helper_providers_max) 12199 return (ENOSPC); 12200 12201 /* 12202 * Check to make sure this isn't a duplicate. 12203 */ 12204 for (i = 0; i < help->dthps_nprovs; i++) { 12205 if (dofhp->dofhp_addr == 12206 help->dthps_provs[i]->dthp_prov.dofhp_addr) 12207 return (EALREADY); 12208 } 12209 12210 hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP); 12211 hprov->dthp_prov = *dofhp; 12212 hprov->dthp_ref = 1; 12213 12214 tmp_nprovs = help->dthps_nprovs; 12215 tmp_provs = help->dthps_provs; 12216 help->dthps_nprovs++; 12217 help->dthps_provs = kmem_zalloc(help->dthps_nprovs * 12218 sizeof (dtrace_helper_provider_t *), KM_SLEEP); 12219 12220 help->dthps_provs[tmp_nprovs] = hprov; 12221 if (tmp_provs != NULL) { 12222 bcopy(tmp_provs, help->dthps_provs, tmp_nprovs * 12223 sizeof (dtrace_helper_provider_t *)); 12224 kmem_free(tmp_provs, tmp_nprovs * 12225 sizeof (dtrace_helper_provider_t *)); 12226 } 12227 12228 return (0); 12229 } 12230 12231 static void 12232 dtrace_helper_provider_remove(dtrace_helper_provider_t *hprov) 12233 { 12234 mutex_enter(&dtrace_lock); 12235 12236 if (--hprov->dthp_ref == 0) { 12237 dof_hdr_t *dof; 12238 mutex_exit(&dtrace_lock); 12239 dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof; 12240 dtrace_dof_destroy(dof); 12241 kmem_free(hprov, sizeof (dtrace_helper_provider_t)); 12242 } else { 12243 mutex_exit(&dtrace_lock); 12244 } 12245 } 12246 12247 static int 12248 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec) 12249 { 12250 uintptr_t daddr = (uintptr_t)dof; 12251 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec; 12252 dof_provider_t *provider; 12253 dof_probe_t *probe; 12254 uint8_t *arg; 12255 char *strtab, *typestr; 12256 dof_stridx_t typeidx; 12257 size_t typesz; 12258 uint_t nprobes, j, k; 12259 12260 ASSERT(sec->dofs_type == DOF_SECT_PROVIDER); 12261 12262 if (sec->dofs_offset & (sizeof (uint_t) - 1)) { 12263 dtrace_dof_error(dof, "misaligned section offset"); 12264 return (-1); 12265 } 12266 12267 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset); 12268 str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab); 12269 prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes); 12270 arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs); 12271 off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs); 12272 12273 if (str_sec == NULL || prb_sec == NULL || 12274 arg_sec == NULL || off_sec == NULL) 12275 return (-1); 12276 12277 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset); 12278 12279 if (provider->dofpv_name >= str_sec->dofs_size || 12280 strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) { 12281 dtrace_dof_error(dof, "invalid provider name"); 12282 return (-1); 12283 } 12284 12285 if (prb_sec->dofs_entsize == 0 || 12286 prb_sec->dofs_entsize > prb_sec->dofs_size) { 12287 dtrace_dof_error(dof, "invalid entry size"); 12288 return (-1); 12289 } 12290 12291 if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) { 12292 dtrace_dof_error(dof, "misaligned entry size"); 12293 return (-1); 12294 } 12295 12296 if (off_sec->dofs_entsize != sizeof (uint32_t)) { 12297 dtrace_dof_error(dof, "invalid entry size"); 12298 return (-1); 12299 } 12300 12301 if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) { 12302 dtrace_dof_error(dof, "misaligned section offset"); 12303 return (-1); 12304 } 12305 12306 if (arg_sec->dofs_entsize != sizeof (uint8_t)) { 12307 dtrace_dof_error(dof, "invalid entry size"); 12308 return (-1); 12309 } 12310 12311 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset); 12312 12313 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize; 12314 12315 /* 12316 * Take a pass through the probes to check for errors. 12317 */ 12318 for (j = 0; j < nprobes; j++) { 12319 probe = (dof_probe_t *)(uintptr_t)(daddr + 12320 prb_sec->dofs_offset + j * prb_sec->dofs_entsize); 12321 12322 if (probe->dofpr_func >= str_sec->dofs_size) { 12323 dtrace_dof_error(dof, "invalid function name"); 12324 return (-1); 12325 } 12326 12327 if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) { 12328 dtrace_dof_error(dof, "function name too long"); 12329 return (-1); 12330 } 12331 12332 if (probe->dofpr_name >= str_sec->dofs_size || 12333 strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) { 12334 dtrace_dof_error(dof, "invalid probe name"); 12335 return (-1); 12336 } 12337 12338 12339 if (probe->dofpr_offidx + probe->dofpr_noffs < 12340 probe->dofpr_offidx || 12341 (probe->dofpr_offidx + probe->dofpr_noffs) * 12342 off_sec->dofs_entsize > off_sec->dofs_size) { 12343 dtrace_dof_error(dof, "invalid probe offset"); 12344 return (-1); 12345 } 12346 12347 if (probe->dofpr_argidx + probe->dofpr_xargc < 12348 probe->dofpr_argidx || 12349 (probe->dofpr_argidx + probe->dofpr_xargc) * 12350 arg_sec->dofs_entsize > arg_sec->dofs_size) { 12351 dtrace_dof_error(dof, "invalid args"); 12352 return (-1); 12353 } 12354 12355 typeidx = probe->dofpr_nargv; 12356 typestr = strtab + probe->dofpr_nargv; 12357 for (k = 0; k < probe->dofpr_nargc; k++) { 12358 if (typeidx >= str_sec->dofs_size) { 12359 dtrace_dof_error(dof, "bad " 12360 "native argument type"); 12361 return (-1); 12362 } 12363 12364 typesz = strlen(typestr) + 1; 12365 if (typesz > DTRACE_ARGTYPELEN) { 12366 dtrace_dof_error(dof, "native " 12367 "argument type too long"); 12368 return (-1); 12369 } 12370 typeidx += typesz; 12371 typestr += typesz; 12372 } 12373 12374 typeidx = probe->dofpr_xargv; 12375 typestr = strtab + probe->dofpr_xargv; 12376 for (k = 0; k < probe->dofpr_xargc; k++) { 12377 if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) { 12378 dtrace_dof_error(dof, "bad " 12379 "native argument index"); 12380 return (-1); 12381 } 12382 12383 if (typeidx >= str_sec->dofs_size) { 12384 dtrace_dof_error(dof, "bad " 12385 "translated argument type"); 12386 return (-1); 12387 } 12388 12389 typesz = strlen(typestr) + 1; 12390 if (typesz > DTRACE_ARGTYPELEN) { 12391 dtrace_dof_error(dof, "translated argument " 12392 "type too long"); 12393 return (-1); 12394 } 12395 12396 typeidx += typesz; 12397 typestr += typesz; 12398 } 12399 } 12400 12401 return (0); 12402 } 12403 12404 static int 12405 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp) 12406 { 12407 dtrace_helpers_t *help; 12408 dtrace_vstate_t *vstate; 12409 dtrace_enabling_t *enab = NULL; 12410 int i, gen, rv, nhelpers = 0, destroy = 1; 12411 12412 ASSERT(MUTEX_HELD(&dtrace_lock)); 12413 12414 if ((help = curproc->p_dtrace_helpers) == NULL) 12415 help = dtrace_helpers_create(curproc); 12416 12417 vstate = &help->dthps_vstate; 12418 12419 if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab, 12420 dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) { 12421 dtrace_dof_destroy(dof); 12422 return (rv); 12423 } 12424 12425 /* 12426 * Now we need to walk through the ECB descriptions in the enabling. 12427 */ 12428 for (i = 0; i < enab->dten_ndesc; i++) { 12429 dtrace_ecbdesc_t *ep = enab->dten_desc[i]; 12430 dtrace_probedesc_t *desc = &ep->dted_probe; 12431 12432 if (strcmp(desc->dtpd_provider, "dtrace") != 0) 12433 continue; 12434 12435 if (strcmp(desc->dtpd_mod, "helper") != 0) 12436 continue; 12437 12438 if (strcmp(desc->dtpd_func, "ustack") != 0) 12439 continue; 12440 12441 if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK, 12442 ep)) != 0) { 12443 /* 12444 * Adding this helper action failed -- we are now going 12445 * to rip out the entire generation and return failure. 