dtrace.c revision 268572
1/* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 * 21 * $FreeBSD: stable/10/sys/cddl/contrib/opensolaris/uts/common/dtrace/dtrace.c 268572 2014-07-12 18:23:35Z pfg $ 22 */ 23 24/* 25 * Copyright 2008 Sun Microsystems, Inc. All rights reserved. 26 * Copyright (c) 2013, Joyent, Inc. All rights reserved. 27 * Copyright (c) 2012 by Delphix. All rights reserved. 28 */ 29 30/* 31 * DTrace - Dynamic Tracing for Solaris 32 * 33 * This is the implementation of the Solaris Dynamic Tracing framework 34 * (DTrace). The user-visible interface to DTrace is described at length in 35 * the "Solaris Dynamic Tracing Guide". The interfaces between the libdtrace 36 * library, the in-kernel DTrace framework, and the DTrace providers are 37 * described in the block comments in the <sys/dtrace.h> header file. The 38 * internal architecture of DTrace is described in the block comments in the 39 * <sys/dtrace_impl.h> header file. The comments contained within the DTrace 40 * implementation very much assume mastery of all of these sources; if one has 41 * an unanswered question about the implementation, one should consult them 42 * first. 43 * 44 * The functions here are ordered roughly as follows: 45 * 46 * - Probe context functions 47 * - Probe hashing functions 48 * - Non-probe context utility functions 49 * - Matching functions 50 * - Provider-to-Framework API functions 51 * - Probe management functions 52 * - DIF object functions 53 * - Format functions 54 * - Predicate functions 55 * - ECB functions 56 * - Buffer functions 57 * - Enabling functions 58 * - DOF functions 59 * - Anonymous enabling functions 60 * - Consumer state functions 61 * - Helper functions 62 * - Hook functions 63 * - Driver cookbook functions 64 * 65 * Each group of functions begins with a block comment labelled the "DTrace 66 * [Group] Functions", allowing one to find each block by searching forward 67 * on capital-f functions. 68 */ 69#include <sys/errno.h> 70#if !defined(sun) 71#include <sys/time.h> 72#endif 73#include <sys/stat.h> 74#include <sys/modctl.h> 75#include <sys/conf.h> 76#include <sys/systm.h> 77#if defined(sun) 78#include <sys/ddi.h> 79#include <sys/sunddi.h> 80#endif 81#include <sys/cpuvar.h> 82#include <sys/kmem.h> 83#if defined(sun) 84#include <sys/strsubr.h> 85#endif 86#include <sys/sysmacros.h> 87#include <sys/dtrace_impl.h> 88#include <sys/atomic.h> 89#include <sys/cmn_err.h> 90#if defined(sun) 91#include <sys/mutex_impl.h> 92#include <sys/rwlock_impl.h> 93#endif 94#include <sys/ctf_api.h> 95#if defined(sun) 96#include <sys/panic.h> 97#include <sys/priv_impl.h> 98#endif 99#include <sys/policy.h> 100#if defined(sun) 101#include <sys/cred_impl.h> 102#include <sys/procfs_isa.h> 103#endif 104#include <sys/taskq.h> 105#if defined(sun) 106#include <sys/mkdev.h> 107#include <sys/kdi.h> 108#endif 109#include <sys/zone.h> 110#include <sys/socket.h> 111#include <netinet/in.h> 112 113/* FreeBSD includes: */ 114#if !defined(sun) 115#include <sys/callout.h> 116#include <sys/ctype.h> 117#include <sys/eventhandler.h> 118#include <sys/limits.h> 119#include <sys/kdb.h> 120#include <sys/kernel.h> 121#include <sys/malloc.h> 122#include <sys/sysctl.h> 123#include <sys/lock.h> 124#include <sys/mutex.h> 125#include <sys/rwlock.h> 126#include <sys/sx.h> 127#include <sys/dtrace_bsd.h> 128#include <netinet/in.h> 129#include "dtrace_cddl.h" 130#include "dtrace_debug.c" 131#endif 132 133/* 134 * DTrace Tunable Variables 135 * 136 * The following variables may be tuned by adding a line to /etc/system that 137 * includes both the name of the DTrace module ("dtrace") and the name of the 138 * variable. For example: 139 * 140 * set dtrace:dtrace_destructive_disallow = 1 141 * 142 * In general, the only variables that one should be tuning this way are those 143 * that affect system-wide DTrace behavior, and for which the default behavior 144 * is undesirable. Most of these variables are tunable on a per-consumer 145 * basis using DTrace options, and need not be tuned on a system-wide basis. 146 * When tuning these variables, avoid pathological values; while some attempt 147 * is made to verify the integrity of these variables, they are not considered 148 * part of the supported interface to DTrace, and they are therefore not 149 * checked comprehensively. Further, these variables should not be tuned 150 * dynamically via "mdb -kw" or other means; they should only be tuned via 151 * /etc/system. 152 */ 153int dtrace_destructive_disallow = 0; 154dtrace_optval_t dtrace_nonroot_maxsize = (16 * 1024 * 1024); 155size_t dtrace_difo_maxsize = (256 * 1024); 156dtrace_optval_t dtrace_dof_maxsize = (8 * 1024 * 1024); 157size_t dtrace_global_maxsize = (16 * 1024); 158size_t dtrace_actions_max = (16 * 1024); 159size_t dtrace_retain_max = 1024; 160dtrace_optval_t dtrace_helper_actions_max = 128; 161dtrace_optval_t dtrace_helper_providers_max = 32; 162dtrace_optval_t dtrace_dstate_defsize = (1 * 1024 * 1024); 163size_t dtrace_strsize_default = 256; 164dtrace_optval_t dtrace_cleanrate_default = 9900990; /* 101 hz */ 165dtrace_optval_t dtrace_cleanrate_min = 200000; /* 5000 hz */ 166dtrace_optval_t dtrace_cleanrate_max = (uint64_t)60 * NANOSEC; /* 1/minute */ 167dtrace_optval_t dtrace_aggrate_default = NANOSEC; /* 1 hz */ 168dtrace_optval_t dtrace_statusrate_default = NANOSEC; /* 1 hz */ 169dtrace_optval_t dtrace_statusrate_max = (hrtime_t)10 * NANOSEC; /* 6/minute */ 170dtrace_optval_t dtrace_switchrate_default = NANOSEC; /* 1 hz */ 171dtrace_optval_t dtrace_nspec_default = 1; 172dtrace_optval_t dtrace_specsize_default = 32 * 1024; 173dtrace_optval_t dtrace_stackframes_default = 20; 174dtrace_optval_t dtrace_ustackframes_default = 20; 175dtrace_optval_t dtrace_jstackframes_default = 50; 176dtrace_optval_t dtrace_jstackstrsize_default = 512; 177int dtrace_msgdsize_max = 128; 178hrtime_t dtrace_chill_max = 500 * (NANOSEC / MILLISEC); /* 500 ms */ 179hrtime_t dtrace_chill_interval = NANOSEC; /* 1000 ms */ 180int dtrace_devdepth_max = 32; 181int dtrace_err_verbose; 182hrtime_t dtrace_deadman_interval = NANOSEC; 183hrtime_t dtrace_deadman_timeout = (hrtime_t)10 * NANOSEC; 184hrtime_t dtrace_deadman_user = (hrtime_t)30 * NANOSEC; 185hrtime_t dtrace_unregister_defunct_reap = (hrtime_t)60 * NANOSEC; 186 187/* 188 * DTrace External Variables 189 * 190 * As dtrace(7D) is a kernel module, any DTrace variables are obviously 191 * available to DTrace consumers via the backtick (`) syntax. One of these, 192 * dtrace_zero, is made deliberately so: it is provided as a source of 193 * well-known, zero-filled memory. While this variable is not documented, 194 * it is used by some translators as an implementation detail. 195 */ 196const char dtrace_zero[256] = { 0 }; /* zero-filled memory */ 197 198/* 199 * DTrace Internal Variables 200 */ 201#if defined(sun) 202static dev_info_t *dtrace_devi; /* device info */ 203#endif 204#if defined(sun) 205static vmem_t *dtrace_arena; /* probe ID arena */ 206static vmem_t *dtrace_minor; /* minor number arena */ 207#else 208static taskq_t *dtrace_taskq; /* task queue */ 209static struct unrhdr *dtrace_arena; /* Probe ID number. */ 210#endif 211static dtrace_probe_t **dtrace_probes; /* array of all probes */ 212static int dtrace_nprobes; /* number of probes */ 213static dtrace_provider_t *dtrace_provider; /* provider list */ 214static dtrace_meta_t *dtrace_meta_pid; /* user-land meta provider */ 215static int dtrace_opens; /* number of opens */ 216static int dtrace_helpers; /* number of helpers */ 217#if defined(sun) 218static void *dtrace_softstate; /* softstate pointer */ 219#endif 220static dtrace_hash_t *dtrace_bymod; /* probes hashed by module */ 221static dtrace_hash_t *dtrace_byfunc; /* probes hashed by function */ 222static dtrace_hash_t *dtrace_byname; /* probes hashed by name */ 223static dtrace_toxrange_t *dtrace_toxrange; /* toxic range array */ 224static int dtrace_toxranges; /* number of toxic ranges */ 225static int dtrace_toxranges_max; /* size of toxic range array */ 226static dtrace_anon_t dtrace_anon; /* anonymous enabling */ 227static kmem_cache_t *dtrace_state_cache; /* cache for dynamic state */ 228static uint64_t dtrace_vtime_references; /* number of vtimestamp refs */ 229static kthread_t *dtrace_panicked; /* panicking thread */ 230static dtrace_ecb_t *dtrace_ecb_create_cache; /* cached created ECB */ 231static dtrace_genid_t dtrace_probegen; /* current probe generation */ 232static dtrace_helpers_t *dtrace_deferred_pid; /* deferred helper list */ 233static dtrace_enabling_t *dtrace_retained; /* list of retained enablings */ 234static dtrace_genid_t dtrace_retained_gen; /* current retained enab gen */ 235static dtrace_dynvar_t dtrace_dynhash_sink; /* end of dynamic hash chains */ 236#if !defined(sun) 237static struct mtx dtrace_unr_mtx; 238MTX_SYSINIT(dtrace_unr_mtx, &dtrace_unr_mtx, "Unique resource identifier", MTX_DEF); 239int dtrace_in_probe; /* non-zero if executing a probe */ 240#if defined(__i386__) || defined(__amd64__) || defined(__mips__) || defined(__powerpc__) 241uintptr_t dtrace_in_probe_addr; /* Address of invop when already in probe */ 242#endif 243static eventhandler_tag dtrace_kld_load_tag; 244static eventhandler_tag dtrace_kld_unload_try_tag; 245#endif 246 247/* 248 * DTrace Locking 249 * DTrace is protected by three (relatively coarse-grained) locks: 250 * 251 * (1) dtrace_lock is required to manipulate essentially any DTrace state, 252 * including enabling state, probes, ECBs, consumer state, helper state, 253 * etc. Importantly, dtrace_lock is _not_ required when in probe context; 254 * probe context is lock-free -- synchronization is handled via the 255 * dtrace_sync() cross call mechanism. 256 * 257 * (2) dtrace_provider_lock is required when manipulating provider state, or 258 * when provider state must be held constant. 259 * 260 * (3) dtrace_meta_lock is required when manipulating meta provider state, or 261 * when meta provider state must be held constant. 262 * 263 * The lock ordering between these three locks is dtrace_meta_lock before 264 * dtrace_provider_lock before dtrace_lock. (In particular, there are 265 * several places where dtrace_provider_lock is held by the framework as it 266 * calls into the providers -- which then call back into the framework, 267 * grabbing dtrace_lock.) 268 * 269 * There are two other locks in the mix: mod_lock and cpu_lock. With respect 270 * to dtrace_provider_lock and dtrace_lock, cpu_lock continues its historical 271 * role as a coarse-grained lock; it is acquired before both of these locks. 272 * With respect to dtrace_meta_lock, its behavior is stranger: cpu_lock must 273 * be acquired _between_ dtrace_meta_lock and any other DTrace locks. 274 * mod_lock is similar with respect to dtrace_provider_lock in that it must be 275 * acquired _between_ dtrace_provider_lock and dtrace_lock. 276 */ 277static kmutex_t dtrace_lock; /* probe state lock */ 278static kmutex_t dtrace_provider_lock; /* provider state lock */ 279static kmutex_t dtrace_meta_lock; /* meta-provider state lock */ 280 281#if !defined(sun) 282/* XXX FreeBSD hacks. */ 283#define cr_suid cr_svuid 284#define cr_sgid cr_svgid 285#define ipaddr_t in_addr_t 286#define mod_modname pathname 287#define vuprintf vprintf 288#define ttoproc(_a) ((_a)->td_proc) 289#define crgetzoneid(_a) 0 290#define NCPU MAXCPU 291#define SNOCD 0 292#define CPU_ON_INTR(_a) 0 293 294#define PRIV_EFFECTIVE (1 << 0) 295#define PRIV_DTRACE_KERNEL (1 << 1) 296#define PRIV_DTRACE_PROC (1 << 2) 297#define PRIV_DTRACE_USER (1 << 3) 298#define PRIV_PROC_OWNER (1 << 4) 299#define PRIV_PROC_ZONE (1 << 5) 300#define PRIV_ALL ~0 301 302SYSCTL_DECL(_debug_dtrace); 303SYSCTL_DECL(_kern_dtrace); 304#endif 305 306#if defined(sun) 307#define curcpu CPU->cpu_id 308#endif 309 310 311/* 312 * DTrace Provider Variables 313 * 314 * These are the variables relating to DTrace as a provider (that is, the 315 * provider of the BEGIN, END, and ERROR probes). 316 */ 317static dtrace_pattr_t dtrace_provider_attr = { 318{ DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON }, 319{ DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN }, 320{ DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN }, 321{ DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON }, 322{ DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON }, 323}; 324 325static void 326dtrace_nullop(void) 327{} 328 329static dtrace_pops_t dtrace_provider_ops = { 330 (void (*)(void *, dtrace_probedesc_t *))dtrace_nullop, 331 (void (*)(void *, modctl_t *))dtrace_nullop, 332 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop, 333 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop, 334 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop, 335 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop, 336 NULL, 337 NULL, 338 NULL, 339 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop 340}; 341 342static dtrace_id_t dtrace_probeid_begin; /* special BEGIN probe */ 343static dtrace_id_t dtrace_probeid_end; /* special END probe */ 344dtrace_id_t dtrace_probeid_error; /* special ERROR probe */ 345 346/* 347 * DTrace Helper Tracing Variables 348 */ 349uint32_t dtrace_helptrace_next = 0; 350uint32_t dtrace_helptrace_nlocals; 351char *dtrace_helptrace_buffer; 352int dtrace_helptrace_bufsize = 512 * 1024; 353 354#ifdef DEBUG 355int dtrace_helptrace_enabled = 1; 356#else 357int dtrace_helptrace_enabled = 0; 358#endif 359 360/* 361 * DTrace Error Hashing 362 * 363 * On DEBUG kernels, DTrace will track the errors that has seen in a hash 364 * table. This is very useful for checking coverage of tests that are 365 * expected to induce DIF or DOF processing errors, and may be useful for 366 * debugging problems in the DIF code generator or in DOF generation . The 367 * error hash may be examined with the ::dtrace_errhash MDB dcmd. 368 */ 369#ifdef DEBUG 370static dtrace_errhash_t dtrace_errhash[DTRACE_ERRHASHSZ]; 371static const char *dtrace_errlast; 372static kthread_t *dtrace_errthread; 373static kmutex_t dtrace_errlock; 374#endif 375 376/* 377 * DTrace Macros and Constants 378 * 379 * These are various macros that are useful in various spots in the 380 * implementation, along with a few random constants that have no meaning 381 * outside of the implementation. There is no real structure to this cpp 382 * mishmash -- but is there ever? 383 */ 384#define DTRACE_HASHSTR(hash, probe) \ 385 dtrace_hash_str(*((char **)((uintptr_t)(probe) + (hash)->dth_stroffs))) 386 387#define DTRACE_HASHNEXT(hash, probe) \ 388 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_nextoffs) 389 390#define DTRACE_HASHPREV(hash, probe) \ 391 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_prevoffs) 392 393#define DTRACE_HASHEQ(hash, lhs, rhs) \ 394 (strcmp(*((char **)((uintptr_t)(lhs) + (hash)->dth_stroffs)), \ 395 *((char **)((uintptr_t)(rhs) + (hash)->dth_stroffs))) == 0) 396 397#define DTRACE_AGGHASHSIZE_SLEW 17 398 399#define DTRACE_V4MAPPED_OFFSET (sizeof (uint32_t) * 3) 400 401/* 402 * The key for a thread-local variable consists of the lower 61 bits of the 403 * t_did, plus the 3 bits of the highest active interrupt above LOCK_LEVEL. 404 * We add DIF_VARIABLE_MAX to t_did to assure that the thread key is never 405 * equal to a variable identifier. This is necessary (but not sufficient) to 406 * assure that global associative arrays never collide with thread-local 407 * variables. To guarantee that they cannot collide, we must also define the 408 * order for keying dynamic variables. That order is: 409 * 410 * [ key0 ] ... [ keyn ] [ variable-key ] [ tls-key ] 411 * 412 * Because the variable-key and the tls-key are in orthogonal spaces, there is 413 * no way for a global variable key signature to match a thread-local key 414 * signature. 415 */ 416#if defined(sun) 417#define DTRACE_TLS_THRKEY(where) { \ 418 uint_t intr = 0; \ 419 uint_t actv = CPU->cpu_intr_actv >> (LOCK_LEVEL + 1); \ 420 for (; actv; actv >>= 1) \ 421 intr++; \ 422 ASSERT(intr < (1 << 3)); \ 423 (where) = ((curthread->t_did + DIF_VARIABLE_MAX) & \ 424 (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \ 425} 426#else 427#define DTRACE_TLS_THRKEY(where) { \ 428 solaris_cpu_t *_c = &solaris_cpu[curcpu]; \ 429 uint_t intr = 0; \ 430 uint_t actv = _c->cpu_intr_actv; \ 431 for (; actv; actv >>= 1) \ 432 intr++; \ 433 ASSERT(intr < (1 << 3)); \ 434 (where) = ((curthread->td_tid + DIF_VARIABLE_MAX) & \ 435 (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \ 436} 437#endif 438 439#define DT_BSWAP_8(x) ((x) & 0xff) 440#define DT_BSWAP_16(x) ((DT_BSWAP_8(x) << 8) | DT_BSWAP_8((x) >> 8)) 441#define DT_BSWAP_32(x) ((DT_BSWAP_16(x) << 16) | DT_BSWAP_16((x) >> 16)) 442#define DT_BSWAP_64(x) ((DT_BSWAP_32(x) << 32) | DT_BSWAP_32((x) >> 32)) 443 444#define DT_MASK_LO 0x00000000FFFFFFFFULL 445 446#define DTRACE_STORE(type, tomax, offset, what) \ 447 *((type *)((uintptr_t)(tomax) + (uintptr_t)offset)) = (type)(what); 448 449#ifndef __x86 450#define DTRACE_ALIGNCHECK(addr, size, flags) \ 451 if (addr & (size - 1)) { \ 452 *flags |= CPU_DTRACE_BADALIGN; \ 453 cpu_core[curcpu].cpuc_dtrace_illval = addr; \ 454 return (0); \ 455 } 456#else 457#define DTRACE_ALIGNCHECK(addr, size, flags) 458#endif 459 460/* 461 * Test whether a range of memory starting at testaddr of size testsz falls 462 * within the range of memory described by addr, sz. We take care to avoid 463 * problems with overflow and underflow of the unsigned quantities, and 464 * disallow all negative sizes. Ranges of size 0 are allowed. 465 */ 466#define DTRACE_INRANGE(testaddr, testsz, baseaddr, basesz) \ 467 ((testaddr) - (baseaddr) < (basesz) && \ 468 (testaddr) + (testsz) - (baseaddr) <= (basesz) && \ 469 (testaddr) + (testsz) >= (testaddr)) 470 471/* 472 * Test whether alloc_sz bytes will fit in the scratch region. We isolate 473 * alloc_sz on the righthand side of the comparison in order to avoid overflow 474 * or underflow in the comparison with it. This is simpler than the INRANGE 475 * check above, because we know that the dtms_scratch_ptr is valid in the 476 * range. Allocations of size zero are allowed. 477 */ 478#define DTRACE_INSCRATCH(mstate, alloc_sz) \ 479 ((mstate)->dtms_scratch_base + (mstate)->dtms_scratch_size - \ 480 (mstate)->dtms_scratch_ptr >= (alloc_sz)) 481 482#define DTRACE_LOADFUNC(bits) \ 483/*CSTYLED*/ \ 484uint##bits##_t \ 485dtrace_load##bits(uintptr_t addr) \ 486{ \ 487 size_t size = bits / NBBY; \ 488 /*CSTYLED*/ \ 489 uint##bits##_t rval; \ 490 int i; \ 491 volatile uint16_t *flags = (volatile uint16_t *) \ 492 &cpu_core[curcpu].cpuc_dtrace_flags; \ 493 \ 494 DTRACE_ALIGNCHECK(addr, size, flags); \ 495 \ 496 for (i = 0; i < dtrace_toxranges; i++) { \ 497 if (addr >= dtrace_toxrange[i].dtt_limit) \ 498 continue; \ 499 \ 500 if (addr + size <= dtrace_toxrange[i].dtt_base) \ 501 continue; \ 502 \ 503 /* \ 504 * This address falls within a toxic region; return 0. \ 505 */ \ 506 *flags |= CPU_DTRACE_BADADDR; \ 507 cpu_core[curcpu].cpuc_dtrace_illval = addr; \ 508 return (0); \ 509 } \ 510 \ 511 *flags |= CPU_DTRACE_NOFAULT; \ 512 /*CSTYLED*/ \ 513 rval = *((volatile uint##bits##_t *)addr); \ 514 *flags &= ~CPU_DTRACE_NOFAULT; \ 515 \ 516 return (!(*flags & CPU_DTRACE_FAULT) ? rval : 0); \ 517} 518 519#ifdef _LP64 520#define dtrace_loadptr dtrace_load64 521#else 522#define dtrace_loadptr dtrace_load32 523#endif 524 525#define DTRACE_DYNHASH_FREE 0 526#define DTRACE_DYNHASH_SINK 1 527#define DTRACE_DYNHASH_VALID 2 528 529#define DTRACE_MATCH_NEXT 0 530#define DTRACE_MATCH_DONE 1 531#define DTRACE_ANCHORED(probe) ((probe)->dtpr_func[0] != '\0') 532#define DTRACE_STATE_ALIGN 64 533 534#define DTRACE_FLAGS2FLT(flags) \ 535 (((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR : \ 536 ((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP : \ 537 ((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO : \ 538 ((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV : \ 539 ((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV : \ 540 ((flags) & CPU_DTRACE_TUPOFLOW) ? DTRACEFLT_TUPOFLOW : \ 541 ((flags) & CPU_DTRACE_BADALIGN) ? DTRACEFLT_BADALIGN : \ 542 ((flags) & CPU_DTRACE_NOSCRATCH) ? DTRACEFLT_NOSCRATCH : \ 543 ((flags) & CPU_DTRACE_BADSTACK) ? DTRACEFLT_BADSTACK : \ 544 DTRACEFLT_UNKNOWN) 545 546#define DTRACEACT_ISSTRING(act) \ 547 ((act)->dta_kind == DTRACEACT_DIFEXPR && \ 548 (act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) 549 550/* Function prototype definitions: */ 551static size_t dtrace_strlen(const char *, size_t); 552static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id); 553static void dtrace_enabling_provide(dtrace_provider_t *); 554static int dtrace_enabling_match(dtrace_enabling_t *, int *); 555static void dtrace_enabling_matchall(void); 556static void dtrace_enabling_reap(void); 557static dtrace_state_t *dtrace_anon_grab(void); 558static uint64_t dtrace_helper(int, dtrace_mstate_t *, 559 dtrace_state_t *, uint64_t, uint64_t); 560static dtrace_helpers_t *dtrace_helpers_create(proc_t *); 561static void dtrace_buffer_drop(dtrace_buffer_t *); 562static int dtrace_buffer_consumed(dtrace_buffer_t *, hrtime_t when); 563static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t, 564 dtrace_state_t *, dtrace_mstate_t *); 565static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t, 566 dtrace_optval_t); 567static int dtrace_ecb_create_enable(dtrace_probe_t *, void *); 568static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *); 569uint16_t dtrace_load16(uintptr_t); 570uint32_t dtrace_load32(uintptr_t); 571uint64_t dtrace_load64(uintptr_t); 572uint8_t dtrace_load8(uintptr_t); 573void dtrace_dynvar_clean(dtrace_dstate_t *); 574dtrace_dynvar_t *dtrace_dynvar(dtrace_dstate_t *, uint_t, dtrace_key_t *, 575 size_t, dtrace_dynvar_op_t, dtrace_mstate_t *, dtrace_vstate_t *); 576uintptr_t dtrace_dif_varstr(uintptr_t, dtrace_state_t *, dtrace_mstate_t *); 577 578/* 579 * DTrace Probe Context Functions 580 * 581 * These functions are called from probe context. Because probe context is 582 * any context in which C may be called, arbitrarily locks may be held, 583 * interrupts may be disabled, we may be in arbitrary dispatched state, etc. 584 * As a result, functions called from probe context may only call other DTrace 585 * support functions -- they may not interact at all with the system at large. 586 * (Note that the ASSERT macro is made probe-context safe by redefining it in 587 * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary 588 * loads are to be performed from probe context, they _must_ be in terms of 589 * the safe dtrace_load*() variants. 590 * 591 * Some functions in this block are not actually called from probe context; 592 * for these functions, there will be a comment above the function reading 593 * "Note: not called from probe context." 594 */ 595void 596dtrace_panic(const char *format, ...) 597{ 598 va_list alist; 599 600 va_start(alist, format); 601 dtrace_vpanic(format, alist); 602 va_end(alist); 603} 604 605int 606dtrace_assfail(const char *a, const char *f, int l) 607{ 608 dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l); 609 610 /* 611 * We just need something here that even the most clever compiler 612 * cannot optimize away. 613 */ 614 return (a[(uintptr_t)f]); 615} 616 617/* 618 * Atomically increment a specified error counter from probe context. 619 */ 620static void 621dtrace_error(uint32_t *counter) 622{ 623 /* 624 * Most counters stored to in probe context are per-CPU counters. 625 * However, there are some error conditions that are sufficiently 626 * arcane that they don't merit per-CPU storage. If these counters 627 * are incremented concurrently on different CPUs, scalability will be 628 * adversely affected -- but we don't expect them to be white-hot in a 629 * correctly constructed enabling... 630 */ 631 uint32_t oval, nval; 632 633 do { 634 oval = *counter; 635 636 if ((nval = oval + 1) == 0) { 637 /* 638 * If the counter would wrap, set it to 1 -- assuring 639 * that the counter is never zero when we have seen 640 * errors. (The counter must be 32-bits because we 641 * aren't guaranteed a 64-bit compare&swap operation.) 642 * To save this code both the infamy of being fingered 643 * by a priggish news story and the indignity of being 644 * the target of a neo-puritan witch trial, we're 645 * carefully avoiding any colorful description of the 646 * likelihood of this condition -- but suffice it to 647 * say that it is only slightly more likely than the 648 * overflow of predicate cache IDs, as discussed in 649 * dtrace_predicate_create(). 650 */ 651 nval = 1; 652 } 653 } while (dtrace_cas32(counter, oval, nval) != oval); 654} 655 656/* 657 * Use the DTRACE_LOADFUNC macro to define functions for each of loading a 658 * uint8_t, a uint16_t, a uint32_t and a uint64_t. 659 */ 660DTRACE_LOADFUNC(8) 661DTRACE_LOADFUNC(16) 662DTRACE_LOADFUNC(32) 663DTRACE_LOADFUNC(64) 664 665static int 666dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate) 667{ 668 if (dest < mstate->dtms_scratch_base) 669 return (0); 670 671 if (dest + size < dest) 672 return (0); 673 674 if (dest + size > mstate->dtms_scratch_ptr) 675 return (0); 676 677 return (1); 678} 679 680static int 681dtrace_canstore_statvar(uint64_t addr, size_t sz, 682 dtrace_statvar_t **svars, int nsvars) 683{ 684 int i; 685 686 for (i = 0; i < nsvars; i++) { 687 dtrace_statvar_t *svar = svars[i]; 688 689 if (svar == NULL || svar->dtsv_size == 0) 690 continue; 691 692 if (DTRACE_INRANGE(addr, sz, svar->dtsv_data, svar->dtsv_size)) 693 return (1); 694 } 695 696 return (0); 697} 698 699/* 700 * Check to see if the address is within a memory region to which a store may 701 * be issued. This includes the DTrace scratch areas, and any DTrace variable 702 * region. The caller of dtrace_canstore() is responsible for performing any 703 * alignment checks that are needed before stores are actually executed. 704 */ 705static int 706dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate, 707 dtrace_vstate_t *vstate) 708{ 709 /* 710 * First, check to see if the address is in scratch space... 711 */ 712 if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base, 713 mstate->dtms_scratch_size)) 714 return (1); 715 716 /* 717 * Now check to see if it's a dynamic variable. This check will pick 718 * up both thread-local variables and any global dynamically-allocated 719 * variables. 720 */ 721 if (DTRACE_INRANGE(addr, sz, (uintptr_t)vstate->dtvs_dynvars.dtds_base, 722 vstate->dtvs_dynvars.dtds_size)) { 723 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars; 724 uintptr_t base = (uintptr_t)dstate->dtds_base + 725 (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t)); 726 uintptr_t chunkoffs; 727 728 /* 729 * Before we assume that we can store here, we need to make 730 * sure that it isn't in our metadata -- storing to our 731 * dynamic variable metadata would corrupt our state. For 732 * the range to not include any dynamic variable metadata, 733 * it must: 734 * 735 * (1) Start above the hash table that is at the base of 736 * the dynamic variable space 737 * 738 * (2) Have a starting chunk offset that is beyond the 739 * dtrace_dynvar_t that is at the base of every chunk 740 * 741 * (3) Not span a chunk boundary 742 * 743 */ 744 if (addr < base) 745 return (0); 746 747 chunkoffs = (addr - base) % dstate->dtds_chunksize; 748 749 if (chunkoffs < sizeof (dtrace_dynvar_t)) 750 return (0); 751 752 if (chunkoffs + sz > dstate->dtds_chunksize) 753 return (0); 754 755 return (1); 756 } 757 758 /* 759 * Finally, check the static local and global variables. These checks 760 * take the longest, so we perform them last. 761 */ 762 if (dtrace_canstore_statvar(addr, sz, 763 vstate->dtvs_locals, vstate->dtvs_nlocals)) 764 return (1); 765 766 if (dtrace_canstore_statvar(addr, sz, 767 vstate->dtvs_globals, vstate->dtvs_nglobals)) 768 return (1); 769 770 return (0); 771} 772 773 774/* 775 * Convenience routine to check to see if the address is within a memory 776 * region in which a load may be issued given the user's privilege level; 777 * if not, it sets the appropriate error flags and loads 'addr' into the 778 * illegal value slot. 779 * 780 * DTrace subroutines (DIF_SUBR_*) should use this helper to implement 781 * appropriate memory access protection. 782 */ 783static int 784dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate, 785 dtrace_vstate_t *vstate) 786{ 787 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval; 788 789 /* 790 * If we hold the privilege to read from kernel memory, then 791 * everything is readable. 792 */ 793 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0) 794 return (1); 795 796 /* 797 * You can obviously read that which you can store. 798 */ 799 if (dtrace_canstore(addr, sz, mstate, vstate)) 800 return (1); 801 802 /* 803 * We're allowed to read from our own string table. 804 */ 805 if (DTRACE_INRANGE(addr, sz, (uintptr_t)mstate->dtms_difo->dtdo_strtab, 806 mstate->dtms_difo->dtdo_strlen)) 807 return (1); 808 809 DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV); 810 *illval = addr; 811 return (0); 812} 813 814/* 815 * Convenience routine to check to see if a given string is within a memory 816 * region in which a load may be issued given the user's privilege level; 817 * this exists so that we don't need to issue unnecessary dtrace_strlen() 818 * calls in the event that the user has all privileges. 819 */ 820static int 821dtrace_strcanload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate, 822 dtrace_vstate_t *vstate) 823{ 824 size_t strsz; 825 826 /* 827 * If we hold the privilege to read from kernel memory, then 828 * everything is readable. 829 */ 830 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0) 831 return (1); 832 833 strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr, sz); 834 if (dtrace_canload(addr, strsz, mstate, vstate)) 835 return (1); 836 837 return (0); 838} 839 840/* 841 * Convenience routine to check to see if a given variable is within a memory 842 * region in which a load may be issued given the user's privilege level. 843 */ 844static int 845dtrace_vcanload(void *src, dtrace_diftype_t *type, dtrace_mstate_t *mstate, 846 dtrace_vstate_t *vstate) 847{ 848 size_t sz; 849 ASSERT(type->dtdt_flags & DIF_TF_BYREF); 850 851 /* 852 * If we hold the privilege to read from kernel memory, then 853 * everything is readable. 854 */ 855 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0) 856 return (1); 857 858 if (type->dtdt_kind == DIF_TYPE_STRING) 859 sz = dtrace_strlen(src, 860 vstate->dtvs_state->dts_options[DTRACEOPT_STRSIZE]) + 1; 861 else 862 sz = type->dtdt_size; 863 864 return (dtrace_canload((uintptr_t)src, sz, mstate, vstate)); 865} 866 867/* 868 * Compare two strings using safe loads. 869 */ 870static int 871dtrace_strncmp(char *s1, char *s2, size_t limit) 872{ 873 uint8_t c1, c2; 874 volatile uint16_t *flags; 875 876 if (s1 == s2 || limit == 0) 877 return (0); 878 879 flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags; 880 881 do { 882 if (s1 == NULL) { 883 c1 = '\0'; 884 } else { 885 c1 = dtrace_load8((uintptr_t)s1++); 886 } 887 888 if (s2 == NULL) { 889 c2 = '\0'; 890 } else { 891 c2 = dtrace_load8((uintptr_t)s2++); 892 } 893 894 if (c1 != c2) 895 return (c1 - c2); 896 } while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT)); 897 898 return (0); 899} 900 901/* 902 * Compute strlen(s) for a string using safe memory accesses. The additional 903 * len parameter is used to specify a maximum length to ensure completion. 904 */ 905static size_t 906dtrace_strlen(const char *s, size_t lim) 907{ 908 uint_t len; 909 910 for (len = 0; len != lim; len++) { 911 if (dtrace_load8((uintptr_t)s++) == '\0') 912 break; 913 } 914 915 return (len); 916} 917 918/* 919 * Check if an address falls within a toxic region. 920 */ 921static int 922dtrace_istoxic(uintptr_t kaddr, size_t size) 923{ 924 uintptr_t taddr, tsize; 925 int i; 926 927 for (i = 0; i < dtrace_toxranges; i++) { 928 taddr = dtrace_toxrange[i].dtt_base; 929 tsize = dtrace_toxrange[i].dtt_limit - taddr; 930 931 if (kaddr - taddr < tsize) { 932 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR); 933 cpu_core[curcpu].cpuc_dtrace_illval = kaddr; 934 return (1); 935 } 936 937 if (taddr - kaddr < size) { 938 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR); 939 cpu_core[curcpu].cpuc_dtrace_illval = taddr; 940 return (1); 941 } 942 } 943 944 return (0); 945} 946 947/* 948 * Copy src to dst using safe memory accesses. The src is assumed to be unsafe 949 * memory specified by the DIF program. The dst is assumed to be safe memory 950 * that we can store to directly because it is managed by DTrace. As with 951 * standard bcopy, overlapping copies are handled properly. 952 */ 953static void 954dtrace_bcopy(const void *src, void *dst, size_t len) 955{ 956 if (len != 0) { 957 uint8_t *s1 = dst; 958 const uint8_t *s2 = src; 959 960 if (s1 <= s2) { 961 do { 962 *s1++ = dtrace_load8((uintptr_t)s2++); 963 } while (--len != 0); 964 } else { 965 s2 += len; 966 s1 += len; 967 968 do { 969 *--s1 = dtrace_load8((uintptr_t)--s2); 970 } while (--len != 0); 971 } 972 } 973} 974 975/* 976 * Copy src to dst using safe memory accesses, up to either the specified 977 * length, or the point that a nul byte is encountered. The src is assumed to 978 * be unsafe memory specified by the DIF program. The dst is assumed to be 979 * safe memory that we can store to directly because it is managed by DTrace. 980 * Unlike dtrace_bcopy(), overlapping regions are not handled. 981 */ 982static void 983dtrace_strcpy(const void *src, void *dst, size_t len) 984{ 985 if (len != 0) { 986 uint8_t *s1 = dst, c; 987 const uint8_t *s2 = src; 988 989 do { 990 *s1++ = c = dtrace_load8((uintptr_t)s2++); 991 } while (--len != 0 && c != '\0'); 992 } 993} 994 995/* 996 * Copy src to dst, deriving the size and type from the specified (BYREF) 997 * variable type. The src is assumed to be unsafe memory specified by the DIF 998 * program. The dst is assumed to be DTrace variable memory that is of the 999 * specified type; we assume that we can store to directly. 1000 */ 1001static void 1002dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type) 1003{ 1004 ASSERT(type->dtdt_flags & DIF_TF_BYREF); 1005 1006 if (type->dtdt_kind == DIF_TYPE_STRING) { 1007 dtrace_strcpy(src, dst, type->dtdt_size); 1008 } else { 1009 dtrace_bcopy(src, dst, type->dtdt_size); 1010 } 1011} 1012 1013/* 1014 * Compare s1 to s2 using safe memory accesses. The s1 data is assumed to be 1015 * unsafe memory specified by the DIF program. The s2 data is assumed to be 1016 * safe memory that we can access directly because it is managed by DTrace. 1017 */ 1018static int 1019dtrace_bcmp(const void *s1, const void *s2, size_t len) 1020{ 1021 volatile uint16_t *flags; 1022 1023 flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags; 1024 1025 if (s1 == s2) 1026 return (0); 1027 1028 if (s1 == NULL || s2 == NULL) 1029 return (1); 1030 1031 if (s1 != s2 && len != 0) { 1032 const uint8_t *ps1 = s1; 1033 const uint8_t *ps2 = s2; 1034 1035 do { 1036 if (dtrace_load8((uintptr_t)ps1++) != *ps2++) 1037 return (1); 1038 } while (--len != 0 && !(*flags & CPU_DTRACE_FAULT)); 1039 } 1040 return (0); 1041} 1042 1043/* 1044 * Zero the specified region using a simple byte-by-byte loop. Note that this 1045 * is for safe DTrace-managed memory only. 1046 */ 1047static void 1048dtrace_bzero(void *dst, size_t len) 1049{ 1050 uchar_t *cp; 1051 1052 for (cp = dst; len != 0; len--) 1053 *cp++ = 0; 1054} 1055 1056static void 1057dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum) 1058{ 1059 uint64_t result[2]; 1060 1061 result[0] = addend1[0] + addend2[0]; 1062 result[1] = addend1[1] + addend2[1] + 1063 (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0); 1064 1065 sum[0] = result[0]; 1066 sum[1] = result[1]; 1067} 1068 1069/* 1070 * Shift the 128-bit value in a by b. If b is positive, shift left. 1071 * If b is negative, shift right. 1072 */ 1073static void 1074dtrace_shift_128(uint64_t *a, int b) 1075{ 1076 uint64_t mask; 1077 1078 if (b == 0) 1079 return; 1080 1081 if (b < 0) { 1082 b = -b; 1083 if (b >= 64) { 1084 a[0] = a[1] >> (b - 64); 1085 a[1] = 0; 1086 } else { 1087 a[0] >>= b; 1088 mask = 1LL << (64 - b); 1089 mask -= 1; 1090 a[0] |= ((a[1] & mask) << (64 - b)); 1091 a[1] >>= b; 1092 } 1093 } else { 1094 if (b >= 64) { 1095 a[1] = a[0] << (b - 64); 1096 a[0] = 0; 1097 } else { 1098 a[1] <<= b; 1099 mask = a[0] >> (64 - b); 1100 a[1] |= mask; 1101 a[0] <<= b; 1102 } 1103 } 1104} 1105 1106/* 1107 * The basic idea is to break the 2 64-bit values into 4 32-bit values, 1108 * use native multiplication on those, and then re-combine into the 1109 * resulting 128-bit value. 1110 * 1111 * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) = 1112 * hi1 * hi2 << 64 + 1113 * hi1 * lo2 << 32 + 1114 * hi2 * lo1 << 32 + 1115 * lo1 * lo2 1116 */ 1117static void 1118dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product) 1119{ 1120 uint64_t hi1, hi2, lo1, lo2; 1121 uint64_t tmp[2]; 1122 1123 hi1 = factor1 >> 32; 1124 hi2 = factor2 >> 32; 1125 1126 lo1 = factor1 & DT_MASK_LO; 1127 lo2 = factor2 & DT_MASK_LO; 1128 1129 product[0] = lo1 * lo2; 1130 product[1] = hi1 * hi2; 1131 1132 tmp[0] = hi1 * lo2; 1133 tmp[1] = 0; 1134 dtrace_shift_128(tmp, 32); 1135 dtrace_add_128(product, tmp, product); 1136 1137 tmp[0] = hi2 * lo1; 1138 tmp[1] = 0; 1139 dtrace_shift_128(tmp, 32); 1140 dtrace_add_128(product, tmp, product); 1141} 1142 1143/* 1144 * This privilege check should be used by actions and subroutines to 1145 * verify that the user credentials of the process that enabled the 1146 * invoking ECB match the target credentials 1147 */ 1148static int 1149dtrace_priv_proc_common_user(dtrace_state_t *state) 1150{ 1151 cred_t *cr, *s_cr = state->dts_cred.dcr_cred; 1152 1153 /* 1154 * We should always have a non-NULL state cred here, since if cred 1155 * is null (anonymous tracing), we fast-path bypass this routine. 