12446 */ 12447 (void) dtrace_helper_destroygen(help->dthps_generation); 12448 dtrace_enabling_destroy(enab); 12449 dtrace_dof_destroy(dof); 12450 dtrace_err = rv; 12451 return (-1); 12452 } 12453 12454 nhelpers++; 12455 } 12456 12457 if (nhelpers < enab->dten_ndesc) 12458 dtrace_dof_error(dof, "unmatched helpers"); 12459 12460 if (dhp != NULL) { 12461 uintptr_t daddr = (uintptr_t)dof; 12462 int err = 0, count = 0; 12463 12464 /* 12465 * Look for helper probes. 12466 */ 12467 for (i = 0; i < dof->dofh_secnum; i++) { 12468 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr + 12469 dof->dofh_secoff + i * dof->dofh_secsize); 12470 12471 if (sec->dofs_type != DOF_SECT_PROVIDER) 12472 continue; 12473 12474 if (dtrace_helper_provider_validate(dof, sec) != 0) { 12475 err = 1; 12476 break; 12477 } 12478 12479 count++; 12480 } 12481 12482 dhp->dofhp_dof = (uint64_t)(uintptr_t)dof; 12483 if (err == 0 && count > 0 && 12484 dtrace_helper_provider_add(dhp) == 0) 12485 destroy = 0; 12486 else 12487 dhp = NULL; 12488 } 12489 12490 gen = help->dthps_generation++; 12491 dtrace_enabling_destroy(enab); 12492 12493 if (dhp != NULL) { 12494 mutex_exit(&dtrace_lock); 12495 dtrace_helper_provider_register(curproc, help, dhp); 12496 mutex_enter(&dtrace_lock); 12497 } 12498 12499 if (destroy) 12500 dtrace_dof_destroy(dof); 12501 12502 return (gen); 12503 } 12504 12505 static dtrace_helpers_t * 12506 dtrace_helpers_create(proc_t *p) 12507 { 12508 dtrace_helpers_t *help; 12509 12510 ASSERT(MUTEX_HELD(&dtrace_lock)); 12511 ASSERT(p->p_dtrace_helpers == NULL); 12512 12513 help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP); 12514 help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) * 12515 DTRACE_NHELPER_ACTIONS, KM_SLEEP); 12516 12517 p->p_dtrace_helpers = help; 12518 dtrace_helpers++; 12519 12520 return (help); 12521 } 12522 12523 static void 12524 dtrace_helpers_destroy(void) 12525 { 12526 dtrace_helpers_t *help; 12527 dtrace_vstate_t *vstate; 12528 proc_t *p = curproc; 12529 int i; 12530 12531 mutex_enter(&dtrace_lock); 12532 12533 ASSERT(p->p_dtrace_helpers != NULL); 12534 ASSERT(dtrace_helpers > 0); 12535 12536 help = p->p_dtrace_helpers; 12537 vstate = &help->dthps_vstate; 12538 12539 /* 12540 * We're now going to lose the help from this process. 12541 */ 12542 p->p_dtrace_helpers = NULL; 12543 dtrace_sync(); 12544 12545 /* 12546 * Destory the helper actions. 12547 */ 12548 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) { 12549 dtrace_helper_action_t *h, *next; 12550 12551 for (h = help->dthps_actions[i]; h != NULL; h = next) { 12552 next = h->dthp_next; 12553 dtrace_helper_destroy(h, vstate); 12554 h = next; 12555 } 12556 } 12557 12558 mutex_exit(&dtrace_lock); 12559 12560 /* 12561 * Destroy the helper providers. 12562 */ 12563 if (help->dthps_nprovs > 0) { 12564 mutex_enter(&dtrace_meta_lock); 12565 if (dtrace_meta_pid != NULL) { 12566 ASSERT(dtrace_deferred_pid == NULL); 12567 12568 for (i = 0; i < help->dthps_nprovs; i++) { 12569 dtrace_helper_remove( 12570 &help->dthps_provs[i]->dthp_prov, p->p_pid); 12571 } 12572 } else { 12573 mutex_enter(&dtrace_lock); 12574 ASSERT(help->dthps_deferred == 0 || 12575 help->dthps_next != NULL || 12576 help->dthps_prev != NULL || 12577 help == dtrace_deferred_pid); 12578 12579 /* 12580 * Remove the helper from the deferred list. 12581 */ 12582 if (help->dthps_next != NULL) 12583 help->dthps_next->dthps_prev = help->dthps_prev; 12584 if (help->dthps_prev != NULL) 12585 help->dthps_prev->dthps_next = help->dthps_next; 12586 if (dtrace_deferred_pid == help) { 12587 dtrace_deferred_pid = help->dthps_next; 12588 ASSERT(help->dthps_prev == NULL); 12589 } 12590 12591 mutex_exit(&dtrace_lock); 12592 } 12593 12594 mutex_exit(&dtrace_meta_lock); 12595 12596 for (i = 0; i < help->dthps_nprovs; i++) { 12597 dtrace_helper_provider_remove(help->dthps_provs[i]); 12598 } 12599 12600 kmem_free(help->dthps_provs, help->dthps_nprovs * 12601 sizeof (dtrace_helper_provider_t *)); 12602 } 12603 12604 mutex_enter(&dtrace_lock); 12605 12606 dtrace_vstate_fini(&help->dthps_vstate); 12607 kmem_free(help->dthps_actions, 12608 sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS); 12609 kmem_free(help, sizeof (dtrace_helpers_t)); 12610 12611 --dtrace_helpers; 12612 mutex_exit(&dtrace_lock); 12613 } 12614 12615 static void 12616 dtrace_helpers_duplicate(proc_t *from, proc_t *to) 12617 { 12618 dtrace_helpers_t *help, *newhelp; 12619 dtrace_helper_action_t *helper, *new, *last; 12620 dtrace_difo_t *dp; 12621 dtrace_vstate_t *vstate; 12622 int i, j, sz, hasprovs = 0; 12623 12624 mutex_enter(&dtrace_lock); 12625 ASSERT(from->p_dtrace_helpers != NULL); 12626 ASSERT(dtrace_helpers > 0); 12627 12628 help = from->p_dtrace_helpers; 12629 newhelp = dtrace_helpers_create(to); 12630 ASSERT(to->p_dtrace_helpers != NULL); 12631 12632 newhelp->dthps_generation = help->dthps_generation; 12633 vstate = &newhelp->dthps_vstate; 12634 12635 /* 12636 * Duplicate the helper actions. 12637 */ 12638 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) { 12639 if ((helper = help->dthps_actions[i]) == NULL) 12640 continue; 12641 12642 for (last = NULL; helper != NULL; helper = helper->dthp_next) { 12643 new = kmem_zalloc(sizeof (dtrace_helper_action_t), 12644 KM_SLEEP); 12645 new->dthp_generation = helper->dthp_generation; 12646 12647 if ((dp = helper->dthp_predicate) != NULL) { 12648 dp = dtrace_difo_duplicate(dp, vstate); 12649 new->dthp_predicate = dp; 12650 } 12651 12652 new->dthp_nactions = helper->dthp_nactions; 12653 sz = sizeof (dtrace_difo_t *) * new->dthp_nactions; 12654 new->dthp_actions = kmem_alloc(sz, KM_SLEEP); 12655 12656 for (j = 0; j < new->dthp_nactions; j++) { 12657 dtrace_difo_t *dp = helper->dthp_actions[j]; 12658 12659 ASSERT(dp != NULL); 12660 dp = dtrace_difo_duplicate(dp, vstate); 12661 new->dthp_actions[j] = dp; 12662 } 12663 12664 if (last != NULL) { 12665 last->dthp_next = new; 12666 } else { 12667 newhelp->dthps_actions[i] = new; 12668 } 12669 12670 last = new; 12671 } 12672 } 12673 12674 /* 12675 * Duplicate the helper providers and register them with the 12676 * DTrace framework. 12677 */ 12678 if (help->dthps_nprovs > 0) { 12679 newhelp->dthps_nprovs = help->dthps_nprovs; 12680 newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs * 12681 sizeof (dtrace_helper_provider_t *), KM_SLEEP); 12682 for (i = 0; i < newhelp->dthps_nprovs; i++) { 12683 newhelp->dthps_provs[i] = help->dthps_provs[i]; 12684 newhelp->dthps_provs[i]->dthp_ref++; 12685 } 12686 12687 hasprovs = 1; 12688 } 12689 12690 mutex_exit(&dtrace_lock); 12691 12692 if (hasprovs) 12693 dtrace_helper_provider_register(to, newhelp, NULL); 12694 } 12695 12696 /* 12697 * DTrace Hook Functions 12698 */ 12699 static void 12700 dtrace_module_loaded(struct modctl *ctl) 12701 { 12702 dtrace_provider_t *prv; 12703 12704 mutex_enter(&dtrace_provider_lock); 12705 mutex_enter(&mod_lock); 12706 12707 ASSERT(ctl->mod_busy); 12708 12709 /* 12710 * We're going to call each providers per-module provide operation 12711 * specifying only this module. 12712 */ 12713 for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next) 12714 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl); 12715 12716 mutex_exit(&mod_lock); 12717 mutex_exit(&dtrace_provider_lock); 12718 12719 /* 12720 * If we have any retained enablings, we need to match against them. 12721 * Enabling probes requires that cpu_lock be held, and we cannot hold 12722 * cpu_lock here -- it is legal for cpu_lock to be held when loading a 12723 * module. (In particular, this happens when loading scheduling 12724 * classes.) So if we have any retained enablings, we need to dispatch 12725 * our task queue to do the match for us. 12726 */ 12727 mutex_enter(&dtrace_lock); 12728 12729 if (dtrace_retained == NULL) { 12730 mutex_exit(&dtrace_lock); 12731 return; 12732 } 12733 12734 (void) taskq_dispatch(dtrace_taskq, 12735 (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP); 12736 12737 mutex_exit(&dtrace_lock); 12738 12739 /* 12740 * And now, for a little heuristic sleaze: in general, we want to 12741 * match modules as soon as they load. However, we cannot guarantee 12742 * this, because it would lead us to the lock ordering violation 12743 * outlined above. The common case, of course, is that cpu_lock is 12744 * _not_ held -- so we delay here for a clock tick, hoping that that's 12745 * long enough for the task queue to do its work. If it's not, it's 12746 * not a serious problem -- it just means that the module that we 12747 * just loaded may not be immediately instrumentable. 12748 */ 12749 delay(1); 12750 } 12751 12752 static void 12753 dtrace_module_unloaded(struct modctl *ctl) 12754 { 12755 dtrace_probe_t template, *probe, *first, *next; 12756 dtrace_provider_t *prov; 12757 12758 template.dtpr_mod = ctl->mod_modname; 12759 12760 mutex_enter(&dtrace_provider_lock); 12761 mutex_enter(&mod_lock); 12762 mutex_enter(&dtrace_lock); 12763 12764 if (dtrace_bymod == NULL) { 12765 /* 12766 * The DTrace module is loaded (obviously) but not attached; 12767 * we don't have any work to do. 12768 */ 12769 mutex_exit(&dtrace_provider_lock); 12770 mutex_exit(&mod_lock); 12771 mutex_exit(&dtrace_lock); 12772 return; 12773 } 12774 12775 for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template); 12776 probe != NULL; probe = probe->dtpr_nextmod) { 12777 if (probe->dtpr_ecb != NULL) { 12778 mutex_exit(&dtrace_provider_lock); 12779 mutex_exit(&mod_lock); 12780 mutex_exit(&dtrace_lock); 12781 12782 /* 12783 * This shouldn't _actually_ be possible -- we're 12784 * unloading a module that has an enabled probe in it. 12785 * (It's normally up to the provider to make sure that 12786 * this can't happen.) However, because dtps_enable() 12787 * doesn't have a failure mode, there can be an 12788 * enable/unload race. Upshot: we don't want to 12789 * assert, but we're not going to disable the 12790 * probe, either. 12791 */ 12792 if (dtrace_err_verbose) { 12793 cmn_err(CE_WARN, "unloaded module '%s' had " 12794 "enabled probes", ctl->mod_modname); 12795 } 12796 12797 return; 12798 } 12799 } 12800 12801 probe = first; 12802 12803 for (first = NULL; probe != NULL; probe = next) { 12804 ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe); 12805 12806 dtrace_probes[probe->dtpr_id - 1] = NULL; 12807 12808 next = probe->dtpr_nextmod; 12809 dtrace_hash_remove(dtrace_bymod, probe); 12810 dtrace_hash_remove(dtrace_byfunc, probe); 12811 dtrace_hash_remove(dtrace_byname, probe); 12812 12813 if (first == NULL) { 12814 first = probe; 12815 probe->dtpr_nextmod = NULL; 12816 } else { 12817 probe->dtpr_nextmod = first; 12818 first = probe; 12819 } 12820 } 12821 12822 /* 12823 * We've removed all of the module's probes from the hash chains and 12824 * from the probe array. Now issue a dtrace_sync() to be sure that 12825 * everyone has cleared out from any probe array processing. 12826 */ 12827 dtrace_sync(); 12828 12829 for (probe = first; probe != NULL; probe = first) { 12830 first = probe->dtpr_nextmod; 12831 prov = probe->dtpr_provider; 12832 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id, 12833 probe->dtpr_arg); 12834 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1); 12835 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1); 12836 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1); 12837 vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1); 12838 kmem_free(probe, sizeof (dtrace_probe_t)); 12839 } 12840 12841 mutex_exit(&dtrace_lock); 12842 mutex_exit(&mod_lock); 12843 mutex_exit(&dtrace_provider_lock); 12844 } 12845 12846 void 12847 dtrace_suspend(void) 12848 { 12849 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend)); 12850 } 12851 12852 void 12853 dtrace_resume(void) 12854 { 12855 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume)); 12856 } 12857 12858 static int 12859 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu) 12860 { 12861 ASSERT(MUTEX_HELD(&cpu_lock)); 12862 mutex_enter(&dtrace_lock); 12863 12864 switch (what) { 12865 case CPU_CONFIG: { 12866 dtrace_state_t *state; 12867 dtrace_optval_t *opt, rs, c; 12868 12869 /* 12870 * For now, we only allocate a new buffer for anonymous state. 12871 */ 12872 if ((state = dtrace_anon.dta_state) == NULL) 12873 break; 12874 12875 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) 12876 break; 12877 12878 opt = state->dts_options; 12879 c = opt[DTRACEOPT_CPU]; 12880 12881 if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu) 12882 break; 12883 12884 /* 12885 * Regardless of what the actual policy is, we're going to 12886 * temporarily set our resize policy to be manual. We're 12887 * also going to temporarily set our CPU option to denote 12888 * the newly configured CPU. 12889 */ 12890 rs = opt[DTRACEOPT_BUFRESIZE]; 12891 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL; 12892 opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu; 12893 12894 (void) dtrace_state_buffers(state); 12895 12896 opt[DTRACEOPT_BUFRESIZE] = rs; 12897 opt[DTRACEOPT_CPU] = c; 12898 12899 break; 12900 } 12901 12902 case CPU_UNCONFIG: 12903 /* 12904 * We don't free the buffer in the CPU_UNCONFIG case. (The 12905 * buffer will be freed when the consumer exits.) 12906 */ 12907 break; 12908 12909 default: 12910 break; 12911 } 12912 12913 mutex_exit(&dtrace_lock); 12914 return (0); 12915 } 12916 12917 static void 12918 dtrace_cpu_setup_initial(processorid_t cpu) 12919 { 12920 (void) dtrace_cpu_setup(CPU_CONFIG, cpu); 12921 } 12922 12923 static void 12924 dtrace_toxrange_add(uintptr_t base, uintptr_t limit) 12925 { 12926 if (dtrace_toxranges >= dtrace_toxranges_max) { 12927 int osize, nsize; 12928 dtrace_toxrange_t *range; 12929 12930 osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t); 12931 12932 if (osize == 0) { 12933 ASSERT(dtrace_toxrange == NULL); 12934 ASSERT(dtrace_toxranges_max == 0); 12935 dtrace_toxranges_max = 1; 12936 } else { 12937 dtrace_toxranges_max <<= 1; 12938 } 12939 12940 nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t); 12941 range = kmem_zalloc(nsize, KM_SLEEP); 12942 12943 if (dtrace_toxrange != NULL) { 12944 ASSERT(osize != 0); 12945 bcopy(dtrace_toxrange, range, osize); 12946 kmem_free(dtrace_toxrange, osize); 12947 } 12948 12949 dtrace_toxrange = range; 12950 } 12951 12952 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == NULL); 12953 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == NULL); 12954 12955 dtrace_toxrange[dtrace_toxranges].dtt_base = base; 12956 