1156 */ 1157 ASSERT(s_cr != NULL); 1158 1159 if ((cr = CRED()) != NULL && 1160 s_cr->cr_uid == cr->cr_uid && 1161 s_cr->cr_uid == cr->cr_ruid && 1162 s_cr->cr_uid == cr->cr_suid && 1163 s_cr->cr_gid == cr->cr_gid && 1164 s_cr->cr_gid == cr->cr_rgid && 1165 s_cr->cr_gid == cr->cr_sgid) 1166 return (1); 1167 1168 return (0); 1169} 1170 1171/* 1172 * This privilege check should be used by actions and subroutines to 1173 * verify that the zone of the process that enabled the invoking ECB 1174 * matches the target credentials 1175 */ 1176static int 1177dtrace_priv_proc_common_zone(dtrace_state_t *state) 1178{ 1179#if defined(sun) 1180 cred_t *cr, *s_cr = state->dts_cred.dcr_cred; 1181 1182 /* 1183 * We should always have a non-NULL state cred here, since if cred 1184 * is null (anonymous tracing), we fast-path bypass this routine. 1185 */ 1186 ASSERT(s_cr != NULL); 1187 1188 if ((cr = CRED()) != NULL && 1189 s_cr->cr_zone == cr->cr_zone) 1190 return (1); 1191 1192 return (0); 1193#else 1194 return (1); 1195#endif 1196} 1197 1198/* 1199 * This privilege check should be used by actions and subroutines to 1200 * verify that the process has not setuid or changed credentials. 1201 */ 1202static int 1203dtrace_priv_proc_common_nocd(void) 1204{ 1205 proc_t *proc; 1206 1207 if ((proc = ttoproc(curthread)) != NULL && 1208 !(proc->p_flag & SNOCD)) 1209 return (1); 1210 1211 return (0); 1212} 1213 1214static int 1215dtrace_priv_proc_destructive(dtrace_state_t *state) 1216{ 1217 int action = state->dts_cred.dcr_action; 1218 1219 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) && 1220 dtrace_priv_proc_common_zone(state) == 0) 1221 goto bad; 1222 1223 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) && 1224 dtrace_priv_proc_common_user(state) == 0) 1225 goto bad; 1226 1227 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) && 1228 dtrace_priv_proc_common_nocd() == 0) 1229 goto bad; 1230 1231 return (1); 1232 1233bad: 1234 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV; 1235 1236 return (0); 1237} 1238 1239static int 1240dtrace_priv_proc_control(dtrace_state_t *state) 1241{ 1242 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL) 1243 return (1); 1244 1245 if (dtrace_priv_proc_common_zone(state) && 1246 dtrace_priv_proc_common_user(state) && 1247 dtrace_priv_proc_common_nocd()) 1248 return (1); 1249 1250 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV; 1251 1252 return (0); 1253} 1254 1255static int 1256dtrace_priv_proc(dtrace_state_t *state) 1257{ 1258 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC) 1259 return (1); 1260 1261 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV; 1262 1263 return (0); 1264} 1265 1266static int 1267dtrace_priv_kernel(dtrace_state_t *state) 1268{ 1269 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL) 1270 return (1); 1271 1272 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV; 1273 1274 return (0); 1275} 1276 1277static int 1278dtrace_priv_kernel_destructive(dtrace_state_t *state) 1279{ 1280 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE) 1281 return (1); 1282 1283 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV; 1284 1285 return (0); 1286} 1287 1288/* 1289 * Note: not called from probe context. This function is called 1290 * asynchronously (and at a regular interval) from outside of probe context to 1291 * clean the dirty dynamic variable lists on all CPUs. Dynamic variable 1292 * cleaning is explained in detail in <sys/dtrace_impl.h>. 1293 */ 1294void 1295dtrace_dynvar_clean(dtrace_dstate_t *dstate) 1296{ 1297 dtrace_dynvar_t *dirty; 1298 dtrace_dstate_percpu_t *dcpu; 1299 int i, work = 0; 1300 1301 for (i = 0; i < NCPU; i++) { 1302 dcpu = &dstate->dtds_percpu[i]; 1303 1304 ASSERT(dcpu->dtdsc_rinsing == NULL); 1305 1306 /* 1307 * If the dirty list is NULL, there is no dirty work to do. 1308 */ 1309 if (dcpu->dtdsc_dirty == NULL) 1310 continue; 1311 1312 /* 1313 * If the clean list is non-NULL, then we're not going to do 1314 * any work for this CPU -- it means that there has not been 1315 * a dtrace_dynvar() allocation on this CPU (or from this CPU) 1316 * since the last time we cleaned house. 1317 */ 1318 if (dcpu->dtdsc_clean != NULL) 1319 continue; 1320 1321 work = 1; 1322 1323 /* 1324 * Atomically move the dirty list aside. 1325 */ 1326 do { 1327 dirty = dcpu->dtdsc_dirty; 1328 1329 /* 1330 * Before we zap the dirty list, set the rinsing list. 1331 * (This allows for a potential assertion in 1332 * dtrace_dynvar(): if a free dynamic variable appears 1333 * on a hash chain, either the dirty list or the 1334 * rinsing list for some CPU must be non-NULL.) 1335 */ 1336 dcpu->dtdsc_rinsing = dirty; 1337 dtrace_membar_producer(); 1338 } while (dtrace_casptr(&dcpu->dtdsc_dirty, 1339 dirty, NULL) != dirty); 1340 } 1341 1342 if (!work) { 1343 /* 1344 * We have no work to do; we can simply return. 1345 */ 1346 return; 1347 } 1348 1349 dtrace_sync(); 1350 1351 for (i = 0; i < NCPU; i++) { 1352 dcpu = &dstate->dtds_percpu[i]; 1353 1354 if (dcpu->dtdsc_rinsing == NULL) 1355 continue; 1356 1357 /* 1358 * We are now guaranteed that no hash chain contains a pointer 1359 * into this dirty list; we can make it clean. 1360 */ 1361 ASSERT(dcpu->dtdsc_clean == NULL); 1362 dcpu->dtdsc_clean = dcpu->dtdsc_rinsing; 1363 dcpu->dtdsc_rinsing = NULL; 1364 } 1365 1366 /* 1367 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make 1368 * sure that all CPUs have seen all of the dtdsc_clean pointers. 1369 * This prevents a race whereby a CPU incorrectly decides that 1370 * the state should be something other than DTRACE_DSTATE_CLEAN 1371 * after dtrace_dynvar_clean() has completed. 1372 */ 1373 dtrace_sync(); 1374 1375 dstate->dtds_state = DTRACE_DSTATE_CLEAN; 1376} 1377 1378/* 1379 * Depending on the value of the op parameter, this function looks-up, 1380 * allocates or deallocates an arbitrarily-keyed dynamic variable. If an 1381 * allocation is requested, this function will return a pointer to a 1382 * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no 1383 * variable can be allocated. If NULL is returned, the appropriate counter 1384 * will be incremented. 1385 */ 1386dtrace_dynvar_t * 1387dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys, 1388 dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op, 1389 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate) 1390{ 1391 uint64_t hashval = DTRACE_DYNHASH_VALID; 1392 dtrace_dynhash_t *hash = dstate->dtds_hash; 1393 dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL; 1394 processorid_t me = curcpu, cpu = me; 1395 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me]; 1396 size_t bucket, ksize; 1397 size_t chunksize = dstate->dtds_chunksize; 1398 uintptr_t kdata, lock, nstate; 1399 uint_t i; 1400 1401 ASSERT(nkeys != 0); 1402 1403 /* 1404 * Hash the key. As with aggregations, we use Jenkins' "One-at-a-time" 1405 * algorithm. For the by-value portions, we perform the algorithm in 1406 * 16-bit chunks (as opposed to 8-bit chunks). This speeds things up a 1407 * bit, and seems to have only a minute effect on distribution. For 1408 * the by-reference data, we perform "One-at-a-time" iterating (safely) 1409 * over each referenced byte. It's painful to do this, but it's much 1410 * better than pathological hash distribution. The efficacy of the 1411 * hashing algorithm (and a comparison with other algorithms) may be 1412 * found by running the ::dtrace_dynstat MDB dcmd. 1413 */ 1414 for (i = 0; i < nkeys; i++) { 1415 if (key[i].dttk_size == 0) { 1416 uint64_t val = key[i].dttk_value; 1417 1418 hashval += (val >> 48) & 0xffff; 1419 hashval += (hashval << 10); 1420 hashval ^= (hashval >> 6); 1421 1422 hashval += (val >> 32) & 0xffff; 1423 hashval += (hashval << 10); 1424 hashval ^= (hashval >> 6); 1425 1426 hashval += (val >> 16) & 0xffff; 1427 hashval += (hashval << 10); 1428 hashval ^= (hashval >> 6); 1429 1430 hashval += val & 0xffff; 1431 hashval += (hashval << 10); 1432 hashval ^= (hashval >> 6); 1433 } else { 1434 /* 1435 * This is incredibly painful, but it beats the hell 1436 * out of the alternative. 1437 */ 1438 uint64_t j, size = key[i].dttk_size; 1439 uintptr_t base = (uintptr_t)key[i].dttk_value; 1440 1441 if (!dtrace_canload(base, size, mstate, vstate)) 1442 break; 1443 1444 for (j = 0; j < size; j++) { 1445 hashval += dtrace_load8(base + j); 1446 hashval += (hashval << 10); 1447 hashval ^= (hashval >> 6); 1448 } 1449 } 1450 } 1451 1452 if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT)) 1453 return (NULL); 1454 1455 hashval += (hashval << 3); 1456 hashval ^= (hashval >> 11); 1457 hashval += (hashval << 15); 1458 1459 /* 1460 * There is a remote chance (ideally, 1 in 2^31) that our hashval 1461 * comes out to be one of our two sentinel hash values. If this 1462 * actually happens, we set the hashval to be a value known to be a 1463 * non-sentinel value. 1464 */ 1465 if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK) 1466 hashval = DTRACE_DYNHASH_VALID; 1467 1468 /* 1469 * Yes, it's painful to do a divide here. If the cycle count becomes 1470 * important here, tricks can be pulled to reduce it. (However, it's 1471 * critical that hash collisions be kept to an absolute minimum; 1472 * they're much more painful than a divide.) It's better to have a 1473 * solution that generates few collisions and still keeps things 1474 * relatively simple. 1475 */ 1476 bucket = hashval % dstate->dtds_hashsize; 1477 1478 if (op == DTRACE_DYNVAR_DEALLOC) { 1479 volatile uintptr_t *lockp = &hash[bucket].dtdh_lock; 1480 1481 for (;;) { 1482 while ((lock = *lockp) & 1) 1483 continue; 1484 1485 if (dtrace_casptr((volatile void *)lockp, 1486 (volatile void *)lock, (volatile void *)(lock + 1)) == (void *)lock) 1487 break; 1488 } 1489 1490 dtrace_membar_producer(); 1491 } 1492 1493top: 1494 prev = NULL; 1495 lock = hash[bucket].dtdh_lock; 1496 1497 dtrace_membar_consumer(); 1498 1499 start = hash[bucket].dtdh_chain; 1500 ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK || 1501 start->dtdv_hashval != DTRACE_DYNHASH_FREE || 1502 op != DTRACE_DYNVAR_DEALLOC)); 1503 1504 for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) { 1505 dtrace_tuple_t *dtuple = &dvar->dtdv_tuple; 1506 dtrace_key_t *dkey = &dtuple->dtt_key[0]; 1507 1508 if (dvar->dtdv_hashval != hashval) { 1509 if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) { 1510 /* 1511 * We've reached the sink, and therefore the 1512 * end of the hash chain; we can kick out of 1513 * the loop knowing that we have seen a valid 1514 * snapshot of state. 1515 */ 1516 ASSERT(dvar->dtdv_next == NULL); 1517 ASSERT(dvar == &dtrace_dynhash_sink); 1518 break; 1519 } 1520 1521 if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) { 1522 /* 1523 * We've gone off the rails: somewhere along 1524 * the line, one of the members of this hash 1525 * chain was deleted. Note that we could also 1526 * detect this by simply letting this loop run 1527 * to completion, as we would eventually hit 1528 * the end of the dirty list. However, we 1529 * want to avoid running the length of the 1530 * dirty list unnecessarily (it might be quite 1531 * long), so we catch this as early as 1532 * possible by detecting the hash marker. In 1533 * this case, we simply set dvar to NULL and 1534 * break; the conditional after the loop will 1535 * send us back to top. 1536 */ 1537 dvar = NULL; 1538 break; 1539 } 1540 1541 goto next; 1542 } 1543 1544 if (dtuple->dtt_nkeys != nkeys) 1545 goto next; 1546 1547 for (i = 0; i < nkeys; i++, dkey++) { 1548 if (dkey->dttk_size != key[i].dttk_size) 1549 goto next; /* size or type mismatch */ 1550 1551 if (dkey->dttk_size != 0) { 1552 if (dtrace_bcmp( 1553 (void *)(uintptr_t)key[i].dttk_value, 1554 (void *)(uintptr_t)dkey->dttk_value, 1555 dkey->dttk_size)) 1556 goto next; 1557 } else { 1558 if (dkey->dttk_value != key[i].dttk_value) 1559 goto next; 1560 } 1561 } 1562 1563 if (op != DTRACE_DYNVAR_DEALLOC) 1564 return (dvar); 1565 1566 ASSERT(dvar->dtdv_next == NULL || 1567 dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE); 1568 1569 if (prev != NULL) { 1570 ASSERT(hash[bucket].dtdh_chain != dvar); 1571 ASSERT(start != dvar); 1572 ASSERT(prev->dtdv_next == dvar); 1573 prev->dtdv_next = dvar->dtdv_next; 1574 } else { 1575 if (dtrace_casptr(&hash[bucket].dtdh_chain, 1576 start, dvar->dtdv_next) != start) { 1577 /* 1578 * We have failed to atomically swing the 1579 * hash table head pointer, presumably because 1580 * of a conflicting allocation on another CPU. 1581 * We need to reread the hash chain and try 1582 * again. 1583 */ 1584 goto top; 1585 } 1586 } 1587 1588 dtrace_membar_producer(); 1589 1590 /* 1591 * Now set the hash value to indicate that it's free. 1592 */ 1593 ASSERT(hash[bucket].dtdh_chain != dvar); 1594 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE; 1595 1596 dtrace_membar_producer(); 1597 1598 /* 1599 * Set the next pointer to point at the dirty list, and 1600 * atomically swing the dirty pointer to the newly freed dvar. 1601 */ 1602 do { 1603 next = dcpu->dtdsc_dirty; 1604 dvar->dtdv_next = next; 1605 } while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next); 1606 1607 /* 1608 * Finally, unlock this hash bucket. 1609 */ 1610 ASSERT(hash[bucket].dtdh_lock == lock); 1611 ASSERT(lock & 1); 1612 hash[bucket].dtdh_lock++; 1613 1614 return (NULL); 1615next: 1616 prev = dvar; 1617 continue; 1618 } 1619 1620 if (dvar == NULL) { 1621 /* 1622 * If dvar is NULL, it is because we went off the rails: 1623 * one of the elements that we traversed in the hash chain 1624 * was deleted while we were traversing it. In this case, 1625 * we assert that we aren't doing a dealloc (deallocs lock 1626 * the hash bucket to prevent themselves from racing with 1627 * one another), and retry the hash chain traversal. 1628 */ 1629 ASSERT(op != DTRACE_DYNVAR_DEALLOC); 1630 goto top; 1631 } 1632 1633 if (op != DTRACE_DYNVAR_ALLOC) { 1634 /* 1635 * If we are not to allocate a new variable, we want to 1636 * return NULL now. Before we return, check that the value 1637 * of the lock word hasn't changed. If it has, we may have 1638 * seen an inconsistent snapshot. 1639 */ 1640 if (op == DTRACE_DYNVAR_NOALLOC) { 1641 if (hash[bucket].dtdh_lock != lock) 1642 goto top; 1643 } else { 1644 ASSERT(op == DTRACE_DYNVAR_DEALLOC); 1645 ASSERT(hash[bucket].dtdh_lock == lock); 1646 ASSERT(lock & 1); 1647 hash[bucket].dtdh_lock++; 1648 } 1649 1650 return (NULL); 1651 } 1652 1653 /* 1654 * We need to allocate a new dynamic variable. The size we need is the 1655 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the 1656 * size of any auxiliary key data (rounded up to 8-byte alignment) plus 1657 * the size of any referred-to data (dsize). We then round the final 1658 * size up to the chunksize for allocation. 1659 */ 1660 for (ksize = 0, i = 0; i < nkeys; i++) 1661 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t)); 1662 1663 /* 1664 * This should be pretty much impossible, but could happen if, say, 1665 * strange DIF specified the tuple. Ideally, this should be an 1666 * assertion and not an error condition -- but that requires that the 1667 * chunksize calculation in dtrace_difo_chunksize() be absolutely 1668 * bullet-proof. (That is, it must not be able to be fooled by 1669 * malicious DIF.) Given the lack of backwards branches in DIF, 1670 * solving this would presumably not amount to solving the Halting 1671 * Problem -- but it still seems awfully hard. 1672 */ 1673 if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) + 1674 ksize + dsize > chunksize) { 1675 dcpu->dtdsc_drops++; 1676 return (NULL); 1677 } 1678 1679 nstate = DTRACE_DSTATE_EMPTY; 1680 1681 do { 1682retry: 1683 free = dcpu->dtdsc_free; 1684 1685 if (free == NULL) { 1686 dtrace_dynvar_t *clean = dcpu->dtdsc_clean; 1687 void *rval; 1688 1689 if (clean == NULL) { 1690 /* 1691 * We're out of dynamic variable space on 1692 * this CPU. Unless we have tried all CPUs, 1693 * we'll try to allocate from a different 1694 * CPU. 1695 */ 1696 switch (dstate->dtds_state) { 1697 case DTRACE_DSTATE_CLEAN: { 1698 void *sp = &dstate->dtds_state; 1699 1700 if (++cpu >= NCPU) 1701 cpu = 0; 1702 1703 if (dcpu->dtdsc_dirty != NULL && 1704 nstate == DTRACE_DSTATE_EMPTY) 1705 nstate = DTRACE_DSTATE_DIRTY; 1706 1707 if (dcpu->dtdsc_rinsing != NULL) 1708 nstate = DTRACE_DSTATE_RINSING; 1709 1710 dcpu = &dstate->dtds_percpu[cpu]; 1711 1712 if (cpu != me) 1713 goto retry; 1714 1715 (void) dtrace_cas32(sp, 1716 DTRACE_DSTATE_CLEAN, nstate); 1717 1718 /* 1719 * To increment the correct bean 1720 * counter, take another lap. 1721 */ 1722 goto retry; 1723 } 1724 1725 case DTRACE_DSTATE_DIRTY: 1726 dcpu->dtdsc_dirty_drops++; 1727 break; 1728 1729 case DTRACE_DSTATE_RINSING: 1730 dcpu->dtdsc_rinsing_drops++; 1731 break; 1732 1733 case DTRACE_DSTATE_EMPTY: 1734 dcpu->dtdsc_drops++; 1735 break; 1736 } 1737 1738 DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP); 1739 return (NULL); 1740 } 1741 1742 /* 1743 * The clean list appears to be non-empty. We want to 1744 * move the clean list to the free list; we start by 1745 * moving the clean pointer aside. 1746 */ 1747 if (dtrace_casptr(&dcpu->dtdsc_clean, 1748 clean, NULL) != clean) { 1749 /* 1750 * We are in one of two situations: 1751 * 1752 * (a) The clean list was switched to the 1753 * free list by another CPU. 1754 * 1755 * (b) The clean list was added to by the 1756 * cleansing cyclic. 1757 * 1758 * In either of these situations, we can 1759 * just reattempt the free list allocation. 1760 */ 1761 goto retry; 1762 } 1763 1764 ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE); 1765 1766 /* 1767 * Now we'll move the clean list to the free list. 1768 * It's impossible for this to fail: the only way 1769 * the free list can be updated is through this 1770 * code path, and only one CPU can own the clean list. 1771 * Thus, it would only be possible for this to fail if 1772 * this code were racing with dtrace_dynvar_clean(). 1773 * (That is, if dtrace_dynvar_clean() updated the clean 1774 * list, and we ended up racing to update the free 1775 * list.) This race is prevented by the dtrace_sync() 1776 * in dtrace_dynvar_clean() -- which flushes the 1777 * owners of the clean lists out before resetting 1778 * the clean lists. 1779 */ 1780 rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean); 1781 ASSERT(rval == NULL); 1782 goto retry; 1783 } 1784 1785 dvar = free; 1786 new_free = dvar->dtdv_next; 1787 } while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free); 1788 1789 /* 1790 * We have now allocated a new chunk. We copy the tuple keys into the 1791 * tuple array and copy any referenced key data into the data space 1792 * following the tuple array. As we do this, we relocate dttk_value 1793 * in the final tuple to point to the key data address in the chunk. 1794 */ 1795 kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys]; 1796 dvar->dtdv_data = (void *)(kdata + ksize); 1797 dvar->dtdv_tuple.dtt_nkeys = nkeys; 1798 1799 for (i = 0; i < nkeys; i++) { 1800 dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i]; 1801 size_t kesize = key[i].dttk_size; 1802 1803 if (kesize != 0) { 1804 dtrace_bcopy( 1805 (const void *)(uintptr_t)key[i].dttk_value, 1806 (void *)kdata, kesize); 1807 dkey->dttk_value = kdata; 1808 kdata += P2ROUNDUP(kesize, sizeof (uint64_t)); 1809 } else { 1810 dkey->dttk_value = key[i].dttk_value; 1811 } 1812 1813 dkey->dttk_size = kesize; 1814 } 1815 1816 ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE); 1817 dvar->dtdv_hashval = hashval; 1818 dvar->dtdv_next = start; 1819 1820 if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start) 1821 return (dvar); 1822 1823 /* 1824 * The cas has failed. Either another CPU is adding an element to 1825 * this hash chain, or another CPU is deleting an element from this 1826 * hash chain. The simplest way to deal with both of these cases 1827 * (though not necessarily the most efficient) is to free our 1828 * allocated block and tail-call ourselves. Note that the free is 1829 * to the dirty list and _not_ to the free list. This is to prevent 1830 * races with allocators, above. 1831 */ 1832 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE; 1833 1834 dtrace_membar_producer(); 1835 1836 do { 1837 free = dcpu->dtdsc_dirty; 1838 dvar->dtdv_next = free; 1839 } while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free); 1840 1841 return (dtrace_dynvar(dstate, nkeys, key, dsize, op, mstate, vstate)); 1842} 1843 1844/*ARGSUSED*/ 1845static void 1846dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg) 1847{ 1848 if ((int64_t)nval < (int64_t)*oval) 1849 *oval = nval; 1850} 1851 1852/*ARGSUSED*/ 1853static void 1854dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg) 1855{ 1856 if ((int64_t)nval > (int64_t)*oval) 1857 *oval = nval; 1858} 1859 1860static void 1861dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr) 1862{ 1863 int i, zero = DTRACE_QUANTIZE_ZEROBUCKET; 1864 int64_t val = (int64_t)nval; 1865 1866 if (val < 0) { 1867 for (i = 0; i < zero; i++) { 1868 if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) { 1869 quanta[i] += incr; 1870 return; 1871 } 1872 } 1873 } else { 1874 for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) { 1875 if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) { 1876 quanta[i - 1] += incr; 1877 return; 1878 } 1879 } 1880 1881 quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr; 1882 return; 1883 } 1884 1885 ASSERT(0); 1886} 1887 1888static void 1889dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr) 1890{ 1891 uint64_t arg = *lquanta++; 1892 int32_t base = DTRACE_LQUANTIZE_BASE(arg); 1893 uint16_t step = DTRACE_LQUANTIZE_STEP(arg); 1894 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg); 1895 int32_t val = (int32_t)nval, level; 1896 1897 ASSERT(step != 0); 1898 ASSERT(levels != 0); 1899 1900 if (val < base) { 1901 /* 1902 * This is an underflow. 1903 */ 1904 lquanta[0] += incr; 1905 return; 1906 } 1907 1908 level = (val - base) / step; 1909 1910 if (level < levels) { 1911 lquanta[level + 1] += incr; 1912 return; 1913 } 1914 1915 /* 1916 * This is an overflow. 1917 */ 1918 lquanta[levels + 1] += incr; 1919} 1920 1921static int 1922dtrace_aggregate_llquantize_bucket(uint16_t factor, uint16_t low, 1923 uint16_t high, uint16_t nsteps, int64_t value) 1924{ 1925 int64_t this = 1, last, next; 1926 int base = 1, order; 1927 1928 ASSERT(factor <= nsteps); 1929 ASSERT(nsteps % factor == 0); 1930 1931 for (order = 0; order < low; order++) 1932 this *= factor; 1933 1934 /* 1935 * If our value is less than our factor taken to the power of the 1936 * low order of magnitude, it goes into the zeroth bucket. 1937 */ 1938 if (value < (last = this)) 1939 return (0); 1940 1941 for (this *= factor; order <= high; order++) { 1942 int nbuckets = this > nsteps ? nsteps : this; 1943 1944 if ((next = this * factor) < this) { 1945 /* 1946 * We should not generally get log/linear quantizations 1947 * with a high magnitude that allows 64-bits to 1948 * overflow, but we nonetheless protect against this 1949 * by explicitly checking for overflow, and clamping 1950 * our value accordingly. 1951 */ 1952 value = this - 1; 1953 } 1954 1955 if (value < this) { 1956 /* 1957 * If our value lies within this order of magnitude, 1958 * determine its position by taking the offset within 1959 * the order of magnitude, dividing by the bucket 1960 * width, and adding to our (accumulated) base. 1961 */ 1962 return (base + (value - last) / (this / nbuckets)); 1963 } 1964 1965 base += nbuckets - (nbuckets / factor); 1966 last = this; 1967 this = next; 1968 } 1969 1970 /* 1971 * Our value is greater than or equal to our factor taken to the 1972 * power of one plus the high magnitude -- return the top bucket. 1973 */ 1974 return (base); 1975} 1976 1977static void 1978dtrace_aggregate_llquantize(uint64_t *llquanta, uint64_t nval, uint64_t incr) 1979{ 1980 uint64_t arg = *llquanta++; 1981 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(arg); 1982 uint16_t low = DTRACE_LLQUANTIZE_LOW(arg); 1983 uint16_t high = DTRACE_LLQUANTIZE_HIGH(arg); 1984 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(arg); 1985 1986 llquanta[dtrace_aggregate_llquantize_bucket(factor, 1987 low, high, nsteps, nval)] += incr; 1988} 1989 1990/*ARGSUSED*/ 1991static void 1992dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg) 1993{ 1994 data[0]++; 1995 data[1] += nval; 1996} 1997 1998/*ARGSUSED*/ 1999static void 2000dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg) 2001{ 2002 int64_t snval = (int64_t)nval; 2003 uint64_t tmp[2]; 2004 2005 data[0]++; 2006 data[1] += nval; 2007 2008 /* 2009 * What we want to say here is: 2010 * 2011 * data[2] += nval * nval; 2012 * 2013 * But given that nval is 64-bit, we could easily overflow, so 2014 * we do this as 128-bit arithmetic. 2015 */ 2016 if (snval < 0) 2017 snval = -snval; 2018 2019 dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp); 2020 dtrace_add_128(data + 2, tmp, data + 2); 2021} 2022 2023/*ARGSUSED*/ 2024static void 2025dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg) 2026{ 2027 *oval = *oval + 1; 2028} 2029 2030/*ARGSUSED*/ 2031static void 2032dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg) 2033{ 2034 *oval += nval; 2035} 2036 2037/* 2038 * Aggregate given the tuple in the principal data buffer, and the aggregating 2039 * action denoted by the specified dtrace_aggregation_t. The aggregation 2040 * buffer is specified as the buf parameter. This routine does not return 2041 * failure; if there is no space in the aggregation buffer, the data will be 2042 * dropped, and a corresponding counter incremented. 2043 */ 2044static void 2045dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf, 2046 intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg) 2047{ 2048 dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec; 2049 uint32_t i, ndx, size, fsize; 2050 uint32_t align = sizeof (uint64_t) - 1; 2051 dtrace_aggbuffer_t *agb; 2052 dtrace_aggkey_t *key; 2053 uint32_t hashval = 0, limit, isstr; 2054 caddr_t tomax, data, kdata; 2055 dtrace_actkind_t action; 2056 dtrace_action_t *act; 2057 uintptr_t offs; 2058 2059 if (buf == NULL) 2060 return; 2061 2062 if (!agg->dtag_hasarg) { 2063 /* 2064 * Currently, only quantize() and lquantize() take additional 2065 * arguments, and they have the same semantics: an increment 2066 * value that defaults to 1 when not present. If additional 2067 * aggregating actions take arguments, the setting of the 2068 * default argument value will presumably have to become more 2069 * sophisticated... 2070 */ 2071 arg = 1; 2072 } 2073 2074 action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION; 2075 size = rec->dtrd_offset - agg->dtag_base; 2076 fsize = size + rec->dtrd_size; 2077 2078 ASSERT(dbuf->dtb_tomax != NULL); 2079 data = dbuf->dtb_tomax + offset + agg->dtag_base; 2080 2081 if ((tomax = buf->dtb_tomax) == NULL) { 2082 dtrace_buffer_drop(buf); 2083 return; 2084 } 2085 2086 /* 2087 * The metastructure is always at the bottom of the buffer. 2088 */ 2089 agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size - 2090 sizeof (dtrace_aggbuffer_t)); 2091 2092 if (buf->dtb_offset == 0) { 2093 /* 2094 * We just kludge up approximately 1/8th of the size to be 2095 * buckets. If this guess ends up being routinely 2096 * off-the-mark, we may need to dynamically readjust this 2097 * based on past performance. 2098 */ 2099 uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t); 2100 2101 if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) < 2102 (uintptr_t)tomax || hashsize == 0) { 2103 /* 2104 * We've been given a ludicrously small buffer; 2105 * increment our drop count and leave. 2106 */ 2107 dtrace_buffer_drop(buf); 2108 return; 2109 } 2110 2111 /* 2112 * And now, a pathetic attempt to try to get a an odd (or 2113 * perchance, a prime) hash size for better hash distribution. 2114 */ 2115 if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3)) 2116 hashsize -= DTRACE_AGGHASHSIZE_SLEW; 2117 2118 agb->dtagb_hashsize = hashsize; 2119 agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb - 2120 agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *)); 2121 agb->dtagb_free = (uintptr_t)agb->dtagb_hash; 2122 2123 for (i = 0; i < agb->dtagb_hashsize; i++) 2124 agb->dtagb_hash[i] = NULL; 2125 } 2126 2127 ASSERT(agg->dtag_first != NULL); 2128 ASSERT(agg->dtag_first->dta_intuple); 2129 2130 /* 2131 * Calculate the hash value based on the key. Note that we _don't_ 2132 * include the aggid in the hashing (but we will store it as part of 2133 * the key). The hashing algorithm is Bob Jenkins' "One-at-a-time" 2134 * algorithm: a simple, quick algorithm that has no known funnels, and 2135 * gets good distribution in practice. The efficacy of the hashing 2136 * algorithm (and a comparison with other algorithms) may be found by 2137 * running the ::dtrace_aggstat MDB dcmd. 2138 */ 2139 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) { 2140 i = act->dta_rec.dtrd_offset - agg->dtag_base; 2141 limit = i + act->dta_rec.dtrd_size; 2142 ASSERT(limit <= size); 2143 isstr = DTRACEACT_ISSTRING(act); 2144 2145 for (; i < limit; i++) { 2146 hashval += data[i]; 2147 hashval += (hashval << 10); 2148 hashval ^= (hashval >> 6); 2149 2150 if (isstr && data[i] == '\0') 2151 break; 2152 } 2153 } 2154 2155 hashval += (hashval << 3); 2156 hashval ^= (hashval >> 11); 2157 hashval += (hashval << 15); 2158 2159 /* 2160 * Yes, the divide here is expensive -- but it's generally the least 2161 * of the performance issues given the amount of data that we iterate 2162 * over to compute hash values, compare data, etc. 2163 */ 2164 ndx = hashval % agb->dtagb_hashsize; 2165 2166 for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) { 2167 ASSERT((caddr_t)key >= tomax); 2168 ASSERT((caddr_t)key < tomax + buf->dtb_size); 2169 2170 if (hashval != key->dtak_hashval || key->dtak_size != size) 2171 continue; 2172 2173 kdata = key->dtak_data; 2174 ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size); 2175 2176 for (act = agg->dtag_first; act->dta_intuple; 2177 act = act->dta_next) { 2178 i = act->dta_rec.dtrd_offset - agg->dtag_base; 2179 limit = i + act->dta_rec.dtrd_size; 2180 ASSERT(limit <= size); 2181 isstr = DTRACEACT_ISSTRING(act); 2182 2183 for (; i < limit; i++) { 2184 if (kdata[i] != data[i]) 2185 goto next; 2186 2187 if (isstr && data[i] == '\0') 2188 break; 2189 } 2190 } 2191 2192 if (action != key->dtak_action) { 2193 /* 2194 * We are aggregating on the same value in the same 2195 * aggregation with two different aggregating actions. 2196 * (This should have been picked up in the compiler, 2197 * so we may be dealing with errant or devious DIF.) 2198 * This is an error condition; we indicate as much, 2199 * and return. 2200 */ 2201 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP); 2202 return; 2203 } 2204 2205 /* 2206 * This is a hit: we need to apply the aggregator to 2207 * the value at this key. 2208 */ 2209 agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg); 2210 return; 2211next: 2212 continue; 2213 } 2214 2215 /* 2216 * We didn't find it. We need to allocate some zero-filled space, 2217 * link it into the hash table appropriately, and apply the aggregator 2218 * to the (zero-filled) value. 2219 */ 2220 offs = buf->dtb_offset; 2221 while (offs & (align - 1)) 2222 offs += sizeof (uint32_t); 2223 2224 /* 2225 * If we don't have enough room to both allocate a new key _and_ 2226 * its associated data, increment the drop count and return. 2227 */ 2228 if ((uintptr_t)tomax + offs + fsize > 2229 agb->dtagb_free - sizeof (dtrace_aggkey_t)) { 2230 dtrace_buffer_drop(buf); 2231 return; 2232 } 2233 2234 /*CONSTCOND*/ 2235 ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1))); 2236 key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t)); 2237 agb->dtagb_free -= sizeof (dtrace_aggkey_t); 2238 2239 key->dtak_data = kdata = tomax + offs; 2240 buf->dtb_offset = offs + fsize; 2241 2242 /* 2243 * Now copy the data across. 2244 */ 2245 *((dtrace_aggid_t *)kdata) = agg->dtag_id; 2246 2247 for (i = sizeof (dtrace_aggid_t); i < size; i++) 2248 kdata[i] = data[i]; 2249 2250 /* 2251 * Because strings are not zeroed out by default, we need to iterate 2252 * looking for actions that store strings, and we need to explicitly 2253 * pad these strings out with zeroes. 2254 */ 2255 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) { 2256 int nul; 2257 2258 if (!DTRACEACT_ISSTRING(act)) 2259 continue; 2260 2261 i = act->dta_rec.dtrd_offset - agg->dtag_base; 2262 limit = i + act->dta_rec.dtrd_size; 2263 ASSERT(limit <= size); 2264 2265 for (nul = 0; i < limit; i++) { 2266 if (nul) { 2267 kdata[i] = '\0'; 2268 continue; 2269 } 2270 2271 if (data[i] != '\0') 2272 continue; 2273 2274 nul = 1; 2275 } 2276 } 2277 2278 for (i = size; i < fsize; i++) 2279 kdata[i] = 0; 2280 2281 key->dtak_hashval = hashval; 2282 key->dtak_size = size; 2283 key->dtak_action = action; 2284 key->dtak_next = agb->dtagb_hash[ndx]; 2285 agb->dtagb_hash[ndx] = key; 2286 2287 /* 2288 * Finally, apply the aggregator. 2289 */ 2290 *((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial; 2291 agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg); 2292} 2293 2294/* 2295 * Given consumer state, this routine finds a speculation in the INACTIVE 2296 * state and transitions it into the ACTIVE state. If there is no speculation 2297 * in the INACTIVE state, 0 is returned. In this case, no error counter is 2298 * incremented -- it is up to the caller to take appropriate action. 2299 */ 2300static int 2301dtrace_speculation(dtrace_state_t *state) 2302{ 2303 int i = 0; 2304 dtrace_speculation_state_t current; 2305 uint32_t *stat = &state->dts_speculations_unavail, count; 2306 2307 while (i < state->dts_nspeculations) { 2308 dtrace_speculation_t *spec = &state->dts_speculations[i]; 2309 2310 current = spec->dtsp_state; 2311 2312 if (current != DTRACESPEC_INACTIVE) { 2313 if (current == DTRACESPEC_COMMITTINGMANY || 2314 current == DTRACESPEC_COMMITTING || 2315 current == DTRACESPEC_DISCARDING) 2316 stat = &state->dts_speculations_busy; 2317 i++; 2318 continue; 2319 } 2320 2321 if (dtrace_cas32((uint32_t *)&spec->dtsp_state, 2322 current, DTRACESPEC_ACTIVE) == current) 2323 return (i + 1); 2324 } 2325 2326 /* 2327 * We couldn't find a speculation. If we found as much as a single 2328 * busy speculation buffer, we'll attribute this failure as "busy" 2329 * instead of "unavail". 2330 */ 2331 do { 2332 count = *stat; 2333 } while (dtrace_cas32(stat, count, count + 1) != count); 2334 2335 return (0); 2336} 2337 2338/* 2339 * This routine commits an active speculation. If the specified speculation 2340 * is not in a valid state to perform a commit(), this routine will silently do 2341 * nothing. The state of the specified speculation is transitioned according 2342 * to the state transition diagram outlined in <sys/dtrace_impl.h> 2343 */ 2344static void 2345dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu, 2346 dtrace_specid_t which) 2347{ 2348 dtrace_speculation_t *spec; 2349 dtrace_buffer_t *src, *dest; 2350 uintptr_t daddr, saddr, dlimit, slimit; 2351 dtrace_speculation_state_t current, new = 0; 2352 intptr_t offs; 2353 uint64_t timestamp; 2354 2355 if (which == 0) 2356 return; 2357 2358 if (which > state->dts_nspeculations) { 2359 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP; 2360 return; 2361 } 2362 2363 spec = &state->dts_speculations[which - 1]; 2364 src = &spec->dtsp_buffer[cpu]; 2365 dest = &state->dts_buffer[cpu]; 2366 2367 do { 2368 current = spec->dtsp_state; 2369 2370 if (current == DTRACESPEC_COMMITTINGMANY) 2371 break; 2372 2373 switch (current) { 2374 case DTRACESPEC_INACTIVE: 2375 case DTRACESPEC_DISCARDING: 2376 return; 2377 2378 case DTRACESPEC_COMMITTING: 2379 /* 2380 * This is only possible if we are (a) commit()'ing 2381 * without having done a prior speculate() on this CPU 2382 * and (b) racing with another commit() on a different 2383 * CPU. There's nothing to do -- we just assert that 2384 * our offset is 0. 2385 */ 2386 ASSERT(src->dtb_offset == 0); 2387 return; 2388 2389 case DTRACESPEC_ACTIVE: 2390 new = DTRACESPEC_COMMITTING; 2391 break; 2392 2393 case DTRACESPEC_ACTIVEONE: 2394 /* 2395 * This speculation is active on one CPU. If our 2396 * buffer offset is non-zero, we know that the one CPU 2397 * must be us. Otherwise, we are committing on a 2398 * different CPU from the speculate(), and we must 2399 * rely on being asynchronously cleaned. 2400 */ 2401 if (src->dtb_offset != 0) { 2402 new = DTRACESPEC_COMMITTING; 2403 break; 2404 } 2405 /*FALLTHROUGH*/ 2406 2407 case DTRACESPEC_ACTIVEMANY: 2408 new = DTRACESPEC_COMMITTINGMANY; 2409 break; 2410 2411 default: 2412 ASSERT(0); 2413 } 2414 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state, 2415 current, new) != current); 2416 2417 /* 2418 * We have set the state to indicate that we are committing this 2419 * speculation. Now reserve the necessary space in the destination 2420 * buffer. 2421 */ 2422 if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset, 2423 sizeof (uint64_t), state, NULL)) < 0) { 2424 dtrace_buffer_drop(dest); 2425 goto out; 2426 } 2427 2428 /* 2429 * We have sufficient space to copy the speculative buffer into the 2430 * primary buffer. First, modify the speculative buffer, filling 2431 * in the timestamp of all entries with the current time. The data 2432 * must have the commit() time rather than the time it was traced, 2433 * so that all entries in the primary buffer are in timestamp order. 2434 */ 2435 timestamp = dtrace_gethrtime(); 2436 saddr = (uintptr_t)src->dtb_tomax; 2437 slimit = saddr + src->dtb_offset; 2438 while (saddr < slimit) { 2439 size_t size; 2440 dtrace_rechdr_t *dtrh = (dtrace_rechdr_t *)saddr; 2441 2442 if (dtrh->dtrh_epid == DTRACE_EPIDNONE) { 2443 saddr += sizeof (dtrace_epid_t); 2444 continue; 2445 } 2446 ASSERT3U(dtrh->dtrh_epid, <=, state->dts_necbs); 2447 size = state->dts_ecbs[dtrh->dtrh_epid - 1]->dte_size; 2448 2449 ASSERT3U(saddr + size, <=, slimit); 2450 ASSERT3U(size, >=, sizeof (dtrace_rechdr_t)); 2451 ASSERT3U(DTRACE_RECORD_LOAD_TIMESTAMP(dtrh), ==, UINT64_MAX); 2452 2453 DTRACE_RECORD_STORE_TIMESTAMP(dtrh, timestamp); 2454 2455 saddr += size; 2456 } 2457 2458 /* 2459 * Copy the buffer across. (Note that this is a 2460 * highly subobtimal bcopy(); in the unlikely event that this becomes 2461 * a serious performance issue, a high-performance DTrace-specific 2462 * bcopy() should obviously be invented.) 2463 */ 2464 daddr = (uintptr_t)dest->dtb_tomax + offs; 2465 dlimit = daddr + src->dtb_offset; 2466 saddr = (uintptr_t)src->dtb_tomax; 2467 2468 /* 2469 * First, the aligned portion. 2470 */ 2471 while (dlimit - daddr >= sizeof (uint64_t)) { 2472 *((uint64_t *)daddr) = *((uint64_t *)saddr); 2473 2474 daddr += sizeof (uint64_t); 2475 saddr += sizeof (uint64_t); 2476 } 2477 2478 /* 2479 * Now any left-over bit... 2480 */ 2481 while (dlimit - daddr) 2482 *((uint8_t *)daddr++) = *((uint8_t *)saddr++); 2483 2484 /* 2485 * Finally, commit the reserved space in the destination buffer. 2486 */ 2487 dest->dtb_offset = offs + src->dtb_offset; 2488 2489out: 2490 /* 2491 * If we're lucky enough to be the only active CPU on this speculation 2492 * buffer, we can just set the state back to DTRACESPEC_INACTIVE. 2493 */ 2494 if (current == DTRACESPEC_ACTIVE || 2495 (current == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) { 2496 uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state, 2497 DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE); 2498 2499 ASSERT(rval == DTRACESPEC_COMMITTING); 2500 } 2501 2502 src->dtb_offset = 0; 2503 src->dtb_xamot_drops += src->dtb_drops; 2504 src->dtb_drops = 0; 2505} 2506 2507/* 2508 * This routine discards an active speculation. If the specified speculation 2509 * is not in a valid state to perform a discard(), this routine will silently 2510 * do nothing. The state of the specified speculation is transitioned 2511 * according to the state transition diagram outlined in <sys/dtrace_impl.h> 2512 */ 2513static void 2514dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu, 2515 dtrace_specid_t which) 2516{ 2517 dtrace_speculation_t *spec; 2518 dtrace_speculation_state_t current, new = 0; 2519 dtrace_buffer_t *buf; 2520 2521 if (which == 0) 2522 return; 2523 2524 if (which > state->dts_nspeculations) { 2525 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP; 2526 return; 2527 } 2528 2529 spec = &state->dts_speculations[which - 1]; 2530 buf = &spec->dtsp_buffer[cpu]; 2531 2532 do { 2533 current = spec->dtsp_state; 2534 2535 switch (current) { 2536 case DTRACESPEC_INACTIVE: 2537 case DTRACESPEC_COMMITTINGMANY: 2538 case DTRACESPEC_COMMITTING: 2539 case DTRACESPEC_DISCARDING: 2540 return; 2541 2542 case DTRACESPEC_ACTIVE: 2543 case DTRACESPEC_ACTIVEMANY: 2544 new = DTRACESPEC_DISCARDING; 2545 break; 2546 2547 case DTRACESPEC_ACTIVEONE: 2548 if (buf->dtb_offset != 0) { 2549 new = DTRACESPEC_INACTIVE; 2550 } else { 2551 new = DTRACESPEC_DISCARDING; 2552 } 2553 break; 2554 2555 default: 2556 ASSERT(0); 2557 } 2558 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state, 2559 current, new) != current); 2560 2561 buf->dtb_offset = 0; 2562 buf->dtb_drops = 0; 2563} 2564 2565/* 2566 * Note: not called from probe context. This function is called 2567 * asynchronously from cross call context to clean any speculations that are 2568 * in the COMMITTINGMANY or DISCARDING states. These speculations may not be 2569 * transitioned back to the INACTIVE state until all CPUs have cleaned the 2570 * speculation. 2571 */ 2572static void 2573dtrace_speculation_clean_here(dtrace_state_t *state) 2574{ 2575 dtrace_icookie_t cookie; 2576 processorid_t cpu = curcpu; 2577 dtrace_buffer_t *dest = &state->dts_buffer[cpu]; 2578 dtrace_specid_t i; 2579 2580 cookie = dtrace_interrupt_disable(); 2581 2582 if (dest->dtb_tomax == NULL) { 2583 dtrace_interrupt_enable(cookie); 2584 return; 2585 } 2586 2587 for (i = 0; i < state->dts_nspeculations; i++) { 2588 dtrace_speculation_t *spec = &state->dts_speculations[i]; 2589 dtrace_buffer_t *src = &spec->dtsp_buffer[cpu]; 2590 2591 if (src->dtb_tomax == NULL) 2592 continue; 2593 2594 if (spec->dtsp_state == DTRACESPEC_DISCARDING) { 2595 src->dtb_offset = 0; 2596 continue; 2597 } 2598 2599 if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY) 2600 continue; 2601 2602 if (src->dtb_offset == 0) 2603 continue; 2604 2605 dtrace_speculation_commit(state, cpu, i + 1); 2606 } 2607 2608 dtrace_interrupt_enable(cookie); 2609} 2610 2611/* 2612 * Note: not called from probe context. This function is called 2613 * asynchronously (and at a regular interval) to clean any speculations that 2614 * are in the COMMITTINGMANY or DISCARDING states. If it discovers that there 2615 * is work to be done, it cross calls all CPUs to perform that work; 2616 * COMMITMANY and DISCARDING speculations may not be transitioned back to the 2617 * INACTIVE state until they have been cleaned by all CPUs. 2618 */ 2619static void 2620dtrace_speculation_clean(dtrace_state_t *state) 2621{ 2622 int work = 0, rv; 2623 dtrace_specid_t i; 2624 2625 for (i = 0; i < state->dts_nspeculations; i++) { 2626 dtrace_speculation_t *spec = &state->dts_speculations[i]; 2627 2628 ASSERT(!spec->dtsp_cleaning); 2629 2630 if (spec->dtsp_state != DTRACESPEC_DISCARDING && 2631 spec->dtsp_state != DTRACESPEC_COMMITTINGMANY) 2632 continue; 2633 2634 work++; 2635 spec->dtsp_cleaning = 1; 2636 } 2637 2638 if (!work) 2639 return; 2640 2641 dtrace_xcall(DTRACE_CPUALL, 2642 (dtrace_xcall_t)dtrace_speculation_clean_here, state); 2643 2644 /* 2645 * We now know that all CPUs have committed or discarded their 2646 * speculation buffers, as appropriate. We can now set the state 2647 * to inactive. 2648 */ 2649 for (i = 0; i < state->dts_nspeculations; i++) { 2650 dtrace_speculation_t *spec = &state->dts_speculations[i]; 2651 dtrace_speculation_state_t current, new; 2652 2653 if (!spec->dtsp_cleaning) 2654 continue; 2655 2656 current = spec->dtsp_state; 2657 ASSERT(current == DTRACESPEC_DISCARDING || 2658 current == DTRACESPEC_COMMITTINGMANY); 2659 2660 new = DTRACESPEC_INACTIVE; 2661 2662 rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, current, new); 2663 ASSERT(rv == current); 2664 spec->dtsp_cleaning = 0; 2665 } 2666} 2667 2668/* 2669 * Called as part of a speculate() to get the speculative buffer associated 2670 * with a given speculation. Returns NULL if the specified speculation is not 2671 * in an ACTIVE state. If the speculation is in the ACTIVEONE state -- and 2672 * the active CPU is not the specified CPU -- the speculation will be 2673 * atomically transitioned into the ACTIVEMANY state. 2674 */ 2675static dtrace_buffer_t * 2676dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid, 2677 dtrace_specid_t which) 2678{ 2679 dtrace_speculation_t *spec; 2680 dtrace_speculation_state_t current, new = 0; 2681 dtrace_buffer_t *buf; 2682 2683 if (which == 0) 2684 return (NULL); 2685 2686 if (which > state->dts_nspeculations) { 2687 cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP; 2688 return (NULL); 2689 } 2690 2691 spec = &state->dts_speculations[which - 1]; 2692 buf = &spec->dtsp_buffer[cpuid]; 2693 2694 do { 2695 current = spec->dtsp_state; 2696 2697 switch (current) { 2698 case DTRACESPEC_INACTIVE: 2699 case DTRACESPEC_COMMITTINGMANY: 2700 case DTRACESPEC_DISCARDING: 2701 return (NULL); 2702 2703 case DTRACESPEC_COMMITTING: 2704 ASSERT(buf->dtb_offset == 0); 2705 return (NULL); 2706 2707 case DTRACESPEC_ACTIVEONE: 2708 /* 2709 * This speculation is currently active on one CPU. 2710 * Check the offset in the buffer; if it's non-zero, 2711 * that CPU must be us (and we leave the state alone). 2712 * If it's zero, assume that we're starting on a new 2713 * CPU -- and change the state to indicate that the 2714 * speculation is active on more than one CPU. 2715 */ 2716 if (buf->dtb_offset != 0) 2717 return (buf); 2718 2719 new = DTRACESPEC_ACTIVEMANY; 2720 break; 2721 2722 case DTRACESPEC_ACTIVEMANY: 2723 return (buf); 2724 2725 case DTRACESPEC_ACTIVE: 2726 new = DTRACESPEC_ACTIVEONE; 2727 break; 2728 2729 default: 2730 ASSERT(0); 2731 } 2732 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state, 2733 current, new) != current); 2734 2735 ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY); 2736 return (buf); 2737} 2738 2739/* 2740 * Return a string. In the event that the user lacks the privilege to access 2741 * arbitrary kernel memory, we copy the string out to scratch memory so that we 2742 * don't fail access checking. 2743 * 2744 * dtrace_dif_variable() uses this routine as a helper for various 2745 * builtin values such as 'execname' and 'probefunc.' 2746 */ 2747uintptr_t 2748dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state, 2749 dtrace_mstate_t *mstate) 2750{ 2751 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE]; 2752 uintptr_t ret; 2753 size_t strsz; 2754 2755 /* 2756 * The easy case: this probe is allowed to read all of memory, so 2757 * we can just return this as a vanilla pointer. 2758 */ 2759 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0) 2760 return (addr); 2761 2762 /* 2763 * This is the tougher case: we copy the string in question from 2764 * kernel memory into scratch memory and return it that way: this 2765 * ensures that we won't trip up when access checking tests the 2766 * BYREF return value. 2767 */ 2768 strsz = dtrace_strlen((char *)addr, size) + 1; 2769 2770 if (mstate->dtms_scratch_ptr + strsz > 2771 mstate->dtms_scratch_base + mstate->dtms_scratch_size) { 2772 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 2773 return (0); 2774 } 2775 2776 dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr, 2777 strsz); 2778 ret = mstate->dtms_scratch_ptr; 2779 mstate->dtms_scratch_ptr += strsz; 2780 return (ret); 2781} 2782 2783/* 2784 * Return a string from a memoy address which is known to have one or 2785 * more concatenated, individually zero terminated, sub-strings. 2786 * In the event that the user lacks the privilege to access 2787 * arbitrary kernel memory, we copy the string out to scratch memory so that we 2788 * don't fail access checking. 2789 * 2790 * dtrace_dif_variable() uses this routine as a helper for various 2791 * builtin values such as 'execargs'. 2792 */ 2793static uintptr_t 2794dtrace_dif_varstrz(uintptr_t addr, size_t strsz, dtrace_state_t *state, 2795 dtrace_mstate_t *mstate) 2796{ 2797 char *p; 2798 size_t i; 2799 uintptr_t ret; 2800 2801 if (mstate->dtms_scratch_ptr + strsz > 2802 mstate->dtms_scratch_base + mstate->dtms_scratch_size) { 2803 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 2804 return (0); 2805 } 2806 2807 dtrace_bcopy((const void *)addr, (void *)mstate->dtms_scratch_ptr, 2808 strsz); 2809 2810 /* Replace sub-string termination characters with a space. */ 2811 for (p = (char *) mstate->dtms_scratch_ptr, i = 0; i < strsz - 1; 2812 p++, i++) 2813 if (*p == '\0') 2814 *p = ' '; 2815 2816 ret = mstate->dtms_scratch_ptr; 2817 mstate->dtms_scratch_ptr += strsz; 2818 return (ret); 2819} 2820 2821/* 2822 * This function implements the DIF emulator's variable lookups. The emulator 2823 * passes a reserved variable identifier and optional built-in array index. 2824 */ 2825static uint64_t 2826dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v, 2827 uint64_t ndx) 2828{ 2829 /* 2830 * If we're accessing one of the uncached arguments, we'll turn this 2831 * into a reference in the args array. 2832 */ 2833 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) { 2834 ndx = v - DIF_VAR_ARG0; 2835 v = DIF_VAR_ARGS; 2836 } 2837 2838 switch (v) { 2839 case DIF_VAR_ARGS: 2840 ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS); 2841 if (ndx >= sizeof (mstate->dtms_arg) / 2842 sizeof (mstate->dtms_arg[0])) { 2843 int aframes = mstate->dtms_probe->dtpr_aframes + 2; 2844 dtrace_provider_t *pv; 2845 uint64_t val; 2846 2847 pv = mstate->dtms_probe->dtpr_provider; 2848 if (pv->dtpv_pops.dtps_getargval != NULL) 2849 val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg, 2850 mstate->dtms_probe->dtpr_id, 2851 mstate->dtms_probe->dtpr_arg, ndx, aframes); 2852 else 2853 val = dtrace_getarg(ndx, aframes); 2854 2855 /* 2856 * This is regrettably required to keep the compiler 2857 * from tail-optimizing the call to dtrace_getarg(). 2858 * The condition always evaluates to true, but the 2859 * compiler has no way of figuring that out a priori. 2860 * (None of this would be necessary if the compiler 2861 * could be relied upon to _always_ tail-optimize 2862 * the call to dtrace_getarg() -- but it can't.) 2863 */ 2864 if (mstate->dtms_probe != NULL) 2865 return (val); 2866 2867 ASSERT(0); 2868 } 2869 2870 return (mstate->dtms_arg[ndx]); 2871 2872#if defined(sun) 2873 case DIF_VAR_UREGS: { 2874 klwp_t *lwp; 2875 2876 if (!dtrace_priv_proc(state)) 2877 return (0); 2878 2879 if ((lwp = curthread->t_lwp) == NULL) { 2880 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR); 2881 cpu_core[curcpu].cpuc_dtrace_illval = NULL; 2882 return (0); 2883 } 2884 2885 return (dtrace_getreg(lwp->lwp_regs, ndx)); 2886 return (0); 2887 } 2888#else 2889 case DIF_VAR_UREGS: { 2890 struct trapframe *tframe; 2891 2892 if (!dtrace_priv_proc(state)) 2893 return (0); 2894 2895 if ((tframe = curthread->td_frame) == NULL) { 2896 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR); 2897 cpu_core[curcpu].cpuc_dtrace_illval = 0; 2898 return (0); 2899 } 2900 2901 return (dtrace_getreg(tframe, ndx)); 2902 } 2903#endif 2904 2905 case DIF_VAR_CURTHREAD: 2906 if (!dtrace_priv_kernel(state)) 2907 return (0); 2908 return ((uint64_t)(uintptr_t)curthread); 2909 2910 case DIF_VAR_TIMESTAMP: 2911 if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) { 2912 mstate->dtms_timestamp = dtrace_gethrtime(); 2913 mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP; 2914 } 2915 return (mstate->dtms_timestamp); 2916 2917 case DIF_VAR_VTIMESTAMP: 2918 ASSERT(dtrace_vtime_references != 0); 2919 return (curthread->t_dtrace_vtime); 2920 2921 case DIF_VAR_WALLTIMESTAMP: 2922 if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) { 2923 mstate->dtms_walltimestamp = dtrace_gethrestime(); 2924 mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP; 2925 } 2926 return (mstate->dtms_walltimestamp); 2927 2928#if defined(sun) 2929 case DIF_VAR_IPL: 2930 if (!dtrace_priv_kernel(state)) 2931 return (0); 2932 if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) { 2933 mstate->dtms_ipl = dtrace_getipl(); 2934 mstate->dtms_present |= DTRACE_MSTATE_IPL; 2935 } 2936 return (mstate->dtms_ipl); 2937#endif 2938 2939 case DIF_VAR_EPID: 2940 ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID); 2941 return (mstate->dtms_epid); 2942 2943 case DIF_VAR_ID: 2944 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE); 2945 return (mstate->dtms_probe->dtpr_id); 2946 2947 case DIF_VAR_STACKDEPTH: 2948 if (!dtrace_priv_kernel(state)) 2949 return (0); 2950 if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) { 2951 int aframes = mstate->dtms_probe->dtpr_aframes + 2; 2952 2953 mstate->dtms_stackdepth = dtrace_getstackdepth(aframes); 2954 mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH; 2955 } 2956 return (mstate->dtms_stackdepth); 2957 2958 case DIF_VAR_USTACKDEPTH: 2959 if (!dtrace_priv_proc(state)) 2960 return (0); 2961 if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) { 2962 /* 2963 * See comment in DIF_VAR_PID. 2964 */ 2965 if (DTRACE_ANCHORED(mstate->dtms_probe) && 2966 CPU_ON_INTR(CPU)) { 2967 mstate->dtms_ustackdepth = 0; 2968 } else { 2969 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT); 2970 mstate->dtms_ustackdepth = 2971 dtrace_getustackdepth(); 2972 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT); 2973 } 2974 mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH; 2975 } 2976 return (mstate->dtms_ustackdepth); 2977 2978 case DIF_VAR_CALLER: 2979 if (!dtrace_priv_kernel(state)) 2980 return (0); 2981 if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) { 2982 int aframes = mstate->dtms_probe->dtpr_aframes + 2; 2983 2984 if (!DTRACE_ANCHORED(mstate->dtms_probe)) { 2985 /* 2986 * If this is an unanchored probe, we are 2987 * required to go through the slow path: 2988 * dtrace_caller() only guarantees correct 2989 * results for anchored probes. 2990 */ 2991 pc_t caller[2] = {0, 0}; 2992 2993 dtrace_getpcstack(caller, 2, aframes, 2994 (uint32_t *)(uintptr_t)mstate->dtms_arg[0]); 2995 mstate->dtms_caller = caller[1]; 2996 } else if ((mstate->dtms_caller = 2997 dtrace_caller(aframes)) == -1) { 2998 /* 2999 * We have failed to do this the quick way; 3000 * we must resort to the slower approach of 3001 * calling dtrace_getpcstack(). 3002 */ 3003 pc_t caller = 0; 3004 3005 dtrace_getpcstack(&caller, 1, aframes, NULL); 3006 mstate->dtms_caller = caller; 3007 } 3008 3009 mstate->dtms_present |= DTRACE_MSTATE_CALLER; 3010 } 3011 return (mstate->dtms_caller); 3012 3013 case DIF_VAR_UCALLER: 3014 if (!dtrace_priv_proc(state)) 3015 return (0); 3016 3017 if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) { 3018 uint64_t ustack[3]; 3019 3020 /* 3021 * dtrace_getupcstack() fills in the first uint64_t 3022 * with the current PID. The second uint64_t will 3023 * be the program counter at user-level. The third 3024 * uint64_t will contain the caller, which is what 3025 * we're after. 3026 */ 3027 ustack[2] = 0; 3028 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT); 3029 dtrace_getupcstack(ustack, 3); 3030 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT); 3031 mstate->dtms_ucaller = ustack[2]; 3032 mstate->dtms_present |= DTRACE_MSTATE_UCALLER; 3033 } 3034 3035 return (mstate->dtms_ucaller); 3036 3037 case DIF_VAR_PROBEPROV: 3038 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE); 3039 return (dtrace_dif_varstr( 3040 (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name, 3041 state, mstate)); 3042 3043 case DIF_VAR_PROBEMOD: 3044 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE); 3045 return (dtrace_dif_varstr( 3046 (uintptr_t)mstate->dtms_probe->dtpr_mod, 3047 state, mstate)); 3048 3049 case DIF_VAR_PROBEFUNC: 3050 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE); 3051 return (dtrace_dif_varstr( 3052 (uintptr_t)mstate->dtms_probe->dtpr_func, 3053 state, mstate)); 3054 3055 case DIF_VAR_PROBENAME: 3056 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE); 3057 return (dtrace_dif_varstr( 3058 (uintptr_t)mstate->dtms_probe->dtpr_name, 3059 state, mstate)); 3060 3061 case DIF_VAR_PID: 3062 if (!dtrace_priv_proc(state)) 3063 return (0); 3064 3065#if defined(sun) 3066 /* 3067 * Note that we are assuming that an unanchored probe is 3068 * always due to a high-level interrupt. (And we're assuming 3069 * that there is only a single high level interrupt.) 3070 */ 3071 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU)) 3072 return (pid0.pid_id); 3073 3074 /* 3075 * It is always safe to dereference one's own t_procp pointer: 3076 * it always points to a valid, allocated proc structure. 3077 * Further, it is always safe to dereference the p_pidp member 3078 * of one's own proc structure. (These are truisms becuase 3079 * threads and processes don't clean up their own state -- 3080 * they leave that task to whomever reaps them.) 3081 */ 3082 return ((uint64_t)curthread->t_procp->p_pidp->pid_id); 3083#else 3084 return ((uint64_t)curproc->p_pid); 3085#endif 3086 3087 case DIF_VAR_PPID: 3088 if (!dtrace_priv_proc(state)) 3089 return (0); 3090 3091#if defined(sun) 3092 /* 3093 * See comment in DIF_VAR_PID. 3094 */ 3095 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU)) 3096 return (pid0.pid_id); 3097 3098 /* 3099 * It is always safe to dereference one's own t_procp pointer: 3100 * it always points to a valid, allocated proc structure. 3101 * (This is true because threads don't clean up their own 3102 * state -- they leave that task to whomever reaps them.) 3103 */ 3104 return ((uint64_t)curthread->t_procp->p_ppid); 3105#else 3106 return ((uint64_t)curproc->p_pptr->p_pid); 3107#endif 3108 3109 case DIF_VAR_TID: 3110#if defined(sun) 3111 /* 3112 * See comment in DIF_VAR_PID. 3113 */ 3114 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU)) 3115 return (0); 3116#endif 3117 3118 return ((uint64_t)curthread->t_tid); 3119 3120 case DIF_VAR_EXECARGS: { 3121 struct pargs *p_args = curthread->td_proc->p_args; 3122 3123 if (p_args == NULL) 3124 return(0); 3125 3126 return (dtrace_dif_varstrz( 3127 (uintptr_t) p_args->ar_args, p_args->ar_length, state, mstate)); 3128 } 3129 3130 case DIF_VAR_EXECNAME: 3131#if defined(sun) 3132 if (!dtrace_priv_proc(state)) 3133 return (0); 3134 3135 /* 3136 * See comment in DIF_VAR_PID. 3137 */ 3138 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU)) 3139 return ((uint64_t)(uintptr_t)p0.p_user.u_comm); 3140 3141 /* 3142 * It is always safe to dereference one's own t_procp pointer: 3143 * it always points to a valid, allocated proc structure. 3144 * (This is true because threads don't clean up their own 3145 * state -- they leave that task to whomever reaps them.) 3146 */ 3147 return (dtrace_dif_varstr( 3148 (uintptr_t)curthread->t_procp->p_user.u_comm, 3149 state, mstate)); 3150#else 3151 return (dtrace_dif_varstr( 3152 (uintptr_t) curthread->td_proc->p_comm, state, mstate)); 3153#endif 3154 3155 case DIF_VAR_ZONENAME: 3156#if defined(sun) 3157 if (!dtrace_priv_proc(state)) 3158 return (0); 3159 3160 /* 3161 * See comment in DIF_VAR_PID. 3162 */ 3163 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU)) 3164 return ((uint64_t)(uintptr_t)p0.p_zone->zone_name); 3165 3166 /* 3167 * It is always safe to dereference one's own t_procp pointer: 3168 * it always points to a valid, allocated proc structure. 3169 * (This is true because threads don't clean up their own 3170 * state -- they leave that task to whomever reaps them.) 3171 */ 3172 return (dtrace_dif_varstr( 3173 (uintptr_t)curthread->t_procp->p_zone->zone_name, 3174 state, mstate)); 3175#else 3176 return (0); 3177#endif 3178 3179 case DIF_VAR_UID: 3180 if (!dtrace_priv_proc(state)) 3181 return (0); 3182 3183#if defined(sun) 3184 /* 3185 * See comment in DIF_VAR_PID. 3186 */ 3187 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU)) 3188 return ((uint64_t)p0.p_cred->cr_uid); 3189#endif 3190 3191 /* 3192 * It is always safe to dereference one's own t_procp pointer: 3193 * it always points to a valid, allocated proc structure. 3194 * (This is true because threads don't clean up their own 3195 * state -- they leave that task to whomever reaps them.) 3196 * 3197 * Additionally, it is safe to dereference one's own process 3198 * credential, since this is never NULL after process birth. 3199 */ 3200 return ((uint64_t)curthread->t_procp->p_cred->cr_uid); 3201 3202 case DIF_VAR_GID: 3203 if (!dtrace_priv_proc(state)) 3204 return (0); 3205 3206#if defined(sun) 3207 /* 3208 * See comment in DIF_VAR_PID. 3209 */ 3210 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU)) 3211 return ((uint64_t)p0.p_cred->cr_gid); 3212#endif 3213 3214 /* 3215 * It is always safe to dereference one's own t_procp pointer: 3216 * it always points to a valid, allocated proc structure. 3217 * (This is true because threads don't clean up their own 3218 * state -- they leave that task to whomever reaps them.) 3219 * 3220 * Additionally, it is safe to dereference one's own process 3221 * credential, since this is never NULL after process birth. 3222 */ 3223 return ((uint64_t)curthread->t_procp->p_cred->cr_gid); 3224 3225 case DIF_VAR_ERRNO: { 3226#if defined(sun) 3227 klwp_t *lwp; 3228 if (!dtrace_priv_proc(state)) 3229 return (0); 3230 3231 /* 3232 * See comment in DIF_VAR_PID. 3233 */ 3234 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU)) 3235 return (0); 3236 3237 /* 3238 * It is always safe to dereference one's own t_lwp pointer in 3239 * the event that this pointer is non-NULL. (This is true 3240 * because threads and lwps don't clean up their own state -- 3241 * they leave that task to whomever reaps them.) 3242 */ 3243 if ((lwp = curthread->t_lwp) == NULL) 3244 return (0); 3245 3246 return ((uint64_t)lwp->lwp_errno); 3247#else 3248 return (curthread->td_errno); 3249#endif 3250 } 3251#if !defined(sun) 3252 case DIF_VAR_CPU: { 3253 return curcpu; 3254 } 3255#endif 3256 default: 3257 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP); 3258 return (0); 3259 } 3260} 3261 3262/* 3263 * Emulate the execution of DTrace ID subroutines invoked by the call opcode. 3264 * Notice that we don't bother validating the proper number of arguments or 3265 * their types in the tuple stack. This isn't needed because all argument 3266 * interpretation is safe because of our load safety -- the worst that can 3267 * happen is that a bogus program can obtain bogus results. 3268 */ 3269static void 3270dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs, 3271 dtrace_key_t *tupregs, int nargs, 3272 dtrace_mstate_t *mstate, dtrace_state_t *state) 3273{ 3274 volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags; 3275 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval; 3276 dtrace_vstate_t *vstate = &state->dts_vstate; 3277 3278#if defined(sun) 3279 union { 3280 mutex_impl_t mi; 3281 uint64_t mx; 3282 } m; 3283 3284 union { 3285 krwlock_t ri; 3286 uintptr_t rw; 3287 } r; 3288#else 3289 struct thread *lowner; 3290 union { 3291 struct lock_object *li; 3292 uintptr_t lx; 3293 } l; 3294#endif 3295 3296 switch (subr) { 3297 case DIF_SUBR_RAND: 3298 regs[rd] = (dtrace_gethrtime() * 2416 + 374441) % 1771875; 3299 break; 3300 3301#if defined(sun) 3302 case DIF_SUBR_MUTEX_OWNED: 3303 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t), 3304 mstate, vstate)) { 3305 regs[rd] = 0; 3306 break; 3307 } 3308 3309 m.mx = dtrace_load64(tupregs[0].dttk_value); 3310 if (MUTEX_TYPE_ADAPTIVE(&m.mi)) 3311 regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER; 3312 else 3313 regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock); 3314 break; 3315 3316 case DIF_SUBR_MUTEX_OWNER: 3317 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t), 3318 mstate, vstate)) { 3319 regs[rd] = 0; 3320 break; 3321 } 3322 3323 m.mx = dtrace_load64(tupregs[0].dttk_value); 3324 if (MUTEX_TYPE_ADAPTIVE(&m.mi) && 3325 MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER) 3326 regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi); 3327 else 3328 regs[rd] = 0; 3329 break; 3330 3331 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE: 3332 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t), 3333 mstate, vstate)) { 3334 regs[rd] = 0; 3335 break; 3336 } 3337 3338 m.mx = dtrace_load64(tupregs[0].dttk_value); 3339 regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi); 3340 break; 3341 3342 case DIF_SUBR_MUTEX_TYPE_SPIN: 3343 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t), 3344 mstate, vstate)) { 3345 regs[rd] = 0; 3346 break; 3347 } 3348 3349 m.mx = dtrace_load64(tupregs[0].dttk_value); 3350 regs[rd] = MUTEX_TYPE_SPIN(&m.mi); 3351 break; 3352 3353 case DIF_SUBR_RW_READ_HELD: { 3354 uintptr_t tmp; 3355 3356 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t), 3357 mstate, vstate)) { 3358 regs[rd] = 0; 3359 break; 3360 } 3361 3362 r.rw = dtrace_loadptr(tupregs[0].dttk_value); 3363 regs[rd] = _RW_READ_HELD(&r.ri, tmp); 3364 break; 3365 } 3366 3367 case DIF_SUBR_RW_WRITE_HELD: 3368 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t), 3369 mstate, vstate)) { 3370 regs[rd] = 0; 3371 break; 3372 } 3373 3374 r.rw = dtrace_loadptr(tupregs[0].dttk_value); 3375 regs[rd] = _RW_WRITE_HELD(&r.ri); 3376 break; 3377 3378 case DIF_SUBR_RW_ISWRITER: 3379 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t), 3380 mstate, vstate)) { 3381 regs[rd] = 0; 3382 break; 3383 } 3384 3385 r.rw = dtrace_loadptr(tupregs[0].dttk_value); 3386 regs[rd] = _RW_ISWRITER(&r.ri); 3387 break; 3388 3389#else 3390 case DIF_SUBR_MUTEX_OWNED: 3391 if (!dtrace_canload(tupregs[0].dttk_value, 3392 sizeof (struct lock_object), mstate, vstate)) { 3393 regs[rd] = 0; 3394 break; 3395 } 3396 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value); 3397 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner); 3398 break; 3399 3400 case DIF_SUBR_MUTEX_OWNER: 3401 if (!dtrace_canload(tupregs[0].dttk_value, 3402 sizeof (struct lock_object), mstate, vstate)) { 3403 regs[rd] = 0; 3404 break; 3405 } 3406 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value); 3407 LOCK_CLASS(l.li)->lc_owner(l.li, &lowner); 3408 regs[rd] = (uintptr_t)lowner; 3409 break; 3410 3411 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE: 3412 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx), 3413 mstate, vstate)) { 3414 regs[rd] = 0; 3415 break; 3416 } 3417 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value); 3418 /* XXX - should be only LC_SLEEPABLE? */ 3419 regs[rd] = (LOCK_CLASS(l.li)->lc_flags & 3420 (LC_SLEEPLOCK | LC_SLEEPABLE)) != 0; 3421 break; 3422 3423 case DIF_SUBR_MUTEX_TYPE_SPIN: 3424 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx), 3425 mstate, vstate)) { 3426 regs[rd] = 0; 3427 break; 3428 } 3429 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value); 3430 regs[rd] = (LOCK_CLASS(l.li)->lc_flags & LC_SPINLOCK) != 0; 3431 break; 3432 3433 case DIF_SUBR_RW_READ_HELD: 3434 case DIF_SUBR_SX_SHARED_HELD: 3435 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t), 3436 mstate, vstate)) { 3437 regs[rd] = 0; 3438 break; 3439 } 3440 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value); 3441 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) && 3442 lowner == NULL; 3443 break; 3444 3445 case DIF_SUBR_RW_WRITE_HELD: 3446 case DIF_SUBR_SX_EXCLUSIVE_HELD: 3447 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t), 3448 mstate, vstate)) { 3449 regs[rd] = 0; 3450 break; 3451 } 3452 l.lx = dtrace_loadptr(tupregs[0].dttk_value); 3453 LOCK_CLASS(l.li)->lc_owner(l.li, &lowner); 3454 regs[rd] = (lowner == curthread); 3455 break; 3456 3457 case DIF_SUBR_RW_ISWRITER: 3458 case DIF_SUBR_SX_ISEXCLUSIVE: 3459 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t), 3460 mstate, vstate)) { 3461 regs[rd] = 0; 3462 break; 3463 } 3464 l.lx = dtrace_loadptr(tupregs[0].dttk_value); 3465 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) && 3466 lowner != NULL; 3467 break; 3468#endif /* ! defined(sun) */ 3469 3470 case DIF_SUBR_BCOPY: { 3471 /* 3472 * We need to be sure that the destination is in the scratch 3473 * region -- no other region is allowed. 3474 */ 3475 uintptr_t src = tupregs[0].dttk_value; 3476 uintptr_t dest = tupregs[1].dttk_value; 3477 size_t size = tupregs[2].dttk_value; 3478 3479 if (!dtrace_inscratch(dest, size, mstate)) { 3480 *flags |= CPU_DTRACE_BADADDR; 3481 *illval = regs[rd]; 3482 break; 3483 } 3484 3485 if (!dtrace_canload(src, size, mstate, vstate)) { 3486 regs[rd] = 0; 3487 break; 3488 } 3489 3490 dtrace_bcopy((void *)src, (void *)dest, size); 3491 break; 3492 } 3493 3494 case DIF_SUBR_ALLOCA: 3495 case DIF_SUBR_COPYIN: { 3496 uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8); 3497 uint64_t size = 3498 tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value; 3499 size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size; 3500 3501 /* 3502 * This action doesn't require any credential checks since 3503 * probes will not activate in user contexts to which the 3504 * enabling user does not have permissions. 3505 */ 3506 3507 /* 3508 * Rounding up the user allocation size could have overflowed 3509 * a large, bogus allocation (like -1ULL) to 0. 3510 */ 3511 if (scratch_size < size || 3512 !DTRACE_INSCRATCH(mstate, scratch_size)) { 3513 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 3514 regs[rd] = 0; 3515 break; 3516 } 3517 3518 if (subr == DIF_SUBR_COPYIN) { 3519 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT); 3520 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags); 3521 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT); 3522 } 3523 3524 mstate->dtms_scratch_ptr += scratch_size; 3525 regs[rd] = dest; 3526 break; 3527 } 3528 3529 case DIF_SUBR_COPYINTO: { 3530 uint64_t size = tupregs[1].dttk_value; 3531 uintptr_t dest = tupregs[2].dttk_value; 3532 3533 /* 3534 * This action doesn't require any credential checks since 3535 * probes will not activate in user contexts to which the 3536 * enabling user does not have permissions. 3537 */ 3538 if (!dtrace_inscratch(dest, size, mstate)) { 3539 *flags |= CPU_DTRACE_BADADDR; 3540 *illval = regs[rd]; 3541 break; 3542 } 3543 3544 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT); 3545 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags); 3546 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT); 3547 break; 3548 } 3549 3550 case DIF_SUBR_COPYINSTR: { 3551 uintptr_t dest = mstate->dtms_scratch_ptr; 3552 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE]; 3553 3554 if (nargs > 1 && tupregs[1].dttk_value < size) 3555 size = tupregs[1].dttk_value + 1; 3556 3557 /* 3558 * This action doesn't require any credential checks since 3559 * probes will not activate in user contexts to which the 3560 * enabling user does not have permissions. 3561 */ 3562 if (!DTRACE_INSCRATCH(mstate, size)) { 3563 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 3564 regs[rd] = 0; 3565 break; 3566 } 3567 3568 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT); 3569 dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags); 3570 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT); 3571 3572 ((char *)dest)[size - 1] = '\0'; 3573 mstate->dtms_scratch_ptr += size; 3574 regs[rd] = dest; 3575 break; 3576 } 3577 3578#if defined(sun) 3579 case DIF_SUBR_MSGSIZE: 3580 case DIF_SUBR_MSGDSIZE: { 3581 uintptr_t baddr = tupregs[0].dttk_value, daddr; 3582 uintptr_t wptr, rptr; 3583 size_t count = 0; 3584 int cont = 0; 3585 3586 while (baddr != 0 && !(*flags & CPU_DTRACE_FAULT)) { 3587 3588 if (!dtrace_canload(baddr, sizeof (mblk_t), mstate, 3589 vstate)) { 3590 regs[rd] = 0; 3591 break; 3592 } 3593 3594 wptr = dtrace_loadptr(baddr + 3595 offsetof(mblk_t, b_wptr)); 3596 3597 rptr = dtrace_loadptr(baddr + 3598 offsetof(mblk_t, b_rptr)); 3599 3600 if (wptr < rptr) { 3601 *flags |= CPU_DTRACE_BADADDR; 3602 *illval = tupregs[0].dttk_value; 3603 break; 3604 } 3605 3606 daddr = dtrace_loadptr(baddr + 3607 offsetof(mblk_t, b_datap)); 3608 3609 baddr = dtrace_loadptr(baddr + 3610 offsetof(mblk_t, b_cont)); 3611 3612 /* 3613 * We want to prevent against denial-of-service here, 3614 * so we're only going to search the list for 3615 * dtrace_msgdsize_max mblks. 3616 */ 3617 if (cont++ > dtrace_msgdsize_max) { 3618 *flags |= CPU_DTRACE_ILLOP; 3619 break; 3620 } 3621 3622 if (subr == DIF_SUBR_MSGDSIZE) { 3623 if (dtrace_load8(daddr + 3624 offsetof(dblk_t, db_type)) != M_DATA) 3625 continue; 3626 } 3627 3628 count += wptr - rptr; 3629 } 3630 3631 if (!(*flags & CPU_DTRACE_FAULT)) 3632 regs[rd] = count; 3633 3634 break; 3635 } 3636#endif 3637 3638 case DIF_SUBR_PROGENYOF: { 3639 pid_t pid = tupregs[0].dttk_value; 3640 proc_t *p; 3641 int rval = 0; 3642 3643 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT); 3644 3645 for (p = curthread->t_procp; p != NULL; p = p->p_parent) { 3646#if defined(sun) 3647 if (p->p_pidp->pid_id == pid) { 3648#else 3649 if (p->p_pid == pid) { 3650#endif 3651 rval = 1; 3652 break; 3653 } 3654 } 3655 3656 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT); 3657 3658 regs[rd] = rval; 3659 break; 3660 } 3661 3662 case DIF_SUBR_SPECULATION: 3663 regs[rd] = dtrace_speculation(state); 3664 break; 3665 3666 case DIF_SUBR_COPYOUT: { 3667 uintptr_t kaddr = tupregs[0].dttk_value; 3668 uintptr_t uaddr = tupregs[1].dttk_value; 3669 uint64_t size = tupregs[2].dttk_value; 3670 3671 if (!dtrace_destructive_disallow && 3672 dtrace_priv_proc_control(state) && 3673 !dtrace_istoxic(kaddr, size)) { 3674 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT); 3675 dtrace_copyout(kaddr, uaddr, size, flags); 3676 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT); 3677 } 3678 break; 3679 } 3680 3681 case DIF_SUBR_COPYOUTSTR: { 3682 uintptr_t kaddr = tupregs[0].dttk_value; 3683 uintptr_t uaddr = tupregs[1].dttk_value; 3684 uint64_t size = tupregs[2].dttk_value; 3685 3686 if (!dtrace_destructive_disallow && 3687 dtrace_priv_proc_control(state) && 3688 !dtrace_istoxic(kaddr, size)) { 3689 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT); 3690 dtrace_copyoutstr(kaddr, uaddr, size, flags); 3691 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT); 3692 } 3693 break; 3694 } 3695 3696 case DIF_SUBR_STRLEN: { 3697 size_t sz; 3698 uintptr_t addr = (uintptr_t)tupregs[0].dttk_value; 3699 sz = dtrace_strlen((char *)addr, 3700 state->dts_options[DTRACEOPT_STRSIZE]); 3701 3702 if (!dtrace_canload(addr, sz + 1, mstate, vstate)) { 3703 regs[rd] = 0; 3704 break; 3705 } 3706 3707 regs[rd] = sz; 3708 3709 break; 3710 } 3711 3712 case DIF_SUBR_STRCHR: 3713 case DIF_SUBR_STRRCHR: { 3714 /* 3715 * We're going to iterate over the string looking for the 3716 * specified character. We will iterate until we have reached 3717 * the string length or we have found the character. If this 3718 * is DIF_SUBR_STRRCHR, we will look for the last occurrence 3719 * of the specified character instead of the first. 3720 */ 3721 uintptr_t saddr = tupregs[0].dttk_value; 3722 uintptr_t addr = tupregs[0].dttk_value; 3723 uintptr_t limit = addr + state->dts_options[DTRACEOPT_STRSIZE]; 3724 char c, target = (char)tupregs[1].dttk_value; 3725 3726 for (regs[rd] = 0; addr < limit; addr++) { 3727 if ((c = dtrace_load8(addr)) == target) { 3728 regs[rd] = addr; 3729 3730 if (subr == DIF_SUBR_STRCHR) 3731 break; 3732 } 3733 3734 if (c == '\0') 3735 break; 3736 } 3737 3738 if (!dtrace_canload(saddr, addr - saddr, mstate, vstate)) { 3739 regs[rd] = 0; 3740 break; 3741 } 3742 3743 break; 3744 } 3745 3746 case DIF_SUBR_STRSTR: 3747 case DIF_SUBR_INDEX: 3748 case DIF_SUBR_RINDEX: { 3749 /* 3750 * We're going to iterate over the string looking for the 3751 * specified string. We will iterate until we have reached 3752 * the string length or we have found the string. (Yes, this 3753 * is done in the most naive way possible -- but considering 3754 * that the string we're searching for is likely to be 3755 * relatively short, the complexity of Rabin-Karp or similar 3756 * hardly seems merited.) 