dtrace_toxrange[dtrace_toxranges].dtt_limit = limit; 12957 dtrace_toxranges++; 12958 } 12959 12960 /* 12961 * DTrace Driver Cookbook Functions 12962 */ 12963 /*ARGSUSED*/ 12964 static int 12965 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd) 12966 { 12967 dtrace_provider_id_t id; 12968 dtrace_state_t *state = NULL; 12969 dtrace_enabling_t *enab; 12970 12971 mutex_enter(&cpu_lock); 12972 mutex_enter(&dtrace_provider_lock); 12973 mutex_enter(&dtrace_lock); 12974 12975 if (ddi_soft_state_init(&dtrace_softstate, sizeof (dtrace_state_t) + 12976 NCPU * sizeof (dtrace_buffer_t), 0) != 0) { 12977 cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state"); 12978 mutex_exit(&cpu_lock); 12979 mutex_exit(&dtrace_provider_lock); 12980 mutex_exit(&dtrace_lock); 12981 return (DDI_FAILURE); 12982 } 12983 12984 if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR, 12985 DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE || 12986 ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR, 12987 DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) { 12988 cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes"); 12989 ddi_remove_minor_node(devi, NULL); 12990 ddi_soft_state_fini(&dtrace_softstate); 12991 mutex_exit(&cpu_lock); 12992 mutex_exit(&dtrace_provider_lock); 12993 mutex_exit(&dtrace_lock); 12994 return (DDI_FAILURE); 12995 } 12996 12997 ddi_report_dev(devi); 12998 dtrace_devi = devi; 12999 13000 dtrace_modload = dtrace_module_loaded; 13001 dtrace_modunload = dtrace_module_unloaded; 13002 dtrace_cpu_init = dtrace_cpu_setup_initial; 13003 dtrace_helpers_cleanup = dtrace_helpers_destroy; 13004 dtrace_helpers_fork = dtrace_helpers_duplicate; 13005 dtrace_cpustart_init = dtrace_suspend; 13006 dtrace_cpustart_fini = dtrace_resume; 13007 dtrace_debugger_init = dtrace_suspend; 13008 dtrace_debugger_fini = dtrace_resume; 13009 dtrace_kreloc_init = dtrace_suspend; 13010 dtrace_kreloc_fini = dtrace_resume; 13011 13012 register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL); 13013 13014 ASSERT(MUTEX_HELD(&cpu_lock)); 13015 13016 dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1, 13017 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER); 13018 dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE, 13019 UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0, 13020 VM_SLEEP | VMC_IDENTIFIER); 13021 dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri, 13022 1, INT_MAX, 0); 13023 13024 dtrace_state_cache = kmem_cache_create("dtrace_state_cache", 13025 sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN, 13026 NULL, NULL, NULL, NULL, NULL, 0); 13027 13028 ASSERT(MUTEX_HELD(&cpu_lock)); 13029 dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod), 13030 offsetof(dtrace_probe_t, dtpr_nextmod), 13031 offsetof(dtrace_probe_t, dtpr_prevmod)); 13032 13033 dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func), 13034 offsetof(dtrace_probe_t, dtpr_nextfunc), 13035 offsetof(dtrace_probe_t, dtpr_prevfunc)); 13036 13037 dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name), 13038 offsetof(dtrace_probe_t, dtpr_nextname), 13039 offsetof(dtrace_probe_t, dtpr_prevname)); 13040 13041 if (dtrace_retain_max < 1) { 13042 cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; " 13043 "setting to 1", dtrace_retain_max); 13044 dtrace_retain_max = 1; 13045 } 13046 13047 /* 13048 * Now discover our toxic ranges. 13049 */ 13050 dtrace_toxic_ranges(dtrace_toxrange_add); 13051 13052 /* 13053 * Before we register ourselves as a provider to our own framework, 13054 * we would like to assert that dtrace_provider is NULL -- but that's 13055 * not true if we were loaded as a dependency of a DTrace provider. 13056 * Once we've registered, we can assert that dtrace_provider is our 13057 * pseudo provider. 13058 */ 13059 (void) dtrace_register("dtrace", &dtrace_provider_attr, 13060 DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id); 13061 13062 ASSERT(dtrace_provider != NULL); 13063 ASSERT((dtrace_provider_id_t)dtrace_provider == id); 13064 13065 dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t) 13066 dtrace_provider, NULL, NULL, "BEGIN", 0, NULL); 13067 dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t) 13068 dtrace_provider, NULL, NULL, "END", 0, NULL); 13069 dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t) 13070 dtrace_provider, NULL, NULL, "ERROR", 1, NULL); 13071 13072 dtrace_anon_property(); 13073 mutex_exit(&cpu_lock); 13074 13075 /* 13076 * If DTrace helper tracing is enabled, we need to allocate the 13077 * trace buffer and initialize the values. 13078 */ 13079 if (dtrace_helptrace_enabled) { 13080 ASSERT(dtrace_helptrace_buffer == NULL); 13081 dtrace_helptrace_buffer = 13082 kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP); 13083 dtrace_helptrace_next = 0; 13084 } 13085 13086 /* 13087 * If there are already providers, we must ask them to provide their 13088 * probes, and then match any anonymous enabling against them. Note 13089 * that there should be no other retained enablings at this time: 13090 * the only retained enablings at this time should be the anonymous 13091 * enabling. 13092 */ 13093 if (dtrace_anon.dta_enabling != NULL) { 13094 ASSERT(dtrace_retained == dtrace_anon.dta_enabling); 13095 13096 dtrace_enabling_provide(NULL); 13097 state = dtrace_anon.dta_state; 13098 13099 /* 13100 * We couldn't hold cpu_lock across the above call to 13101 * dtrace_enabling_provide(), but we must hold it to actually 13102 * enable the probes. We have to drop all of our locks, pick 13103 * up cpu_lock, and regain our locks before matching the 13104 * retained anonymous enabling. 13105 */ 13106 mutex_exit(&dtrace_lock); 13107 mutex_exit(&dtrace_provider_lock); 13108 13109 mutex_enter(&cpu_lock); 13110 mutex_enter(&dtrace_provider_lock); 13111 mutex_enter(&dtrace_lock); 13112 13113 if ((enab = dtrace_anon.dta_enabling) != NULL) 13114 (void) dtrace_enabling_match(enab, NULL); 13115 13116 mutex_exit(&cpu_lock); 13117 } 13118 13119 mutex_exit(&dtrace_lock); 13120 mutex_exit(&dtrace_provider_lock); 13121 13122 if (state != NULL) { 13123 /* 13124 * If we created any anonymous state, set it going now. 13125 */ 13126 (void) dtrace_state_go(state, &dtrace_anon.dta_beganon); 13127 } 13128 13129 return (DDI_SUCCESS); 13130 } 13131 13132 /*ARGSUSED*/ 13133 static int 13134 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p) 13135 { 13136 dtrace_state_t *state; 13137 uint32_t priv; 13138 uid_t uid; 13139 13140 if (getminor(*devp) == DTRACEMNRN_HELPER) 13141 return (0); 13142 13143 /* 13144 * If this wasn't an open with the "helper" minor, then it must be 13145 * the "dtrace" minor. 13146 */ 13147 ASSERT(getminor(*devp) == DTRACEMNRN_DTRACE); 13148 13149 /* 13150 * If no DTRACE_PRIV_* bits are set in the credential, then the 13151 * caller lacks sufficient permission to do anything with DTrace. 13152 */ 13153 dtrace_cred2priv(cred_p, &priv, &uid); 13154 if (priv == DTRACE_PRIV_NONE) 13155 return (EACCES); 13156 13157 /* 13158 * Ask all providers to provide all their probes. 13159 */ 13160 mutex_enter(&dtrace_provider_lock); 13161 dtrace_probe_provide(NULL, NULL); 13162 mutex_exit(&dtrace_provider_lock); 13163 13164 mutex_enter(&cpu_lock); 13165 mutex_enter(&dtrace_lock); 13166 dtrace_opens++; 13167 dtrace_membar_producer(); 13168 13169 /* 13170 * If the kernel debugger is active (that is, if the kernel debugger 13171 * modified text in some way), we won't allow the open. 13172 */ 13173 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) { 13174 dtrace_opens--; 13175 mutex_exit(&cpu_lock); 13176 mutex_exit(&dtrace_lock); 13177 return (EBUSY); 13178 } 13179 13180 state = dtrace_state_create(devp, cred_p); 13181 mutex_exit(&cpu_lock); 13182 13183 if (state == NULL) { 13184 if (--dtrace_opens == 0) 13185 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE); 13186 mutex_exit(&dtrace_lock); 13187 return (EAGAIN); 13188 } 13189 13190 mutex_exit(&dtrace_lock); 13191 13192 return (0); 13193 } 13194 13195 /*ARGSUSED*/ 13196 static int 13197 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p) 13198 { 13199 minor_t minor = getminor(dev); 13200 dtrace_state_t *state; 13201 13202 if (minor == DTRACEMNRN_HELPER) 13203 return (0); 13204 13205 state = ddi_get_soft_state(dtrace_softstate, minor); 13206 13207 mutex_enter(&cpu_lock); 13208 mutex_enter(&dtrace_lock); 13209 13210 if (state->dts_anon) { 13211 /* 13212 * There is anonymous state. Destroy that first. 13213 */ 13214 ASSERT(dtrace_anon.dta_state == NULL); 13215 dtrace_state_destroy(state->dts_anon); 13216 } 13217 13218 dtrace_state_destroy(state); 13219 ASSERT(dtrace_opens > 0); 13220 if (--dtrace_opens == 0) 13221 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE); 13222 13223 mutex_exit(&dtrace_lock); 13224 mutex_exit(&cpu_lock); 13225 13226 return (0); 13227 } 13228 13229 /*ARGSUSED*/ 13230 static int 13231 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv) 13232 { 13233 int rval; 13234 dof_helper_t help, *dhp = NULL; 13235 13236 switch (cmd) { 13237 case DTRACEHIOC_ADDDOF: 13238 if (copyin((void *)arg, &help, sizeof (help)) != 0) { 13239 dtrace_dof_error(NULL, "failed to copyin DOF helper"); 13240 return (EFAULT); 13241 } 13242 13243 dhp = &help; 13244 arg = (intptr_t)help.dofhp_dof; 13245 /*FALLTHROUGH*/ 13246 13247 case DTRACEHIOC_ADD: { 13248 dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval); 13249 13250 if (dof == NULL) 13251 return (rval); 13252 13253 mutex_enter(&dtrace_lock); 13254 dtrace_err = 0; 13255 13256 /* 13257 * dtrace_helper_slurp() takes responsibility for the dof -- 13258 * it may free it now or it may save it and free it later. 13259 */ 13260 if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) { 13261 *rv = rval; 13262 rval = 0; 13263 } else { 13264 rval = EINVAL; 13265 } 13266 13267 mutex_exit(&dtrace_lock); 13268 return (rval); 13269 } 13270 13271 case DTRACEHIOC_REMOVE: { 13272 mutex_enter(&dtrace_lock); 13273 rval = dtrace_helper_destroygen(arg); 13274 mutex_exit(&dtrace_lock); 13275 13276 return (rval); 13277 } 13278 13279 default: 13280 break; 13281 } 13282 13283 return (ENOTTY); 13284 } 13285 13286 /*ARGSUSED*/ 13287 static int 13288 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv) 13289 { 13290 minor_t minor = getminor(dev); 13291 dtrace_state_t *state; 13292 int rval; 13293 13294 if (minor == DTRACEMNRN_HELPER) 13295 return (dtrace_ioctl_helper(cmd, arg, rv)); 13296 13297 state = ddi_get_soft_state(dtrace_softstate, minor); 13298 13299 if (state->dts_anon) { 13300 ASSERT(dtrace_anon.dta_state == NULL); 13301 state = state->dts_anon; 13302 } 13303 13304 switch (cmd) { 13305 case DTRACEIOC_PROVIDER: { 13306 dtrace_providerdesc_t pvd; 13307 dtrace_provider_t *pvp; 13308 13309 if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0) 13310 return (EFAULT); 13311 13312 pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0'; 13313 mutex_enter(&dtrace_provider_lock); 13314 13315 for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) { 13316 if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0) 13317 break; 13318 } 13319 13320 mutex_exit(&dtrace_provider_lock); 13321 13322 if (pvp == NULL) 13323 return (ESRCH); 13324 13325 bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t)); 13326 bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t)); 13327 if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0) 13328 return (EFAULT); 13329 13330 return (0); 13331 } 13332 13333 case DTRACEIOC_EPROBE: { 13334 dtrace_eprobedesc_t epdesc; 13335 dtrace_ecb_t *ecb; 13336 dtrace_action_t *act; 13337 void *buf; 13338 size_t size; 13339 uintptr_t dest; 13340 int nrecs; 13341 13342 if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0) 13343 return (EFAULT); 13344 13345 mutex_enter(&dtrace_lock); 13346 13347 if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) { 13348 mutex_exit(&dtrace_lock); 13349 return (EINVAL); 13350 } 13351 13352 if (ecb->dte_probe == NULL) { 13353 mutex_exit(&dtrace_lock); 13354 return (EINVAL); 13355 } 13356 13357 epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id; 13358 epdesc.dtepd_uarg = ecb->dte_uarg; 13359 epdesc.dtepd_size = ecb->dte_size; 13360 13361 nrecs = epdesc.dtepd_nrecs; 13362 epdesc.dtepd_nrecs = 0; 13363 for (act = ecb->dte_action; act != NULL; act = act->dta_next) { 13364 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple) 13365 continue; 13366 13367 epdesc.dtepd_nrecs++; 13368 } 13369 13370 /* 13371 * Now that we have the size, we need to allocate a temporary 13372 * buffer in which to store the complete description. We need 13373 * the temporary buffer to be able to drop dtrace_lock() 13374 * across the copyout(), below. 13375 */ 13376 size = sizeof (dtrace_eprobedesc_t) + 13377 (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t)); 13378 13379 buf = kmem_alloc(size, KM_SLEEP); 13380 dest = (uintptr_t)buf; 13381 13382 bcopy(&epdesc, (void *)dest, sizeof (epdesc)); 13383 dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]); 13384 13385 for (act = ecb->dte_action; act != NULL; act = act->dta_next) { 13386 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple) 13387 continue; 13388 13389 if (nrecs-- == 0) 13390 break; 13391 13392 bcopy(&act->dta_rec, (void *)dest, 13393 sizeof (dtrace_recdesc_t)); 13394 dest += sizeof (dtrace_recdesc_t); 13395 } 13396 13397 mutex_exit(&dtrace_lock); 13398 13399 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) { 13400 kmem_free(buf, size); 13401 return (EFAULT); 13402 } 13403 13404 kmem_free(buf, size); 13405 return (0); 13406 } 13407 13408 case DTRACEIOC_AGGDESC: { 13409 dtrace_aggdesc_t aggdesc; 13410 dtrace_action_t *act; 13411 dtrace_aggregation_t *agg; 13412 int nrecs; 13413 uint32_t offs; 13414 dtrace_recdesc_t *lrec; 13415 void *buf; 13416 size_t size; 13417 uintptr_t dest; 13418 13419 if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0) 13420 return (EFAULT); 13421 13422 mutex_enter(&dtrace_lock); 13423 13424 if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) { 13425 mutex_exit(&dtrace_lock); 13426 return (EINVAL); 13427 } 13428 13429 aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid; 13430 13431 nrecs = aggdesc.dtagd_nrecs; 13432 aggdesc.dtagd_nrecs = 0; 13433 13434 offs = agg->dtag_base; 13435 lrec = &agg->dtag_action.dta_rec; 13436 aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs; 13437 13438 for (act = agg->dtag_first; ; act = act->dta_next) { 13439 ASSERT(act->dta_intuple || 13440 DTRACEACT_ISAGG(act->dta_kind)); 13441 13442 /* 13443 * If this action has a record size of zero, it 13444 * denotes an argument to the aggregating action. 