3757 */ 3758 char *addr = (char *)(uintptr_t)tupregs[0].dttk_value; 3759 char *substr = (char *)(uintptr_t)tupregs[1].dttk_value; 3760 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE]; 3761 size_t len = dtrace_strlen(addr, size); 3762 size_t sublen = dtrace_strlen(substr, size); 3763 char *limit = addr + len, *orig = addr; 3764 int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1; 3765 int inc = 1; 3766 3767 regs[rd] = notfound; 3768 3769 if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) { 3770 regs[rd] = 0; 3771 break; 3772 } 3773 3774 if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate, 3775 vstate)) { 3776 regs[rd] = 0; 3777 break; 3778 } 3779 3780 /* 3781 * strstr() and index()/rindex() have similar semantics if 3782 * both strings are the empty string: strstr() returns a 3783 * pointer to the (empty) string, and index() and rindex() 3784 * both return index 0 (regardless of any position argument). 3785 */ 3786 if (sublen == 0 && len == 0) { 3787 if (subr == DIF_SUBR_STRSTR) 3788 regs[rd] = (uintptr_t)addr; 3789 else 3790 regs[rd] = 0; 3791 break; 3792 } 3793 3794 if (subr != DIF_SUBR_STRSTR) { 3795 if (subr == DIF_SUBR_RINDEX) { 3796 limit = orig - 1; 3797 addr += len; 3798 inc = -1; 3799 } 3800 3801 /* 3802 * Both index() and rindex() take an optional position 3803 * argument that denotes the starting position. 3804 */ 3805 if (nargs == 3) { 3806 int64_t pos = (int64_t)tupregs[2].dttk_value; 3807 3808 /* 3809 * If the position argument to index() is 3810 * negative, Perl implicitly clamps it at 3811 * zero. This semantic is a little surprising 3812 * given the special meaning of negative 3813 * positions to similar Perl functions like 3814 * substr(), but it appears to reflect a 3815 * notion that index() can start from a 3816 * negative index and increment its way up to 3817 * the string. Given this notion, Perl's 3818 * rindex() is at least self-consistent in 3819 * that it implicitly clamps positions greater 3820 * than the string length to be the string 3821 * length. Where Perl completely loses 3822 * coherence, however, is when the specified 3823 * substring is the empty string (""). In 3824 * this case, even if the position is 3825 * negative, rindex() returns 0 -- and even if 3826 * the position is greater than the length, 3827 * index() returns the string length. These 3828 * semantics violate the notion that index() 3829 * should never return a value less than the 3830 * specified position and that rindex() should 3831 * never return a value greater than the 3832 * specified position. (One assumes that 3833 * these semantics are artifacts of Perl's 3834 * implementation and not the results of 3835 * deliberate design -- it beggars belief that 3836 * even Larry Wall could desire such oddness.) 3837 * While in the abstract one would wish for 3838 * consistent position semantics across 3839 * substr(), index() and rindex() -- or at the 3840 * very least self-consistent position 3841 * semantics for index() and rindex() -- we 3842 * instead opt to keep with the extant Perl 3843 * semantics, in all their broken glory. (Do 3844 * we have more desire to maintain Perl's 3845 * semantics than Perl does? Probably.) 3846 */ 3847 if (subr == DIF_SUBR_RINDEX) { 3848 if (pos < 0) { 3849 if (sublen == 0) 3850 regs[rd] = 0; 3851 break; 3852 } 3853 3854 if (pos > len) 3855 pos = len; 3856 } else { 3857 if (pos < 0) 3858 pos = 0; 3859 3860 if (pos >= len) { 3861 if (sublen == 0) 3862 regs[rd] = len; 3863 break; 3864 } 3865 } 3866 3867 addr = orig + pos; 3868 } 3869 } 3870 3871 for (regs[rd] = notfound; addr != limit; addr += inc) { 3872 if (dtrace_strncmp(addr, substr, sublen) == 0) { 3873 if (subr != DIF_SUBR_STRSTR) { 3874 /* 3875 * As D index() and rindex() are 3876 * modeled on Perl (and not on awk), 3877 * we return a zero-based (and not a 3878 * one-based) index. (For you Perl 3879 * weenies: no, we're not going to add 3880 * $[ -- and shouldn't you be at a con 3881 * or something?) 3882 */ 3883 regs[rd] = (uintptr_t)(addr - orig); 3884 break; 3885 } 3886 3887 ASSERT(subr == DIF_SUBR_STRSTR); 3888 regs[rd] = (uintptr_t)addr; 3889 break; 3890 } 3891 } 3892 3893 break; 3894 } 3895 3896 case DIF_SUBR_STRTOK: { 3897 uintptr_t addr = tupregs[0].dttk_value; 3898 uintptr_t tokaddr = tupregs[1].dttk_value; 3899 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE]; 3900 uintptr_t limit, toklimit = tokaddr + size; 3901 uint8_t c = 0, tokmap[32]; /* 256 / 8 */ 3902 char *dest = (char *)mstate->dtms_scratch_ptr; 3903 int i; 3904 3905 /* 3906 * Check both the token buffer and (later) the input buffer, 3907 * since both could be non-scratch addresses. 3908 */ 3909 if (!dtrace_strcanload(tokaddr, size, mstate, vstate)) { 3910 regs[rd] = 0; 3911 break; 3912 } 3913 3914 if (!DTRACE_INSCRATCH(mstate, size)) { 3915 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 3916 regs[rd] = 0; 3917 break; 3918 } 3919 3920 if (addr == 0) { 3921 /* 3922 * If the address specified is NULL, we use our saved 3923 * strtok pointer from the mstate. Note that this 3924 * means that the saved strtok pointer is _only_ 3925 * valid within multiple enablings of the same probe -- 3926 * it behaves like an implicit clause-local variable. 3927 */ 3928 addr = mstate->dtms_strtok; 3929 } else { 3930 /* 3931 * If the user-specified address is non-NULL we must 3932 * access check it. This is the only time we have 3933 * a chance to do so, since this address may reside 3934 * in the string table of this clause-- future calls 3935 * (when we fetch addr from mstate->dtms_strtok) 3936 * would fail this access check. 3937 */ 3938 if (!dtrace_strcanload(addr, size, mstate, vstate)) { 3939 regs[rd] = 0; 3940 break; 3941 } 3942 } 3943 3944 /* 3945 * First, zero the token map, and then process the token 3946 * string -- setting a bit in the map for every character 3947 * found in the token string. 3948 */ 3949 for (i = 0; i < sizeof (tokmap); i++) 3950 tokmap[i] = 0; 3951 3952 for (; tokaddr < toklimit; tokaddr++) { 3953 if ((c = dtrace_load8(tokaddr)) == '\0') 3954 break; 3955 3956 ASSERT((c >> 3) < sizeof (tokmap)); 3957 tokmap[c >> 3] |= (1 << (c & 0x7)); 3958 } 3959 3960 for (limit = addr + size; addr < limit; addr++) { 3961 /* 3962 * We're looking for a character that is _not_ contained 3963 * in the token string. 3964 */ 3965 if ((c = dtrace_load8(addr)) == '\0') 3966 break; 3967 3968 if (!(tokmap[c >> 3] & (1 << (c & 0x7)))) 3969 break; 3970 } 3971 3972 if (c == '\0') { 3973 /* 3974 * We reached the end of the string without finding 3975 * any character that was not in the token string. 3976 * We return NULL in this case, and we set the saved 3977 * address to NULL as well. 3978 */ 3979 regs[rd] = 0; 3980 mstate->dtms_strtok = 0; 3981 break; 3982 } 3983 3984 /* 3985 * From here on, we're copying into the destination string. 3986 */ 3987 for (i = 0; addr < limit && i < size - 1; addr++) { 3988 if ((c = dtrace_load8(addr)) == '\0') 3989 break; 3990 3991 if (tokmap[c >> 3] & (1 << (c & 0x7))) 3992 break; 3993 3994 ASSERT(i < size); 3995 dest[i++] = c; 3996 } 3997 3998 ASSERT(i < size); 3999 dest[i] = '\0'; 4000 regs[rd] = (uintptr_t)dest; 4001 mstate->dtms_scratch_ptr += size; 4002 mstate->dtms_strtok = addr; 4003 break; 4004 } 4005 4006 case DIF_SUBR_SUBSTR: { 4007 uintptr_t s = tupregs[0].dttk_value; 4008 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE]; 4009 char *d = (char *)mstate->dtms_scratch_ptr; 4010 int64_t index = (int64_t)tupregs[1].dttk_value; 4011 int64_t remaining = (int64_t)tupregs[2].dttk_value; 4012 size_t len = dtrace_strlen((char *)s, size); 4013 int64_t i = 0; 4014 4015 if (!dtrace_canload(s, len + 1, mstate, vstate)) { 4016 regs[rd] = 0; 4017 break; 4018 } 4019 4020 if (!DTRACE_INSCRATCH(mstate, size)) { 4021 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 4022 regs[rd] = 0; 4023 break; 4024 } 4025 4026 if (nargs <= 2) 4027 remaining = (int64_t)size; 4028 4029 if (index < 0) { 4030 index += len; 4031 4032 if (index < 0 && index + remaining > 0) { 4033 remaining += index; 4034 index = 0; 4035 } 4036 } 4037 4038 if (index >= len || index < 0) { 4039 remaining = 0; 4040 } else if (remaining < 0) { 4041 remaining += len - index; 4042 } else if (index + remaining > size) { 4043 remaining = size - index; 4044 } 4045 4046 for (i = 0; i < remaining; i++) { 4047 if ((d[i] = dtrace_load8(s + index + i)) == '\0') 4048 break; 4049 } 4050 4051 d[i] = '\0'; 4052 4053 mstate->dtms_scratch_ptr += size; 4054 regs[rd] = (uintptr_t)d; 4055 break; 4056 } 4057 4058 case DIF_SUBR_TOUPPER: 4059 case DIF_SUBR_TOLOWER: { 4060 uintptr_t s = tupregs[0].dttk_value; 4061 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE]; 4062 char *dest = (char *)mstate->dtms_scratch_ptr, c; 4063 size_t len = dtrace_strlen((char *)s, size); 4064 char lower, upper, convert; 4065 int64_t i; 4066 4067 if (subr == DIF_SUBR_TOUPPER) { 4068 lower = 'a'; 4069 upper = 'z'; 4070 convert = 'A'; 4071 } else { 4072 lower = 'A'; 4073 upper = 'Z'; 4074 convert = 'a'; 4075 } 4076 4077 if (!dtrace_canload(s, len + 1, mstate, vstate)) { 4078 regs[rd] = 0; 4079 break; 4080 } 4081 4082 if (!DTRACE_INSCRATCH(mstate, size)) { 4083 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 4084 regs[rd] = 0; 4085 break; 4086 } 4087 4088 for (i = 0; i < size - 1; i++) { 4089 if ((c = dtrace_load8(s + i)) == '\0') 4090 break; 4091 4092 if (c >= lower && c <= upper) 4093 c = convert + (c - lower); 4094 4095 dest[i] = c; 4096 } 4097 4098 ASSERT(i < size); 4099 dest[i] = '\0'; 4100 regs[rd] = (uintptr_t)dest; 4101 mstate->dtms_scratch_ptr += size; 4102 break; 4103 } 4104 4105#if defined(sun) 4106 case DIF_SUBR_GETMAJOR: 4107#ifdef _LP64 4108 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64; 4109#else 4110 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ; 4111#endif 4112 break; 4113 4114 case DIF_SUBR_GETMINOR: 4115#ifdef _LP64 4116 regs[rd] = tupregs[0].dttk_value & MAXMIN64; 4117#else 4118 regs[rd] = tupregs[0].dttk_value & MAXMIN; 4119#endif 4120 break; 4121 4122 case DIF_SUBR_DDI_PATHNAME: { 4123 /* 4124 * This one is a galactic mess. We are going to roughly 4125 * emulate ddi_pathname(), but it's made more complicated 4126 * by the fact that we (a) want to include the minor name and 4127 * (b) must proceed iteratively instead of recursively. 4128 */ 4129 uintptr_t dest = mstate->dtms_scratch_ptr; 4130 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE]; 4131 char *start = (char *)dest, *end = start + size - 1; 4132 uintptr_t daddr = tupregs[0].dttk_value; 4133 int64_t minor = (int64_t)tupregs[1].dttk_value; 4134 char *s; 4135 int i, len, depth = 0; 4136 4137 /* 4138 * Due to all the pointer jumping we do and context we must 4139 * rely upon, we just mandate that the user must have kernel 4140 * read privileges to use this routine. 4141 */ 4142 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) { 4143 *flags |= CPU_DTRACE_KPRIV; 4144 *illval = daddr; 4145 regs[rd] = 0; 4146 } 4147 4148 if (!DTRACE_INSCRATCH(mstate, size)) { 4149 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 4150 regs[rd] = 0; 4151 break; 4152 } 4153 4154 *end = '\0'; 4155 4156 /* 4157 * We want to have a name for the minor. In order to do this, 4158 * we need to walk the minor list from the devinfo. We want 4159 * to be sure that we don't infinitely walk a circular list, 4160 * so we check for circularity by sending a scout pointer 4161 * ahead two elements for every element that we iterate over; 4162 * if the list is circular, these will ultimately point to the 4163 * same element. You may recognize this little trick as the 4164 * answer to a stupid interview question -- one that always 4165 * seems to be asked by those who had to have it laboriously 4166 * explained to them, and who can't even concisely describe 4167 * the conditions under which one would be forced to resort to 4168 * this technique. Needless to say, those conditions are 4169 * found here -- and probably only here. Is this the only use 4170 * of this infamous trick in shipping, production code? If it 4171 * isn't, it probably should be... 4172 */ 4173 if (minor != -1) { 4174 uintptr_t maddr = dtrace_loadptr(daddr + 4175 offsetof(struct dev_info, devi_minor)); 4176 4177 uintptr_t next = offsetof(struct ddi_minor_data, next); 4178 uintptr_t name = offsetof(struct ddi_minor_data, 4179 d_minor) + offsetof(struct ddi_minor, name); 4180 uintptr_t dev = offsetof(struct ddi_minor_data, 4181 d_minor) + offsetof(struct ddi_minor, dev); 4182 uintptr_t scout; 4183 4184 if (maddr != NULL) 4185 scout = dtrace_loadptr(maddr + next); 4186 4187 while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) { 4188 uint64_t m; 4189#ifdef _LP64 4190 m = dtrace_load64(maddr + dev) & MAXMIN64; 4191#else 4192 m = dtrace_load32(maddr + dev) & MAXMIN; 4193#endif 4194 if (m != minor) { 4195 maddr = dtrace_loadptr(maddr + next); 4196 4197 if (scout == NULL) 4198 continue; 4199 4200 scout = dtrace_loadptr(scout + next); 4201 4202 if (scout == NULL) 4203 continue; 4204 4205 scout = dtrace_loadptr(scout + next); 4206 4207 if (scout == NULL) 4208 continue; 4209 4210 if (scout == maddr) { 4211 *flags |= CPU_DTRACE_ILLOP; 4212 break; 4213 } 4214 4215 continue; 4216 } 4217 4218 /* 4219 * We have the minor data. Now we need to 4220 * copy the minor's name into the end of the 4221 * pathname. 4222 */ 4223 s = (char *)dtrace_loadptr(maddr + name); 4224 len = dtrace_strlen(s, size); 4225 4226 if (*flags & CPU_DTRACE_FAULT) 4227 break; 4228 4229 if (len != 0) { 4230 if ((end -= (len + 1)) < start) 4231 break; 4232 4233 *end = ':'; 4234 } 4235 4236 for (i = 1; i <= len; i++) 4237 end[i] = dtrace_load8((uintptr_t)s++); 4238 break; 4239 } 4240 } 4241 4242 while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) { 4243 ddi_node_state_t devi_state; 4244 4245 devi_state = dtrace_load32(daddr + 4246 offsetof(struct dev_info, devi_node_state)); 4247 4248 if (*flags & CPU_DTRACE_FAULT) 4249 break; 4250 4251 if (devi_state >= DS_INITIALIZED) { 4252 s = (char *)dtrace_loadptr(daddr + 4253 offsetof(struct dev_info, devi_addr)); 4254 len = dtrace_strlen(s, size); 4255 4256 if (*flags & CPU_DTRACE_FAULT) 4257 break; 4258 4259 if (len != 0) { 4260 if ((end -= (len + 1)) < start) 4261 break; 4262 4263 *end = '@'; 4264 } 4265 4266 for (i = 1; i <= len; i++) 4267 end[i] = dtrace_load8((uintptr_t)s++); 4268 } 4269 4270 /* 4271 * Now for the node name... 4272 */ 4273 s = (char *)dtrace_loadptr(daddr + 4274 offsetof(struct dev_info, devi_node_name)); 4275 4276 daddr = dtrace_loadptr(daddr + 4277 offsetof(struct dev_info, devi_parent)); 4278 4279 /* 4280 * If our parent is NULL (that is, if we're the root 4281 * node), we're going to use the special path 4282 * "devices". 4283 */ 4284 if (daddr == 0) 4285 s = "devices"; 4286 4287 len = dtrace_strlen(s, size); 4288 if (*flags & CPU_DTRACE_FAULT) 4289 break; 4290 4291 if ((end -= (len + 1)) < start) 4292 break; 4293 4294 for (i = 1; i <= len; i++) 4295 end[i] = dtrace_load8((uintptr_t)s++); 4296 *end = '/'; 4297 4298 if (depth++ > dtrace_devdepth_max) { 4299 *flags |= CPU_DTRACE_ILLOP; 4300 break; 4301 } 4302 } 4303 4304 if (end < start) 4305 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 4306 4307 if (daddr == 0) { 4308 regs[rd] = (uintptr_t)end; 4309 mstate->dtms_scratch_ptr += size; 4310 } 4311 4312 break; 4313 } 4314#endif 4315 4316 case DIF_SUBR_STRJOIN: { 4317 char *d = (char *)mstate->dtms_scratch_ptr; 4318 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE]; 4319 uintptr_t s1 = tupregs[0].dttk_value; 4320 uintptr_t s2 = tupregs[1].dttk_value; 4321 int i = 0; 4322 4323 if (!dtrace_strcanload(s1, size, mstate, vstate) || 4324 !dtrace_strcanload(s2, size, mstate, vstate)) { 4325 regs[rd] = 0; 4326 break; 4327 } 4328 4329 if (!DTRACE_INSCRATCH(mstate, size)) { 4330 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 4331 regs[rd] = 0; 4332 break; 4333 } 4334 4335 for (;;) { 4336 if (i >= size) { 4337 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 4338 regs[rd] = 0; 4339 break; 4340 } 4341 4342 if ((d[i++] = dtrace_load8(s1++)) == '\0') { 4343 i--; 4344 break; 4345 } 4346 } 4347 4348 for (;;) { 4349 if (i >= size) { 4350 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 4351 regs[rd] = 0; 4352 break; 4353 } 4354 4355 if ((d[i++] = dtrace_load8(s2++)) == '\0') 4356 break; 4357 } 4358 4359 if (i < size) { 4360 mstate->dtms_scratch_ptr += i; 4361 regs[rd] = (uintptr_t)d; 4362 } 4363 4364 break; 4365 } 4366 4367 case DIF_SUBR_LLTOSTR: { 4368 int64_t i = (int64_t)tupregs[0].dttk_value; 4369 uint64_t val, digit; 4370 uint64_t size = 65; /* enough room for 2^64 in binary */ 4371 char *end = (char *)mstate->dtms_scratch_ptr + size - 1; 4372 int base = 10; 4373 4374 if (nargs > 1) { 4375 if ((base = tupregs[1].dttk_value) <= 1 || 4376 base > ('z' - 'a' + 1) + ('9' - '0' + 1)) { 4377 *flags |= CPU_DTRACE_ILLOP; 4378 break; 4379 } 4380 } 4381 4382 val = (base == 10 && i < 0) ? i * -1 : i; 4383 4384 if (!DTRACE_INSCRATCH(mstate, size)) { 4385 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 4386 regs[rd] = 0; 4387 break; 4388 } 4389 4390 for (*end-- = '\0'; val; val /= base) { 4391 if ((digit = val % base) <= '9' - '0') { 4392 *end-- = '0' + digit; 4393 } else { 4394 *end-- = 'a' + (digit - ('9' - '0') - 1); 4395 } 4396 } 4397 4398 if (i == 0 && base == 16) 4399 *end-- = '0'; 4400 4401 if (base == 16) 4402 *end-- = 'x'; 4403 4404 if (i == 0 || base == 8 || base == 16) 4405 *end-- = '0'; 4406 4407 if (i < 0 && base == 10) 4408 *end-- = '-'; 4409 4410 regs[rd] = (uintptr_t)end + 1; 4411 mstate->dtms_scratch_ptr += size; 4412 break; 4413 } 4414 4415 case DIF_SUBR_HTONS: 4416 case DIF_SUBR_NTOHS: 4417#if BYTE_ORDER == BIG_ENDIAN 4418 regs[rd] = (uint16_t)tupregs[0].dttk_value; 4419#else 4420 regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value); 4421#endif 4422 break; 4423 4424 4425 case DIF_SUBR_HTONL: 4426 case DIF_SUBR_NTOHL: 4427#if BYTE_ORDER == BIG_ENDIAN 4428 regs[rd] = (uint32_t)tupregs[0].dttk_value; 4429#else 4430 regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value); 4431#endif 4432 break; 4433 4434 4435 case DIF_SUBR_HTONLL: 4436 case DIF_SUBR_NTOHLL: 4437#if BYTE_ORDER == BIG_ENDIAN 4438 regs[rd] = (uint64_t)tupregs[0].dttk_value; 4439#else 4440 regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value); 4441#endif 4442 break; 4443 4444 4445 case DIF_SUBR_DIRNAME: 4446 case DIF_SUBR_BASENAME: { 4447 char *dest = (char *)mstate->dtms_scratch_ptr; 4448 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE]; 4449 uintptr_t src = tupregs[0].dttk_value; 4450 int i, j, len = dtrace_strlen((char *)src, size); 4451 int lastbase = -1, firstbase = -1, lastdir = -1; 4452 int start, end; 4453 4454 if (!dtrace_canload(src, len + 1, mstate, vstate)) { 4455 regs[rd] = 0; 4456 break; 4457 } 4458 4459 if (!DTRACE_INSCRATCH(mstate, size)) { 4460 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 4461 regs[rd] = 0; 4462 break; 4463 } 4464 4465 /* 4466 * The basename and dirname for a zero-length string is 4467 * defined to be "." 4468 */ 4469 if (len == 0) { 4470 len = 1; 4471 src = (uintptr_t)"."; 4472 } 4473 4474 /* 4475 * Start from the back of the string, moving back toward the 4476 * front until we see a character that isn't a slash. That 4477 * character is the last character in the basename. 4478 */ 4479 for (i = len - 1; i >= 0; i--) { 4480 if (dtrace_load8(src + i) != '/') 4481 break; 4482 } 4483 4484 if (i >= 0) 4485 lastbase = i; 4486 4487 /* 4488 * Starting from the last character in the basename, move 4489 * towards the front until we find a slash. The character 4490 * that we processed immediately before that is the first 4491 * character in the basename. 4492 */ 4493 for (; i >= 0; i--) { 4494 if (dtrace_load8(src + i) == '/') 4495 break; 4496 } 4497 4498 if (i >= 0) 4499 firstbase = i + 1; 4500 4501 /* 4502 * Now keep going until we find a non-slash character. That 4503 * character is the last character in the dirname. 4504 */ 4505 for (; i >= 0; i--) { 4506 if (dtrace_load8(src + i) != '/') 4507 break; 4508 } 4509 4510 if (i >= 0) 4511 lastdir = i; 4512 4513 ASSERT(!(lastbase == -1 && firstbase != -1)); 4514 ASSERT(!(firstbase == -1 && lastdir != -1)); 4515 4516 if (lastbase == -1) { 4517 /* 4518 * We didn't find a non-slash character. We know that 4519 * the length is non-zero, so the whole string must be 4520 * slashes. In either the dirname or the basename 4521 * case, we return '/'. 4522 */ 4523 ASSERT(firstbase == -1); 4524 firstbase = lastbase = lastdir = 0; 4525 } 4526 4527 if (firstbase == -1) { 4528 /* 4529 * The entire string consists only of a basename 4530 * component. If we're looking for dirname, we need 4531 * to change our string to be just "."; if we're 4532 * looking for a basename, we'll just set the first 4533 * character of the basename to be 0. 4534 */ 4535 if (subr == DIF_SUBR_DIRNAME) { 4536 ASSERT(lastdir == -1); 4537 src = (uintptr_t)"."; 4538 lastdir = 0; 4539 } else { 4540 firstbase = 0; 4541 } 4542 } 4543 4544 if (subr == DIF_SUBR_DIRNAME) { 4545 if (lastdir == -1) { 4546 /* 4547 * We know that we have a slash in the name -- 4548 * or lastdir would be set to 0, above. And 4549 * because lastdir is -1, we know that this 4550 * slash must be the first character. (That 4551 * is, the full string must be of the form 4552 * "/basename".) In this case, the last 4553 * character of the directory name is 0. 4554 */ 4555 lastdir = 0; 4556 } 4557 4558 start = 0; 4559 end = lastdir; 4560 } else { 4561 ASSERT(subr == DIF_SUBR_BASENAME); 4562 ASSERT(firstbase != -1 && lastbase != -1); 4563 start = firstbase; 4564 end = lastbase; 4565 } 4566 4567 for (i = start, j = 0; i <= end && j < size - 1; i++, j++) 4568 dest[j] = dtrace_load8(src + i); 4569 4570 dest[j] = '\0'; 4571 regs[rd] = (uintptr_t)dest; 4572 mstate->dtms_scratch_ptr += size; 4573 break; 4574 } 4575 4576 case DIF_SUBR_CLEANPATH: { 4577 char *dest = (char *)mstate->dtms_scratch_ptr, c; 4578 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE]; 4579 uintptr_t src = tupregs[0].dttk_value; 4580 int i = 0, j = 0; 4581 4582 if (!dtrace_strcanload(src, size, mstate, vstate)) { 4583 regs[rd] = 0; 4584 break; 4585 } 4586 4587 if (!DTRACE_INSCRATCH(mstate, size)) { 4588 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 4589 regs[rd] = 0; 4590 break; 4591 } 4592 4593 /* 4594 * Move forward, loading each character. 4595 */ 4596 do { 4597 c = dtrace_load8(src + i++); 4598next: 4599 if (j + 5 >= size) /* 5 = strlen("/..c\0") */ 4600 break; 4601 4602 if (c != '/') { 4603 dest[j++] = c; 4604 continue; 4605 } 4606 4607 c = dtrace_load8(src + i++); 4608 4609 if (c == '/') { 4610 /* 4611 * We have two slashes -- we can just advance 4612 * to the next character. 4613 */ 4614 goto next; 4615 } 4616 4617 if (c != '.') { 4618 /* 4619 * This is not "." and it's not ".." -- we can 4620 * just store the "/" and this character and 4621 * drive on. 4622 */ 4623 dest[j++] = '/'; 4624 dest[j++] = c; 4625 continue; 4626 } 4627 4628 c = dtrace_load8(src + i++); 4629 4630 if (c == '/') { 4631 /* 4632 * This is a "/./" component. We're not going 4633 * to store anything in the destination buffer; 4634 * we're just going to go to the next component. 4635 */ 4636 goto next; 4637 } 4638 4639 if (c != '.') { 4640 /* 4641 * This is not ".." -- we can just store the 4642 * "/." and this character and continue 4643 * processing. 4644 */ 4645 dest[j++] = '/'; 4646 dest[j++] = '.'; 4647 dest[j++] = c; 4648 continue; 4649 } 4650 4651 c = dtrace_load8(src + i++); 4652 4653 if (c != '/' && c != '\0') { 4654 /* 4655 * This is not ".." -- it's "..[mumble]". 4656 * We'll store the "/.." and this character 4657 * and continue processing. 4658 */ 4659 dest[j++] = '/'; 4660 dest[j++] = '.'; 4661 dest[j++] = '.'; 4662 dest[j++] = c; 4663 continue; 4664 } 4665 4666 /* 4667 * This is "/../" or "/..\0". We need to back up 4668 * our destination pointer until we find a "/". 4669 */ 4670 i--; 4671 while (j != 0 && dest[--j] != '/') 4672 continue; 4673 4674 if (c == '\0') 4675 dest[++j] = '/'; 4676 } while (c != '\0'); 4677 4678 dest[j] = '\0'; 4679 regs[rd] = (uintptr_t)dest; 4680 mstate->dtms_scratch_ptr += size; 4681 break; 4682 } 4683 4684 case DIF_SUBR_INET_NTOA: 4685 case DIF_SUBR_INET_NTOA6: 4686 case DIF_SUBR_INET_NTOP: { 4687 size_t size; 4688 int af, argi, i; 4689 char *base, *end; 4690 4691 if (subr == DIF_SUBR_INET_NTOP) { 4692 af = (int)tupregs[0].dttk_value; 4693 argi = 1; 4694 } else { 4695 af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6; 4696 argi = 0; 4697 } 4698 4699 if (af == AF_INET) { 4700 ipaddr_t ip4; 4701 uint8_t *ptr8, val; 4702 4703 /* 4704 * Safely load the IPv4 address. 4705 */ 4706 ip4 = dtrace_load32(tupregs[argi].dttk_value); 4707 4708 /* 4709 * Check an IPv4 string will fit in scratch. 4710 */ 4711 size = INET_ADDRSTRLEN; 4712 if (!DTRACE_INSCRATCH(mstate, size)) { 4713 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 4714 regs[rd] = 0; 4715 break; 4716 } 4717 base = (char *)mstate->dtms_scratch_ptr; 4718 end = (char *)mstate->dtms_scratch_ptr + size - 1; 4719 4720 /* 4721 * Stringify as a dotted decimal quad. 4722 */ 4723 *end-- = '\0'; 4724 ptr8 = (uint8_t *)&ip4; 4725 for (i = 3; i >= 0; i--) { 4726 val = ptr8[i]; 4727 4728 if (val == 0) { 4729 *end-- = '0'; 4730 } else { 4731 for (; val; val /= 10) { 4732 *end-- = '0' + (val % 10); 4733 } 4734 } 4735 4736 if (i > 0) 4737 *end-- = '.'; 4738 } 4739 ASSERT(end + 1 >= base); 4740 4741 } else if (af == AF_INET6) { 4742 struct in6_addr ip6; 4743 int firstzero, tryzero, numzero, v6end; 4744 uint16_t val; 4745 const char digits[] = "0123456789abcdef"; 4746 4747 /* 4748 * Stringify using RFC 1884 convention 2 - 16 bit 4749 * hexadecimal values with a zero-run compression. 4750 * Lower case hexadecimal digits are used. 4751 * eg, fe80::214:4fff:fe0b:76c8. 4752 * The IPv4 embedded form is returned for inet_ntop, 4753 * just the IPv4 string is returned for inet_ntoa6. 4754 */ 4755 4756 /* 4757 * Safely load the IPv6 address. 4758 */ 4759 dtrace_bcopy( 4760 (void *)(uintptr_t)tupregs[argi].dttk_value, 4761 (void *)(uintptr_t)&ip6, sizeof (struct in6_addr)); 4762 4763 /* 4764 * Check an IPv6 string will fit in scratch. 4765 */ 4766 size = INET6_ADDRSTRLEN; 4767 if (!DTRACE_INSCRATCH(mstate, size)) { 4768 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 4769 regs[rd] = 0; 4770 break; 4771 } 4772 base = (char *)mstate->dtms_scratch_ptr; 4773 end = (char *)mstate->dtms_scratch_ptr + size - 1; 4774 *end-- = '\0'; 4775 4776 /* 4777 * Find the longest run of 16 bit zero values 4778 * for the single allowed zero compression - "::". 4779 */ 4780 firstzero = -1; 4781 tryzero = -1; 4782 numzero = 1; 4783 for (i = 0; i < sizeof (struct in6_addr); i++) { 4784#if defined(sun) 4785 if (ip6._S6_un._S6_u8[i] == 0 && 4786#else 4787 if (ip6.__u6_addr.__u6_addr8[i] == 0 && 4788#endif 4789 tryzero == -1 && i % 2 == 0) { 4790 tryzero = i; 4791 continue; 4792 } 4793 4794 if (tryzero != -1 && 4795#if defined(sun) 4796 (ip6._S6_un._S6_u8[i] != 0 || 4797#else 4798 (ip6.__u6_addr.__u6_addr8[i] != 0 || 4799#endif 4800 i == sizeof (struct in6_addr) - 1)) { 4801 4802 if (i - tryzero <= numzero) { 4803 tryzero = -1; 4804 continue; 4805 } 4806 4807 firstzero = tryzero; 4808 numzero = i - i % 2 - tryzero; 4809 tryzero = -1; 4810 4811#if defined(sun) 4812 if (ip6._S6_un._S6_u8[i] == 0 && 4813#else 4814 if (ip6.__u6_addr.__u6_addr8[i] == 0 && 4815#endif 4816 i == sizeof (struct in6_addr) - 1) 4817 numzero += 2; 4818 } 4819 } 4820 ASSERT(firstzero + numzero <= sizeof (struct in6_addr)); 4821 4822 /* 4823 * Check for an IPv4 embedded address. 4824 */ 4825 v6end = sizeof (struct in6_addr) - 2; 4826 if (IN6_IS_ADDR_V4MAPPED(&ip6) || 4827 IN6_IS_ADDR_V4COMPAT(&ip6)) { 4828 for (i = sizeof (struct in6_addr) - 1; 4829 i >= DTRACE_V4MAPPED_OFFSET; i--) { 4830 ASSERT(end >= base); 4831 4832#if defined(sun) 4833 val = ip6._S6_un._S6_u8[i]; 4834#else 4835 val = ip6.__u6_addr.__u6_addr8[i]; 4836#endif 4837 4838 if (val == 0) { 4839 *end-- = '0'; 4840 } else { 4841 for (; val; val /= 10) { 4842 *end-- = '0' + val % 10; 4843 } 4844 } 4845 4846 if (i > DTRACE_V4MAPPED_OFFSET) 4847 *end-- = '.'; 4848 } 4849 4850 if (subr == DIF_SUBR_INET_NTOA6) 4851 goto inetout; 4852 4853 /* 4854 * Set v6end to skip the IPv4 address that 4855 * we have already stringified. 4856 */ 4857 v6end = 10; 4858 } 4859 4860 /* 4861 * Build the IPv6 string by working through the 4862 * address in reverse. 4863 */ 4864 for (i = v6end; i >= 0; i -= 2) { 4865 ASSERT(end >= base); 4866 4867 if (i == firstzero + numzero - 2) { 4868 *end-- = ':'; 4869 *end-- = ':'; 4870 i -= numzero - 2; 4871 continue; 4872 } 4873 4874 if (i < 14 && i != firstzero - 2) 4875 *end-- = ':'; 4876 4877#if defined(sun) 4878 val = (ip6._S6_un._S6_u8[i] << 8) + 4879 ip6._S6_un._S6_u8[i + 1]; 4880#else 4881 val = (ip6.__u6_addr.__u6_addr8[i] << 8) + 4882 ip6.__u6_addr.__u6_addr8[i + 1]; 4883#endif 4884 4885 if (val == 0) { 4886 *end-- = '0'; 4887 } else { 4888 for (; val; val /= 16) { 4889 *end-- = digits[val % 16]; 4890 } 4891 } 4892 } 4893 ASSERT(end + 1 >= base); 4894 4895 } else { 4896 /* 4897 * The user didn't use AH_INET or AH_INET6. 4898 */ 4899 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP); 4900 regs[rd] = 0; 4901 break; 4902 } 4903 4904inetout: regs[rd] = (uintptr_t)end + 1; 4905 mstate->dtms_scratch_ptr += size; 4906 break; 4907 } 4908 4909 case DIF_SUBR_MEMREF: { 4910 uintptr_t size = 2 * sizeof(uintptr_t); 4911 uintptr_t *memref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t)); 4912 size_t scratch_size = ((uintptr_t) memref - mstate->dtms_scratch_ptr) + size; 4913 4914 /* address and length */ 4915 memref[0] = tupregs[0].dttk_value; 4916 memref[1] = tupregs[1].dttk_value; 4917 4918 regs[rd] = (uintptr_t) memref; 4919 mstate->dtms_scratch_ptr += scratch_size; 4920 break; 4921 } 4922 4923 case DIF_SUBR_TYPEREF: { 4924 uintptr_t size = 4 * sizeof(uintptr_t); 4925 uintptr_t *typeref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t)); 4926 size_t scratch_size = ((uintptr_t) typeref - mstate->dtms_scratch_ptr) + size; 4927 4928 /* address, num_elements, type_str, type_len */ 4929 typeref[0] = tupregs[0].dttk_value; 4930 typeref[1] = tupregs[1].dttk_value; 4931 typeref[2] = tupregs[2].dttk_value; 4932 typeref[3] = tupregs[3].dttk_value; 4933 4934 regs[rd] = (uintptr_t) typeref; 4935 mstate->dtms_scratch_ptr += scratch_size; 4936 break; 4937 } 4938 } 4939} 4940 4941/* 4942 * Emulate the execution of DTrace IR instructions specified by the given 4943 * DIF object. This function is deliberately void of assertions as all of 4944 * the necessary checks are handled by a call to dtrace_difo_validate(). 4945 */ 4946static uint64_t 4947dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate, 4948 dtrace_vstate_t *vstate, dtrace_state_t *state) 4949{ 4950 const dif_instr_t *text = difo->dtdo_buf; 4951 const uint_t textlen = difo->dtdo_len; 4952 const char *strtab = difo->dtdo_strtab; 4953 const uint64_t *inttab = difo->dtdo_inttab; 4954 4955 uint64_t rval = 0; 4956 dtrace_statvar_t *svar; 4957 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars; 4958 dtrace_difv_t *v; 4959 volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags; 4960 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval; 4961 4962 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */ 4963 uint64_t regs[DIF_DIR_NREGS]; 4964 uint64_t *tmp; 4965 4966 uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0; 4967 int64_t cc_r; 4968 uint_t pc = 0, id, opc = 0; 4969 uint8_t ttop = 0; 4970 dif_instr_t instr; 4971 uint_t r1, r2, rd; 4972 4973 /* 4974 * We stash the current DIF object into the machine state: we need it 4975 * for subsequent access checking. 4976 */ 4977 mstate->dtms_difo = difo; 4978 4979 regs[DIF_REG_R0] = 0; /* %r0 is fixed at zero */ 4980 4981 while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) { 4982 opc = pc; 4983 4984 instr = text[pc++]; 4985 r1 = DIF_INSTR_R1(instr); 4986 r2 = DIF_INSTR_R2(instr); 4987 rd = DIF_INSTR_RD(instr); 4988 4989 switch (DIF_INSTR_OP(instr)) { 4990 case DIF_OP_OR: 4991 regs[rd] = regs[r1] | regs[r2]; 4992 break; 4993 case DIF_OP_XOR: 4994 regs[rd] = regs[r1] ^ regs[r2]; 4995 break; 4996 case DIF_OP_AND: 4997 regs[rd] = regs[r1] & regs[r2]; 4998 break; 4999 case DIF_OP_SLL: 5000 regs[rd] = regs[r1] << regs[r2]; 5001 break; 5002 case DIF_OP_SRL: 5003 regs[rd] = regs[r1] >> regs[r2]; 5004 break; 5005 case DIF_OP_SUB: 5006 regs[rd] = regs[r1] - regs[r2]; 5007 break; 5008 case DIF_OP_ADD: 5009 regs[rd] = regs[r1] + regs[r2]; 5010 break; 5011 case DIF_OP_MUL: 5012 regs[rd] = regs[r1] * regs[r2]; 5013 break; 5014 case DIF_OP_SDIV: 5015 if (regs[r2] == 0) { 5016 regs[rd] = 0; 5017 *flags |= CPU_DTRACE_DIVZERO; 5018 } else { 5019 regs[rd] = (int64_t)regs[r1] / 5020 (int64_t)regs[r2]; 5021 } 5022 break; 5023 5024 case DIF_OP_UDIV: 5025 if (regs[r2] == 0) { 5026 regs[rd] = 0; 5027 *flags |= CPU_DTRACE_DIVZERO; 5028 } else { 5029 regs[rd] = regs[r1] / regs[r2]; 5030 } 5031 break; 5032 5033 case DIF_OP_SREM: 5034 if (regs[r2] == 0) { 5035 regs[rd] = 0; 5036 *flags |= CPU_DTRACE_DIVZERO; 5037 } else { 5038 regs[rd] = (int64_t)regs[r1] % 5039 (int64_t)regs[r2]; 5040 } 5041 break; 5042 5043 case DIF_OP_UREM: 5044 if (regs[r2] == 0) { 5045 regs[rd] = 0; 5046 *flags |= CPU_DTRACE_DIVZERO; 5047 } else { 5048 regs[rd] = regs[r1] % regs[r2]; 5049 } 5050 break; 5051 5052 case DIF_OP_NOT: 5053 regs[rd] = ~regs[r1]; 5054 break; 5055 case DIF_OP_MOV: 5056 regs[rd] = regs[r1]; 5057 break; 5058 case DIF_OP_CMP: 5059 cc_r = regs[r1] - regs[r2]; 5060 cc_n = cc_r < 0; 5061 cc_z = cc_r == 0; 5062 cc_v = 0; 5063 cc_c = regs[r1] < regs[r2]; 5064 break; 5065 case DIF_OP_TST: 5066 cc_n = cc_v = cc_c = 0; 5067 cc_z = regs[r1] == 0; 5068 break; 5069 case DIF_OP_BA: 5070 pc = DIF_INSTR_LABEL(instr); 5071 break; 5072 case DIF_OP_BE: 5073 if (cc_z) 5074 pc = DIF_INSTR_LABEL(instr); 5075 break; 5076 case DIF_OP_BNE: 5077 if (cc_z == 0) 5078 pc = DIF_INSTR_LABEL(instr); 5079 break; 5080 case DIF_OP_BG: 5081 if ((cc_z | (cc_n ^ cc_v)) == 0) 5082 pc = DIF_INSTR_LABEL(instr); 5083 break; 5084 case DIF_OP_BGU: 5085 if ((cc_c | cc_z) == 0) 5086 pc = DIF_INSTR_LABEL(instr); 5087 break; 5088 case DIF_OP_BGE: 5089 if ((cc_n ^ cc_v) == 0) 5090 pc = DIF_INSTR_LABEL(instr); 5091 break; 5092 case DIF_OP_BGEU: 5093 if (cc_c == 0) 5094 pc = DIF_INSTR_LABEL(instr); 5095 break; 5096 case DIF_OP_BL: 5097 if (cc_n ^ cc_v) 5098 pc = DIF_INSTR_LABEL(instr); 5099 break; 5100 case DIF_OP_BLU: 5101 if (cc_c) 5102 pc = DIF_INSTR_LABEL(instr); 5103 break; 5104 case DIF_OP_BLE: 5105 if (cc_z | (cc_n ^ cc_v)) 5106 pc = DIF_INSTR_LABEL(instr); 5107 break; 5108 case DIF_OP_BLEU: 5109 if (cc_c | cc_z) 5110 pc = DIF_INSTR_LABEL(instr); 5111 break; 5112 case DIF_OP_RLDSB: 5113 if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) { 5114 *flags |= CPU_DTRACE_KPRIV; 5115 *illval = regs[r1]; 5116 break; 5117 } 5118 /*FALLTHROUGH*/ 5119 case DIF_OP_LDSB: 5120 regs[rd] = (int8_t)dtrace_load8(regs[r1]); 5121 break; 5122 case DIF_OP_RLDSH: 5123 if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) { 5124 *flags |= CPU_DTRACE_KPRIV; 5125 *illval = regs[r1]; 5126 break; 5127 } 5128 /*FALLTHROUGH*/ 5129 case DIF_OP_LDSH: 5130 regs[rd] = (int16_t)dtrace_load16(regs[r1]); 5131 break; 5132 case DIF_OP_RLDSW: 5133 if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) { 5134 *flags |= CPU_DTRACE_KPRIV; 5135 *illval = regs[r1]; 5136 break; 5137 } 5138 /*FALLTHROUGH*/ 5139 case DIF_OP_LDSW: 5140 regs[rd] = (int32_t)dtrace_load32(regs[r1]); 5141 break; 5142 case DIF_OP_RLDUB: 5143 if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) { 5144 *flags |= CPU_DTRACE_KPRIV; 5145 *illval = regs[r1]; 5146 break; 5147 } 5148 /*FALLTHROUGH*/ 5149 case DIF_OP_LDUB: 5150 regs[rd] = dtrace_load8(regs[r1]); 5151 break; 5152 case DIF_OP_RLDUH: 5153 if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) { 5154 *flags |= CPU_DTRACE_KPRIV; 5155 *illval = regs[r1]; 5156 break; 5157 } 5158 /*FALLTHROUGH*/ 5159 case DIF_OP_LDUH: 5160 regs[rd] = dtrace_load16(regs[r1]); 5161 break; 5162 case DIF_OP_RLDUW: 5163 if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) { 5164 *flags |= CPU_DTRACE_KPRIV; 5165 *illval = regs[r1]; 5166 break; 5167 } 5168 /*FALLTHROUGH*/ 5169 case DIF_OP_LDUW: 5170 regs[rd] = dtrace_load32(regs[r1]); 5171 break; 5172 case DIF_OP_RLDX: 5173 if (!dtrace_canstore(regs[r1], 8, mstate, vstate)) { 5174 *flags |= CPU_DTRACE_KPRIV; 5175 *illval = regs[r1]; 5176 break; 5177 } 5178 /*FALLTHROUGH*/ 5179 case DIF_OP_LDX: 5180 regs[rd] = dtrace_load64(regs[r1]); 5181 break; 5182 case DIF_OP_ULDSB: 5183 regs[rd] = (int8_t) 5184 dtrace_fuword8((void *)(uintptr_t)regs[r1]); 5185 break; 5186 case DIF_OP_ULDSH: 5187 regs[rd] = (int16_t) 5188 dtrace_fuword16((void *)(uintptr_t)regs[r1]); 5189 break; 5190 case DIF_OP_ULDSW: 5191 regs[rd] = (int32_t) 5192 dtrace_fuword32((void *)(uintptr_t)regs[r1]); 5193 break; 5194 case DIF_OP_ULDUB: 5195 regs[rd] = 5196 dtrace_fuword8((void *)(uintptr_t)regs[r1]); 5197 break; 5198 case DIF_OP_ULDUH: 5199 regs[rd] = 5200 dtrace_fuword16((void *)(uintptr_t)regs[r1]); 5201 break; 5202 case DIF_OP_ULDUW: 5203 regs[rd] = 5204 dtrace_fuword32((void *)(uintptr_t)regs[r1]); 5205 break; 5206 case DIF_OP_ULDX: 5207 regs[rd] = 5208 dtrace_fuword64((void *)(uintptr_t)regs[r1]); 5209 break; 5210 case DIF_OP_RET: 5211 rval = regs[rd]; 5212 pc = textlen; 5213 break; 5214 case DIF_OP_NOP: 5215 break; 5216 case DIF_OP_SETX: 5217 regs[rd] = inttab[DIF_INSTR_INTEGER(instr)]; 5218 break; 5219 case DIF_OP_SETS: 5220 regs[rd] = (uint64_t)(uintptr_t) 5221 (strtab + DIF_INSTR_STRING(instr)); 5222 break; 5223 case DIF_OP_SCMP: { 5224 size_t sz = state->dts_options[DTRACEOPT_STRSIZE]; 5225 uintptr_t s1 = regs[r1]; 5226 uintptr_t s2 = regs[r2]; 5227 5228 if (s1 != 0 && 5229 !dtrace_strcanload(s1, sz, mstate, vstate)) 5230 break; 5231 if (s2 != 0 && 5232 !dtrace_strcanload(s2, sz, mstate, vstate)) 5233 break; 5234 5235 cc_r = dtrace_strncmp((char *)s1, (char *)s2, sz); 5236 5237 cc_n = cc_r < 0; 5238 cc_z = cc_r == 0; 5239 cc_v = cc_c = 0; 5240 break; 5241 } 5242 case DIF_OP_LDGA: 5243 regs[rd] = dtrace_dif_variable(mstate, state, 5244 r1, regs[r2]); 5245 break; 5246 case DIF_OP_LDGS: 5247 id = DIF_INSTR_VAR(instr); 5248 5249 if (id >= DIF_VAR_OTHER_UBASE) { 5250 uintptr_t a; 5251 5252 id -= DIF_VAR_OTHER_UBASE; 5253 svar = vstate->dtvs_globals[id]; 5254 ASSERT(svar != NULL); 5255 v = &svar->dtsv_var; 5256 5257 if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) { 5258 regs[rd] = svar->dtsv_data; 5259 break; 5260 } 5261 5262 a = (uintptr_t)svar->dtsv_data; 5263 5264 if (*(uint8_t *)a == UINT8_MAX) { 5265 /* 5266 * If the 0th byte is set to UINT8_MAX 5267 * then this is to be treated as a 5268 * reference to a NULL variable. 