13445 * Because the presence of this record doesn't (or 13446 * shouldn't) affect the way the data is interpreted, 13447 * we don't copy it out to save user-level the 13448 * confusion of dealing with a zero-length record. 13449 */ 13450 if (act->dta_rec.dtrd_size == 0) { 13451 ASSERT(agg->dtag_hasarg); 13452 continue; 13453 } 13454 13455 aggdesc.dtagd_nrecs++; 13456 13457 if (act == &agg->dtag_action) 13458 break; 13459 } 13460 13461 /* 13462 * Now that we have the size, we need to allocate a temporary 13463 * buffer in which to store the complete description. We need 13464 * the temporary buffer to be able to drop dtrace_lock() 13465 * across the copyout(), below. 13466 */ 13467 size = sizeof (dtrace_aggdesc_t) + 13468 (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t)); 13469 13470 buf = kmem_alloc(size, KM_SLEEP); 13471 dest = (uintptr_t)buf; 13472 13473 bcopy(&aggdesc, (void *)dest, sizeof (aggdesc)); 13474 dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]); 13475 13476 for (act = agg->dtag_first; ; act = act->dta_next) { 13477 dtrace_recdesc_t rec = act->dta_rec; 13478 13479 /* 13480 * See the comment in the above loop for why we pass 13481 * over zero-length records. 13482 */ 13483 if (rec.dtrd_size == 0) { 13484 ASSERT(agg->dtag_hasarg); 13485 continue; 13486 } 13487 13488 if (nrecs-- == 0) 13489 break; 13490 13491 rec.dtrd_offset -= offs; 13492 bcopy(&rec, (void *)dest, sizeof (rec)); 13493 dest += sizeof (dtrace_recdesc_t); 13494 13495 if (act == &agg->dtag_action) 13496 break; 13497 } 13498 13499 mutex_exit(&dtrace_lock); 13500 13501 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) { 13502 kmem_free(buf, size); 13503 return (EFAULT); 13504 } 13505 13506 kmem_free(buf, size); 13507 return (0); 13508 } 13509 13510 case DTRACEIOC_ENABLE: { 13511 dof_hdr_t *dof; 13512 dtrace_enabling_t *enab = NULL; 13513 dtrace_vstate_t *vstate; 13514 int err = 0; 13515 13516 *rv = 0; 13517 13518 /* 13519 * If a NULL argument has been passed, we take this as our 13520 * cue to reevaluate our enablings. 13521 */ 13522 if (arg == NULL) { 13523 mutex_enter(&cpu_lock); 13524 mutex_enter(&dtrace_lock); 13525 err = dtrace_enabling_matchstate(state, rv); 13526 mutex_exit(&dtrace_lock); 13527 mutex_exit(&cpu_lock); 13528 13529 return (err); 13530 } 13531 13532 if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL) 13533 return (rval); 13534 13535 mutex_enter(&cpu_lock); 13536 mutex_enter(&dtrace_lock); 13537 vstate = &state->dts_vstate; 13538 13539 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) { 13540 mutex_exit(&dtrace_lock); 13541 mutex_exit(&cpu_lock); 13542 dtrace_dof_destroy(dof); 13543 return (EBUSY); 13544 } 13545 13546 if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) { 13547 mutex_exit(&dtrace_lock); 13548 mutex_exit(&cpu_lock); 13549 dtrace_dof_destroy(dof); 13550 return (EINVAL); 13551 } 13552 13553 if ((rval = dtrace_dof_options(dof, state)) != 0) { 13554 dtrace_enabling_destroy(enab); 13555 mutex_exit(&dtrace_lock); 13556 mutex_exit(&cpu_lock); 13557 dtrace_dof_destroy(dof); 13558 return (rval); 13559 } 13560 13561 if ((err = dtrace_enabling_match(enab, rv)) == 0) { 13562 err = dtrace_enabling_retain(enab); 13563 } else { 13564 dtrace_enabling_destroy(enab); 13565 } 13566 13567 mutex_exit(&cpu_lock); 13568 mutex_exit(&dtrace_lock); 13569 dtrace_dof_destroy(dof); 13570 13571 return (err); 13572 } 13573 13574 case DTRACEIOC_REPLICATE: { 13575 dtrace_repldesc_t desc; 13576 dtrace_probedesc_t *match = &desc.dtrpd_match; 13577 dtrace_probedesc_t *create = &desc.dtrpd_create; 13578 int err; 13579 13580 if (copyin((void *)arg, &desc, sizeof (desc)) != 0) 13581 return (EFAULT); 13582 13583 match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0'; 13584 match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0'; 13585 match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0'; 13586 match->dtpd_name[DTRACE_NAMELEN - 1] = '\0'; 13587 13588 create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0'; 13589 create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0'; 13590 create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0'; 13591 create->dtpd_name[DTRACE_NAMELEN - 1] = '\0'; 13592 13593 mutex_enter(&dtrace_lock); 13594 err = dtrace_enabling_replicate(state, match, create); 13595 mutex_exit(&dtrace_lock); 13596 13597 return (err); 13598 } 13599 13600 case DTRACEIOC_PROBEMATCH: 13601 case DTRACEIOC_PROBES: { 13602 dtrace_probe_t *probe = NULL; 13603 dtrace_probedesc_t desc; 13604 dtrace_probekey_t pkey; 13605 dtrace_id_t i; 13606 int m = 0; 13607 uint32_t priv; 13608 uid_t uid; 13609 13610 if (copyin((void *)arg, &desc, sizeof (desc)) != 0) 13611 return (EFAULT); 13612 13613 desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0'; 13614 desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0'; 13615 desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0'; 13616 desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0'; 13617 13618 /* 13619 * Before we attempt to match this probe, we want to give 13620 * all providers the opportunity to provide it. 13621 */ 13622 if (desc.dtpd_id == DTRACE_IDNONE) { 13623 mutex_enter(&dtrace_provider_lock); 13624 dtrace_probe_provide(&desc, NULL); 13625 mutex_exit(&dtrace_provider_lock); 13626 desc.dtpd_id++; 13627 } 13628 13629 if (cmd == DTRACEIOC_PROBEMATCH) { 13630 dtrace_probekey(&desc, &pkey); 13631 pkey.dtpk_id = DTRACE_IDNONE; 13632 } 13633 13634 uid = crgetuid(cr); 13635 dtrace_cred2priv(cr, &priv, &uid); 13636 13637 mutex_enter(&dtrace_lock); 13638 13639 if (cmd == DTRACEIOC_PROBEMATCH) { 13640 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) { 13641 if ((probe = dtrace_probes[i - 1]) != NULL && 13642 (m = dtrace_match_probe(probe, &pkey, 13643 priv, uid)) != 0) 13644 break; 13645 } 13646 13647 if (m < 0) { 13648 mutex_exit(&dtrace_lock); 13649 return (EINVAL); 13650 } 13651 13652 } else { 13653 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) { 13654 if ((probe = dtrace_probes[i - 1]) != NULL && 13655 dtrace_match_priv(probe, priv, uid)) 13656 break; 13657 } 13658 } 13659 13660 if (probe == NULL) { 13661 mutex_exit(&dtrace_lock); 13662 return (ESRCH); 13663 } 13664 13665 dtrace_probe_description(probe, &desc); 13666 mutex_exit(&dtrace_lock); 13667 13668 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0) 13669 return (EFAULT); 13670 13671 return (0); 13672 } 13673 13674 case DTRACEIOC_PROBEARG: { 13675 dtrace_argdesc_t desc; 13676 dtrace_probe_t *probe; 13677 dtrace_provider_t *prov; 13678 13679 if (copyin((void *)arg, &desc, sizeof (desc)) != 0) 13680 return (EFAULT); 13681 13682 if (desc.dtargd_id == DTRACE_IDNONE) 13683 return (EINVAL); 13684 13685 if (desc.dtargd_ndx == DTRACE_ARGNONE) 13686 return (EINVAL); 13687 13688 mutex_enter(&dtrace_provider_lock); 13689 mutex_enter(&mod_lock); 13690 mutex_enter(&dtrace_lock); 13691 13692 if (desc.dtargd_id > dtrace_nprobes) { 13693 mutex_exit(&dtrace_lock); 13694 mutex_exit(&mod_lock); 13695 mutex_exit(&dtrace_provider_lock); 13696 return (EINVAL); 13697 } 13698 13699 if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) { 13700 mutex_exit(&dtrace_lock); 13701 mutex_exit(&mod_lock); 13702 mutex_exit(&dtrace_provider_lock); 13703 return (EINVAL); 13704 } 13705 13706 mutex_exit(&dtrace_lock); 13707 13708 prov = probe->dtpr_provider; 13709 13710 if (prov->dtpv_pops.dtps_getargdesc == NULL) { 13711 /* 13712 * There isn't any typed information for this probe. 13713 * Set the argument number to DTRACE_ARGNONE. 13714 */ 13715 desc.dtargd_ndx = DTRACE_ARGNONE; 13716 } else { 13717 desc.dtargd_native[0] = '\0'; 13718 desc.dtargd_xlate[0] = '\0'; 13719 desc.dtargd_mapping = desc.dtargd_ndx; 13720 13721 prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg, 13722 probe->dtpr_id, probe->dtpr_arg, &desc); 13723 } 13724 13725 mutex_exit(&mod_lock); 13726 mutex_exit(&dtrace_provider_lock); 13727 13728 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0) 13729 return (EFAULT); 13730 13731 return (0); 13732 } 13733 13734 case DTRACEIOC_GO: { 13735 processorid_t cpuid; 13736 rval = dtrace_state_go(state, &cpuid); 13737 13738 if (rval != 0) 13739 return (rval); 13740 13741 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0) 13742 return (EFAULT); 13743 13744 return (0); 13745 } 13746 13747 case DTRACEIOC_STOP: { 13748 processorid_t cpuid; 13749 13750 mutex_enter(&dtrace_lock); 13751 rval = dtrace_state_stop(state, &cpuid); 13752 mutex_exit(&dtrace_lock); 13753 13754 if (rval != 0) 13755 return (rval); 13756 13757 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0) 13758 return (EFAULT); 13759 13760 return (0); 13761 } 13762 13763 case DTRACEIOC_DOFGET: { 13764 dof_hdr_t hdr, *dof; 13765 uint64_t len; 13766 13767 if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0) 13768 return (EFAULT); 13769 13770 mutex_enter(&dtrace_lock); 13771 dof = dtrace_dof_create(state); 13772 mutex_exit(&dtrace_lock); 13773 13774 len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz); 13775 rval = copyout(dof, (void *)arg, len); 13776 dtrace_dof_destroy(dof); 13777 13778 return (rval == 0 ? 0 : EFAULT); 13779 } 13780 13781 case DTRACEIOC_AGGSNAP: 13782 case DTRACEIOC_BUFSNAP: { 13783 dtrace_bufdesc_t desc; 13784 caddr_t cached; 13785 dtrace_buffer_t *buf; 13786 13787 if (copyin((void *)arg, &desc, sizeof (desc)) != 0) 13788 return (EFAULT); 13789 13790 if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU) 13791 return (EINVAL); 13792 13793 mutex_enter(&dtrace_lock); 13794 13795 if (cmd == DTRACEIOC_BUFSNAP) { 13796 buf = &state->dts_buffer[desc.dtbd_cpu]; 13797 } else { 13798 buf = &state->dts_aggbuffer[desc.dtbd_cpu]; 13799 } 13800 13801 if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) { 13802 size_t sz = buf->dtb_offset; 13803 13804 if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) { 13805 mutex_exit(&dtrace_lock); 13806 return (EBUSY); 13807 } 13808 13809 /* 13810 * If this buffer has already been consumed, we're 13811 * going to indicate that there's nothing left here 13812 * to consume. 13813 */ 13814 if (buf->dtb_flags & DTRACEBUF_CONSUMED) { 13815 mutex_exit(&dtrace_lock); 13816 13817 desc.dtbd_size = 0; 13818 desc.dtbd_drops = 0; 13819 desc.dtbd_errors = 0; 13820 desc.dtbd_oldest = 0; 13821 sz = sizeof (desc); 13822 13823 if (copyout(&desc, (void *)arg, sz) != 0) 13824 return (EFAULT); 13825 13826 return (0); 13827 } 13828 13829 /* 13830 * If this is a ring buffer that has wrapped, we want 13831 * to copy the whole thing out. 13832 */ 13833 if (buf->dtb_flags & DTRACEBUF_WRAPPED) { 13834 dtrace_buffer_polish(buf); 13835 sz = buf->dtb_size; 13836 } 13837 13838 if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) { 13839 mutex_exit(&dtrace_lock); 13840 return (EFAULT); 13841 } 13842 13843 desc.dtbd_size = sz; 13844 desc.dtbd_drops = buf->dtb_drops; 13845 desc.dtbd_errors = buf->dtb_errors; 13846 desc.dtbd_oldest = buf->dtb_xamot_offset; 13847 13848 mutex_exit(&dtrace_lock); 13849 13850 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0) 13851 return (EFAULT); 13852 13853 buf->dtb_flags |= DTRACEBUF_CONSUMED; 13854 13855 return (0); 13856 } 13857 13858 if (buf->dtb_tomax == NULL) { 13859 ASSERT(buf->dtb_xamot == NULL); 13860 mutex_exit(&dtrace_lock); 13861 return (ENOENT); 13862 } 13863 13864 cached = buf->dtb_tomax; 13865 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH)); 13866 13867 dtrace_xcall(desc.dtbd_cpu, 13868 (dtrace_xcall_t)dtrace_buffer_switch, buf); 13869 13870 state->dts_errors += buf->dtb_xamot_errors; 13871 13872 /* 13873 * If the buffers did not actually switch, then the cross call 13874 * did not take place -- presumably because the given CPU is 13875 * not in the ready set. If this is the case, we'll return 13876 * ENOENT. 13877 */ 13878 if (buf->dtb_tomax == cached) { 13879 ASSERT(buf->dtb_xamot != cached); 13880 mutex_exit(&dtrace_lock); 13881 return (ENOENT); 13882 } 13883 13884 ASSERT(cached == buf->dtb_xamot); 13885 13886 /* 13887 * We have our snapshot; now copy it out. 13888 */ 13889 if (copyout(buf->dtb_xamot, desc.dtbd_data, 13890 buf->dtb_xamot_offset) != 0) { 13891 mutex_exit(&dtrace_lock); 13892 return (EFAULT); 13893 } 13894 13895 desc.dtbd_size = buf->dtb_xamot_offset; 13896 desc.dtbd_drops = buf->dtb_xamot_drops; 13897 desc.dtbd_errors = buf->dtb_xamot_errors; 13898 desc.dtbd_oldest = 0; 13899 13900 mutex_exit(&dtrace_lock); 13901 13902 /* 13903 * Finally, copy out the buffer description. 13904 */ 13905 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0) 13906 return (EFAULT); 13907 13908 return (0); 13909 } 13910 13911 case DTRACEIOC_CONF: { 13912 dtrace_conf_t conf; 13913 13914 bzero(&conf, sizeof (conf)); 13915 conf.dtc_difversion = DIF_VERSION; 13916 conf.dtc_difintregs = DIF_DIR_NREGS; 13917 conf.dtc_diftupregs = DIF_DTR_NREGS; 13918 conf.dtc_ctfmodel = CTF_MODEL_NATIVE; 13919 13920 if (copyout(&conf, (void *)arg, sizeof (conf)) != 0) 13921 return (EFAULT); 13922 13923 return (0); 13924 } 13925 13926 case DTRACEIOC_STATUS: { 13927 dtrace_status_t stat; 13928 dtrace_dstate_t *dstate; 13929 int i, j; 13930 uint64_t nerrs; 13931 13932 /* 13933 * See the comment in dtrace_state_deadman() for the reason 13934 * for setting dts_laststatus to INT64_MAX before setting 13935 * it to the correct value. 13936 */ 13937 state->dts_laststatus = INT64_MAX; 13938 dtrace_membar_producer(); 13939 state->dts_laststatus = dtrace_gethrtime(); 13940 13941 bzero(&stat, sizeof (stat)); 13942 13943 mutex_enter(&dtrace_lock); 13944 13945 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) { 13946 mutex_exit(&dtrace_lock); 13947 return (ENOENT); 13948 } 13949 13950 if (state->dts_activity == DTRACE_ACTIVITY_DRAINING) 13951 stat.dtst_exiting = 1; 13952 13953 nerrs = state->dts_errors; 13954 dstate = &state->dts_vstate.dtvs_dynvars; 13955 13956 for (i = 0; i < NCPU; i++) { 13957 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i]; 13958 13959 stat.dtst_dyndrops += dcpu->dtdsc_drops; 13960 stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops; 13961 stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops; 13962 13963 if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL) 13964 stat.dtst_filled++; 13965 13966 nerrs += state->dts_buffer[i].dtb_errors; 13967 13968 for (j = 0; j < state->dts_nspeculations; j++) { 13969 dtrace_speculation_t *spec; 13970 dtrace_buffer_t *buf; 13971 13972 spec = &state->dts_speculations[j]; 13973 buf = &spec->dtsp_buffer[i]; 13974 stat.dtst_specdrops += buf->dtb_xamot_drops; 13975 } 13976 } 13977 13978 stat.dtst_specdrops_busy = state->dts_speculations_busy; 13979 stat.dtst_specdrops_unavail = state->dts_speculations_unavail; 