5269 */ 5270 regs[rd] = 0; 5271 } else { 5272 regs[rd] = a + sizeof (uint64_t); 5273 } 5274 5275 break; 5276 } 5277 5278 regs[rd] = dtrace_dif_variable(mstate, state, id, 0); 5279 break; 5280 5281 case DIF_OP_STGS: 5282 id = DIF_INSTR_VAR(instr); 5283 5284 ASSERT(id >= DIF_VAR_OTHER_UBASE); 5285 id -= DIF_VAR_OTHER_UBASE; 5286 5287 svar = vstate->dtvs_globals[id]; 5288 ASSERT(svar != NULL); 5289 v = &svar->dtsv_var; 5290 5291 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) { 5292 uintptr_t a = (uintptr_t)svar->dtsv_data; 5293 5294 ASSERT(a != 0); 5295 ASSERT(svar->dtsv_size != 0); 5296 5297 if (regs[rd] == 0) { 5298 *(uint8_t *)a = UINT8_MAX; 5299 break; 5300 } else { 5301 *(uint8_t *)a = 0; 5302 a += sizeof (uint64_t); 5303 } 5304 if (!dtrace_vcanload( 5305 (void *)(uintptr_t)regs[rd], &v->dtdv_type, 5306 mstate, vstate)) 5307 break; 5308 5309 dtrace_vcopy((void *)(uintptr_t)regs[rd], 5310 (void *)a, &v->dtdv_type); 5311 break; 5312 } 5313 5314 svar->dtsv_data = regs[rd]; 5315 break; 5316 5317 case DIF_OP_LDTA: 5318 /* 5319 * There are no DTrace built-in thread-local arrays at 5320 * present. This opcode is saved for future work. 5321 */ 5322 *flags |= CPU_DTRACE_ILLOP; 5323 regs[rd] = 0; 5324 break; 5325 5326 case DIF_OP_LDLS: 5327 id = DIF_INSTR_VAR(instr); 5328 5329 if (id < DIF_VAR_OTHER_UBASE) { 5330 /* 5331 * For now, this has no meaning. 5332 */ 5333 regs[rd] = 0; 5334 break; 5335 } 5336 5337 id -= DIF_VAR_OTHER_UBASE; 5338 5339 ASSERT(id < vstate->dtvs_nlocals); 5340 ASSERT(vstate->dtvs_locals != NULL); 5341 5342 svar = vstate->dtvs_locals[id]; 5343 ASSERT(svar != NULL); 5344 v = &svar->dtsv_var; 5345 5346 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) { 5347 uintptr_t a = (uintptr_t)svar->dtsv_data; 5348 size_t sz = v->dtdv_type.dtdt_size; 5349 5350 sz += sizeof (uint64_t); 5351 ASSERT(svar->dtsv_size == NCPU * sz); 5352 a += curcpu * sz; 5353 5354 if (*(uint8_t *)a == UINT8_MAX) { 5355 /* 5356 * If the 0th byte is set to UINT8_MAX 5357 * then this is to be treated as a 5358 * reference to a NULL variable. 5359 */ 5360 regs[rd] = 0; 5361 } else { 5362 regs[rd] = a + sizeof (uint64_t); 5363 } 5364 5365 break; 5366 } 5367 5368 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t)); 5369 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data; 5370 regs[rd] = tmp[curcpu]; 5371 break; 5372 5373 case DIF_OP_STLS: 5374 id = DIF_INSTR_VAR(instr); 5375 5376 ASSERT(id >= DIF_VAR_OTHER_UBASE); 5377 id -= DIF_VAR_OTHER_UBASE; 5378 ASSERT(id < vstate->dtvs_nlocals); 5379 5380 ASSERT(vstate->dtvs_locals != NULL); 5381 svar = vstate->dtvs_locals[id]; 5382 ASSERT(svar != NULL); 5383 v = &svar->dtsv_var; 5384 5385 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) { 5386 uintptr_t a = (uintptr_t)svar->dtsv_data; 5387 size_t sz = v->dtdv_type.dtdt_size; 5388 5389 sz += sizeof (uint64_t); 5390 ASSERT(svar->dtsv_size == NCPU * sz); 5391 a += curcpu * sz; 5392 5393 if (regs[rd] == 0) { 5394 *(uint8_t *)a = UINT8_MAX; 5395 break; 5396 } else { 5397 *(uint8_t *)a = 0; 5398 a += sizeof (uint64_t); 5399 } 5400 5401 if (!dtrace_vcanload( 5402 (void *)(uintptr_t)regs[rd], &v->dtdv_type, 5403 mstate, vstate)) 5404 break; 5405 5406 dtrace_vcopy((void *)(uintptr_t)regs[rd], 5407 (void *)a, &v->dtdv_type); 5408 break; 5409 } 5410 5411 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t)); 5412 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data; 5413 tmp[curcpu] = regs[rd]; 5414 break; 5415 5416 case DIF_OP_LDTS: { 5417 dtrace_dynvar_t *dvar; 5418 dtrace_key_t *key; 5419 5420 id = DIF_INSTR_VAR(instr); 5421 ASSERT(id >= DIF_VAR_OTHER_UBASE); 5422 id -= DIF_VAR_OTHER_UBASE; 5423 v = &vstate->dtvs_tlocals[id]; 5424 5425 key = &tupregs[DIF_DTR_NREGS]; 5426 key[0].dttk_value = (uint64_t)id; 5427 key[0].dttk_size = 0; 5428 DTRACE_TLS_THRKEY(key[1].dttk_value); 5429 key[1].dttk_size = 0; 5430 5431 dvar = dtrace_dynvar(dstate, 2, key, 5432 sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC, 5433 mstate, vstate); 5434 5435 if (dvar == NULL) { 5436 regs[rd] = 0; 5437 break; 5438 } 5439 5440 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) { 5441 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data; 5442 } else { 5443 regs[rd] = *((uint64_t *)dvar->dtdv_data); 5444 } 5445 5446 break; 5447 } 5448 5449 case DIF_OP_STTS: { 5450 dtrace_dynvar_t *dvar; 5451 dtrace_key_t *key; 5452 5453 id = DIF_INSTR_VAR(instr); 5454 ASSERT(id >= DIF_VAR_OTHER_UBASE); 5455 id -= DIF_VAR_OTHER_UBASE; 5456 5457 key = &tupregs[DIF_DTR_NREGS]; 5458 key[0].dttk_value = (uint64_t)id; 5459 key[0].dttk_size = 0; 5460 DTRACE_TLS_THRKEY(key[1].dttk_value); 5461 key[1].dttk_size = 0; 5462 v = &vstate->dtvs_tlocals[id]; 5463 5464 dvar = dtrace_dynvar(dstate, 2, key, 5465 v->dtdv_type.dtdt_size > sizeof (uint64_t) ? 5466 v->dtdv_type.dtdt_size : sizeof (uint64_t), 5467 regs[rd] ? DTRACE_DYNVAR_ALLOC : 5468 DTRACE_DYNVAR_DEALLOC, mstate, vstate); 5469 5470 /* 5471 * Given that we're storing to thread-local data, 5472 * we need to flush our predicate cache. 5473 */ 5474 curthread->t_predcache = 0; 5475 5476 if (dvar == NULL) 5477 break; 5478 5479 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) { 5480 if (!dtrace_vcanload( 5481 (void *)(uintptr_t)regs[rd], 5482 &v->dtdv_type, mstate, vstate)) 5483 break; 5484 5485 dtrace_vcopy((void *)(uintptr_t)regs[rd], 5486 dvar->dtdv_data, &v->dtdv_type); 5487 } else { 5488 *((uint64_t *)dvar->dtdv_data) = regs[rd]; 5489 } 5490 5491 break; 5492 } 5493 5494 case DIF_OP_SRA: 5495 regs[rd] = (int64_t)regs[r1] >> regs[r2]; 5496 break; 5497 5498 case DIF_OP_CALL: 5499 dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd, 5500 regs, tupregs, ttop, mstate, state); 5501 break; 5502 5503 case DIF_OP_PUSHTR: 5504 if (ttop == DIF_DTR_NREGS) { 5505 *flags |= CPU_DTRACE_TUPOFLOW; 5506 break; 5507 } 5508 5509 if (r1 == DIF_TYPE_STRING) { 5510 /* 5511 * If this is a string type and the size is 0, 5512 * we'll use the system-wide default string 5513 * size. Note that we are _not_ looking at 5514 * the value of the DTRACEOPT_STRSIZE option; 5515 * had this been set, we would expect to have 5516 * a non-zero size value in the "pushtr". 5517 */ 5518 tupregs[ttop].dttk_size = 5519 dtrace_strlen((char *)(uintptr_t)regs[rd], 5520 regs[r2] ? regs[r2] : 5521 dtrace_strsize_default) + 1; 5522 } else { 5523 tupregs[ttop].dttk_size = regs[r2]; 5524 } 5525 5526 tupregs[ttop++].dttk_value = regs[rd]; 5527 break; 5528 5529 case DIF_OP_PUSHTV: 5530 if (ttop == DIF_DTR_NREGS) { 5531 *flags |= CPU_DTRACE_TUPOFLOW; 5532 break; 5533 } 5534 5535 tupregs[ttop].dttk_value = regs[rd]; 5536 tupregs[ttop++].dttk_size = 0; 5537 break; 5538 5539 case DIF_OP_POPTS: 5540 if (ttop != 0) 5541 ttop--; 5542 break; 5543 5544 case DIF_OP_FLUSHTS: 5545 ttop = 0; 5546 break; 5547 5548 case DIF_OP_LDGAA: 5549 case DIF_OP_LDTAA: { 5550 dtrace_dynvar_t *dvar; 5551 dtrace_key_t *key = tupregs; 5552 uint_t nkeys = ttop; 5553 5554 id = DIF_INSTR_VAR(instr); 5555 ASSERT(id >= DIF_VAR_OTHER_UBASE); 5556 id -= DIF_VAR_OTHER_UBASE; 5557 5558 key[nkeys].dttk_value = (uint64_t)id; 5559 key[nkeys++].dttk_size = 0; 5560 5561 if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) { 5562 DTRACE_TLS_THRKEY(key[nkeys].dttk_value); 5563 key[nkeys++].dttk_size = 0; 5564 v = &vstate->dtvs_tlocals[id]; 5565 } else { 5566 v = &vstate->dtvs_globals[id]->dtsv_var; 5567 } 5568 5569 dvar = dtrace_dynvar(dstate, nkeys, key, 5570 v->dtdv_type.dtdt_size > sizeof (uint64_t) ? 5571 v->dtdv_type.dtdt_size : sizeof (uint64_t), 5572 DTRACE_DYNVAR_NOALLOC, mstate, vstate); 5573 5574 if (dvar == NULL) { 5575 regs[rd] = 0; 5576 break; 5577 } 5578 5579 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) { 5580 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data; 5581 } else { 5582 regs[rd] = *((uint64_t *)dvar->dtdv_data); 5583 } 5584 5585 break; 5586 } 5587 5588 case DIF_OP_STGAA: 5589 case DIF_OP_STTAA: { 5590 dtrace_dynvar_t *dvar; 5591 dtrace_key_t *key = tupregs; 5592 uint_t nkeys = ttop; 5593 5594 id = DIF_INSTR_VAR(instr); 5595 ASSERT(id >= DIF_VAR_OTHER_UBASE); 5596 id -= DIF_VAR_OTHER_UBASE; 5597 5598 key[nkeys].dttk_value = (uint64_t)id; 5599 key[nkeys++].dttk_size = 0; 5600 5601 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) { 5602 DTRACE_TLS_THRKEY(key[nkeys].dttk_value); 5603 key[nkeys++].dttk_size = 0; 5604 v = &vstate->dtvs_tlocals[id]; 5605 } else { 5606 v = &vstate->dtvs_globals[id]->dtsv_var; 5607 } 5608 5609 dvar = dtrace_dynvar(dstate, nkeys, key, 5610 v->dtdv_type.dtdt_size > sizeof (uint64_t) ? 5611 v->dtdv_type.dtdt_size : sizeof (uint64_t), 5612 regs[rd] ? DTRACE_DYNVAR_ALLOC : 5613 DTRACE_DYNVAR_DEALLOC, mstate, vstate); 5614 5615 if (dvar == NULL) 5616 break; 5617 5618 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) { 5619 if (!dtrace_vcanload( 5620 (void *)(uintptr_t)regs[rd], &v->dtdv_type, 5621 mstate, vstate)) 5622 break; 5623 5624 dtrace_vcopy((void *)(uintptr_t)regs[rd], 5625 dvar->dtdv_data, &v->dtdv_type); 5626 } else { 5627 *((uint64_t *)dvar->dtdv_data) = regs[rd]; 5628 } 5629 5630 break; 5631 } 5632 5633 case DIF_OP_ALLOCS: { 5634 uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8); 5635 size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1]; 5636 5637 /* 5638 * Rounding up the user allocation size could have 5639 * overflowed large, bogus allocations (like -1ULL) to 5640 * 0. 5641 */ 5642 if (size < regs[r1] || 5643 !DTRACE_INSCRATCH(mstate, size)) { 5644 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 5645 regs[rd] = 0; 5646 break; 5647 } 5648 5649 dtrace_bzero((void *) mstate->dtms_scratch_ptr, size); 5650 mstate->dtms_scratch_ptr += size; 5651 regs[rd] = ptr; 5652 break; 5653 } 5654 5655 case DIF_OP_COPYS: 5656 if (!dtrace_canstore(regs[rd], regs[r2], 5657 mstate, vstate)) { 5658 *flags |= CPU_DTRACE_BADADDR; 5659 *illval = regs[rd]; 5660 break; 5661 } 5662 5663 if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate)) 5664 break; 5665 5666 dtrace_bcopy((void *)(uintptr_t)regs[r1], 5667 (void *)(uintptr_t)regs[rd], (size_t)regs[r2]); 5668 break; 5669 5670 case DIF_OP_STB: 5671 if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) { 5672 *flags |= CPU_DTRACE_BADADDR; 5673 *illval = regs[rd]; 5674 break; 5675 } 5676 *((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1]; 5677 break; 5678 5679 case DIF_OP_STH: 5680 if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) { 5681 *flags |= CPU_DTRACE_BADADDR; 5682 *illval = regs[rd]; 5683 break; 5684 } 5685 if (regs[rd] & 1) { 5686 *flags |= CPU_DTRACE_BADALIGN; 5687 *illval = regs[rd]; 5688 break; 5689 } 5690 *((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1]; 5691 break; 5692 5693 case DIF_OP_STW: 5694 if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) { 5695 *flags |= CPU_DTRACE_BADADDR; 5696 *illval = regs[rd]; 5697 break; 5698 } 5699 if (regs[rd] & 3) { 5700 *flags |= CPU_DTRACE_BADALIGN; 5701 *illval = regs[rd]; 5702 break; 5703 } 5704 *((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1]; 5705 break; 5706 5707 case DIF_OP_STX: 5708 if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) { 5709 *flags |= CPU_DTRACE_BADADDR; 5710 *illval = regs[rd]; 5711 break; 5712 } 5713 if (regs[rd] & 7) { 5714 *flags |= CPU_DTRACE_BADALIGN; 5715 *illval = regs[rd]; 5716 break; 5717 } 5718 *((uint64_t *)(uintptr_t)regs[rd]) = regs[r1]; 5719 break; 5720 } 5721 } 5722 5723 if (!(*flags & CPU_DTRACE_FAULT)) 5724 return (rval); 5725 5726 mstate->dtms_fltoffs = opc * sizeof (dif_instr_t); 5727 mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS; 5728 5729 return (0); 5730} 5731 5732static void 5733dtrace_action_breakpoint(dtrace_ecb_t *ecb) 5734{ 5735 dtrace_probe_t *probe = ecb->dte_probe; 5736 dtrace_provider_t *prov = probe->dtpr_provider; 5737 char c[DTRACE_FULLNAMELEN + 80], *str; 5738 char *msg = "dtrace: breakpoint action at probe "; 5739 char *ecbmsg = " (ecb "; 5740 uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4)); 5741 uintptr_t val = (uintptr_t)ecb; 5742 int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0; 5743 5744 if (dtrace_destructive_disallow) 5745 return; 5746 5747 /* 5748 * It's impossible to be taking action on the NULL probe. 5749 */ 5750 ASSERT(probe != NULL); 5751 5752 /* 5753 * This is a poor man's (destitute man's?) sprintf(): we want to 5754 * print the provider name, module name, function name and name of 5755 * the probe, along with the hex address of the ECB with the breakpoint 5756 * action -- all of which we must place in the character buffer by 5757 * hand. 5758 */ 5759 while (*msg != '\0') 5760 c[i++] = *msg++; 5761 5762 for (str = prov->dtpv_name; *str != '\0'; str++) 5763 c[i++] = *str; 5764 c[i++] = ':'; 5765 5766 for (str = probe->dtpr_mod; *str != '\0'; str++) 5767 c[i++] = *str; 5768 c[i++] = ':'; 5769 5770 for (str = probe->dtpr_func; *str != '\0'; str++) 5771 c[i++] = *str; 5772 c[i++] = ':'; 5773 5774 for (str = probe->dtpr_name; *str != '\0'; str++) 5775 c[i++] = *str; 5776 5777 while (*ecbmsg != '\0') 5778 c[i++] = *ecbmsg++; 5779 5780 while (shift >= 0) { 5781 mask = (uintptr_t)0xf << shift; 5782 5783 if (val >= ((uintptr_t)1 << shift)) 5784 c[i++] = "0123456789abcdef"[(val & mask) >> shift]; 5785 shift -= 4; 5786 } 5787 5788 c[i++] = ')'; 5789 c[i] = '\0'; 5790 5791#if defined(sun) 5792 debug_enter(c); 5793#else 5794 kdb_enter(KDB_WHY_DTRACE, "breakpoint action"); 5795#endif 5796} 5797 5798static void 5799dtrace_action_panic(dtrace_ecb_t *ecb) 5800{ 5801 dtrace_probe_t *probe = ecb->dte_probe; 5802 5803 /* 5804 * It's impossible to be taking action on the NULL probe. 5805 */ 5806 ASSERT(probe != NULL); 5807 5808 if (dtrace_destructive_disallow) 5809 return; 5810 5811 if (dtrace_panicked != NULL) 5812 return; 5813 5814 if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL) 5815 return; 5816 5817 /* 5818 * We won the right to panic. (We want to be sure that only one 5819 * thread calls panic() from dtrace_probe(), and that panic() is 5820 * called exactly once.) 5821 */ 5822 dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)", 5823 probe->dtpr_provider->dtpv_name, probe->dtpr_mod, 5824 probe->dtpr_func, probe->dtpr_name, (void *)ecb); 5825} 5826 5827static void 5828dtrace_action_raise(uint64_t sig) 5829{ 5830 if (dtrace_destructive_disallow) 5831 return; 5832 5833 if (sig >= NSIG) { 5834 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP); 5835 return; 5836 } 5837 5838#if defined(sun) 5839 /* 5840 * raise() has a queue depth of 1 -- we ignore all subsequent 5841 * invocations of the raise() action. 5842 */ 5843 if (curthread->t_dtrace_sig == 0) 5844 curthread->t_dtrace_sig = (uint8_t)sig; 5845 5846 curthread->t_sig_check = 1; 5847 aston(curthread); 5848#else 5849 struct proc *p = curproc; 5850 PROC_LOCK(p); 5851 kern_psignal(p, sig); 5852 PROC_UNLOCK(p); 5853#endif 5854} 5855 5856static void 5857dtrace_action_stop(void) 5858{ 5859 if (dtrace_destructive_disallow) 5860 return; 5861 5862#if defined(sun) 5863 if (!curthread->t_dtrace_stop) { 5864 curthread->t_dtrace_stop = 1; 5865 curthread->t_sig_check = 1; 5866 aston(curthread); 5867 } 5868#else 5869 struct proc *p = curproc; 5870 PROC_LOCK(p); 5871 kern_psignal(p, SIGSTOP); 5872 PROC_UNLOCK(p); 5873#endif 5874} 5875 5876static void 5877dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val) 5878{ 5879 hrtime_t now; 5880 volatile uint16_t *flags; 5881#if defined(sun) 5882 cpu_t *cpu = CPU; 5883#else 5884 cpu_t *cpu = &solaris_cpu[curcpu]; 5885#endif 5886 5887 if (dtrace_destructive_disallow) 5888 return; 5889 5890 flags = (volatile uint16_t *)&cpu_core[cpu->cpu_id].cpuc_dtrace_flags; 5891 5892 now = dtrace_gethrtime(); 5893 5894 if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) { 5895 /* 5896 * We need to advance the mark to the current time. 5897 */ 5898 cpu->cpu_dtrace_chillmark = now; 5899 cpu->cpu_dtrace_chilled = 0; 5900 } 5901 5902 /* 5903 * Now check to see if the requested chill time would take us over 5904 * the maximum amount of time allowed in the chill interval. (Or 5905 * worse, if the calculation itself induces overflow.) 5906 */ 5907 if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max || 5908 cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) { 5909 *flags |= CPU_DTRACE_ILLOP; 5910 return; 5911 } 5912 5913 while (dtrace_gethrtime() - now < val) 5914 continue; 5915 5916 /* 5917 * Normally, we assure that the value of the variable "timestamp" does 5918 * not change within an ECB. The presence of chill() represents an 5919 * exception to this rule, however. 5920 */ 5921 mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP; 5922 cpu->cpu_dtrace_chilled += val; 5923} 5924 5925static void 5926dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state, 5927 uint64_t *buf, uint64_t arg) 5928{ 5929 int nframes = DTRACE_USTACK_NFRAMES(arg); 5930 int strsize = DTRACE_USTACK_STRSIZE(arg); 5931 uint64_t *pcs = &buf[1], *fps; 5932 char *str = (char *)&pcs[nframes]; 5933 int size, offs = 0, i, j; 5934 uintptr_t old = mstate->dtms_scratch_ptr, saved; 5935 uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags; 5936 char *sym; 5937 5938 /* 5939 * Should be taking a faster path if string space has not been 5940 * allocated. 5941 */ 5942 ASSERT(strsize != 0); 5943 5944 /* 5945 * We will first allocate some temporary space for the frame pointers. 5946 */ 5947 fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8); 5948 size = (uintptr_t)fps - mstate->dtms_scratch_ptr + 5949 (nframes * sizeof (uint64_t)); 5950 5951 if (!DTRACE_INSCRATCH(mstate, size)) { 5952 /* 5953 * Not enough room for our frame pointers -- need to indicate 5954 * that we ran out of scratch space. 5955 */ 5956 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH); 5957 return; 5958 } 5959 5960 mstate->dtms_scratch_ptr += size; 5961 saved = mstate->dtms_scratch_ptr; 5962 5963 /* 5964 * Now get a stack with both program counters and frame pointers. 5965 */ 5966 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT); 5967 dtrace_getufpstack(buf, fps, nframes + 1); 5968 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT); 5969 5970 /* 5971 * If that faulted, we're cooked. 5972 */ 5973 if (*flags & CPU_DTRACE_FAULT) 5974 goto out; 5975 5976 /* 5977 * Now we want to walk up the stack, calling the USTACK helper. For 5978 * each iteration, we restore the scratch pointer. 5979 */ 5980 for (i = 0; i < nframes; i++) { 5981 mstate->dtms_scratch_ptr = saved; 5982 5983 if (offs >= strsize) 5984 break; 5985 5986 sym = (char *)(uintptr_t)dtrace_helper( 5987 DTRACE_HELPER_ACTION_USTACK, 5988 mstate, state, pcs[i], fps[i]); 5989 5990 /* 5991 * If we faulted while running the helper, we're going to 5992 * clear the fault and null out the corresponding string. 5993 */ 5994 if (*flags & CPU_DTRACE_FAULT) { 5995 *flags &= ~CPU_DTRACE_FAULT; 5996 str[offs++] = '\0'; 5997 continue; 5998 } 5999 6000 if (sym == NULL) { 6001 str[offs++] = '\0'; 6002 continue; 6003 } 6004 6005 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT); 6006 6007 /* 6008 * Now copy in the string that the helper returned to us. 6009 */ 6010 for (j = 0; offs + j < strsize; j++) { 6011 if ((str[offs + j] = sym[j]) == '\0') 6012 break; 6013 } 6014 6015 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT); 6016 6017 offs += j + 1; 6018 } 6019 6020 if (offs >= strsize) { 6021 /* 6022 * If we didn't have room for all of the strings, we don't 6023 * abort processing -- this needn't be a fatal error -- but we 6024 * still want to increment a counter (dts_stkstroverflows) to 6025 * allow this condition to be warned about. (If this is from 6026 * a jstack() action, it is easily tuned via jstackstrsize.) 6027 */ 6028 dtrace_error(&state->dts_stkstroverflows); 6029 } 6030 6031 while (offs < strsize) 6032 str[offs++] = '\0'; 6033 6034out: 6035 mstate->dtms_scratch_ptr = old; 6036} 6037 6038/* 6039 * If you're looking for the epicenter of DTrace, you just found it. This 6040 * is the function called by the provider to fire a probe -- from which all 6041 * subsequent probe-context DTrace activity emanates. 6042 */ 6043void 6044dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1, 6045 uintptr_t arg2, uintptr_t arg3, uintptr_t arg4) 6046{ 6047 processorid_t cpuid; 6048 dtrace_icookie_t cookie; 6049 dtrace_probe_t *probe; 6050 dtrace_mstate_t mstate; 6051 dtrace_ecb_t *ecb; 6052 dtrace_action_t *act; 6053 intptr_t offs; 6054 size_t size; 6055 int vtime, onintr; 6056 volatile uint16_t *flags; 6057 hrtime_t now; 6058 6059 if (panicstr != NULL) 6060 return; 6061 6062#if defined(sun) 6063 /* 6064 * Kick out immediately if this CPU is still being born (in which case 6065 * curthread will be set to -1) or the current thread can't allow 6066 * probes in its current context. 6067 */ 6068 if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE)) 6069 return; 6070#endif 6071 6072 cookie = dtrace_interrupt_disable(); 6073 probe = dtrace_probes[id - 1]; 6074 cpuid = curcpu; 6075 onintr = CPU_ON_INTR(CPU); 6076 6077 if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE && 6078 probe->dtpr_predcache == curthread->t_predcache) { 6079 /* 6080 * We have hit in the predicate cache; we know that 6081 * this predicate would evaluate to be false. 6082 */ 6083 dtrace_interrupt_enable(cookie); 6084 return; 6085 } 6086 6087#if defined(sun) 6088 if (panic_quiesce) { 6089#else 6090 if (panicstr != NULL) { 6091#endif 6092 /* 6093 * We don't trace anything if we're panicking. 6094 */ 6095 dtrace_interrupt_enable(cookie); 6096 return; 6097 } 6098 6099 now = dtrace_gethrtime(); 6100 vtime = dtrace_vtime_references != 0; 6101 6102 if (vtime && curthread->t_dtrace_start) 6103 curthread->t_dtrace_vtime += now - curthread->t_dtrace_start; 6104 6105 mstate.dtms_difo = NULL; 6106 mstate.dtms_probe = probe; 6107 mstate.dtms_strtok = 0; 6108 mstate.dtms_arg[0] = arg0; 6109 mstate.dtms_arg[1] = arg1; 6110 mstate.dtms_arg[2] = arg2; 6111 mstate.dtms_arg[3] = arg3; 6112 mstate.dtms_arg[4] = arg4; 6113 6114 flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags; 6115 6116 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) { 6117 dtrace_predicate_t *pred = ecb->dte_predicate; 6118 dtrace_state_t *state = ecb->dte_state; 6119 dtrace_buffer_t *buf = &state->dts_buffer[cpuid]; 6120 dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid]; 6121 dtrace_vstate_t *vstate = &state->dts_vstate; 6122 dtrace_provider_t *prov = probe->dtpr_provider; 6123 uint64_t tracememsize = 0; 6124 int committed = 0; 6125 caddr_t tomax; 6126 6127 /* 6128 * A little subtlety with the following (seemingly innocuous) 6129 * declaration of the automatic 'val': by looking at the 6130 * code, you might think that it could be declared in the 6131 * action processing loop, below. (That is, it's only used in 6132 * the action processing loop.) However, it must be declared 6133 * out of that scope because in the case of DIF expression 6134 * arguments to aggregating actions, one iteration of the 6135 * action loop will use the last iteration's value. 6136 */ 6137 uint64_t val = 0; 6138 6139 mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE; 6140 *flags &= ~CPU_DTRACE_ERROR; 6141 6142 if (prov == dtrace_provider) { 6143 /* 6144 * If dtrace itself is the provider of this probe, 6145 * we're only going to continue processing the ECB if 6146 * arg0 (the dtrace_state_t) is equal to the ECB's 6147 * creating state. (This prevents disjoint consumers 6148 * from seeing one another's metaprobes.) 6149 */ 6150 if (arg0 != (uint64_t)(uintptr_t)state) 6151 continue; 6152 } 6153 6154 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) { 6155 /* 6156 * We're not currently active. If our provider isn't 6157 * the dtrace pseudo provider, we're not interested. 6158 */ 6159 if (prov != dtrace_provider) 6160 continue; 6161 6162 /* 6163 * Now we must further check if we are in the BEGIN 6164 * probe. If we are, we will only continue processing 6165 * if we're still in WARMUP -- if one BEGIN enabling 6166 * has invoked the exit() action, we don't want to 6167 * evaluate subsequent BEGIN enablings. 6168 */ 6169 if (probe->dtpr_id == dtrace_probeid_begin && 6170 state->dts_activity != DTRACE_ACTIVITY_WARMUP) { 6171 ASSERT(state->dts_activity == 6172 DTRACE_ACTIVITY_DRAINING); 6173 continue; 6174 } 6175 } 6176 6177 if (ecb->dte_cond) { 6178 /* 6179 * If the dte_cond bits indicate that this 6180 * consumer is only allowed to see user-mode firings 6181 * of this probe, call the provider's dtps_usermode() 6182 * entry point to check that the probe was fired 6183 * while in a user context. Skip this ECB if that's 6184 * not the case. 6185 */ 6186 if ((ecb->dte_cond & DTRACE_COND_USERMODE) && 6187 prov->dtpv_pops.dtps_usermode(prov->dtpv_arg, 6188 probe->dtpr_id, probe->dtpr_arg) == 0) 6189 continue; 6190 6191#if defined(sun) 6192 /* 6193 * This is more subtle than it looks. We have to be 6194 * absolutely certain that CRED() isn't going to 6195 * change out from under us so it's only legit to 6196 * examine that structure if we're in constrained 6197 * situations. Currently, the only times we'll this 6198 * check is if a non-super-user has enabled the 6199 * profile or syscall providers -- providers that 6200 * allow visibility of all processes. For the 6201 * profile case, the check above will ensure that 6202 * we're examining a user context. 6203 */ 6204 if (ecb->dte_cond & DTRACE_COND_OWNER) { 6205 cred_t *cr; 6206 cred_t *s_cr = 6207 ecb->dte_state->dts_cred.dcr_cred; 6208 proc_t *proc; 6209 6210 ASSERT(s_cr != NULL); 6211 6212 if ((cr = CRED()) == NULL || 6213 s_cr->cr_uid != cr->cr_uid || 6214 s_cr->cr_uid != cr->cr_ruid || 6215 s_cr->cr_uid != cr->cr_suid || 6216 s_cr->cr_gid != cr->cr_gid || 6217 s_cr->cr_gid != cr->cr_rgid || 6218 s_cr->cr_gid != cr->cr_sgid || 6219 (proc = ttoproc(curthread)) == NULL || 6220 (proc->p_flag & SNOCD)) 6221 continue; 6222 } 6223 6224 if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) { 6225 cred_t *cr; 6226 cred_t *s_cr = 6227 ecb->dte_state->dts_cred.dcr_cred; 6228 6229 ASSERT(s_cr != NULL); 6230 6231 if ((cr = CRED()) == NULL || 6232 s_cr->cr_zone->zone_id != 6233 cr->cr_zone->zone_id) 6234 continue; 6235 } 6236#endif 6237 } 6238 6239 if (now - state->dts_alive > dtrace_deadman_timeout) { 6240 /* 6241 * We seem to be dead. Unless we (a) have kernel 6242 * destructive permissions (b) have explicitly enabled 6243 * destructive actions and (c) destructive actions have 6244 * not been disabled, we're going to transition into 6245 * the KILLED state, from which no further processing 6246 * on this state will be performed. 6247 */ 6248 if (!dtrace_priv_kernel_destructive(state) || 6249 !state->dts_cred.dcr_destructive || 6250 dtrace_destructive_disallow) { 6251 void *activity = &state->dts_activity; 6252 dtrace_activity_t current; 6253 6254 do { 6255 current = state->dts_activity; 6256 } while (dtrace_cas32(activity, current, 6257 DTRACE_ACTIVITY_KILLED) != current); 6258 6259 continue; 6260 } 6261 } 6262 6263 if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed, 6264 ecb->dte_alignment, state, &mstate)) < 0) 6265 continue; 6266 6267 tomax = buf->dtb_tomax; 6268 ASSERT(tomax != NULL); 6269 6270 if (ecb->dte_size != 0) { 6271 dtrace_rechdr_t dtrh; 6272 if (!(mstate.dtms_present & DTRACE_MSTATE_TIMESTAMP)) { 6273 mstate.dtms_timestamp = dtrace_gethrtime(); 6274 mstate.dtms_present |= DTRACE_MSTATE_TIMESTAMP; 6275 } 6276 ASSERT3U(ecb->dte_size, >=, sizeof (dtrace_rechdr_t)); 6277 dtrh.dtrh_epid = ecb->dte_epid; 6278 DTRACE_RECORD_STORE_TIMESTAMP(&dtrh, 6279 mstate.dtms_timestamp); 6280 *((dtrace_rechdr_t *)(tomax + offs)) = dtrh; 6281 } 6282 6283 mstate.dtms_epid = ecb->dte_epid; 6284 mstate.dtms_present |= DTRACE_MSTATE_EPID; 6285 6286 if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) 6287 mstate.dtms_access = DTRACE_ACCESS_KERNEL; 6288 else 6289 mstate.dtms_access = 0; 6290 6291 if (pred != NULL) { 6292 dtrace_difo_t *dp = pred->dtp_difo; 6293 int rval; 6294 6295 rval = dtrace_dif_emulate(dp, &mstate, vstate, state); 6296 6297 if (!(*flags & CPU_DTRACE_ERROR) && !rval) { 6298 dtrace_cacheid_t cid = probe->dtpr_predcache; 6299 6300 if (cid != DTRACE_CACHEIDNONE && !onintr) { 6301 /* 6302 * Update the predicate cache... 6303 */ 6304 ASSERT(cid == pred->dtp_cacheid); 6305 curthread->t_predcache = cid; 6306 } 6307 6308 continue; 6309 } 6310 } 6311 6312 for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) && 6313 act != NULL; act = act->dta_next) { 6314 size_t valoffs; 6315 dtrace_difo_t *dp; 6316 dtrace_recdesc_t *rec = &act->dta_rec; 6317 6318 size = rec->dtrd_size; 6319 valoffs = offs + rec->dtrd_offset; 6320 6321 if (DTRACEACT_ISAGG(act->dta_kind)) { 6322 uint64_t v = 0xbad; 6323 dtrace_aggregation_t *agg; 6324 6325 agg = (dtrace_aggregation_t *)act; 6326 6327 if ((dp = act->dta_difo) != NULL) 6328 v = dtrace_dif_emulate(dp, 6329 &mstate, vstate, state); 6330 6331 if (*flags & CPU_DTRACE_ERROR) 6332 continue; 6333 6334 /* 6335 * Note that we always pass the expression 6336 * value from the previous iteration of the 6337 * action loop. This value will only be used 6338 * if there is an expression argument to the 6339 * aggregating action, denoted by the 6340 * dtag_hasarg field. 6341 */ 6342 dtrace_aggregate(agg, buf, 6343 offs, aggbuf, v, val); 6344 continue; 6345 } 6346 6347 switch (act->dta_kind) { 6348 case DTRACEACT_STOP: 6349 if (dtrace_priv_proc_destructive(state)) 6350 dtrace_action_stop(); 6351 continue; 6352 6353 case DTRACEACT_BREAKPOINT: 6354 if (dtrace_priv_kernel_destructive(state)) 6355 dtrace_action_breakpoint(ecb); 6356 continue; 6357 6358 case DTRACEACT_PANIC: 6359 if (dtrace_priv_kernel_destructive(state)) 6360 dtrace_action_panic(ecb); 6361 continue; 6362 6363 case DTRACEACT_STACK: 6364 if (!dtrace_priv_kernel(state)) 6365 continue; 6366 6367 dtrace_getpcstack((pc_t *)(tomax + valoffs), 6368 size / sizeof (pc_t), probe->dtpr_aframes, 6369 DTRACE_ANCHORED(probe) ? NULL : 6370 (uint32_t *)arg0); 6371 continue; 6372 6373 case DTRACEACT_JSTACK: 6374 case DTRACEACT_USTACK: 6375 if (!dtrace_priv_proc(state)) 6376 continue; 6377 6378 /* 6379 * See comment in DIF_VAR_PID. 6380 */ 6381 if (DTRACE_ANCHORED(mstate.dtms_probe) && 6382 CPU_ON_INTR(CPU)) { 6383 int depth = DTRACE_USTACK_NFRAMES( 6384 rec->dtrd_arg) + 1; 6385 6386 dtrace_bzero((void *)(tomax + valoffs), 6387 DTRACE_USTACK_STRSIZE(rec->dtrd_arg) 6388 + depth * sizeof (uint64_t)); 6389 6390 continue; 6391 } 6392 6393 if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 && 6394 curproc->p_dtrace_helpers != NULL) { 6395 /* 6396 * This is the slow path -- we have 6397 * allocated string space, and we're 6398 * getting the stack of a process that 6399 * has helpers. Call into a separate 6400 * routine to perform this processing. 6401 */ 6402 dtrace_action_ustack(&mstate, state, 6403 (uint64_t *)(tomax + valoffs), 6404 rec->dtrd_arg); 6405 continue; 6406 } 6407 6408 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT); 6409 dtrace_getupcstack((uint64_t *) 6410 (tomax + valoffs), 6411 DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1); 6412 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT); 6413 continue; 6414 6415 default: 6416 break; 6417 } 6418 6419 dp = act->dta_difo; 6420 ASSERT(dp != NULL); 6421 6422 val = dtrace_dif_emulate(dp, &mstate, vstate, state); 6423 6424 if (*flags & CPU_DTRACE_ERROR) 6425 continue; 6426 6427 switch (act->dta_kind) { 6428 case DTRACEACT_SPECULATE: { 6429 dtrace_rechdr_t *dtrh; 6430 6431 ASSERT(buf == &state->dts_buffer[cpuid]); 6432 buf = dtrace_speculation_buffer(state, 6433 cpuid, val); 6434 6435 if (buf == NULL) { 6436 *flags |= CPU_DTRACE_DROP; 6437 continue; 6438 } 6439 6440 offs = dtrace_buffer_reserve(buf, 6441 ecb->dte_needed, ecb->dte_alignment, 6442 state, NULL); 6443 6444 if (offs < 0) { 6445 *flags |= CPU_DTRACE_DROP; 6446 continue; 6447 } 6448 6449 tomax = buf->dtb_tomax; 6450 ASSERT(tomax != NULL); 6451 6452 if (ecb->dte_size == 0) 6453 continue; 6454 6455 ASSERT3U(ecb->dte_size, >=, 6456 sizeof (dtrace_rechdr_t)); 6457 dtrh = ((void *)(tomax + offs)); 6458 dtrh->dtrh_epid = ecb->dte_epid; 6459 /* 6460 * When the speculation is committed, all of 6461 * the records in the speculative buffer will 6462 * have their timestamps set to the commit 6463 * time. Until then, it is set to a sentinel 6464 * value, for debugability. 6465 */ 6466 DTRACE_RECORD_STORE_TIMESTAMP(dtrh, UINT64_MAX); 6467 continue; 6468 } 6469 6470 case DTRACEACT_PRINTM: { 6471 /* The DIF returns a 'memref'. */ 6472 uintptr_t *memref = (uintptr_t *)(uintptr_t) val; 6473 6474 /* Get the size from the memref. */ 6475 size = memref[1]; 6476 6477 /* 6478 * Check if the size exceeds the allocated 6479 * buffer size. 6480 */ 6481 if (size + sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) { 6482 /* Flag a drop! */ 6483 *flags |= CPU_DTRACE_DROP; 6484 continue; 6485 } 6486 6487 /* Store the size in the buffer first. */ 6488 DTRACE_STORE(uintptr_t, tomax, 6489 valoffs, size); 6490 6491 /* 6492 * Offset the buffer address to the start 6493 * of the data. 6494 */ 6495 valoffs += sizeof(uintptr_t); 6496 6497 /* 6498 * Reset to the memory address rather than 6499 * the memref array, then let the BYREF 6500 * code below do the work to store the 6501 * memory data in the buffer. 6502 */ 6503 val = memref[0]; 6504 break; 6505 } 6506 6507 case DTRACEACT_PRINTT: { 6508 /* The DIF returns a 'typeref'. */ 6509 uintptr_t *typeref = (uintptr_t *)(uintptr_t) val; 6510 char c = '\0' + 1; 6511 size_t s; 6512 6513 /* 6514 * Get the type string length and round it 6515 * up so that the data that follows is 6516 * aligned for easy access. 6517 */ 6518 size_t typs = strlen((char *) typeref[2]) + 1; 6519 typs = roundup(typs, sizeof(uintptr_t)); 6520 6521 /* 6522 *Get the size from the typeref using the 6523 * number of elements and the type size. 6524 */ 6525 size = typeref[1] * typeref[3]; 6526 6527 /* 6528 * Check if the size exceeds the allocated 6529 * buffer size. 6530 */ 6531 if (size + typs + 2 * sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) { 6532 /* Flag a drop! */ 6533 *flags |= CPU_DTRACE_DROP; 6534 6535 } 6536 6537 /* Store the size in the buffer first. */ 6538 DTRACE_STORE(uintptr_t, tomax, 6539 valoffs, size); 6540 valoffs += sizeof(uintptr_t); 6541 6542 /* Store the type size in the buffer. */ 6543 DTRACE_STORE(uintptr_t, tomax, 6544 valoffs, typeref[3]); 6545 valoffs += sizeof(uintptr_t); 6546 6547 val = typeref[2]; 6548 6549 for (s = 0; s < typs; s++) { 6550 if (c != '\0') 6551 c = dtrace_load8(val++); 6552 6553 DTRACE_STORE(uint8_t, tomax, 6554 valoffs++, c); 6555 } 6556 6557 /* 6558 * Reset to the memory address rather than 6559 * the typeref array, then let the BYREF 6560 * code below do the work to store the 6561 * memory data in the buffer. 6562 */ 6563 val = typeref[0]; 6564 break; 6565 } 6566 6567 case DTRACEACT_CHILL: 6568 if (dtrace_priv_kernel_destructive(state)) 6569 dtrace_action_chill(&mstate, val); 6570 continue; 6571 6572 case DTRACEACT_RAISE: 6573 if (dtrace_priv_proc_destructive(state)) 6574 dtrace_action_raise(val); 6575 continue; 6576 6577 case DTRACEACT_COMMIT: 6578 ASSERT(!committed); 6579 6580 /* 6581 * We need to commit our buffer state. 6582 */ 6583 if (ecb->dte_size) 6584 buf->dtb_offset = offs + ecb->dte_size; 6585 buf = &state->dts_buffer[cpuid]; 6586 dtrace_speculation_commit(state, cpuid, val); 6587 committed = 1; 6588 continue; 6589 6590 case DTRACEACT_DISCARD: 6591 dtrace_speculation_discard(state, cpuid, val); 6592 continue; 6593 6594 case DTRACEACT_DIFEXPR: 6595 case DTRACEACT_LIBACT: 6596 case DTRACEACT_PRINTF: 6597 case DTRACEACT_PRINTA: 6598 case DTRACEACT_SYSTEM: 6599 case DTRACEACT_FREOPEN: 6600 case DTRACEACT_TRACEMEM: 6601 break; 6602 6603 case DTRACEACT_TRACEMEM_DYNSIZE: 6604 tracememsize = val; 6605 break; 6606 6607 case DTRACEACT_SYM: 6608 case DTRACEACT_MOD: 6609 if (!dtrace_priv_kernel(state)) 6610 continue; 6611 break; 6612 6613 case DTRACEACT_USYM: 6614 case DTRACEACT_UMOD: 6615 case DTRACEACT_UADDR: { 6616#if defined(sun) 6617 struct pid *pid = curthread->t_procp->p_pidp; 6618#endif 6619 6620 if (!dtrace_priv_proc(state)) 6621 continue; 6622 6623 DTRACE_STORE(uint64_t, tomax, 6624#if defined(sun) 6625 valoffs, (uint64_t)pid->pid_id); 6626#else 6627 valoffs, (uint64_t) curproc->p_pid); 6628#endif 6629 DTRACE_STORE(uint64_t, tomax, 6630 valoffs + sizeof (uint64_t), val); 6631 6632 continue; 6633 } 6634 6635 case DTRACEACT_EXIT: { 6636 /* 6637 * For the exit action, we are going to attempt 6638 * to atomically set our activity to be 6639 * draining. If this fails (either because 6640 * another CPU has beat us to the exit action, 6641 * or because our current activity is something 6642 * other than ACTIVE or WARMUP), we will 6643 * continue. This assures that the exit action 6644 * can be successfully recorded at most once 6645 * when we're in the ACTIVE state. If we're 6646 * encountering the exit() action while in 6647 * COOLDOWN, however, we want to honor the new 6648 * status code. (We know that we're the only 6649 * thread in COOLDOWN, so there is no race.) 6650 */ 6651 void *activity = &state->dts_activity; 6652 dtrace_activity_t current = state->dts_activity; 6653 6654 if (current == DTRACE_ACTIVITY_COOLDOWN) 6655 break; 6656 6657 if (current != DTRACE_ACTIVITY_WARMUP) 6658 current = DTRACE_ACTIVITY_ACTIVE; 6659 6660 if (dtrace_cas32(activity, current, 6661 DTRACE_ACTIVITY_DRAINING) != current) { 6662 *flags |= CPU_DTRACE_DROP; 6663 continue; 6664 } 6665 6666 break; 6667 } 6668 6669 default: 6670 ASSERT(0); 6671 } 6672 6673 if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF) { 6674 uintptr_t end = valoffs + size; 6675 6676 if (tracememsize != 0 && 6677 valoffs + tracememsize < end) { 6678 end = valoffs + tracememsize; 6679 tracememsize = 0; 6680 } 6681 6682 if (!dtrace_vcanload((void *)(uintptr_t)val, 6683 &dp->dtdo_rtype, &mstate, vstate)) 6684 continue; 6685 6686 /* 6687 * If this is a string, we're going to only 6688 * load until we find the zero byte -- after 6689 * which we'll store zero bytes. 6690 */ 6691 if (dp->dtdo_rtype.dtdt_kind == 6692 DIF_TYPE_STRING) { 6693 char c = '\0' + 1; 6694 int intuple = act->dta_intuple; 6695 size_t s; 6696 6697 for (s = 0; s < size; s++) { 6698 if (c != '\0') 6699 c = dtrace_load8(val++); 6700 6701 DTRACE_STORE(uint8_t, tomax, 6702 valoffs++, c); 6703 6704 if (c == '\0' && intuple) 6705 break; 6706 } 6707 6708 continue; 6709 } 6710 6711 while (valoffs < end) { 6712 DTRACE_STORE(uint8_t, tomax, valoffs++, 6713 dtrace_load8(val++)); 6714 } 6715 6716 continue; 6717 } 6718 6719 switch (size) { 6720 case 0: 6721 break; 6722 6723 case sizeof (uint8_t): 6724 DTRACE_STORE(uint8_t, tomax, valoffs, val); 6725 break; 6726 case sizeof (uint16_t): 6727 DTRACE_STORE(uint16_t, tomax, valoffs, val); 6728 break; 6729 case sizeof (uint32_t): 6730 DTRACE_STORE(uint32_t, tomax, valoffs, val); 6731 break; 6732 case sizeof (uint64_t): 6733 DTRACE_STORE(uint64_t, tomax, valoffs, val); 6734 break; 6735 default: 6736 /* 6737 * Any other size should have been returned by 6738 * reference, not by value. 6739 */ 6740 ASSERT(0); 6741 break; 6742 } 6743 } 6744 6745 if (*flags & CPU_DTRACE_DROP) 6746 continue; 6747 6748 if (*flags & CPU_DTRACE_FAULT) { 6749 int ndx; 6750 dtrace_action_t *err; 6751 6752 buf->dtb_errors++; 6753 6754 if (probe->dtpr_id == dtrace_probeid_error) { 6755 /* 6756 * There's nothing we can do -- we had an 6757 * error on the error probe. We bump an 6758 * error counter to at least indicate that 6759 * this condition happened. 6760 */ 6761 dtrace_error(&state->dts_dblerrors); 6762 continue; 6763 } 6764 6765 if (vtime) { 6766 /* 6767 * Before recursing on dtrace_probe(), we 6768 * need to explicitly clear out our start 6769 * time to prevent it from being accumulated 6770 * into t_dtrace_vtime. 6771 */ 6772 curthread->t_dtrace_start = 0; 6773 } 6774 6775 /* 6776 * Iterate over the actions to figure out which action 6777 * we were processing when we experienced the error. 6778 * Note that act points _past_ the faulting action; if 6779 * act is ecb->dte_action, the fault was in the 6780 * predicate, if it's ecb->dte_action->dta_next it's 6781 * in action #1, and so on. 6782 */ 6783 for (err = ecb->dte_action, ndx = 0; 6784 err != act; err = err->dta_next, ndx++) 6785 continue; 6786 6787 dtrace_probe_error(state, ecb->dte_epid, ndx, 6788 (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ? 6789 mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags), 6790 cpu_core[cpuid].cpuc_dtrace_illval); 6791 6792 continue; 6793 } 6794 6795 if (!committed) 6796 buf->dtb_offset = offs + ecb->dte_size; 6797 } 6798 6799 if (vtime) 6800 curthread->t_dtrace_start = dtrace_gethrtime(); 6801 6802 dtrace_interrupt_enable(cookie); 6803} 6804 6805/* 6806 * DTrace Probe Hashing Functions 6807 * 6808 * The functions in this section (and indeed, the functions in remaining 6809 * sections) are not _called_ from probe context. (Any exceptions to this are 6810 * marked with a "Note:".) Rather, they are called from elsewhere in the 6811 * DTrace framework to look-up probes in, add probes to and remove probes from 6812 * the DTrace probe hashes. (Each probe is hashed by each element of the 6813 * probe tuple -- allowing for fast lookups, regardless of what was 6814 * specified.) 6815 */ 6816static uint_t 6817dtrace_hash_str(const char *p) 6818{ 6819 unsigned int g; 6820 uint_t hval = 0; 6821 6822 while (*p) { 6823 hval = (hval << 4) + *p++; 6824 if ((g = (hval & 0xf0000000)) != 0) 6825 hval ^= g >> 24; 6826 hval &= ~g; 6827 } 6828 return (hval); 6829} 6830 6831static dtrace_hash_t * 6832dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs) 6833{ 6834 dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP); 6835 6836 hash->dth_stroffs = stroffs; 6837 hash->dth_nextoffs = nextoffs; 6838 hash->dth_prevoffs = prevoffs; 6839 6840 hash->dth_size = 1; 6841 hash->dth_mask = hash->dth_size - 1; 6842 6843 hash->dth_tab = kmem_zalloc(hash->dth_size * 6844 sizeof (dtrace_hashbucket_t *), KM_SLEEP); 6845 6846 return (hash); 6847} 6848 6849static void 6850dtrace_hash_destroy(dtrace_hash_t *hash) 6851{ 6852#ifdef DEBUG 6853 int i; 6854 6855 for (i = 0; i < hash->dth_size; i++) 6856 ASSERT(hash->dth_tab[i] == NULL); 6857#endif 6858 6859 kmem_free(hash->dth_tab, 6860 hash->dth_size * sizeof (dtrace_hashbucket_t *)); 6861 kmem_free(hash, sizeof (dtrace_hash_t)); 6862} 6863 6864static void 6865dtrace_hash_resize(dtrace_hash_t *hash) 6866{ 6867 int size = hash->dth_size, i, ndx; 6868 int new_size = hash->dth_size << 1; 6869 int new_mask = new_size - 1; 6870 dtrace_hashbucket_t **new_tab, *bucket, *next; 6871 6872 ASSERT((new_size & new_mask) == 0); 6873 6874 new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP); 6875 6876 for (i = 0; i < size; i++) { 6877 for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) { 6878 dtrace_probe_t *probe = bucket->dthb_chain; 6879 6880 ASSERT(probe != NULL); 6881 ndx = DTRACE_HASHSTR(hash, probe) & new_mask; 6882 6883 next = bucket->dthb_next; 6884 bucket->dthb_next = new_tab[ndx]; 6885 new_tab[ndx] = bucket; 6886 } 6887 } 6888 6889 kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *)); 6890 hash->dth_tab = new_tab; 6891 hash->dth_size = new_size; 6892 hash->dth_mask = new_mask; 6893} 6894 6895static void 6896dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new) 6897{ 6898 int hashval = DTRACE_HASHSTR(hash, new); 6899 int ndx = hashval & hash->dth_mask; 6900 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx]; 6901 dtrace_probe_t **nextp, **prevp; 6902 6903 for (; bucket != NULL; bucket = bucket->dthb_next) { 6904 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new)) 6905 goto add; 6906 } 6907 6908 if ((hash->dth_nbuckets >> 1) > hash->dth_size) { 6909 dtrace_hash_resize(hash); 6910 dtrace_hash_add(hash, new); 6911 return; 6912 } 6913 6914 bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP); 6915 bucket->dthb_next = hash->dth_tab[ndx]; 6916 hash->dth_tab[ndx] = bucket; 6917 hash->dth_nbuckets++; 6918 6919add: 6920 nextp = DTRACE_HASHNEXT(hash, new); 6921 ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL); 6922 *nextp = bucket->dthb_chain; 6923 6924 if (bucket->dthb_chain != NULL) { 6925 prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain); 6926 ASSERT(*prevp == NULL); 6927 *prevp = new; 6928 } 6929 6930 bucket->dthb_chain = new; 6931 bucket->dthb_len++; 6932} 6933 6934static dtrace_probe_t * 6935dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template) 6936{ 6937 int hashval = DTRACE_HASHSTR(hash, template); 6938 int ndx = hashval & hash->dth_mask; 6939 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx]; 6940 6941 for (; bucket != NULL; bucket = bucket->dthb_next) { 6942 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template)) 6943 return (bucket->dthb_chain); 6944 } 6945 6946 return (NULL); 6947} 6948 6949static int 6950dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template) 6951{ 6952 int hashval = DTRACE_HASHSTR(hash, template); 6953 int ndx = hashval & hash->dth_mask; 6954 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx]; 6955 6956 for (; bucket != NULL; bucket = bucket->dthb_next) { 6957 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template)) 6958 return (bucket->dthb_len); 6959 } 6960 6961 return (0); 6962} 6963 6964static void 6965dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe) 6966{ 6967 int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask; 6968 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx]; 6969 6970 dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe); 6971 dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe); 6972 6973 /* 6974 * Find the bucket that we're removing this probe from. 6975 */ 6976 for (; bucket != NULL; bucket = bucket->dthb_next) { 6977 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe)) 6978 break; 6979 } 6980 6981 ASSERT(bucket != NULL); 6982 6983 if (*prevp == NULL) { 6984 if (*nextp == NULL) { 6985 /* 6986 * The removed probe was the only probe on this 6987 * bucket; we need to remove the bucket. 6988 */ 6989 dtrace_hashbucket_t *b = hash->dth_tab[ndx]; 6990 6991 ASSERT(bucket->dthb_chain == probe); 6992 ASSERT(b != NULL); 6993 6994 if (b == bucket) { 6995 hash->dth_tab[ndx] = bucket->dthb_next; 6996 } else { 6997 while (b->dthb_next != bucket) 6998 b = b->dthb_next; 6999 b->dthb_next = bucket->dthb_next; 7000 } 7001 7002 ASSERT(hash->dth_nbuckets > 0); 7003 hash->dth_nbuckets--; 7004 kmem_free(bucket, sizeof (dtrace_hashbucket_t)); 7005 return; 7006 } 7007 7008 bucket->dthb_chain = *nextp; 7009 } else { 7010 *(DTRACE_HASHNEXT(hash, *prevp)) = *nextp; 7011 } 7012 7013 if (*nextp != NULL) 7014 *(DTRACE_HASHPREV(hash, *nextp)) = *prevp; 7015} 7016 7017/* 7018 * DTrace Utility Functions 7019 * 7020 * These are random utility functions that are _not_ called from probe context. 7021 */ 7022static int 7023dtrace_badattr(const dtrace_attribute_t *a) 7024{ 7025 return (a->dtat_name > DTRACE_STABILITY_MAX || 7026 a->dtat_data > DTRACE_STABILITY_MAX || 7027 a->dtat_class > DTRACE_CLASS_MAX); 7028} 7029 7030/* 7031 * Return a duplicate copy of a string. If the specified string is NULL, 7032 * this function returns a zero-length string. 7033 */ 7034static char * 7035dtrace_strdup(const char *str) 7036{ 7037 char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP); 7038 7039 if (str != NULL) 7040 (void) strcpy(new, str); 7041 7042 return (new); 7043} 7044 7045#define DTRACE_ISALPHA(c) \ 7046 (((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z')) 7047 7048static int 7049dtrace_badname(const char *s) 7050{ 7051 char c; 7052 7053 if (s == NULL || (c = *s++) == '\0') 7054 return (0); 7055 7056 if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.') 7057 return (1); 7058 7059 while ((c = *s++) != '\0') { 7060 if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') && 7061 c != '-' && c != '_' && c != '.' && c != '`') 7062 return (1); 7063 } 7064 7065 return (0); 7066} 7067 7068static void 7069dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp) 7070{ 7071 uint32_t priv; 7072 7073#if defined(sun) 7074 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) { 7075 /* 7076 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter. 7077 */ 7078 priv = DTRACE_PRIV_ALL; 7079 } else { 7080 *uidp = crgetuid(cr); 7081 *zoneidp = crgetzoneid(cr); 7082 7083 priv = 0; 7084 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) 7085 priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER; 7086 else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) 7087 priv |= DTRACE_PRIV_USER; 7088 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) 7089 priv |= DTRACE_PRIV_PROC; 7090 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) 7091 priv |= DTRACE_PRIV_OWNER; 7092 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) 7093 priv |= DTRACE_PRIV_ZONEOWNER; 7094 } 7095#else 7096 priv = DTRACE_PRIV_ALL; 7097#endif 7098 7099 *privp = priv; 7100} 7101 7102#ifdef DTRACE_ERRDEBUG 7103static void 7104dtrace_errdebug(const char *str) 7105{ 7106 int hval = dtrace_hash_str(str) % DTRACE_ERRHASHSZ; 7107 int occupied = 0; 7108 7109 mutex_enter(&dtrace_errlock); 7110 dtrace_errlast = str; 7111 dtrace_errthread = curthread; 7112 7113 while (occupied++ < DTRACE_ERRHASHSZ) { 7114 if (dtrace_errhash[hval].dter_msg == str) { 7115 dtrace_errhash[hval].dter_count++; 7116 goto out; 7117 } 7118 7119 if (dtrace_errhash[hval].dter_msg != NULL) { 7120 hval = (hval + 1) % DTRACE_ERRHASHSZ; 7121 continue; 7122 } 7123 7124 dtrace_errhash[hval].dter_msg = str; 7125 dtrace_errhash[hval].dter_count = 1; 7126 goto out; 7127 } 7128 7129 panic("dtrace: undersized error hash"); 7130out: 7131 mutex_exit(&dtrace_errlock); 7132} 7133#endif 7134 7135/* 7136 * DTrace Matching Functions 7137 * 7138 * These functions are used to match groups of probes, given some elements of 7139 * a probe tuple, or some globbed expressions for elements of a probe tuple. 7140 */ 7141static int 7142dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid, 7143 zoneid_t zoneid) 7144{ 7145 if (priv != DTRACE_PRIV_ALL) { 7146 uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags; 7147 uint32_t match = priv & ppriv; 7148 7149 /* 7150 * No PRIV_DTRACE_* privileges... 7151 */ 7152 if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER | 7153 DTRACE_PRIV_KERNEL)) == 0) 7154 return (0); 7155 7156 /* 7157 * No matching bits, but there were bits to match... 7158 */ 7159 if (match == 0 && ppriv != 0) 7160 return (0); 7161 7162 /* 7163 * Need to have permissions to the process, but don't... 7164 */ 7165 if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 && 7166 uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) { 7167 return (0); 7168 } 7169 7170 /* 7171 * Need to be in the same zone unless we possess the 7172 * privilege to examine all zones. 7173 */ 7174 if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 && 7175 zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) { 7176 return (0); 7177 } 7178 } 7179 7180 return (1); 7181} 7182 7183/* 7184 * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which 7185 * consists of input pattern strings and an ops-vector to evaluate them. 7186 * This function returns >0 for match, 0 for no match, and <0 for error. 7187 */ 7188static int 7189dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp, 7190 uint32_t priv, uid_t uid, zoneid_t zoneid) 7191{ 7192 dtrace_provider_t *pvp = prp->dtpr_provider; 7193 int rv; 7194 7195 if (pvp->dtpv_defunct) 7196 return (0); 7197 7198 if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0) 7199 return (rv); 7200 7201 if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0) 7202 return (rv); 7203 7204 if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0) 7205 return (rv); 7206 7207 if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0) 7208 return (rv); 7209 7210 if (dtrace_match_priv(prp, priv, uid, zoneid) == 0) 7211 return (0); 7212 7213 return (rv); 7214} 7215 7216/* 7217 * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN) 7218 * interface for matching a glob pattern 'p' to an input string 's'. Unlike 7219 * libc's version, the kernel version only applies to 8-bit ASCII strings. 7220 * In addition, all of the recursion cases except for '*' matching have been 7221 * unwound. For '*', we still implement recursive evaluation, but a depth 7222 * counter is maintained and matching is aborted if we recurse too deep. 7223 * The function returns 0 if no match, >0 if match, and <0 if recursion error. 7224 */ 7225static int 7226dtrace_match_glob(const char *s, const char *p, int depth) 7227{ 7228 const char *olds; 7229 char s1, c; 7230 int gs; 7231 7232 if (depth > DTRACE_PROBEKEY_MAXDEPTH) 7233 return (-1); 7234 7235 if (s == NULL) 7236 s = ""; /* treat NULL as empty string */ 7237 7238top: 7239 olds = s; 7240 s1 = *s++; 7241 7242 if (p == NULL) 7243 return (0); 7244 7245 if ((c = *p++) == '\0') 7246 return (s1 == '\0'); 7247 7248 switch (c) { 7249 case '[': { 7250 int ok = 0, notflag = 0; 7251 char lc = '\0'; 7252 7253 if (s1 == '\0') 7254 return (0); 7255 7256 if (*p == '!') { 7257 notflag = 1; 7258 p++; 7259 } 7260 7261 if ((c = *p++) == '\0') 7262 return (0); 7263 7264 do { 7265 if (c == '-' && lc != '\0' && *p != ']') { 7266 if ((c = *p++) == '\0') 7267 return (0); 7268 if (c == '\\' && (c = *p++) == '\0') 7269 return (0); 7270 7271 if (notflag) { 7272 if (s1 < lc || s1 > c) 7273 ok++; 7274 else 7275 return (0); 7276 } else if (lc <= s1 && s1 <= c) 7277 ok++; 7278 7279 } else if (c == '\\' && (c = *p++) == '\0') 7280 return (0); 7281 7282 lc = c; /* save left-hand 'c' for next iteration */ 7283 7284 if (notflag) { 7285 if (s1 != c) 7286 ok++; 7287 else 7288 return (0); 7289 } else if (s1 == c) 7290 ok++; 7291 7292 if ((c = *p++) == '\0') 7293 return (0); 7294 7295 } while (c != ']'); 7296 7297 if (ok) 7298 goto top; 7299 7300 return (0); 7301 } 7302 7303 case '\\': 7304 if ((c = *p++) == '\0') 7305 return (0); 7306 /*FALLTHRU*/ 7307 7308 default: 7309 if (c != s1) 7310 return (0); 7311 /*FALLTHRU*/ 7312 7313 case '?': 7314 if (s1 != '\0') 7315 goto top; 7316 return (0); 7317 7318 case '*': 7319 while (*p == '*') 7320 p++; /* consecutive *'s are identical to a single one */ 7321 7322 if (*p == '\0') 7323 return (1); 7324 7325 for (s = olds; *s != '\0'; s++) { 7326 if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0) 7327 return (gs); 7328 } 7329 7330 return (0); 7331 } 7332} 7333 7334/*ARGSUSED*/ 7335static int 7336dtrace_match_string(const char *s, const char *p, int depth) 7337{ 7338 return (s != NULL && strcmp(s, p) == 0); 7339} 7340 7341/*ARGSUSED*/ 7342static int 7343dtrace_match_nul(const char *s, const char *p, int depth) 7344{ 7345 return (1); /* always match the empty pattern */ 7346} 7347 7348/*ARGSUSED*/ 7349static int 7350dtrace_match_nonzero(const char *s, const char *p, int depth) 7351{ 7352 return (s != NULL && s[0] != '\0'); 7353} 7354 7355static int 7356dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid, 7357 zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg) 7358{ 7359 dtrace_probe_t template, *probe; 7360 dtrace_hash_t *hash = NULL; 7361 int len, best = INT_MAX, nmatched = 0; 7362 dtrace_id_t i; 7363 7364 ASSERT(MUTEX_HELD(&dtrace_lock)); 7365 7366 /* 7367 * If the probe ID is specified in the key, just lookup by ID and 7368 * invoke the match callback once if a matching probe is found. 7369 */ 7370 if (pkp->dtpk_id != DTRACE_IDNONE) { 7371 if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL && 7372 dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) { 7373 (void) (*matched)(probe, arg); 7374 nmatched++; 7375 } 7376 return (nmatched); 7377 } 7378 7379 template.dtpr_mod = (char *)pkp->dtpk_mod; 7380 template.dtpr_func = (char *)pkp->dtpk_func; 7381 template.dtpr_name = (char *)pkp->dtpk_name; 7382 7383 /* 7384 * We want to find the most distinct of the module name, function 7385 * name, and name. So for each one that is not a glob pattern or 7386 * empty string, we perform a lookup in the corresponding hash and 7387 * use the hash table with the fewest collisions to do our search. 7388 */ 7389 if (pkp->dtpk_mmatch == &dtrace_match_string && 7390 (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) { 7391 best = len; 7392 hash = dtrace_bymod; 7393 } 7394 7395 if (pkp->dtpk_fmatch == &dtrace_match_string && 7396 (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) { 7397 best = len; 7398 hash = dtrace_byfunc; 7399 } 7400 7401 if (pkp->dtpk_nmatch == &dtrace_match_string && 7402 (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) { 7403 best = len; 7404 hash = dtrace_byname; 7405 } 7406 7407 /* 7408 * If we did not select a hash table, iterate over every probe and 7409 * invoke our callback for each one that matches our input probe key. 7410 */ 7411 if (hash == NULL) { 7412 for (i = 0; i < dtrace_nprobes; i++) { 7413 if ((probe = dtrace_probes[i]) == NULL || 7414 dtrace_match_probe(probe, pkp, priv, uid, 7415 zoneid) <= 0) 7416 continue; 7417 7418 nmatched++; 7419 7420 if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT) 7421 break; 7422 } 7423 7424 return (nmatched); 7425 } 7426 7427 /* 7428 * If we selected a hash table, iterate over each probe of the same key 7429 * name and invoke the callback for every probe that matches the other 7430 * attributes of our input probe key. 7431 */ 7432 for (probe = dtrace_hash_lookup(hash, &template); probe != NULL; 7433 probe = *(DTRACE_HASHNEXT(hash, probe))) { 7434 7435 if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0) 7436 continue; 7437 7438 nmatched++; 7439 7440 if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT) 7441 break; 7442 } 7443 7444 return (nmatched); 7445} 7446 7447/* 7448 * Return the function pointer dtrace_probecmp() should use to compare the 7449 * specified pattern with a string. For NULL or empty patterns, we select 7450 * dtrace_match_nul(). For glob pattern strings, we use dtrace_match_glob(). 7451 * For non-empty non-glob strings, we use dtrace_match_string(). 7452 */ 7453static dtrace_probekey_f * 7454dtrace_probekey_func(const char *p) 7455{ 7456 char c; 7457 7458 if (p == NULL || *p == '\0') 7459 return (&dtrace_match_nul); 7460 7461 while ((c = *p++) != '\0') { 7462 if (c == '[' || c == '?' || c == '*' || c == '\\') 7463 return (&dtrace_match_glob); 7464 } 7465 7466 return (&dtrace_match_string); 7467} 7468 7469/* 7470 * Build a probe comparison key for use with dtrace_match_probe() from the 7471 * given probe description. By convention, a null key only matches anchored 7472 * probes: if each field is the empty string, reset dtpk_fmatch to 7473 * dtrace_match_nonzero(). 7474 */ 7475static void 7476dtrace_probekey(dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp) 7477{ 7478 pkp->dtpk_prov = pdp->dtpd_provider; 7479 pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider); 7480 7481 pkp->dtpk_mod = pdp->dtpd_mod; 7482 pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod); 7483 7484 pkp->dtpk_func = pdp->dtpd_func; 7485 pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func); 7486 7487 pkp->dtpk_name = pdp->dtpd_name; 7488 pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name); 7489 7490 pkp->dtpk_id = pdp->dtpd_id; 7491 7492 if (pkp->dtpk_id == DTRACE_IDNONE && 7493 pkp->dtpk_pmatch == &dtrace_match_nul && 7494 pkp->dtpk_mmatch == &dtrace_match_nul && 7495 pkp->dtpk_fmatch == &dtrace_match_nul && 7496 pkp->dtpk_nmatch == &dtrace_match_nul) 7497 pkp->dtpk_fmatch = &dtrace_match_nonzero; 7498} 7499 7500/* 7501 * DTrace Provider-to-Framework API Functions 7502 * 7503 * These functions implement much of the Provider-to-Framework API, as 7504 * described in <sys/dtrace.h>. The parts of the API not in this section are 7505 * the functions in the API for probe management (found below), and 7506 * dtrace_probe() itself (found above). 7507 */ 7508 7509/* 7510 * Register the calling provider with the DTrace framework. This should 7511 * generally be called by DTrace providers in their attach(9E) entry point. 7512 */ 7513int 7514dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv, 7515 cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp) 7516{ 7517 dtrace_provider_t *provider; 7518 7519 if (name == NULL || pap == NULL || pops == NULL || idp == NULL) { 7520 cmn_err(CE_WARN, "failed to register provider '%s': invalid " 7521 "arguments", name ? name : "<NULL>"); 7522 return (EINVAL); 7523 } 7524 7525 if (name[0] == '\0' || dtrace_badname(name)) { 7526 cmn_err(CE_WARN, "failed to register provider '%s': invalid " 7527 "provider name", name); 7528 return (EINVAL); 7529 } 7530 7531 if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) || 7532 pops->dtps_enable == NULL || pops->dtps_disable == NULL || 7533 pops->dtps_destroy == NULL || 7534 ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) { 7535 cmn_err(CE_WARN, "failed to register provider '%s': invalid " 7536 "provider ops", name); 7537 return (EINVAL); 7538 } 7539 7540 if (dtrace_badattr(&pap->dtpa_provider) || 7541 dtrace_badattr(&pap->dtpa_mod) || 7542 dtrace_badattr(&pap->dtpa_func) || 7543 dtrace_badattr(&pap->dtpa_name) || 7544 dtrace_badattr(&pap->dtpa_args)) { 7545 cmn_err(CE_WARN, "failed to register provider '%s': invalid " 7546 "provider attributes", name); 7547 return (EINVAL); 7548 } 7549 7550 if (priv & ~DTRACE_PRIV_ALL) { 7551 cmn_err(CE_WARN, "failed to register provider '%s': invalid " 7552 "privilege attributes", name); 7553 return (EINVAL); 7554 } 7555 7556 if ((priv & DTRACE_PRIV_KERNEL) && 7557 (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) && 7558 pops->dtps_usermode == NULL) { 7559 cmn_err(CE_WARN, "failed to register provider '%s': need " 7560 "dtps_usermode() op for given privilege attributes", name); 7561 return (EINVAL); 7562 } 7563 7564 provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP); 7565 provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP); 7566 (void) strcpy(provider->dtpv_name, name); 7567 7568 provider->dtpv_attr = *pap; 7569 provider->dtpv_priv.dtpp_flags = priv; 7570 if (cr != NULL) { 7571 provider->dtpv_priv.dtpp_uid = crgetuid(cr); 7572 provider->dtpv_priv.dtpp_zoneid = crgetzoneid(cr); 7573 } 7574 provider->dtpv_pops = *pops; 7575 7576 if (pops->dtps_provide == NULL) { 7577 ASSERT(pops->dtps_provide_module != NULL); 7578 provider->dtpv_pops.dtps_provide = 7579 (void (*)(void *, dtrace_probedesc_t *))dtrace_nullop; 7580 } 7581 7582 if (pops->dtps_provide_module == NULL) { 7583 ASSERT(pops->dtps_provide != NULL); 7584 provider->dtpv_pops.dtps_provide_module = 7585 (void (*)(void *, modctl_t *))dtrace_nullop; 7586 } 7587 7588 if (pops->dtps_suspend == NULL) { 7589 ASSERT(pops->dtps_resume == NULL); 7590 provider->dtpv_pops.dtps_suspend = 7591 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop; 7592 provider->dtpv_pops.dtps_resume = 7593 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop; 7594 } 7595 7596 provider->dtpv_arg = arg; 7597 *idp = (dtrace_provider_id_t)provider; 7598 7599 if (pops == &dtrace_provider_ops) { 7600 ASSERT(MUTEX_HELD(&dtrace_provider_lock)); 7601 ASSERT(MUTEX_HELD(&dtrace_lock)); 7602 ASSERT(dtrace_anon.dta_enabling == NULL); 7603 7604 /* 7605 * We make sure that the DTrace provider is at the head of 7606 * the provider chain. 7607 */ 7608 provider->dtpv_next = dtrace_provider; 7609 dtrace_provider = provider; 7610 return (0); 7611 } 7612 7613 mutex_enter(&dtrace_provider_lock); 7614 mutex_enter(&dtrace_lock); 7615 7616 /* 7617 * If there is at least one provider registered, we'll add this 7618 * provider after the first provider. 7619 */ 7620 if (dtrace_provider != NULL) { 7621 provider->dtpv_next = dtrace_provider->dtpv_next; 7622 dtrace_provider->dtpv_next = provider; 7623 } else { 7624 dtrace_provider = provider; 7625 } 7626 7627 if (dtrace_retained != NULL) { 7628 dtrace_enabling_provide(provider); 7629 7630 /* 7631 * Now we need to call dtrace_enabling_matchall() -- which 7632 * will acquire cpu_lock and dtrace_lock. We therefore need 7633 * to drop all of our locks before calling into it... 7634 */ 7635 mutex_exit(&dtrace_lock); 7636 mutex_exit(&dtrace_provider_lock); 7637 dtrace_enabling_matchall(); 7638 7639 return (0); 7640 } 7641 7642 mutex_exit(&dtrace_lock); 7643 mutex_exit(&dtrace_provider_lock); 7644 7645 return (0); 7646} 7647 7648/* 7649 * Unregister the specified provider from the DTrace framework. This should 7650 * generally be called by DTrace providers in their detach(9E) entry point. 7651 */ 7652int 7653dtrace_unregister(dtrace_provider_id_t id) 7654{ 7655 dtrace_provider_t *old = (dtrace_provider_t *)id; 7656 dtrace_provider_t *prev = NULL; 7657 int i, self = 0, noreap = 0; 7658 dtrace_probe_t *probe, *first = NULL; 7659 7660 if (old->dtpv_pops.dtps_enable == 7661 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop) { 7662 /* 7663 * If DTrace itself is the provider, we're called with locks 7664 * already held. 7665 */ 7666 ASSERT(old == dtrace_provider); 7667#if defined(sun) 7668 ASSERT(dtrace_devi != NULL); 7669#endif 7670 ASSERT(MUTEX_HELD(&dtrace_provider_lock)); 7671 ASSERT(MUTEX_HELD(&dtrace_lock)); 7672 self = 1; 7673 7674 if (dtrace_provider->dtpv_next != NULL) { 7675 /* 7676 * There's another provider here; return failure. 7677 */ 7678 return (EBUSY); 7679 } 7680 } else { 7681 mutex_enter(&dtrace_provider_lock); 7682#if defined(sun) 7683 mutex_enter(&mod_lock); 7684#endif 7685 mutex_enter(&dtrace_lock); 7686 } 7687 7688 /* 7689 * If anyone has /dev/dtrace open, or if there are anonymous enabled 7690 * probes, we refuse to let providers slither away, unless this 7691 * provider has already been explicitly invalidated. 7692 */ 7693 if (!old->dtpv_defunct && 7694 (dtrace_opens || (dtrace_anon.dta_state != NULL && 7695 dtrace_anon.dta_state->dts_necbs > 0))) { 7696 if (!self) { 7697 mutex_exit(&dtrace_lock); 7698#if defined(sun) 7699 mutex_exit(&mod_lock); 7700#endif 7701 mutex_exit(&dtrace_provider_lock); 7702 } 7703 return (EBUSY); 7704 } 7705 7706 /* 7707 * Attempt to destroy the probes associated with this provider. 7708 */ 7709 for (i = 0; i < dtrace_nprobes; i++) { 7710 if ((probe = dtrace_probes[i]) == NULL) 7711 continue; 7712 7713 if (probe->dtpr_provider != old) 7714 continue; 7715 7716 if (probe->dtpr_ecb == NULL) 7717 continue; 7718 7719 /* 7720 * If we are trying to unregister a defunct provider, and the 7721 * provider was made defunct within the interval dictated by 7722 * dtrace_unregister_defunct_reap, we'll (asynchronously) 7723 * attempt to reap our enablings. To denote that the provider 7724 * should reattempt to unregister itself at some point in the 7725 * future, we will return a differentiable error code (EAGAIN 7726 * instead of EBUSY) in this case. 7727 */ 7728 if (dtrace_gethrtime() - old->dtpv_defunct > 7729 dtrace_unregister_defunct_reap) 7730 noreap = 1; 7731 7732 if (!self) { 7733 mutex_exit(&dtrace_lock); 7734#if defined(sun) 7735 mutex_exit(&mod_lock); 7736#endif 7737 mutex_exit(&dtrace_provider_lock); 7738 } 7739 7740 if (noreap) 7741 return (EBUSY); 7742 7743 (void) taskq_dispatch(dtrace_taskq, 7744 (task_func_t *)dtrace_enabling_reap, NULL, TQ_SLEEP); 7745 7746 return (EAGAIN); 7747 } 7748 7749 /* 7750 * All of the probes for this provider are disabled; we can safely 7751 * remove all of them from their hash chains and from the probe array. 7752 */ 7753 for (i = 0; i < dtrace_nprobes; i++) { 7754 if ((probe = dtrace_probes[i]) == NULL) 7755 continue; 7756 7757 if (probe->dtpr_provider != old) 7758 continue; 7759 7760 dtrace_probes[i] = NULL; 7761 7762 dtrace_hash_remove(dtrace_bymod, probe); 7763 dtrace_hash_remove(dtrace_byfunc, probe); 7764 dtrace_hash_remove(dtrace_byname, probe); 7765 7766 if (first == NULL) { 7767 first = probe; 7768 probe->dtpr_nextmod = NULL; 7769 } else { 7770 probe->dtpr_nextmod = first; 7771 first = probe; 7772 } 7773 } 7774 7775 /* 7776 * The provider's probes have been removed from the hash chains and 7777 * from the probe array. Now issue a dtrace_sync() to be sure that 7778 * everyone has cleared out from any probe array processing. 7779 */ 7780 dtrace_sync(); 7781 7782 for (probe = first; probe != NULL; probe = first) { 7783 first = probe->dtpr_nextmod; 7784 7785 old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id, 7786 probe->dtpr_arg); 7787 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1); 7788 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1); 7789 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1); 7790#if defined(sun) 7791 vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1); 7792#else 7793 free_unr(dtrace_arena, probe->dtpr_id); 7794#endif 7795 kmem_free(probe, sizeof (dtrace_probe_t)); 7796 } 7797 7798 if ((prev = dtrace_provider) == old) { 7799#if defined(sun) 7800 ASSERT(self || dtrace_devi == NULL); 7801 ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL); 7802#endif 7803 dtrace_provider = old->dtpv_next; 7804 } else { 7805 while (prev != NULL && prev->dtpv_next != old) 7806 prev = prev->dtpv_next; 7807 7808 if (prev == NULL) { 7809 panic("attempt to unregister non-existent " 7810 "dtrace provider %p\n", (void *)id); 7811 } 7812 7813 prev->dtpv_next = old->dtpv_next; 7814 } 7815 7816 if (!self) { 7817 mutex_exit(&dtrace_lock); 7818#if defined(sun) 7819 mutex_exit(&mod_lock); 7820#endif 7821 mutex_exit(&dtrace_provider_lock); 7822 } 7823 7824 kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1); 7825 kmem_free(old, sizeof (dtrace_provider_t)); 7826 7827 return (0); 7828} 7829 7830/* 7831 * Invalidate the specified provider. All subsequent probe lookups for the 7832 * specified provider will fail, but its probes will not be removed. 7833 */ 7834void 7835dtrace_invalidate(dtrace_provider_id_t id) 7836{ 7837 dtrace_provider_t *pvp = (dtrace_provider_t *)id; 7838 7839 ASSERT(pvp->dtpv_pops.dtps_enable != 7840 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop); 7841 7842 mutex_enter(&dtrace_provider_lock); 7843 mutex_enter(&dtrace_lock); 7844 7845 pvp->dtpv_defunct = dtrace_gethrtime(); 7846 7847 mutex_exit(&dtrace_lock); 7848 mutex_exit(&dtrace_provider_lock); 7849} 7850 7851/* 7852 * Indicate whether or not DTrace has attached. 7853 */ 7854int 7855dtrace_attached(void) 7856{ 7857 /* 7858 * dtrace_provider will be non-NULL iff the DTrace driver has 7859 * attached. (It's non-NULL because DTrace is always itself a 7860 * provider.) 7861 */ 7862 return (dtrace_provider != NULL); 7863} 7864 7865/* 7866 * Remove all the unenabled probes for the given provider. This function is 7867 * not unlike dtrace_unregister(), except that it doesn't remove the provider 7868 * -- just as many of its associated probes as it can. 7869 */ 7870int 7871dtrace_condense(dtrace_provider_id_t id) 7872{ 7873 dtrace_provider_t *prov = (dtrace_provider_t *)id; 7874 int i; 7875 dtrace_probe_t *probe; 7876 7877 /* 7878 * Make sure this isn't the dtrace provider itself. 7879 */ 7880 ASSERT(prov->dtpv_pops.dtps_enable != 7881 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop); 7882 7883 mutex_enter(&dtrace_provider_lock); 7884 mutex_enter(&dtrace_lock); 7885 7886 /* 7887 * Attempt to destroy the probes associated with this provider. 7888 */ 7889 for (i = 0; i < dtrace_nprobes; i++) { 7890 if ((probe = dtrace_probes[i]) == NULL) 7891 continue; 7892 7893 if (probe->dtpr_provider != prov) 7894 continue; 7895 7896 if (probe->dtpr_ecb != NULL) 7897 continue; 7898 7899 dtrace_probes[i] = NULL; 7900 7901 dtrace_hash_remove(dtrace_bymod, probe); 7902 dtrace_hash_remove(dtrace_byfunc, probe); 7903 dtrace_hash_remove(dtrace_byname, probe); 7904 7905 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1, 7906 probe->dtpr_arg); 7907 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1); 7908 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1); 7909 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1); 7910 kmem_free(probe, sizeof (dtrace_probe_t)); 7911#if defined(sun) 7912 vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1); 7913#else 7914 free_unr(dtrace_arena, i + 1); 7915#endif 7916 } 7917 7918 mutex_exit(&dtrace_lock); 7919 mutex_exit(&dtrace_provider_lock); 7920 7921 return (0); 7922} 7923 7924/* 7925 * DTrace Probe Management Functions 7926 * 7927 * The functions in this section perform the DTrace probe management, 7928 * including functions to create probes, look-up probes, and call into the 7929 * providers to request that probes be provided. Some of these functions are 7930 * in the Provider-to-Framework API; these functions can be identified by the 7931 * fact that they are not declared "static". 7932 */ 7933 7934/* 7935 * Create a probe with the specified module name, function name, and name. 7936 */ 7937dtrace_id_t 7938dtrace_probe_create(dtrace_provider_id_t prov, const char *mod, 7939 const char *func, const char *name, int aframes, void *arg) 7940{ 7941 dtrace_probe_t *probe, **probes; 7942 dtrace_provider_t *provider = (dtrace_provider_t *)prov; 7943 dtrace_id_t id; 7944 7945 if (provider == dtrace_provider) { 7946 ASSERT(MUTEX_HELD(&dtrace_lock)); 7947 } else { 7948 mutex_enter(&dtrace_lock); 7949 } 7950 7951#if defined(sun) 7952 id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1, 7953 VM_BESTFIT | VM_SLEEP); 7954#else 7955 id = alloc_unr(dtrace_arena); 7956#endif 7957 probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP); 7958 7959 probe->dtpr_id = id; 7960 probe->dtpr_gen = dtrace_probegen++; 7961 probe->dtpr_mod = dtrace_strdup(mod); 7962 probe->dtpr_func = dtrace_strdup(func); 7963 probe->dtpr_name = dtrace_strdup(name); 7964 probe->dtpr_arg = arg; 7965 probe->dtpr_aframes = aframes; 7966 probe->dtpr_provider = provider; 7967 7968 dtrace_hash_add(dtrace_bymod, probe); 7969 dtrace_hash_add(dtrace_byfunc, probe); 7970 dtrace_hash_add(dtrace_byname, probe); 7971 7972 if (id - 1 >= dtrace_nprobes) { 7973 size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *); 7974 size_t nsize = osize << 1; 7975 7976 if (nsize == 0) { 7977 ASSERT(osize == 0); 7978 ASSERT(dtrace_probes == NULL); 7979 nsize = sizeof (dtrace_probe_t *); 7980 } 7981 7982 probes = kmem_zalloc(nsize, KM_SLEEP); 7983 7984 if (dtrace_probes == NULL) { 7985 ASSERT(osize == 0); 7986 dtrace_probes = probes; 7987 dtrace_nprobes = 1; 7988 } else { 7989 dtrace_probe_t **oprobes = dtrace_probes; 7990 7991 bcopy(oprobes, probes, osize); 7992 dtrace_membar_producer(); 7993 dtrace_probes = probes; 7994 7995 dtrace_sync(); 7996 7997 /* 7998 * All CPUs are now seeing the new probes array; we can 7999 * safely free the old array. 