13980 stat.dtst_stkstroverflows = state->dts_stkstroverflows; 13981 stat.dtst_dblerrors = state->dts_dblerrors; 13982 stat.dtst_killed = 13983 (state->dts_activity == DTRACE_ACTIVITY_KILLED); 13984 stat.dtst_errors = nerrs; 13985 13986 mutex_exit(&dtrace_lock); 13987 13988 if (copyout(&stat, (void *)arg, sizeof (stat)) != 0) 13989 return (EFAULT); 13990 13991 return (0); 13992 } 13993 13994 case DTRACEIOC_FORMAT: { 13995 dtrace_fmtdesc_t fmt; 13996 char *str; 13997 int len; 13998 13999 if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0) 14000 return (EFAULT); 14001 14002 mutex_enter(&dtrace_lock); 14003 14004 if (fmt.dtfd_format == 0 || 14005 fmt.dtfd_format > state->dts_nformats) { 14006 mutex_exit(&dtrace_lock); 14007 return (EINVAL); 14008 } 14009 14010 /* 14011 * Format strings are allocated contiguously and they are 14012 * never freed; if a format index is less than the number 14013 * of formats, we can assert that the format map is non-NULL 14014 * and that the format for the specified index is non-NULL. 14015 */ 14016 ASSERT(state->dts_formats != NULL); 14017 str = state->dts_formats[fmt.dtfd_format - 1]; 14018 ASSERT(str != NULL); 14019 14020 len = strlen(str) + 1; 14021 14022 if (len > fmt.dtfd_length) { 14023 fmt.dtfd_length = len; 14024 14025 if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) { 14026 mutex_exit(&dtrace_lock); 14027 return (EINVAL); 14028 } 14029 } else { 14030 if (copyout(str, fmt.dtfd_string, len) != 0) { 14031 mutex_exit(&dtrace_lock); 14032 return (EINVAL); 14033 } 14034 } 14035 14036 mutex_exit(&dtrace_lock); 14037 return (0); 14038 } 14039 14040 default: 14041 break; 14042 } 14043 14044 return (ENOTTY); 14045 } 14046 14047 /*ARGSUSED*/ 14048 static int 14049 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd) 14050 { 14051 dtrace_state_t *state; 14052 14053 switch (cmd) { 14054 case DDI_DETACH: 14055 break; 14056 14057 case DDI_SUSPEND: 14058 return (DDI_SUCCESS); 14059 14060 default: 14061 return (DDI_FAILURE); 14062 } 14063 14064 mutex_enter(&cpu_lock); 14065 mutex_enter(&dtrace_provider_lock); 14066 mutex_enter(&dtrace_lock); 14067 14068 ASSERT(dtrace_opens == 0); 14069 14070 if (dtrace_helpers > 0) { 14071 mutex_exit(&dtrace_provider_lock); 14072 mutex_exit(&dtrace_lock); 14073 mutex_exit(&cpu_lock); 14074 return (DDI_FAILURE); 14075 } 14076 14077 if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) { 14078 mutex_exit(&dtrace_provider_lock); 14079 mutex_exit(&dtrace_lock); 14080 mutex_exit(&cpu_lock); 14081 return (DDI_FAILURE); 14082 } 14083 14084 dtrace_provider = NULL; 14085 14086 if ((state = dtrace_anon_grab()) != NULL) { 14087 /* 14088 * If there were ECBs on this state, the provider should 14089 * have not been allowed to detach; assert that there is 14090 * none. 14091 */ 14092 ASSERT(state->dts_necbs == 0); 14093 dtrace_state_destroy(state); 14094 14095 /* 14096 * If we're being detached with anonymous state, we need to 14097 * indicate to the kernel debugger that DTrace is now inactive. 14098 */ 14099 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE); 14100 } 14101 14102 bzero(&dtrace_anon, sizeof (dtrace_anon_t)); 14103 unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL); 14104 dtrace_cpu_init = NULL; 14105 dtrace_helpers_cleanup = NULL; 14106 dtrace_helpers_fork = NULL; 14107 dtrace_cpustart_init = NULL; 14108 dtrace_cpustart_fini = NULL; 14109 dtrace_debugger_init = NULL; 14110 dtrace_debugger_fini = NULL; 14111 dtrace_kreloc_init = NULL; 14112 dtrace_kreloc_fini = NULL; 14113 dtrace_modload = NULL; 14114 dtrace_modunload = NULL; 14115 14116 mutex_exit(&cpu_lock); 14117 14118 if (dtrace_helptrace_enabled) { 14119 kmem_free(dtrace_helptrace_buffer, dtrace_helptrace_bufsize); 14120 dtrace_helptrace_buffer = NULL; 14121 } 14122 14123 kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *)); 14124 dtrace_probes = NULL; 14125 dtrace_nprobes = 0; 14126 14127 dtrace_hash_destroy(dtrace_bymod); 14128 dtrace_hash_destroy(dtrace_byfunc); 14129 dtrace_hash_destroy(dtrace_byname); 14130 dtrace_bymod = NULL; 14131 dtrace_byfunc = NULL; 14132 dtrace_byname = NULL; 14133 14134 kmem_cache_destroy(dtrace_state_cache); 14135 vmem_destroy(dtrace_minor); 14136 vmem_destroy(dtrace_arena); 14137 14138 if (dtrace_toxrange != NULL) { 14139 kmem_free(dtrace_toxrange, 14140 dtrace_toxranges_max * sizeof (dtrace_toxrange_t)); 14141 dtrace_toxrange = NULL; 14142 dtrace_toxranges = 0; 14143 dtrace_toxranges_max = 0; 14144 } 14145 14146 ddi_remove_minor_node(dtrace_devi, NULL); 14147 dtrace_devi = NULL; 14148 14149 ddi_soft_state_fini(&dtrace_softstate); 14150 14151 ASSERT(dtrace_vtime_references == 0); 14152 ASSERT(dtrace_opens == 0); 14153 ASSERT(dtrace_retained == NULL); 14154 14155 mutex_exit(&dtrace_lock); 14156 mutex_exit(&dtrace_provider_lock); 14157 14158 /* 14159 * We don't destroy the task queue until after we have dropped our 14160 * locks (taskq_destroy() may block on running tasks). To prevent 14161 * attempting to do work after we have effectively detached but before 14162 * the task queue has been destroyed, all tasks dispatched via the 14163 * task queue must check that DTrace is still attached before 14164 * performing any operation. 14165 */ 14166 taskq_destroy(dtrace_taskq); 14167 dtrace_taskq = NULL; 14168 14169 return (DDI_SUCCESS); 14170 } 14171 14172 /*ARGSUSED*/ 14173 static int 14174 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result) 14175 { 14176 int error; 14177 14178 switch (infocmd) { 14179 case DDI_INFO_DEVT2DEVINFO: 14180 *result = (void *)dtrace_devi; 14181 error = DDI_SUCCESS; 14182 break; 14183 case DDI_INFO_DEVT2INSTANCE: 14184 *result = (void *)0; 14185 error = DDI_SUCCESS; 14186 break; 14187 default: 14188 error = DDI_FAILURE; 14189 } 14190 return (error); 14191 } 14192 14193 static struct cb_ops dtrace_cb_ops = { 14194 dtrace_open, /* open */ 14195 dtrace_close, /* close */ 14196 nulldev, /* strategy */ 14197 nulldev, /* print */ 14198 nodev, /* dump */ 14199 nodev, /* read */ 14200 nodev, /* write */ 14201 dtrace_ioctl, /* ioctl */ 14202 nodev, /* devmap */ 14203 nodev, /* mmap */ 14204 nodev, /* segmap */ 14205 nochpoll, /* poll */ 14206 ddi_prop_op, /* cb_prop_op */ 14207 0, /* streamtab */ 14208 D_NEW | D_MP /* Driver compatibility flag */ 14209 }; 14210 14211 static struct dev_ops dtrace_ops = { 14212 DEVO_REV, /* devo_rev */ 14213 0, /* refcnt */ 14214 dtrace_info, /* get_dev_info */ 14215 nulldev, /* identify */ 14216 nulldev, /* probe */ 14217 dtrace_attach, /* attach */ 14218 dtrace_detach, /* detach */ 14219 nodev, /* reset */ 14220 &dtrace_cb_ops, /* driver operations */ 14221 NULL, /* bus operations */ 14222 nodev /* dev power */ 14223 }; 14224 14225 static struct modldrv modldrv = { 14226 &mod_driverops, /* module type (this is a pseudo driver) */ 14227 "Dynamic Tracing", /* name of module */ 14228 &dtrace_ops, /* driver ops */ 14229 }; 14230 14231 static struct modlinkage modlinkage = { 14232 MODREV_1, 14233 (void *)&modldrv, 14234 NULL 14235 }; 14236 14237 int 14238 _init(void) 14239 { 14240 return (mod_install(&modlinkage)); 14241 } 14242 14243 int 14244 _info(struct modinfo *modinfop) 14245 { 14246 return (mod_info(&modlinkage, modinfop)); 14247 } 14248 14249 int 14250 _fini(void) 14251 { 14252 return (mod_remove(&modlinkage)); 14253 } 14254