8000 */ 8001 kmem_free(oprobes, osize); 8002 dtrace_nprobes <<= 1; 8003 } 8004 8005 ASSERT(id - 1 < dtrace_nprobes); 8006 } 8007 8008 ASSERT(dtrace_probes[id - 1] == NULL); 8009 dtrace_probes[id - 1] = probe; 8010 8011 if (provider != dtrace_provider) 8012 mutex_exit(&dtrace_lock); 8013 8014 return (id); 8015} 8016 8017static dtrace_probe_t * 8018dtrace_probe_lookup_id(dtrace_id_t id) 8019{ 8020 ASSERT(MUTEX_HELD(&dtrace_lock)); 8021 8022 if (id == 0 || id > dtrace_nprobes) 8023 return (NULL); 8024 8025 return (dtrace_probes[id - 1]); 8026} 8027 8028static int 8029dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg) 8030{ 8031 *((dtrace_id_t *)arg) = probe->dtpr_id; 8032 8033 return (DTRACE_MATCH_DONE); 8034} 8035 8036/* 8037 * Look up a probe based on provider and one or more of module name, function 8038 * name and probe name. 8039 */ 8040dtrace_id_t 8041dtrace_probe_lookup(dtrace_provider_id_t prid, char *mod, 8042 char *func, char *name) 8043{ 8044 dtrace_probekey_t pkey; 8045 dtrace_id_t id; 8046 int match; 8047 8048 pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name; 8049 pkey.dtpk_pmatch = &dtrace_match_string; 8050 pkey.dtpk_mod = mod; 8051 pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul; 8052 pkey.dtpk_func = func; 8053 pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul; 8054 pkey.dtpk_name = name; 8055 pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul; 8056 pkey.dtpk_id = DTRACE_IDNONE; 8057 8058 mutex_enter(&dtrace_lock); 8059 match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0, 8060 dtrace_probe_lookup_match, &id); 8061 mutex_exit(&dtrace_lock); 8062 8063 ASSERT(match == 1 || match == 0); 8064 return (match ? id : 0); 8065} 8066 8067/* 8068 * Returns the probe argument associated with the specified probe. 8069 */ 8070void * 8071dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid) 8072{ 8073 dtrace_probe_t *probe; 8074 void *rval = NULL; 8075 8076 mutex_enter(&dtrace_lock); 8077 8078 if ((probe = dtrace_probe_lookup_id(pid)) != NULL && 8079 probe->dtpr_provider == (dtrace_provider_t *)id) 8080 rval = probe->dtpr_arg; 8081 8082 mutex_exit(&dtrace_lock); 8083 8084 return (rval); 8085} 8086 8087/* 8088 * Copy a probe into a probe description. 8089 */ 8090static void 8091dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp) 8092{ 8093 bzero(pdp, sizeof (dtrace_probedesc_t)); 8094 pdp->dtpd_id = prp->dtpr_id; 8095 8096 (void) strncpy(pdp->dtpd_provider, 8097 prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1); 8098 8099 (void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1); 8100 (void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1); 8101 (void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1); 8102} 8103 8104/* 8105 * Called to indicate that a probe -- or probes -- should be provided by a 8106 * specfied provider. If the specified description is NULL, the provider will 8107 * be told to provide all of its probes. (This is done whenever a new 8108 * consumer comes along, or whenever a retained enabling is to be matched.) If 8109 * the specified description is non-NULL, the provider is given the 8110 * opportunity to dynamically provide the specified probe, allowing providers 8111 * to support the creation of probes on-the-fly. (So-called _autocreated_ 8112 * probes.) If the provider is NULL, the operations will be applied to all 8113 * providers; if the provider is non-NULL the operations will only be applied 8114 * to the specified provider. The dtrace_provider_lock must be held, and the 8115 * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation 8116 * will need to grab the dtrace_lock when it reenters the framework through 8117 * dtrace_probe_lookup(), dtrace_probe_create(), etc. 8118 */ 8119static void 8120dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv) 8121{ 8122#if defined(sun) 8123 modctl_t *ctl; 8124#endif 8125 int all = 0; 8126 8127 ASSERT(MUTEX_HELD(&dtrace_provider_lock)); 8128 8129 if (prv == NULL) { 8130 all = 1; 8131 prv = dtrace_provider; 8132 } 8133 8134 do { 8135 /* 8136 * First, call the blanket provide operation. 8137 */ 8138 prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc); 8139 8140#if defined(sun) 8141 /* 8142 * Now call the per-module provide operation. We will grab 8143 * mod_lock to prevent the list from being modified. Note 8144 * that this also prevents the mod_busy bits from changing. 8145 * (mod_busy can only be changed with mod_lock held.) 8146 */ 8147 mutex_enter(&mod_lock); 8148 8149 ctl = &modules; 8150 do { 8151 if (ctl->mod_busy || ctl->mod_mp == NULL) 8152 continue; 8153 8154 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl); 8155 8156 } while ((ctl = ctl->mod_next) != &modules); 8157 8158 mutex_exit(&mod_lock); 8159#endif 8160 } while (all && (prv = prv->dtpv_next) != NULL); 8161} 8162 8163#if defined(sun) 8164/* 8165 * Iterate over each probe, and call the Framework-to-Provider API function 8166 * denoted by offs. 8167 */ 8168static void 8169dtrace_probe_foreach(uintptr_t offs) 8170{ 8171 dtrace_provider_t *prov; 8172 void (*func)(void *, dtrace_id_t, void *); 8173 dtrace_probe_t *probe; 8174 dtrace_icookie_t cookie; 8175 int i; 8176 8177 /* 8178 * We disable interrupts to walk through the probe array. This is 8179 * safe -- the dtrace_sync() in dtrace_unregister() assures that we 8180 * won't see stale data. 8181 */ 8182 cookie = dtrace_interrupt_disable(); 8183 8184 for (i = 0; i < dtrace_nprobes; i++) { 8185 if ((probe = dtrace_probes[i]) == NULL) 8186 continue; 8187 8188 if (probe->dtpr_ecb == NULL) { 8189 /* 8190 * This probe isn't enabled -- don't call the function. 8191 */ 8192 continue; 8193 } 8194 8195 prov = probe->dtpr_provider; 8196 func = *((void(**)(void *, dtrace_id_t, void *)) 8197 ((uintptr_t)&prov->dtpv_pops + offs)); 8198 8199 func(prov->dtpv_arg, i + 1, probe->dtpr_arg); 8200 } 8201 8202 dtrace_interrupt_enable(cookie); 8203} 8204#endif 8205 8206static int 8207dtrace_probe_enable(dtrace_probedesc_t *desc, dtrace_enabling_t *enab) 8208{ 8209 dtrace_probekey_t pkey; 8210 uint32_t priv; 8211 uid_t uid; 8212 zoneid_t zoneid; 8213 8214 ASSERT(MUTEX_HELD(&dtrace_lock)); 8215 dtrace_ecb_create_cache = NULL; 8216 8217 if (desc == NULL) { 8218 /* 8219 * If we're passed a NULL description, we're being asked to 8220 * create an ECB with a NULL probe. 8221 */ 8222 (void) dtrace_ecb_create_enable(NULL, enab); 8223 return (0); 8224 } 8225 8226 dtrace_probekey(desc, &pkey); 8227 dtrace_cred2priv(enab->dten_vstate->dtvs_state->dts_cred.dcr_cred, 8228 &priv, &uid, &zoneid); 8229 8230 return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable, 8231 enab)); 8232} 8233 8234/* 8235 * DTrace Helper Provider Functions 8236 */ 8237static void 8238dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr) 8239{ 8240 attr->dtat_name = DOF_ATTR_NAME(dofattr); 8241 attr->dtat_data = DOF_ATTR_DATA(dofattr); 8242 attr->dtat_class = DOF_ATTR_CLASS(dofattr); 8243} 8244 8245static void 8246dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov, 8247 const dof_provider_t *dofprov, char *strtab) 8248{ 8249 hprov->dthpv_provname = strtab + dofprov->dofpv_name; 8250 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider, 8251 dofprov->dofpv_provattr); 8252 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod, 8253 dofprov->dofpv_modattr); 8254 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func, 8255 dofprov->dofpv_funcattr); 8256 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name, 8257 dofprov->dofpv_nameattr); 8258 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args, 8259 dofprov->dofpv_argsattr); 8260} 8261 8262static void 8263dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid) 8264{ 8265 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof; 8266 dof_hdr_t *dof = (dof_hdr_t *)daddr; 8267 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec; 8268 dof_provider_t *provider; 8269 dof_probe_t *probe; 8270 uint32_t *off, *enoff; 8271 uint8_t *arg; 8272 char *strtab; 8273 uint_t i, nprobes; 8274 dtrace_helper_provdesc_t dhpv; 8275 dtrace_helper_probedesc_t dhpb; 8276 dtrace_meta_t *meta = dtrace_meta_pid; 8277 dtrace_mops_t *mops = &meta->dtm_mops; 8278 void *parg; 8279 8280 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset); 8281 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff + 8282 provider->dofpv_strtab * dof->dofh_secsize); 8283 prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff + 8284 provider->dofpv_probes * dof->dofh_secsize); 8285 arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff + 8286 provider->dofpv_prargs * dof->dofh_secsize); 8287 off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff + 8288 provider->dofpv_proffs * dof->dofh_secsize); 8289 8290 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset); 8291 off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset); 8292 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset); 8293 enoff = NULL; 8294 8295 /* 8296 * See dtrace_helper_provider_validate(). 8297 */ 8298 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 && 8299 provider->dofpv_prenoffs != DOF_SECT_NONE) { 8300 enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff + 8301 provider->dofpv_prenoffs * dof->dofh_secsize); 8302 enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset); 8303 } 8304 8305 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize; 8306 8307 /* 8308 * Create the provider. 8309 */ 8310 dtrace_dofprov2hprov(&dhpv, provider, strtab); 8311 8312 if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL) 8313 return; 8314 8315 meta->dtm_count++; 8316 8317 /* 8318 * Create the probes. 8319 */ 8320 for (i = 0; i < nprobes; i++) { 8321 probe = (dof_probe_t *)(uintptr_t)(daddr + 8322 prb_sec->dofs_offset + i * prb_sec->dofs_entsize); 8323 8324 dhpb.dthpb_mod = dhp->dofhp_mod; 8325 dhpb.dthpb_func = strtab + probe->dofpr_func; 8326 dhpb.dthpb_name = strtab + probe->dofpr_name; 8327 dhpb.dthpb_base = probe->dofpr_addr; 8328 dhpb.dthpb_offs = off + probe->dofpr_offidx; 8329 dhpb.dthpb_noffs = probe->dofpr_noffs; 8330 if (enoff != NULL) { 8331 dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx; 8332 dhpb.dthpb_nenoffs = probe->dofpr_nenoffs; 8333 } else { 8334 dhpb.dthpb_enoffs = NULL; 8335 dhpb.dthpb_nenoffs = 0; 8336 } 8337 dhpb.dthpb_args = arg + probe->dofpr_argidx; 8338 dhpb.dthpb_nargc = probe->dofpr_nargc; 8339 dhpb.dthpb_xargc = probe->dofpr_xargc; 8340 dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv; 8341 dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv; 8342 8343 mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb); 8344 } 8345} 8346 8347static void 8348dtrace_helper_provide(dof_helper_t *dhp, pid_t pid) 8349{ 8350 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof; 8351 dof_hdr_t *dof = (dof_hdr_t *)daddr; 8352 int i; 8353 8354 ASSERT(MUTEX_HELD(&dtrace_meta_lock)); 8355 8356 for (i = 0; i < dof->dofh_secnum; i++) { 8357 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr + 8358 dof->dofh_secoff + i * dof->dofh_secsize); 8359 8360 if (sec->dofs_type != DOF_SECT_PROVIDER) 8361 continue; 8362 8363 dtrace_helper_provide_one(dhp, sec, pid); 8364 } 8365 8366 /* 8367 * We may have just created probes, so we must now rematch against 8368 * any retained enablings. Note that this call will acquire both 8369 * cpu_lock and dtrace_lock; the fact that we are holding 8370 * dtrace_meta_lock now is what defines the ordering with respect to 8371 * these three locks. 8372 */ 8373 dtrace_enabling_matchall(); 8374} 8375 8376static void 8377dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid) 8378{ 8379 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof; 8380 dof_hdr_t *dof = (dof_hdr_t *)daddr; 8381 dof_sec_t *str_sec; 8382 dof_provider_t *provider; 8383 char *strtab; 8384 dtrace_helper_provdesc_t dhpv; 8385 dtrace_meta_t *meta = dtrace_meta_pid; 8386 dtrace_mops_t *mops = &meta->dtm_mops; 8387 8388 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset); 8389 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff + 8390 provider->dofpv_strtab * dof->dofh_secsize); 8391 8392 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset); 8393 8394 /* 8395 * Create the provider. 8396 */ 8397 dtrace_dofprov2hprov(&dhpv, provider, strtab); 8398 8399 mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid); 8400 8401 meta->dtm_count--; 8402} 8403 8404static void 8405dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid) 8406{ 8407 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof; 8408 dof_hdr_t *dof = (dof_hdr_t *)daddr; 8409 int i; 8410 8411 ASSERT(MUTEX_HELD(&dtrace_meta_lock)); 8412 8413 for (i = 0; i < dof->dofh_secnum; i++) { 8414 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr + 8415 dof->dofh_secoff + i * dof->dofh_secsize); 8416 8417 if (sec->dofs_type != DOF_SECT_PROVIDER) 8418 continue; 8419 8420 dtrace_helper_provider_remove_one(dhp, sec, pid); 8421 } 8422} 8423 8424/* 8425 * DTrace Meta Provider-to-Framework API Functions 8426 * 8427 * These functions implement the Meta Provider-to-Framework API, as described 8428 * in <sys/dtrace.h>. 8429 */ 8430int 8431dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg, 8432 dtrace_meta_provider_id_t *idp) 8433{ 8434 dtrace_meta_t *meta; 8435 dtrace_helpers_t *help, *next; 8436 int i; 8437 8438 *idp = DTRACE_METAPROVNONE; 8439 8440 /* 8441 * We strictly don't need the name, but we hold onto it for 8442 * debuggability. All hail error queues! 8443 */ 8444 if (name == NULL) { 8445 cmn_err(CE_WARN, "failed to register meta-provider: " 8446 "invalid name"); 8447 return (EINVAL); 8448 } 8449 8450 if (mops == NULL || 8451 mops->dtms_create_probe == NULL || 8452 mops->dtms_provide_pid == NULL || 8453 mops->dtms_remove_pid == NULL) { 8454 cmn_err(CE_WARN, "failed to register meta-register %s: " 8455 "invalid ops", name); 8456 return (EINVAL); 8457 } 8458 8459 meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP); 8460 meta->dtm_mops = *mops; 8461 meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP); 8462 (void) strcpy(meta->dtm_name, name); 8463 meta->dtm_arg = arg; 8464 8465 mutex_enter(&dtrace_meta_lock); 8466 mutex_enter(&dtrace_lock); 8467 8468 if (dtrace_meta_pid != NULL) { 8469 mutex_exit(&dtrace_lock); 8470 mutex_exit(&dtrace_meta_lock); 8471 cmn_err(CE_WARN, "failed to register meta-register %s: " 8472 "user-land meta-provider exists", name); 8473 kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1); 8474 kmem_free(meta, sizeof (dtrace_meta_t)); 8475 return (EINVAL); 8476 } 8477 8478 dtrace_meta_pid = meta; 8479 *idp = (dtrace_meta_provider_id_t)meta; 8480 8481 /* 8482 * If there are providers and probes ready to go, pass them 8483 * off to the new meta provider now. 8484 */ 8485 8486 help = dtrace_deferred_pid; 8487 dtrace_deferred_pid = NULL; 8488 8489 mutex_exit(&dtrace_lock); 8490 8491 while (help != NULL) { 8492 for (i = 0; i < help->dthps_nprovs; i++) { 8493 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov, 8494 help->dthps_pid); 8495 } 8496 8497 next = help->dthps_next; 8498 help->dthps_next = NULL; 8499 help->dthps_prev = NULL; 8500 help->dthps_deferred = 0; 8501 help = next; 8502 } 8503 8504 mutex_exit(&dtrace_meta_lock); 8505 8506 return (0); 8507} 8508 8509int 8510dtrace_meta_unregister(dtrace_meta_provider_id_t id) 8511{ 8512 dtrace_meta_t **pp, *old = (dtrace_meta_t *)id; 8513 8514 mutex_enter(&dtrace_meta_lock); 8515 mutex_enter(&dtrace_lock); 8516 8517 if (old == dtrace_meta_pid) { 8518 pp = &dtrace_meta_pid; 8519 } else { 8520 panic("attempt to unregister non-existent " 8521 "dtrace meta-provider %p\n", (void *)old); 8522 } 8523 8524 if (old->dtm_count != 0) { 8525 mutex_exit(&dtrace_lock); 8526 mutex_exit(&dtrace_meta_lock); 8527 return (EBUSY); 8528 } 8529 8530 *pp = NULL; 8531 8532 mutex_exit(&dtrace_lock); 8533 mutex_exit(&dtrace_meta_lock); 8534 8535 kmem_free(old->dtm_name, strlen(old->dtm_name) + 1); 8536 kmem_free(old, sizeof (dtrace_meta_t)); 8537 8538 return (0); 8539} 8540 8541 8542/* 8543 * DTrace DIF Object Functions 8544 */ 8545static int 8546dtrace_difo_err(uint_t pc, const char *format, ...) 8547{ 8548 if (dtrace_err_verbose) { 8549 va_list alist; 8550 8551 (void) uprintf("dtrace DIF object error: [%u]: ", pc); 8552 va_start(alist, format); 8553 (void) vuprintf(format, alist); 8554 va_end(alist); 8555 } 8556 8557#ifdef DTRACE_ERRDEBUG 8558 dtrace_errdebug(format); 8559#endif 8560 return (1); 8561} 8562 8563/* 8564 * Validate a DTrace DIF object by checking the IR instructions. The following 8565 * rules are currently enforced by dtrace_difo_validate(): 8566 * 8567 * 1. Each instruction must have a valid opcode 8568 * 2. Each register, string, variable, or subroutine reference must be valid 8569 * 3. No instruction can modify register %r0 (must be zero) 8570 * 4. All instruction reserved bits must be set to zero 8571 * 5. The last instruction must be a "ret" instruction 8572 * 6. All branch targets must reference a valid instruction _after_ the branch 8573 */ 8574static int 8575dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs, 8576 cred_t *cr) 8577{ 8578 int err = 0, i; 8579 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err; 8580 int kcheckload; 8581 uint_t pc; 8582 8583 kcheckload = cr == NULL || 8584 (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0; 8585 8586 dp->dtdo_destructive = 0; 8587 8588 for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) { 8589 dif_instr_t instr = dp->dtdo_buf[pc]; 8590 8591 uint_t r1 = DIF_INSTR_R1(instr); 8592 uint_t r2 = DIF_INSTR_R2(instr); 8593 uint_t rd = DIF_INSTR_RD(instr); 8594 uint_t rs = DIF_INSTR_RS(instr); 8595 uint_t label = DIF_INSTR_LABEL(instr); 8596 uint_t v = DIF_INSTR_VAR(instr); 8597 uint_t subr = DIF_INSTR_SUBR(instr); 8598 uint_t type = DIF_INSTR_TYPE(instr); 8599 uint_t op = DIF_INSTR_OP(instr); 8600 8601 switch (op) { 8602 case DIF_OP_OR: 8603 case DIF_OP_XOR: 8604 case DIF_OP_AND: 8605 case DIF_OP_SLL: 8606 case DIF_OP_SRL: 8607 case DIF_OP_SRA: 8608 case DIF_OP_SUB: 8609 case DIF_OP_ADD: 8610 case DIF_OP_MUL: 8611 case DIF_OP_SDIV: 8612 case DIF_OP_UDIV: 8613 case DIF_OP_SREM: 8614 case DIF_OP_UREM: 8615 case DIF_OP_COPYS: 8616 if (r1 >= nregs) 8617 err += efunc(pc, "invalid register %u\n", r1); 8618 if (r2 >= nregs) 8619 err += efunc(pc, "invalid register %u\n", r2); 8620 if (rd >= nregs) 8621 err += efunc(pc, "invalid register %u\n", rd); 8622 if (rd == 0) 8623 err += efunc(pc, "cannot write to %r0\n"); 8624 break; 8625 case DIF_OP_NOT: 8626 case DIF_OP_MOV: 8627 case DIF_OP_ALLOCS: 8628 if (r1 >= nregs) 8629 err += efunc(pc, "invalid register %u\n", r1); 8630 if (r2 != 0) 8631 err += efunc(pc, "non-zero reserved bits\n"); 8632 if (rd >= nregs) 8633 err += efunc(pc, "invalid register %u\n", rd); 8634 if (rd == 0) 8635 err += efunc(pc, "cannot write to %r0\n"); 8636 break; 8637 case DIF_OP_LDSB: 8638 case DIF_OP_LDSH: 8639 case DIF_OP_LDSW: 8640 case DIF_OP_LDUB: 8641 case DIF_OP_LDUH: 8642 case DIF_OP_LDUW: 8643 case DIF_OP_LDX: 8644 if (r1 >= nregs) 8645 err += efunc(pc, "invalid register %u\n", r1); 8646 if (r2 != 0) 8647 err += efunc(pc, "non-zero reserved bits\n"); 8648 if (rd >= nregs) 8649 err += efunc(pc, "invalid register %u\n", rd); 8650 if (rd == 0) 8651 err += efunc(pc, "cannot write to %r0\n"); 8652 if (kcheckload) 8653 dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op + 8654 DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd); 8655 break; 8656 case DIF_OP_RLDSB: 8657 case DIF_OP_RLDSH: 8658 case DIF_OP_RLDSW: 8659 case DIF_OP_RLDUB: 8660 case DIF_OP_RLDUH: 8661 case DIF_OP_RLDUW: 8662 case DIF_OP_RLDX: 8663 if (r1 >= nregs) 8664 err += efunc(pc, "invalid register %u\n", r1); 8665 if (r2 != 0) 8666 err += efunc(pc, "non-zero reserved bits\n"); 8667 if (rd >= nregs) 8668 err += efunc(pc, "invalid register %u\n", rd); 8669 if (rd == 0) 8670 err += efunc(pc, "cannot write to %r0\n"); 8671 break; 8672 case DIF_OP_ULDSB: 8673 case DIF_OP_ULDSH: 8674 case DIF_OP_ULDSW: 8675 case DIF_OP_ULDUB: 8676 case DIF_OP_ULDUH: 8677 case DIF_OP_ULDUW: 8678 case DIF_OP_ULDX: 8679 if (r1 >= nregs) 8680 err += efunc(pc, "invalid register %u\n", r1); 8681 if (r2 != 0) 8682 err += efunc(pc, "non-zero reserved bits\n"); 8683 if (rd >= nregs) 8684 err += efunc(pc, "invalid register %u\n", rd); 8685 if (rd == 0) 8686 err += efunc(pc, "cannot write to %r0\n"); 8687 break; 8688 case DIF_OP_STB: 8689 case DIF_OP_STH: 8690 case DIF_OP_STW: 8691 case DIF_OP_STX: 8692 if (r1 >= nregs) 8693 err += efunc(pc, "invalid register %u\n", r1); 8694 if (r2 != 0) 8695 err += efunc(pc, "non-zero reserved bits\n"); 8696 if (rd >= nregs) 8697 err += efunc(pc, "invalid register %u\n", rd); 8698 if (rd == 0) 8699 err += efunc(pc, "cannot write to 0 address\n"); 8700 break; 8701 case DIF_OP_CMP: 8702 case DIF_OP_SCMP: 8703 if (r1 >= nregs) 8704 err += efunc(pc, "invalid register %u\n", r1); 8705 if (r2 >= nregs) 8706 err += efunc(pc, "invalid register %u\n", r2); 8707 if (rd != 0) 8708 err += efunc(pc, "non-zero reserved bits\n"); 8709 break; 8710 case DIF_OP_TST: 8711 if (r1 >= nregs) 8712 err += efunc(pc, "invalid register %u\n", r1); 8713 if (r2 != 0 || rd != 0) 8714 err += efunc(pc, "non-zero reserved bits\n"); 8715 break; 8716 case DIF_OP_BA: 8717 case DIF_OP_BE: 8718 case DIF_OP_BNE: 8719 case DIF_OP_BG: 8720 case DIF_OP_BGU: 8721 case DIF_OP_BGE: 8722 case DIF_OP_BGEU: 8723 case DIF_OP_BL: 8724 case DIF_OP_BLU: 8725 case DIF_OP_BLE: 8726 case DIF_OP_BLEU: 8727 if (label >= dp->dtdo_len) { 8728 err += efunc(pc, "invalid branch target %u\n", 8729 label); 8730 } 8731 if (label <= pc) { 8732 err += efunc(pc, "backward branch to %u\n", 8733 label); 8734 } 8735 break; 8736 case DIF_OP_RET: 8737 if (r1 != 0 || r2 != 0) 8738 err += efunc(pc, "non-zero reserved bits\n"); 8739 if (rd >= nregs) 8740 err += efunc(pc, "invalid register %u\n", rd); 8741 break; 8742 case DIF_OP_NOP: 8743 case DIF_OP_POPTS: 8744 case DIF_OP_FLUSHTS: 8745 if (r1 != 0 || r2 != 0 || rd != 0) 8746 err += efunc(pc, "non-zero reserved bits\n"); 8747 break; 8748 case DIF_OP_SETX: 8749 if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) { 8750 err += efunc(pc, "invalid integer ref %u\n", 8751 DIF_INSTR_INTEGER(instr)); 8752 } 8753 if (rd >= nregs) 8754 err += efunc(pc, "invalid register %u\n", rd); 8755 if (rd == 0) 8756 err += efunc(pc, "cannot write to %r0\n"); 8757 break; 8758 case DIF_OP_SETS: 8759 if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) { 8760 err += efunc(pc, "invalid string ref %u\n", 8761 DIF_INSTR_STRING(instr)); 8762 } 8763 if (rd >= nregs) 8764 err += efunc(pc, "invalid register %u\n", rd); 8765 if (rd == 0) 8766 err += efunc(pc, "cannot write to %r0\n"); 8767 break; 8768 case DIF_OP_LDGA: 8769 case DIF_OP_LDTA: 8770 if (r1 > DIF_VAR_ARRAY_MAX) 8771 err += efunc(pc, "invalid array %u\n", r1); 8772 if (r2 >= nregs) 8773 err += efunc(pc, "invalid register %u\n", r2); 8774 if (rd >= nregs) 8775 err += efunc(pc, "invalid register %u\n", rd); 8776 if (rd == 0) 8777 err += efunc(pc, "cannot write to %r0\n"); 8778 break; 8779 case DIF_OP_LDGS: 8780 case DIF_OP_LDTS: 8781 case DIF_OP_LDLS: 8782 case DIF_OP_LDGAA: 8783 case DIF_OP_LDTAA: 8784 if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX) 8785 err += efunc(pc, "invalid variable %u\n", v); 8786 if (rd >= nregs) 8787 err += efunc(pc, "invalid register %u\n", rd); 8788 if (rd == 0) 8789 err += efunc(pc, "cannot write to %r0\n"); 8790 break; 8791 case DIF_OP_STGS: 8792 case DIF_OP_STTS: 8793 case DIF_OP_STLS: 8794 case DIF_OP_STGAA: 8795 case DIF_OP_STTAA: 8796 if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX) 8797 err += efunc(pc, "invalid variable %u\n", v); 8798 if (rs >= nregs) 8799 err += efunc(pc, "invalid register %u\n", rd); 8800 break; 8801 case DIF_OP_CALL: 8802 if (subr > DIF_SUBR_MAX) 8803 err += efunc(pc, "invalid subr %u\n", subr); 8804 if (rd >= nregs) 8805 err += efunc(pc, "invalid register %u\n", rd); 8806 if (rd == 0) 8807 err += efunc(pc, "cannot write to %r0\n"); 8808 8809 if (subr == DIF_SUBR_COPYOUT || 8810 subr == DIF_SUBR_COPYOUTSTR) { 8811 dp->dtdo_destructive = 1; 8812 } 8813 break; 8814 case DIF_OP_PUSHTR: 8815 if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF) 8816 err += efunc(pc, "invalid ref type %u\n", type); 8817 if (r2 >= nregs) 8818 err += efunc(pc, "invalid register %u\n", r2); 8819 if (rs >= nregs) 8820 err += efunc(pc, "invalid register %u\n", rs); 8821 break; 8822 case DIF_OP_PUSHTV: 8823 if (type != DIF_TYPE_CTF) 8824 err += efunc(pc, "invalid val type %u\n", type); 8825 if (r2 >= nregs) 8826 err += efunc(pc, "invalid register %u\n", r2); 8827 if (rs >= nregs) 8828 err += efunc(pc, "invalid register %u\n", rs); 8829 break; 8830 default: 8831 err += efunc(pc, "invalid opcode %u\n", 8832 DIF_INSTR_OP(instr)); 8833 } 8834 } 8835 8836 if (dp->dtdo_len != 0 && 8837 DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) { 8838 err += efunc(dp->dtdo_len - 1, 8839 "expected 'ret' as last DIF instruction\n"); 8840 } 8841 8842 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) { 8843 /* 8844 * If we're not returning by reference, the size must be either 8845 * 0 or the size of one of the base types. 8846 */ 8847 switch (dp->dtdo_rtype.dtdt_size) { 8848 case 0: 8849 case sizeof (uint8_t): 8850 case sizeof (uint16_t): 8851 case sizeof (uint32_t): 8852 case sizeof (uint64_t): 8853 break; 8854 8855 default: 8856 err += efunc(dp->dtdo_len - 1, "bad return size\n"); 8857 } 8858 } 8859 8860 for (i = 0; i < dp->dtdo_varlen && err == 0; i++) { 8861 dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL; 8862 dtrace_diftype_t *vt, *et; 8863 uint_t id, ndx; 8864 8865 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL && 8866 v->dtdv_scope != DIFV_SCOPE_THREAD && 8867 v->dtdv_scope != DIFV_SCOPE_LOCAL) { 8868 err += efunc(i, "unrecognized variable scope %d\n", 8869 v->dtdv_scope); 8870 break; 8871 } 8872 8873 if (v->dtdv_kind != DIFV_KIND_ARRAY && 8874 v->dtdv_kind != DIFV_KIND_SCALAR) { 8875 err += efunc(i, "unrecognized variable type %d\n", 8876 v->dtdv_kind); 8877 break; 8878 } 8879 8880 if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) { 8881 err += efunc(i, "%d exceeds variable id limit\n", id); 8882 break; 8883 } 8884 8885 if (id < DIF_VAR_OTHER_UBASE) 8886 continue; 8887 8888 /* 8889 * For user-defined variables, we need to check that this 8890 * definition is identical to any previous definition that we 8891 * encountered. 8892 */ 8893 ndx = id - DIF_VAR_OTHER_UBASE; 8894 8895 switch (v->dtdv_scope) { 8896 case DIFV_SCOPE_GLOBAL: 8897 if (ndx < vstate->dtvs_nglobals) { 8898 dtrace_statvar_t *svar; 8899 8900 if ((svar = vstate->dtvs_globals[ndx]) != NULL) 8901 existing = &svar->dtsv_var; 8902 } 8903 8904 break; 8905 8906 case DIFV_SCOPE_THREAD: 8907 if (ndx < vstate->dtvs_ntlocals) 8908 existing = &vstate->dtvs_tlocals[ndx]; 8909 break; 8910 8911 case DIFV_SCOPE_LOCAL: 8912 if (ndx < vstate->dtvs_nlocals) { 8913 dtrace_statvar_t *svar; 8914 8915 if ((svar = vstate->dtvs_locals[ndx]) != NULL) 8916 existing = &svar->dtsv_var; 8917 } 8918 8919 break; 8920 } 8921 8922 vt = &v->dtdv_type; 8923 8924 if (vt->dtdt_flags & DIF_TF_BYREF) { 8925 if (vt->dtdt_size == 0) { 8926 err += efunc(i, "zero-sized variable\n"); 8927 break; 8928 } 8929 8930 if (v->dtdv_scope == DIFV_SCOPE_GLOBAL && 8931 vt->dtdt_size > dtrace_global_maxsize) { 8932 err += efunc(i, "oversized by-ref global\n"); 8933 break; 8934 } 8935 } 8936 8937 if (existing == NULL || existing->dtdv_id == 0) 8938 continue; 8939 8940 ASSERT(existing->dtdv_id == v->dtdv_id); 8941 ASSERT(existing->dtdv_scope == v->dtdv_scope); 8942 8943 if (existing->dtdv_kind != v->dtdv_kind) 8944 err += efunc(i, "%d changed variable kind\n", id); 8945 8946 et = &existing->dtdv_type; 8947 8948 if (vt->dtdt_flags != et->dtdt_flags) { 8949 err += efunc(i, "%d changed variable type flags\n", id); 8950 break; 8951 } 8952 8953 if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) { 8954 err += efunc(i, "%d changed variable type size\n", id); 8955 break; 8956 } 8957 } 8958 8959 return (err); 8960} 8961 8962/* 8963 * Validate a DTrace DIF object that it is to be used as a helper. Helpers 8964 * are much more constrained than normal DIFOs. Specifically, they may 8965 * not: 8966 * 8967 * 1. Make calls to subroutines other than copyin(), copyinstr() or 8968 * miscellaneous string routines 8969 * 2. Access DTrace variables other than the args[] array, and the 8970 * curthread, pid, ppid, tid, execname, zonename, uid and gid variables. 8971 * 3. Have thread-local variables. 8972 * 4. Have dynamic variables. 8973 */ 8974static int 8975dtrace_difo_validate_helper(dtrace_difo_t *dp) 8976{ 8977 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err; 8978 int err = 0; 8979 uint_t pc; 8980 8981 for (pc = 0; pc < dp->dtdo_len; pc++) { 8982 dif_instr_t instr = dp->dtdo_buf[pc]; 8983 8984 uint_t v = DIF_INSTR_VAR(instr); 8985 uint_t subr = DIF_INSTR_SUBR(instr); 8986 uint_t op = DIF_INSTR_OP(instr); 8987 8988 switch (op) { 8989 case DIF_OP_OR: 8990 case DIF_OP_XOR: 8991 case DIF_OP_AND: 8992 case DIF_OP_SLL: 8993 case DIF_OP_SRL: 8994 case DIF_OP_SRA: 8995 case DIF_OP_SUB: 8996 case DIF_OP_ADD: 8997 case DIF_OP_MUL: 8998 case DIF_OP_SDIV: 8999 case DIF_OP_UDIV: 9000 case DIF_OP_SREM: 9001 case DIF_OP_UREM: 9002 case DIF_OP_COPYS: 9003 case DIF_OP_NOT: 9004 case DIF_OP_MOV: 9005 case DIF_OP_RLDSB: 9006 case DIF_OP_RLDSH: 9007 case DIF_OP_RLDSW: 9008 case DIF_OP_RLDUB: 9009 case DIF_OP_RLDUH: 9010 case DIF_OP_RLDUW: 9011 case DIF_OP_RLDX: 9012 case DIF_OP_ULDSB: 9013 case DIF_OP_ULDSH: 9014 case DIF_OP_ULDSW: 9015 case DIF_OP_ULDUB: 9016 case DIF_OP_ULDUH: 9017 case DIF_OP_ULDUW: 9018 case DIF_OP_ULDX: 9019 case DIF_OP_STB: 9020 case DIF_OP_STH: 9021 case DIF_OP_STW: 9022 case DIF_OP_STX: 9023 case DIF_OP_ALLOCS: 9024 case DIF_OP_CMP: 9025 case DIF_OP_SCMP: 9026 case DIF_OP_TST: 9027 case DIF_OP_BA: 9028 case DIF_OP_BE: 9029 case DIF_OP_BNE: 9030 case DIF_OP_BG: 9031 case DIF_OP_BGU: 9032 case DIF_OP_BGE: 9033 case DIF_OP_BGEU: 9034 case DIF_OP_BL: 9035 case DIF_OP_BLU: 9036 case DIF_OP_BLE: 9037 case DIF_OP_BLEU: 9038 case DIF_OP_RET: 9039 case DIF_OP_NOP: 9040 case DIF_OP_POPTS: 9041 case DIF_OP_FLUSHTS: 9042 case DIF_OP_SETX: 9043 case DIF_OP_SETS: 9044 case DIF_OP_LDGA: 9045 case DIF_OP_LDLS: 9046 case DIF_OP_STGS: 9047 case DIF_OP_STLS: 9048 case DIF_OP_PUSHTR: 9049 case DIF_OP_PUSHTV: 9050 break; 9051 9052 case DIF_OP_LDGS: 9053 if (v >= DIF_VAR_OTHER_UBASE) 9054 break; 9055 9056 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) 9057 break; 9058 9059 if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID || 9060 v == DIF_VAR_PPID || v == DIF_VAR_TID || 9061 v == DIF_VAR_EXECARGS || 9062 v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME || 9063 v == DIF_VAR_UID || v == DIF_VAR_GID) 9064 break; 9065 9066 err += efunc(pc, "illegal variable %u\n", v); 9067 break; 9068 9069 case DIF_OP_LDTA: 9070 case DIF_OP_LDTS: 9071 case DIF_OP_LDGAA: 9072 case DIF_OP_LDTAA: 9073 err += efunc(pc, "illegal dynamic variable load\n"); 9074 break; 9075 9076 case DIF_OP_STTS: 9077 case DIF_OP_STGAA: 9078 case DIF_OP_STTAA: 9079 err += efunc(pc, "illegal dynamic variable store\n"); 9080 break; 9081 9082 case DIF_OP_CALL: 9083 if (subr == DIF_SUBR_ALLOCA || 9084 subr == DIF_SUBR_BCOPY || 9085 subr == DIF_SUBR_COPYIN || 9086 subr == DIF_SUBR_COPYINTO || 9087 subr == DIF_SUBR_COPYINSTR || 9088 subr == DIF_SUBR_INDEX || 9089 subr == DIF_SUBR_INET_NTOA || 9090 subr == DIF_SUBR_INET_NTOA6 || 9091 subr == DIF_SUBR_INET_NTOP || 9092 subr == DIF_SUBR_LLTOSTR || 9093 subr == DIF_SUBR_RINDEX || 9094 subr == DIF_SUBR_STRCHR || 9095 subr == DIF_SUBR_STRJOIN || 9096 subr == DIF_SUBR_STRRCHR || 9097 subr == DIF_SUBR_STRSTR || 9098 subr == DIF_SUBR_HTONS || 9099 subr == DIF_SUBR_HTONL || 9100 subr == DIF_SUBR_HTONLL || 9101 subr == DIF_SUBR_NTOHS || 9102 subr == DIF_SUBR_NTOHL || 9103 subr == DIF_SUBR_NTOHLL || 9104 subr == DIF_SUBR_MEMREF || 9105 subr == DIF_SUBR_TYPEREF) 9106 break; 9107 9108 err += efunc(pc, "invalid subr %u\n", subr); 9109 break; 9110 9111 default: 9112 err += efunc(pc, "invalid opcode %u\n", 9113 DIF_INSTR_OP(instr)); 9114 } 9115 } 9116 9117 return (err); 9118} 9119 9120/* 9121 * Returns 1 if the expression in the DIF object can be cached on a per-thread 9122 * basis; 0 if not. 9123 */ 9124static int 9125dtrace_difo_cacheable(dtrace_difo_t *dp) 9126{ 9127 int i; 9128 9129 if (dp == NULL) 9130 return (0); 9131 9132 for (i = 0; i < dp->dtdo_varlen; i++) { 9133 dtrace_difv_t *v = &dp->dtdo_vartab[i]; 9134 9135 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL) 9136 continue; 9137 9138 switch (v->dtdv_id) { 9139 case DIF_VAR_CURTHREAD: 9140 case DIF_VAR_PID: 9141 case DIF_VAR_TID: 9142 case DIF_VAR_EXECARGS: 9143 case DIF_VAR_EXECNAME: 9144 case DIF_VAR_ZONENAME: 9145 break; 9146 9147 default: 9148 return (0); 9149 } 9150 } 9151 9152 /* 9153 * This DIF object may be cacheable. Now we need to look for any 9154 * array loading instructions, any memory loading instructions, or 9155 * any stores to thread-local variables. 9156 */ 9157 for (i = 0; i < dp->dtdo_len; i++) { 9158 uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]); 9159 9160 if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) || 9161 (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) || 9162 (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) || 9163 op == DIF_OP_LDGA || op == DIF_OP_STTS) 9164 return (0); 9165 } 9166 9167 return (1); 9168} 9169 9170static void 9171dtrace_difo_hold(dtrace_difo_t *dp) 9172{ 9173 int i; 9174 9175 ASSERT(MUTEX_HELD(&dtrace_lock)); 9176 9177 dp->dtdo_refcnt++; 9178 ASSERT(dp->dtdo_refcnt != 0); 9179 9180 /* 9181 * We need to check this DIF object for references to the variable 9182 * DIF_VAR_VTIMESTAMP. 9183 */ 9184 for (i = 0; i < dp->dtdo_varlen; i++) { 9185 dtrace_difv_t *v = &dp->dtdo_vartab[i]; 9186 9187 if (v->dtdv_id != DIF_VAR_VTIMESTAMP) 9188 continue; 9189 9190 if (dtrace_vtime_references++ == 0) 9191 dtrace_vtime_enable(); 9192 } 9193} 9194 9195/* 9196 * This routine calculates the dynamic variable chunksize for a given DIF 9197 * object. The calculation is not fool-proof, and can probably be tricked by 9198 * malicious DIF -- but it works for all compiler-generated DIF. Because this 9199 * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail 9200 * if a dynamic variable size exceeds the chunksize. 9201 */ 9202static void 9203dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate) 9204{ 9205 uint64_t sval = 0; 9206 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */ 9207 const dif_instr_t *text = dp->dtdo_buf; 9208 uint_t pc, srd = 0; 9209 uint_t ttop = 0; 9210 size_t size, ksize; 9211 uint_t id, i; 9212 9213 for (pc = 0; pc < dp->dtdo_len; pc++) { 9214 dif_instr_t instr = text[pc]; 9215 uint_t op = DIF_INSTR_OP(instr); 9216 uint_t rd = DIF_INSTR_RD(instr); 9217 uint_t r1 = DIF_INSTR_R1(instr); 9218 uint_t nkeys = 0; 9219 uchar_t scope = 0; 9220 9221 dtrace_key_t *key = tupregs; 9222 9223 switch (op) { 9224 case DIF_OP_SETX: 9225 sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)]; 9226 srd = rd; 9227 continue; 9228 9229 case DIF_OP_STTS: 9230 key = &tupregs[DIF_DTR_NREGS]; 9231 key[0].dttk_size = 0; 9232 key[1].dttk_size = 0; 9233 nkeys = 2; 9234 scope = DIFV_SCOPE_THREAD; 9235 break; 9236 9237 case DIF_OP_STGAA: 9238 case DIF_OP_STTAA: 9239 nkeys = ttop; 9240 9241 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) 9242 key[nkeys++].dttk_size = 0; 9243 9244 key[nkeys++].dttk_size = 0; 9245 9246 if (op == DIF_OP_STTAA) { 9247 scope = DIFV_SCOPE_THREAD; 9248 } else { 9249 scope = DIFV_SCOPE_GLOBAL; 9250 } 9251 9252 break; 9253 9254 case DIF_OP_PUSHTR: 9255 if (ttop == DIF_DTR_NREGS) 9256 return; 9257 9258 if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) { 9259 /* 9260 * If the register for the size of the "pushtr" 9261 * is %r0 (or the value is 0) and the type is 9262 * a string, we'll use the system-wide default 9263 * string size. 9264 */ 9265 tupregs[ttop++].dttk_size = 9266 dtrace_strsize_default; 9267 } else { 9268 if (srd == 0) 9269 return; 9270 9271 tupregs[ttop++].dttk_size = sval; 9272 } 9273 9274 break; 9275 9276 case DIF_OP_PUSHTV: 9277 if (ttop == DIF_DTR_NREGS) 9278 return; 9279 9280 tupregs[ttop++].dttk_size = 0; 9281 break; 9282 9283 case DIF_OP_FLUSHTS: 9284 ttop = 0; 9285 break; 9286 9287 case DIF_OP_POPTS: 9288 if (ttop != 0) 9289 ttop--; 9290 break; 9291 } 9292 9293 sval = 0; 9294 srd = 0; 9295 9296 if (nkeys == 0) 9297 continue; 9298 9299 /* 9300 * We have a dynamic variable allocation; calculate its size. 9301 */ 9302 for (ksize = 0, i = 0; i < nkeys; i++) 9303 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t)); 9304 9305 size = sizeof (dtrace_dynvar_t); 9306 size += sizeof (dtrace_key_t) * (nkeys - 1); 9307 size += ksize; 9308 9309 /* 9310 * Now we need to determine the size of the stored data. 9311 */ 9312 id = DIF_INSTR_VAR(instr); 9313 9314 for (i = 0; i < dp->dtdo_varlen; i++) { 9315 dtrace_difv_t *v = &dp->dtdo_vartab[i]; 9316 9317 if (v->dtdv_id == id && v->dtdv_scope == scope) { 9318 size += v->dtdv_type.dtdt_size; 9319 break; 9320 } 9321 } 9322 9323 if (i == dp->dtdo_varlen) 9324 return; 9325 9326 /* 9327 * We have the size. If this is larger than the chunk size 9328 * for our dynamic variable state, reset the chunk size. 9329 */ 9330 size = P2ROUNDUP(size, sizeof (uint64_t)); 9331 9332 if (size > vstate->dtvs_dynvars.dtds_chunksize) 9333 vstate->dtvs_dynvars.dtds_chunksize = size; 9334 } 9335} 9336 9337static void 9338dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate) 9339{ 9340 int i, oldsvars, osz, nsz, otlocals, ntlocals; 9341 uint_t id; 9342 9343 ASSERT(MUTEX_HELD(&dtrace_lock)); 9344 ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0); 9345 9346 for (i = 0; i < dp->dtdo_varlen; i++) { 9347 dtrace_difv_t *v = &dp->dtdo_vartab[i]; 9348 dtrace_statvar_t *svar, ***svarp = NULL; 9349 size_t dsize = 0; 9350 uint8_t scope = v->dtdv_scope; 9351 int *np = NULL; 9352 9353 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE) 9354 continue; 9355 9356 id -= DIF_VAR_OTHER_UBASE; 9357 9358 switch (scope) { 9359 case DIFV_SCOPE_THREAD: 9360 while (id >= (otlocals = vstate->dtvs_ntlocals)) { 9361 dtrace_difv_t *tlocals; 9362 9363 if ((ntlocals = (otlocals << 1)) == 0) 9364 ntlocals = 1; 9365 9366 osz = otlocals * sizeof (dtrace_difv_t); 9367 nsz = ntlocals * sizeof (dtrace_difv_t); 9368 9369 tlocals = kmem_zalloc(nsz, KM_SLEEP); 9370 9371 if (osz != 0) { 9372 bcopy(vstate->dtvs_tlocals, 9373 tlocals, osz); 9374 kmem_free(vstate->dtvs_tlocals, osz); 9375 } 9376 9377 vstate->dtvs_tlocals = tlocals; 9378 vstate->dtvs_ntlocals = ntlocals; 9379 } 9380 9381 vstate->dtvs_tlocals[id] = *v; 9382 continue; 9383 9384 case DIFV_SCOPE_LOCAL: 9385 np = &vstate->dtvs_nlocals; 9386 svarp = &vstate->dtvs_locals; 9387 9388 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) 9389 dsize = NCPU * (v->dtdv_type.dtdt_size + 9390 sizeof (uint64_t)); 9391 else 9392 dsize = NCPU * sizeof (uint64_t); 9393 9394 break; 9395 9396 case DIFV_SCOPE_GLOBAL: 9397 np = &vstate->dtvs_nglobals; 9398 svarp = &vstate->dtvs_globals; 9399 9400 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) 9401 dsize = v->dtdv_type.dtdt_size + 9402 sizeof (uint64_t); 9403 9404 break; 9405 9406 default: 9407 ASSERT(0); 9408 } 9409 9410 while (id >= (oldsvars = *np)) { 9411 dtrace_statvar_t **statics; 9412 int newsvars, oldsize, newsize; 9413 9414 if ((newsvars = (oldsvars << 1)) == 0) 9415 newsvars = 1; 9416 9417 oldsize = oldsvars * sizeof (dtrace_statvar_t *); 9418 newsize = newsvars * sizeof (dtrace_statvar_t *); 9419 9420 statics = kmem_zalloc(newsize, KM_SLEEP); 9421 9422 if (oldsize != 0) { 9423 bcopy(*svarp, statics, oldsize); 9424 kmem_free(*svarp, oldsize); 9425 } 9426 9427 *svarp = statics; 9428 *np = newsvars; 9429 } 9430 9431 if ((svar = (*svarp)[id]) == NULL) { 9432 svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP); 9433 svar->dtsv_var = *v; 9434 9435 if ((svar->dtsv_size = dsize) != 0) { 9436 svar->dtsv_data = (uint64_t)(uintptr_t) 9437 kmem_zalloc(dsize, KM_SLEEP); 9438 } 9439 9440 (*svarp)[id] = svar; 9441 } 9442 9443 svar->dtsv_refcnt++; 9444 } 9445 9446 dtrace_difo_chunksize(dp, vstate); 9447 dtrace_difo_hold(dp); 9448} 9449 9450static dtrace_difo_t * 9451dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate) 9452{ 9453 dtrace_difo_t *new; 9454 size_t sz; 9455 9456 ASSERT(dp->dtdo_buf != NULL); 9457 ASSERT(dp->dtdo_refcnt != 0); 9458 9459 new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP); 9460 9461 ASSERT(dp->dtdo_buf != NULL); 9462 sz = dp->dtdo_len * sizeof (dif_instr_t); 9463 new->dtdo_buf = kmem_alloc(sz, KM_SLEEP); 9464 bcopy(dp->dtdo_buf, new->dtdo_buf, sz); 9465 new->dtdo_len = dp->dtdo_len; 9466 9467 if (dp->dtdo_strtab != NULL) { 9468 ASSERT(dp->dtdo_strlen != 0); 9469 new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP); 9470 bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen); 9471 new->dtdo_strlen = dp->dtdo_strlen; 9472 } 9473 9474 if (dp->dtdo_inttab != NULL) { 9475 ASSERT(dp->dtdo_intlen != 0); 9476 sz = dp->dtdo_intlen * sizeof (uint64_t); 9477 new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP); 9478 bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz); 9479 new->dtdo_intlen = dp->dtdo_intlen; 9480 } 9481 9482 if (dp->dtdo_vartab != NULL) { 9483 ASSERT(dp->dtdo_varlen != 0); 9484 sz = dp->dtdo_varlen * sizeof (dtrace_difv_t); 9485 new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP); 9486 bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz); 9487 new->dtdo_varlen = dp->dtdo_varlen; 9488 } 9489 9490 dtrace_difo_init(new, vstate); 9491 return (new); 9492} 9493 9494static void 9495dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate) 9496{ 9497 int i; 9498 9499 ASSERT(dp->dtdo_refcnt == 0); 9500 9501 for (i = 0; i < dp->dtdo_varlen; i++) { 9502 dtrace_difv_t *v = &dp->dtdo_vartab[i]; 9503 dtrace_statvar_t *svar, **svarp = NULL; 9504 uint_t id; 9505 uint8_t scope = v->dtdv_scope; 9506 int *np = NULL; 9507 9508 switch (scope) { 9509 case DIFV_SCOPE_THREAD: 9510 continue; 9511 9512 case DIFV_SCOPE_LOCAL: 9513 np = &vstate->dtvs_nlocals; 9514 svarp = vstate->dtvs_locals; 9515 break; 9516 9517 case DIFV_SCOPE_GLOBAL: 9518 np = &vstate->dtvs_nglobals; 9519 svarp = vstate->dtvs_globals; 9520 break; 9521 9522 default: 9523 ASSERT(0); 9524 } 9525 9526 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE) 9527 continue; 9528 9529 id -= DIF_VAR_OTHER_UBASE; 9530 ASSERT(id < *np); 9531 9532 svar = svarp[id]; 9533 ASSERT(svar != NULL); 9534 ASSERT(svar->dtsv_refcnt > 0); 9535 9536 if (--svar->dtsv_refcnt > 0) 9537 continue; 9538 9539 if (svar->dtsv_size != 0) { 9540 ASSERT(svar->dtsv_data != 0); 9541 kmem_free((void *)(uintptr_t)svar->dtsv_data, 9542 svar->dtsv_size); 9543 } 9544 9545 kmem_free(svar, sizeof (dtrace_statvar_t)); 9546 svarp[id] = NULL; 9547 } 9548 9549 if (dp->dtdo_buf != NULL) 9550 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t)); 9551 if (dp->dtdo_inttab != NULL) 9552 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t)); 9553 if (dp->dtdo_strtab != NULL) 9554 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen); 9555 if (dp->dtdo_vartab != NULL) 9556 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t)); 9557 9558 kmem_free(dp, sizeof (dtrace_difo_t)); 9559} 9560 9561static void 9562dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate) 9563{ 9564 int i; 9565 9566 ASSERT(MUTEX_HELD(&dtrace_lock)); 9567 ASSERT(dp->dtdo_refcnt != 0); 9568 9569 for (i = 0; i < dp->dtdo_varlen; i++) { 9570 dtrace_difv_t *v = &dp->dtdo_vartab[i]; 9571 9572 if (v->dtdv_id != DIF_VAR_VTIMESTAMP) 9573 continue; 9574 9575 ASSERT(dtrace_vtime_references > 0); 9576 if (--dtrace_vtime_references == 0) 9577 dtrace_vtime_disable(); 9578 } 9579 9580 if (--dp->dtdo_refcnt == 0) 9581 dtrace_difo_destroy(dp, vstate); 9582} 9583 9584/* 9585 * DTrace Format Functions 9586 */ 9587static uint16_t 9588dtrace_format_add(dtrace_state_t *state, char *str) 9589{ 9590 char *fmt, **new; 9591 uint16_t ndx, len = strlen(str) + 1; 9592 9593 fmt = kmem_zalloc(len, KM_SLEEP); 9594 bcopy(str, fmt, len); 9595 9596 for (ndx = 0; ndx < state->dts_nformats; ndx++) { 9597 if (state->dts_formats[ndx] == NULL) { 9598 state->dts_formats[ndx] = fmt; 9599 return (ndx + 1); 9600 } 9601 } 9602 9603 if (state->dts_nformats == USHRT_MAX) { 9604 /* 9605 * This is only likely if a denial-of-service attack is being 9606 * attempted. As such, it's okay to fail silently here. 9607 */ 9608 kmem_free(fmt, len); 9609 return (0); 9610 } 9611 9612 /* 9613 * For simplicity, we always resize the formats array to be exactly the 9614 * number of formats. 9615 */ 9616 ndx = state->dts_nformats++; 9617 new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP); 9618 9619 if (state->dts_formats != NULL) { 9620 ASSERT(ndx != 0); 9621 bcopy(state->dts_formats, new, ndx * sizeof (char *)); 9622 kmem_free(state->dts_formats, ndx * sizeof (char *)); 9623 } 9624 9625 state->dts_formats = new; 9626 state->dts_formats[ndx] = fmt; 9627 9628 return (ndx + 1); 9629} 9630 9631static void 9632dtrace_format_remove(dtrace_state_t *state, uint16_t format) 9633{ 9634 char *fmt; 9635 9636 ASSERT(state->dts_formats != NULL); 9637 ASSERT(format <= state->dts_nformats); 9638 ASSERT(state->dts_formats[format - 1] != NULL); 9639 9640 fmt = state->dts_formats[format - 1]; 9641 kmem_free(fmt, strlen(fmt) + 1); 9642 state->dts_formats[format - 1] = NULL; 9643} 9644 9645static void 9646dtrace_format_destroy(dtrace_state_t *state) 9647{ 9648 int i; 9649 9650 if (state->dts_nformats == 0) { 9651 ASSERT(state->dts_formats == NULL); 9652 return; 9653 } 9654 9655 ASSERT(state->dts_formats != NULL); 9656 9657 for (i = 0; i < state->dts_nformats; i++) { 9658 char *fmt = state->dts_formats[i]; 9659 9660 if (fmt == NULL) 9661 continue; 9662 9663 kmem_free(fmt, strlen(fmt) + 1); 9664 } 9665 9666 kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *)); 9667 state->dts_nformats = 0; 9668 state->dts_formats = NULL; 9669} 9670 9671/* 9672 * DTrace Predicate Functions 9673 */ 9674static dtrace_predicate_t * 9675dtrace_predicate_create(dtrace_difo_t *dp) 9676{ 9677 dtrace_predicate_t *pred; 9678 9679 ASSERT(MUTEX_HELD(&dtrace_lock)); 9680 ASSERT(dp->dtdo_refcnt != 0); 9681 9682 pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP); 9683 pred->dtp_difo = dp; 9684 pred->dtp_refcnt = 1; 9685 9686 if (!dtrace_difo_cacheable(dp)) 9687 return (pred); 9688 9689 if (dtrace_predcache_id == DTRACE_CACHEIDNONE) { 9690 /* 9691 * This is only theoretically possible -- we have had 2^32 9692 * cacheable predicates on this machine. We cannot allow any 9693 * more predicates to become cacheable: as unlikely as it is, 9694 * there may be a thread caching a (now stale) predicate cache 9695 * ID. (N.B.: the temptation is being successfully resisted to 9696 * have this cmn_err() "Holy shit -- we executed this code!") 9697 */ 9698 return (pred); 9699 } 9700 9701 pred->dtp_cacheid = dtrace_predcache_id++; 9702 9703 return (pred); 9704} 9705 9706static void 9707dtrace_predicate_hold(dtrace_predicate_t *pred) 9708{ 9709 ASSERT(MUTEX_HELD(&dtrace_lock)); 9710 ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0); 9711 ASSERT(pred->dtp_refcnt > 0); 9712 9713 pred->dtp_refcnt++; 9714} 9715 9716static void 9717dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate) 9718{ 9719 dtrace_difo_t *dp = pred->dtp_difo; 9720 9721 ASSERT(MUTEX_HELD(&dtrace_lock)); 9722 ASSERT(dp != NULL && dp->dtdo_refcnt != 0); 9723 ASSERT(pred->dtp_refcnt > 0); 9724 9725 if (--pred->dtp_refcnt == 0) { 9726 dtrace_difo_release(pred->dtp_difo, vstate); 9727 kmem_free(pred, sizeof (dtrace_predicate_t)); 9728 } 9729} 9730 9731/* 9732 * DTrace Action Description Functions 9733 */ 9734static dtrace_actdesc_t * 9735dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple, 9736 uint64_t uarg, uint64_t arg) 9737{ 9738 dtrace_actdesc_t *act; 9739 9740#if defined(sun) 9741 ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL && 9742 arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA)); 9743#endif 9744 9745 act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP); 9746 act->dtad_kind = kind; 9747 act->dtad_ntuple = ntuple; 9748 act->dtad_uarg = uarg; 9749 act->dtad_arg = arg; 9750 act->dtad_refcnt = 1; 9751 9752 return (act); 9753} 9754 9755static void 9756dtrace_actdesc_hold(dtrace_actdesc_t *act) 9757{ 9758 ASSERT(act->dtad_refcnt >= 1); 9759 act->dtad_refcnt++; 9760} 9761 9762static void 9763dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate) 9764{ 9765 dtrace_actkind_t kind = act->dtad_kind; 9766 dtrace_difo_t *dp; 9767 9768 ASSERT(act->dtad_refcnt >= 1); 9769 9770 if (--act->dtad_refcnt != 0) 9771 return; 9772 9773 if ((dp = act->dtad_difo) != NULL) 9774 dtrace_difo_release(dp, vstate); 9775 9776 if (DTRACEACT_ISPRINTFLIKE(kind)) { 9777 char *str = (char *)(uintptr_t)act->dtad_arg; 9778 9779#if defined(sun) 9780 ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) || 9781 (str == NULL && act->dtad_kind == DTRACEACT_PRINTA)); 9782#endif 9783 9784 if (str != NULL) 9785 kmem_free(str, strlen(str) + 1); 9786 } 9787 9788 kmem_free(act, sizeof (dtrace_actdesc_t)); 9789} 9790 9791/* 9792 * DTrace ECB Functions 9793 */ 9794static dtrace_ecb_t * 9795dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe) 9796{ 9797 dtrace_ecb_t *ecb; 9798 dtrace_epid_t epid; 9799 9800 ASSERT(MUTEX_HELD(&dtrace_lock)); 9801 9802 ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP); 9803 ecb->dte_predicate = NULL; 9804 ecb->dte_probe = probe; 9805 9806 /* 9807 * The default size is the size of the default action: recording 9808 * the header. 9809 */ 9810 ecb->dte_size = ecb->dte_needed = sizeof (dtrace_rechdr_t); 9811 ecb->dte_alignment = sizeof (dtrace_epid_t); 9812 9813 epid = state->dts_epid++; 9814 9815 if (epid - 1 >= state->dts_necbs) { 9816 dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs; 9817 int necbs = state->dts_necbs << 1; 9818 9819 ASSERT(epid == state->dts_necbs + 1); 9820 9821 if (necbs == 0) { 9822 ASSERT(oecbs == NULL); 9823 necbs = 1; 9824 } 9825 9826 ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP); 9827 9828 if (oecbs != NULL) 9829 bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs)); 9830 9831 dtrace_membar_producer(); 9832 state->dts_ecbs = ecbs; 9833 9834 if (oecbs != NULL) { 9835 /* 9836 * If this state is active, we must dtrace_sync() 9837 * before we can free the old dts_ecbs array: we're 9838 * coming in hot, and there may be active ring 9839 * buffer processing (which indexes into the dts_ecbs 9840 * array) on another CPU. 9841 */ 9842 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) 9843 dtrace_sync(); 9844 9845 kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs)); 9846 } 9847 9848 dtrace_membar_producer(); 9849 state->dts_necbs = necbs; 9850 } 9851 9852 ecb->dte_state = state; 9853 9854 ASSERT(state->dts_ecbs[epid - 1] == NULL); 9855 dtrace_membar_producer(); 9856 state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb; 9857 9858 return (ecb); 9859} 9860 9861static void 9862dtrace_ecb_enable(dtrace_ecb_t *ecb) 9863{ 9864 dtrace_probe_t *probe = ecb->dte_probe; 9865 9866 ASSERT(MUTEX_HELD(&cpu_lock)); 9867 ASSERT(MUTEX_HELD(&dtrace_lock)); 9868 ASSERT(ecb->dte_next == NULL); 9869 9870 if (probe == NULL) { 9871 /* 9872 * This is the NULL probe -- there's nothing to do. 9873 */ 9874 return; 9875 } 9876 9877 if (probe->dtpr_ecb == NULL) { 9878 dtrace_provider_t *prov = probe->dtpr_provider; 9879 9880 /* 9881 * We're the first ECB on this probe. 9882 */ 9883 probe->dtpr_ecb = probe->dtpr_ecb_last = ecb; 9884 9885 if (ecb->dte_predicate != NULL) 9886 probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid; 9887 9888 prov->dtpv_pops.dtps_enable(prov->dtpv_arg, 9889 probe->dtpr_id, probe->dtpr_arg); 9890 } else { 9891 /* 9892 * This probe is already active. Swing the last pointer to 9893 * point to the new ECB, and issue a dtrace_sync() to assure 9894 * that all CPUs have seen the change. 9895 */ 9896 ASSERT(probe->dtpr_ecb_last != NULL); 9897 probe->dtpr_ecb_last->dte_next = ecb; 9898 probe->dtpr_ecb_last = ecb; 9899 probe->dtpr_predcache = 0; 9900 9901 dtrace_sync(); 9902 } 9903} 9904 9905static void 9906dtrace_ecb_resize(dtrace_ecb_t *ecb) 9907{ 9908 dtrace_action_t *act; 9909 uint32_t curneeded = UINT32_MAX; 9910 uint32_t aggbase = UINT32_MAX; 9911 9912 /* 9913 * If we record anything, we always record the dtrace_rechdr_t. (And 9914 * we always record it first.) 9915 */ 9916 ecb->dte_size = sizeof (dtrace_rechdr_t); 9917 ecb->dte_alignment = sizeof (dtrace_epid_t); 9918 9919 for (act = ecb->dte_action; act != NULL; act = act->dta_next) { 9920 dtrace_recdesc_t *rec = &act->dta_rec; 9921 ASSERT(rec->dtrd_size > 0 || rec->dtrd_alignment == 1); 9922 9923 ecb->dte_alignment = MAX(ecb->dte_alignment, 9924 rec->dtrd_alignment); 9925 9926 if (DTRACEACT_ISAGG(act->dta_kind)) { 9927 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act; 9928 9929 ASSERT(rec->dtrd_size != 0); 9930 ASSERT(agg->dtag_first != NULL); 9931 ASSERT(act->dta_prev->dta_intuple); 9932 ASSERT(aggbase != UINT32_MAX); 9933 ASSERT(curneeded != UINT32_MAX); 9934 9935 agg->dtag_base = aggbase; 9936 9937 curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment); 9938 rec->dtrd_offset = curneeded; 9939 curneeded += rec->dtrd_size; 9940 ecb->dte_needed = MAX(ecb->dte_needed, curneeded); 9941 9942 aggbase = UINT32_MAX; 9943 curneeded = UINT32_MAX; 9944 } else if (act->dta_intuple) { 9945 if (curneeded == UINT32_MAX) { 9946 /* 9947 * This is the first record in a tuple. Align 9948 * curneeded to be at offset 4 in an 8-byte 9949 * aligned block. 9950 */ 9951 ASSERT(act->dta_prev == NULL || 9952 !act->dta_prev->dta_intuple); 9953 ASSERT3U(aggbase, ==, UINT32_MAX); 9954 curneeded = P2PHASEUP(ecb->dte_size, 9955 sizeof (uint64_t), sizeof (dtrace_aggid_t)); 9956 9957 aggbase = curneeded - sizeof (dtrace_aggid_t); 9958 ASSERT(IS_P2ALIGNED(aggbase, 9959 sizeof (uint64_t))); 9960 } 9961 curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment); 9962 rec->dtrd_offset = curneeded; 9963 curneeded += rec->dtrd_size; 9964 } else { 9965 /* tuples must be followed by an aggregation */ 9966 ASSERT(act->dta_prev == NULL || 9967 !act->dta_prev->dta_intuple); 9968 9969 ecb->dte_size = P2ROUNDUP(ecb->dte_size, 9970 rec->dtrd_alignment); 9971 rec->dtrd_offset = ecb->dte_size; 9972 ecb->dte_size += rec->dtrd_size; 9973 ecb->dte_needed = MAX(ecb->dte_needed, ecb->dte_size); 9974 } 9975 } 9976 9977 if ((act = ecb->dte_action) != NULL && 9978 !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) && 9979 ecb->dte_size == sizeof (dtrace_rechdr_t)) { 9980 /* 9981 * If the size is still sizeof (dtrace_rechdr_t), then all 9982 * actions store no data; set the size to 0. 9983 */ 9984 ecb->dte_size = 0; 9985 } 9986 9987 ecb->dte_size = P2ROUNDUP(ecb->dte_size, sizeof (dtrace_epid_t)); 9988 ecb->dte_needed = P2ROUNDUP(ecb->dte_needed, (sizeof (dtrace_epid_t))); 9989 ecb->dte_state->dts_needed = MAX(ecb->dte_state->dts_needed, 9990 ecb->dte_needed); 9991} 9992 9993static dtrace_action_t * 9994dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc) 9995{ 9996 dtrace_aggregation_t *agg; 9997 size_t size = sizeof (uint64_t); 9998 int ntuple = desc->dtad_ntuple; 9999 dtrace_action_t *act; 10000 dtrace_recdesc_t *frec; 10001 dtrace_aggid_t aggid; 10002 dtrace_state_t *state = ecb->dte_state; 10003 10004 agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP); 10005 agg->dtag_ecb = ecb; 10006 10007 ASSERT(DTRACEACT_ISAGG(desc->dtad_kind)); 10008 10009 switch (desc->dtad_kind) { 10010 case DTRACEAGG_MIN: 10011 agg->dtag_initial = INT64_MAX; 10012 agg->dtag_aggregate = dtrace_aggregate_min; 10013 break; 10014 10015 case DTRACEAGG_MAX: 10016 agg->dtag_initial = INT64_MIN; 10017 agg->dtag_aggregate = dtrace_aggregate_max; 10018 break; 10019 10020 case DTRACEAGG_COUNT: 10021 agg->dtag_aggregate = dtrace_aggregate_count; 10022 break; 10023 10024 case DTRACEAGG_QUANTIZE: 10025 agg->dtag_aggregate = dtrace_aggregate_quantize; 10026 size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) * 10027 sizeof (uint64_t); 10028 break; 10029 10030 case DTRACEAGG_LQUANTIZE: { 10031 uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg); 10032 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg); 10033 10034 agg->dtag_initial = desc->dtad_arg; 10035 agg->dtag_aggregate = dtrace_aggregate_lquantize; 10036 10037 if (step == 0 || levels == 0) 10038 goto err; 10039 10040 size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t); 10041 break; 10042 } 10043 10044 case DTRACEAGG_LLQUANTIZE: { 10045 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(desc->dtad_arg); 10046 uint16_t low = DTRACE_LLQUANTIZE_LOW(desc->dtad_arg); 10047 uint16_t high = DTRACE_LLQUANTIZE_HIGH(desc->dtad_arg); 10048 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(desc->dtad_arg); 10049 int64_t v; 10050 10051 agg->dtag_initial = desc->dtad_arg; 10052 agg->dtag_aggregate = dtrace_aggregate_llquantize; 10053 10054 if (factor < 2 || low >= high || nsteps < factor) 10055 goto err; 10056 10057 /* 10058 * Now check that the number of steps evenly divides a power 10059 * of the factor. (This assures both integer bucket size and 10060 * linearity within each magnitude.) 10061 */ 10062 for (v = factor; v < nsteps; v *= factor) 10063 continue; 10064 10065 if ((v % nsteps) || (nsteps % factor)) 10066 goto err; 10067 10068 size = (dtrace_aggregate_llquantize_bucket(factor, 10069 low, high, nsteps, INT64_MAX) + 2) * sizeof (uint64_t); 10070 break; 10071 } 10072 10073 case DTRACEAGG_AVG: 10074 agg->dtag_aggregate = dtrace_aggregate_avg; 10075 size = sizeof (uint64_t) * 2; 10076 break; 10077 10078 case DTRACEAGG_STDDEV: 10079 agg->dtag_aggregate = dtrace_aggregate_stddev; 10080 size = sizeof (uint64_t) * 4; 10081 break; 10082 10083 case DTRACEAGG_SUM: 10084 agg->dtag_aggregate = dtrace_aggregate_sum; 10085 break; 10086 10087 default: 10088 goto err; 10089 } 10090 10091 agg->dtag_action.dta_rec.dtrd_size = size; 10092 10093 if (ntuple == 0) 10094 goto err; 10095 10096 /* 10097 * We must make sure that we have enough actions for the n-tuple. 10098 */ 10099 for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) { 10100 if (DTRACEACT_ISAGG(act->dta_kind)) 10101 break; 10102 10103 if (--ntuple == 0) { 10104 /* 10105 * This is the action with which our n-tuple begins. 10106 */ 10107 agg->dtag_first = act; 10108 goto success; 10109 } 10110 } 10111 10112 /* 10113 * This n-tuple is short by ntuple elements. Return failure. 10114 */ 10115 ASSERT(ntuple != 0); 10116err: 10117 kmem_free(agg, sizeof (dtrace_aggregation_t)); 10118 return (NULL); 10119 10120success: 10121 /* 10122 * If the last action in the tuple has a size of zero, it's actually 10123 * an expression argument for the aggregating action. 10124 */ 10125 ASSERT(ecb->dte_action_last != NULL); 10126 act = ecb->dte_action_last; 10127 10128 if (act->dta_kind == DTRACEACT_DIFEXPR) { 10129 ASSERT(act->dta_difo != NULL); 10130 10131 if (act->dta_difo->dtdo_rtype.dtdt_size == 0) 10132 agg->dtag_hasarg = 1; 10133 } 10134 10135 /* 10136 * We need to allocate an id for this aggregation. 10137 */ 10138#if defined(sun) 10139 aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1, 10140 VM_BESTFIT | VM_SLEEP); 10141#else 10142 aggid = alloc_unr(state->dts_aggid_arena); 10143#endif 10144 10145 if (aggid - 1 >= state->dts_naggregations) { 10146 dtrace_aggregation_t **oaggs = state->dts_aggregations; 10147 dtrace_aggregation_t **aggs; 10148 int naggs = state->dts_naggregations << 1; 10149 int onaggs = state->dts_naggregations; 10150 10151 ASSERT(aggid == state->dts_naggregations + 1); 10152 10153 if (naggs == 0) { 10154 ASSERT(oaggs == NULL); 10155 naggs = 1; 10156 } 10157 10158 aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP); 10159 10160 if (oaggs != NULL) { 10161 bcopy(oaggs, aggs, onaggs * sizeof (*aggs)); 10162 kmem_free(oaggs, onaggs * sizeof (*aggs)); 10163 } 10164 10165 state->dts_aggregations = aggs; 10166 state->dts_naggregations = naggs; 10167 } 10168 10169 ASSERT(state->dts_aggregations[aggid - 1] == NULL); 10170 state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg; 10171 10172 frec = &agg->dtag_first->dta_rec; 10173 if (frec->dtrd_alignment < sizeof (dtrace_aggid_t)) 10174 frec->dtrd_alignment = sizeof (dtrace_aggid_t); 10175 10176 for (act = agg->dtag_first; act != NULL; act = act->dta_next) { 10177 ASSERT(!act->dta_intuple); 10178 act->dta_intuple = 1; 10179 } 10180 10181 return (&agg->dtag_action); 10182} 10183 10184static void 10185dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act) 10186{ 10187 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act; 10188 dtrace_state_t *state = ecb->dte_state; 10189 dtrace_aggid_t aggid = agg->dtag_id; 10190 10191 ASSERT(DTRACEACT_ISAGG(act->dta_kind)); 10192#if defined(sun) 10193 vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1); 10194#else 10195 free_unr(state->dts_aggid_arena, aggid); 10196#endif 10197 10198 ASSERT(state->dts_aggregations[aggid - 1] == agg); 10199 state->dts_aggregations[aggid - 1] = NULL; 10200 10201 kmem_free(agg, sizeof (dtrace_aggregation_t)); 10202} 10203 10204static int 10205dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc) 10206{ 10207 dtrace_action_t *action, *last; 10208 dtrace_difo_t *dp = desc->dtad_difo; 10209 uint32_t size = 0, align = sizeof (uint8_t), mask; 10210 uint16_t format = 0; 10211 dtrace_recdesc_t *rec; 10212 dtrace_state_t *state = ecb->dte_state; 10213 dtrace_optval_t *opt = state->dts_options, nframes = 0, strsize; 10214 uint64_t arg = desc->dtad_arg; 10215 10216 ASSERT(MUTEX_HELD(&dtrace_lock)); 10217 ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1); 10218 10219 if (DTRACEACT_ISAGG(desc->dtad_kind)) { 10220 /* 10221 * If this is an aggregating action, there must be neither 10222 * a speculate nor a commit on the action chain. 10223 */ 10224 dtrace_action_t *act; 10225 10226 for (act = ecb->dte_action; act != NULL; act = act->dta_next) { 10227 if (act->dta_kind == DTRACEACT_COMMIT) 10228 return (EINVAL); 10229 10230 if (act->dta_kind == DTRACEACT_SPECULATE) 10231 return (EINVAL); 10232 } 10233 10234 action = dtrace_ecb_aggregation_create(ecb, desc); 10235 10236 if (action == NULL) 10237 return (EINVAL); 10238 } else { 10239 if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) || 10240 (desc->dtad_kind == DTRACEACT_DIFEXPR && 10241 dp != NULL && dp->dtdo_destructive)) { 10242 state->dts_destructive = 1; 10243 } 10244 10245 switch (desc->dtad_kind) { 10246 case DTRACEACT_PRINTF: 10247 case DTRACEACT_PRINTA: 10248 case DTRACEACT_SYSTEM: 10249 case DTRACEACT_FREOPEN: 10250 case DTRACEACT_DIFEXPR: 10251 /* 10252 * We know that our arg is a string -- turn it into a 10253 * format. 10254 */ 10255 if (arg == 0) { 10256 ASSERT(desc->dtad_kind == DTRACEACT_PRINTA || 10257 desc->dtad_kind == DTRACEACT_DIFEXPR); 10258 format = 0; 10259 } else { 10260 ASSERT(arg != 0); 10261#if defined(sun) 10262 ASSERT(arg > KERNELBASE); 10263#endif 10264 format = dtrace_format_add(state, 10265 (char *)(uintptr_t)arg); 10266 } 10267 10268 /*FALLTHROUGH*/ 10269 case DTRACEACT_LIBACT: 10270 case DTRACEACT_TRACEMEM: 10271 case DTRACEACT_TRACEMEM_DYNSIZE: 10272 if (dp == NULL) 10273 return (EINVAL); 10274 10275 if ((size = dp->dtdo_rtype.dtdt_size) != 0) 10276 break; 10277 10278 if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) { 10279 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) 10280 return (EINVAL); 10281 10282 size = opt[DTRACEOPT_STRSIZE]; 10283 } 10284 10285 break; 10286 10287 case DTRACEACT_STACK: 10288 if ((nframes = arg) == 0) { 10289 nframes = opt[DTRACEOPT_STACKFRAMES]; 10290 ASSERT(nframes > 0); 10291 arg = nframes; 10292 } 10293 10294 size = nframes * sizeof (pc_t); 10295 break; 10296 10297 case DTRACEACT_JSTACK: 10298 if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0) 10299 strsize = opt[DTRACEOPT_JSTACKSTRSIZE]; 10300 10301 if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) 10302 nframes = opt[DTRACEOPT_JSTACKFRAMES]; 10303 10304 arg = DTRACE_USTACK_ARG(nframes, strsize); 10305 10306 /*FALLTHROUGH*/ 10307 case DTRACEACT_USTACK: 10308 if (desc->dtad_kind != DTRACEACT_JSTACK && 10309 (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) { 10310 strsize = DTRACE_USTACK_STRSIZE(arg); 10311 nframes = opt[DTRACEOPT_USTACKFRAMES]; 10312 ASSERT(nframes > 0); 10313 arg = DTRACE_USTACK_ARG(nframes, strsize); 10314 } 10315 10316 /* 10317 * Save a slot for the pid. 10318 */ 10319 size = (nframes + 1) * sizeof (uint64_t); 10320 size += DTRACE_USTACK_STRSIZE(arg); 10321 size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t))); 10322 10323 break; 10324 10325 case DTRACEACT_SYM: 10326 case DTRACEACT_MOD: 10327 if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) != 10328 sizeof (uint64_t)) || 10329 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) 10330 return (EINVAL); 10331 break; 10332 10333 case DTRACEACT_USYM: 10334 case DTRACEACT_UMOD: 10335 case DTRACEACT_UADDR: 10336 if (dp == NULL || 10337 (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) || 10338 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) 10339 return (EINVAL); 10340 10341 /* 10342 * We have a slot for the pid, plus a slot for the 10343 * argument. To keep things simple (aligned with 10344 * bitness-neutral sizing), we store each as a 64-bit 10345 * quantity. 10346 */ 10347 size = 2 * sizeof (uint64_t); 10348 break; 10349 10350 case DTRACEACT_STOP: 10351 case DTRACEACT_BREAKPOINT: 10352 case DTRACEACT_PANIC: 10353 break; 10354 10355 case DTRACEACT_CHILL: 10356 case DTRACEACT_DISCARD: 10357 case DTRACEACT_RAISE: 10358 if (dp == NULL) 10359 return (EINVAL); 10360 break; 10361 10362 case DTRACEACT_EXIT: 10363 if (dp == NULL || 10364 (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) || 10365 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) 10366 return (EINVAL); 10367 break; 10368 10369 case DTRACEACT_SPECULATE: 10370 if (ecb->dte_size > sizeof (dtrace_rechdr_t)) 10371 return (EINVAL); 10372 10373 if (dp == NULL) 10374 return (EINVAL); 10375 10376 state->dts_speculates = 1; 10377 break; 10378 10379 case DTRACEACT_PRINTM: 10380 size = dp->dtdo_rtype.dtdt_size; 10381 break; 10382 10383 case DTRACEACT_PRINTT: 10384 size = dp->dtdo_rtype.dtdt_size; 10385 break; 10386 10387 case DTRACEACT_COMMIT: { 10388 dtrace_action_t *act = ecb->dte_action; 10389 10390 for (; act != NULL; act = act->dta_next) { 10391 if (act->dta_kind == DTRACEACT_COMMIT) 10392 return (EINVAL); 10393 } 10394 10395 if (dp == NULL) 10396 return (EINVAL); 10397 break; 10398 } 10399 10400 default: 10401 return (EINVAL); 10402 } 10403 10404 if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) { 10405 /* 10406 * If this is a data-storing action or a speculate, 10407 * we must be sure that there isn't a commit on the 10408 * action chain. 10409 */ 10410 dtrace_action_t *act = ecb->dte_action; 10411 10412 for (; act != NULL; act = act->dta_next) { 10413 if (act->dta_kind == DTRACEACT_COMMIT) 10414 return (EINVAL); 10415 } 10416 } 10417 10418 action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP); 10419 action->dta_rec.dtrd_size = size; 10420 } 10421 10422 action->dta_refcnt = 1; 10423 rec = &action->dta_rec; 10424 size = rec->dtrd_size; 10425 10426 for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) { 10427 if (!(size & mask)) { 10428 align = mask + 1; 10429 break; 10430 } 10431 } 10432 10433 action->dta_kind = desc->dtad_kind; 10434 10435 if ((action->dta_difo = dp) != NULL) 10436 dtrace_difo_hold(dp); 10437 10438 rec->dtrd_action = action->dta_kind; 10439 rec->dtrd_arg = arg; 10440 rec->dtrd_uarg = desc->dtad_uarg; 10441 rec->dtrd_alignment = (uint16_t)align; 10442 rec->dtrd_format = format; 10443 10444 if ((last = ecb->dte_action_last) != NULL) { 10445 ASSERT(ecb->dte_action != NULL); 10446 action->dta_prev = last; 10447 last->dta_next = action; 10448 } else { 10449 ASSERT(ecb->dte_action == NULL); 10450 ecb->dte_action = action; 10451 } 10452 10453 ecb->dte_action_last = action; 10454 10455 return (0); 10456} 10457 10458static void 10459dtrace_ecb_action_remove(dtrace_ecb_t *ecb) 10460{ 10461 dtrace_action_t *act = ecb->dte_action, *next; 10462 dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate; 10463 dtrace_difo_t *dp; 10464 uint16_t format; 10465 10466 if (act != NULL && act->dta_refcnt > 1) { 10467 ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1); 10468 act->dta_refcnt--; 10469 } else { 10470 for (; act != NULL; act = next) { 10471 next = act->dta_next; 10472 ASSERT(next != NULL || act == ecb->dte_action_last); 10473 ASSERT(act->dta_refcnt == 1); 10474 10475 if ((format = act->dta_rec.dtrd_format) != 0) 10476 dtrace_format_remove(ecb->dte_state, format); 10477 10478 if ((dp = act->dta_difo) != NULL) 10479 dtrace_difo_release(dp, vstate); 10480 10481 if (DTRACEACT_ISAGG(act->dta_kind)) { 10482 dtrace_ecb_aggregation_destroy(ecb, act); 10483 } else { 10484 kmem_free(act, sizeof (dtrace_action_t)); 10485 } 10486 } 10487 } 10488 10489 ecb->dte_action = NULL; 10490 ecb->dte_action_last = NULL; 10491 ecb->dte_size = 0; 10492} 10493 10494static void 10495dtrace_ecb_disable(dtrace_ecb_t *ecb) 10496{ 10497 /* 10498 * We disable the ECB by removing it from its probe. 10499 */ 10500 dtrace_ecb_t *pecb, *prev = NULL; 10501 dtrace_probe_t *probe = ecb->dte_probe; 10502 10503 ASSERT(MUTEX_HELD(&dtrace_lock)); 10504 10505 if (probe == NULL) { 10506 /* 10507 * This is the NULL probe; there is nothing to disable. 10508 */ 10509 return; 10510 } 10511 10512 for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) { 10513 if (pecb == ecb) 10514 break; 10515 prev = pecb; 10516 } 10517 10518 ASSERT(pecb != NULL); 10519 10520 if (prev == NULL) { 10521 probe->dtpr_ecb = ecb->dte_next; 10522 } else { 10523 prev->dte_next = ecb->dte_next; 10524 } 10525 10526 if (ecb == probe->dtpr_ecb_last) { 10527 ASSERT(ecb->dte_next == NULL); 10528 probe->dtpr_ecb_last = prev; 10529 } 10530 10531 /* 10532 * The ECB has been disconnected from the probe; now sync to assure 10533 * that all CPUs have seen the change before returning. 10534 */ 10535 dtrace_sync(); 10536 10537 if (probe->dtpr_ecb == NULL) { 10538 /* 10539 * That was the last ECB on the probe; clear the predicate 10540 * cache ID for the probe, disable it and sync one more time 10541 * to assure that we'll never hit it again. 10542 */ 10543 dtrace_provider_t *prov = probe->dtpr_provider; 10544 10545 ASSERT(ecb->dte_next == NULL); 10546 ASSERT(probe->dtpr_ecb_last == NULL); 10547 probe->dtpr_predcache = DTRACE_CACHEIDNONE; 10548 prov->dtpv_pops.dtps_disable(prov->dtpv_arg, 10549 probe->dtpr_id, probe->dtpr_arg); 10550 dtrace_sync(); 10551 } else { 10552 /* 10553 * There is at least one ECB remaining on the probe. If there 10554 * is _exactly_ one, set the probe's predicate cache ID to be 10555 * the predicate cache ID of the remaining ECB. 10556 */ 10557 ASSERT(probe->dtpr_ecb_last != NULL); 10558 ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE); 10559 10560 if (probe->dtpr_ecb == probe->dtpr_ecb_last) { 10561 dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate; 10562 10563 ASSERT(probe->dtpr_ecb->dte_next == NULL); 10564 10565 if (p != NULL) 10566 probe->dtpr_predcache = p->dtp_cacheid; 10567 } 10568 10569 ecb->dte_next = NULL; 10570 } 10571} 10572 10573static void 10574dtrace_ecb_destroy(dtrace_ecb_t *ecb) 10575{ 10576 dtrace_state_t *state = ecb->dte_state; 10577 dtrace_vstate_t *vstate = &state->dts_vstate; 10578 dtrace_predicate_t *pred; 10579 dtrace_epid_t epid = ecb->dte_epid; 10580 10581 ASSERT(MUTEX_HELD(&dtrace_lock)); 10582 ASSERT(ecb->dte_next == NULL); 10583 ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb); 10584 10585 if ((pred = ecb->dte_predicate) != NULL) 10586 dtrace_predicate_release(pred, vstate); 10587 10588 dtrace_ecb_action_remove(ecb); 10589 10590 ASSERT(state->dts_ecbs[epid - 1] == ecb); 10591 state->dts_ecbs[epid - 1] = NULL; 10592 10593 kmem_free(ecb, sizeof (dtrace_ecb_t)); 10594} 10595 10596static dtrace_ecb_t * 10597dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe, 10598 dtrace_enabling_t *enab) 10599{ 10600 dtrace_ecb_t *ecb; 10601 dtrace_predicate_t *pred; 10602 dtrace_actdesc_t *act; 10603 dtrace_provider_t *prov; 10604 dtrace_ecbdesc_t *desc = enab->dten_current; 10605 10606 ASSERT(MUTEX_HELD(&dtrace_lock)); 10607 ASSERT(state != NULL); 10608 10609 ecb = dtrace_ecb_add(state, probe); 10610 ecb->dte_uarg = desc->dted_uarg; 10611 10612 if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) { 10613 dtrace_predicate_hold(pred); 10614 ecb->dte_predicate = pred; 10615 } 10616 10617 if (probe != NULL) { 10618 /* 10619 * If the provider shows more leg than the consumer is old 10620 * enough to see, we need to enable the appropriate implicit 10621 * predicate bits to prevent the ecb from activating at 10622 * revealing times. 10623 * 10624 * Providers specifying DTRACE_PRIV_USER at register time 10625 * are stating that they need the /proc-style privilege 10626 * model to be enforced, and this is what DTRACE_COND_OWNER 10627 * and DTRACE_COND_ZONEOWNER will then do at probe time. 10628 */ 10629 prov = probe->dtpr_provider; 10630 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) && 10631 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER)) 10632 ecb->dte_cond |= DTRACE_COND_OWNER; 10633 10634 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) && 10635 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER)) 10636 ecb->dte_cond |= DTRACE_COND_ZONEOWNER; 10637 10638 /* 10639 * If the provider shows us kernel innards and the user 10640 * is lacking sufficient privilege, enable the 10641 * DTRACE_COND_USERMODE implicit predicate. 10642 */ 10643 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) && 10644 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL)) 10645 ecb->dte_cond |= DTRACE_COND_USERMODE; 10646 } 10647 10648 if (dtrace_ecb_create_cache != NULL) { 10649 /* 10650 * If we have a cached ecb, we'll use its action list instead 10651 * of creating our own (saving both time and space). 10652 */ 10653 dtrace_ecb_t *cached = dtrace_ecb_create_cache; 10654 dtrace_action_t *act = cached->dte_action; 10655 10656 if (act != NULL) { 10657 ASSERT(act->dta_refcnt > 0); 10658 act->dta_refcnt++; 10659 ecb->dte_action = act; 10660 ecb->dte_action_last = cached->dte_action_last; 10661 ecb->dte_needed = cached->dte_needed; 10662 ecb->dte_size = cached->dte_size; 10663 ecb->dte_alignment = cached->dte_alignment; 10664 } 10665 10666 return (ecb); 10667 } 10668 10669 for (act = desc->dted_action; act != NULL; act = act->dtad_next) { 10670 if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) { 10671 dtrace_ecb_destroy(ecb); 10672 return (NULL); 10673 } 10674 } 10675 10676 dtrace_ecb_resize(ecb); 10677 10678 return (dtrace_ecb_create_cache = ecb); 10679} 10680 10681static int 10682dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg) 10683{ 10684 dtrace_ecb_t *ecb; 10685 dtrace_enabling_t *enab = arg; 10686 dtrace_state_t *state = enab->dten_vstate->dtvs_state; 10687 10688 ASSERT(state != NULL); 10689 10690 if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) { 10691 /* 10692 * This probe was created in a generation for which this 10693 * enabling has previously created ECBs; we don't want to 10694 * enable it again, so just kick out. 10695 */ 10696 return (DTRACE_MATCH_NEXT); 10697 } 10698 10699 if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL) 10700 return (DTRACE_MATCH_DONE); 10701 10702 dtrace_ecb_enable(ecb); 10703 return (DTRACE_MATCH_NEXT); 10704} 10705 10706static dtrace_ecb_t * 10707dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id) 10708{ 10709 dtrace_ecb_t *ecb; 10710 10711 ASSERT(MUTEX_HELD(&dtrace_lock)); 10712 10713 if (id == 0 || id > state->dts_necbs) 10714 return (NULL); 10715 10716 ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL); 10717 ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id); 10718 10719 return (state->dts_ecbs[id - 1]); 10720} 10721 10722static dtrace_aggregation_t * 10723dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id) 10724{ 10725 dtrace_aggregation_t *agg; 10726 10727 ASSERT(MUTEX_HELD(&dtrace_lock)); 10728 10729 if (id == 0 || id > state->dts_naggregations) 10730 return (NULL); 10731 10732 ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL); 10733 ASSERT((agg = state->dts_aggregations[id - 1]) == NULL || 10734 agg->dtag_id == id); 10735 10736 return (state->dts_aggregations[id - 1]); 10737} 10738 10739/* 10740 * DTrace Buffer Functions 10741 * 10742 * The following functions manipulate DTrace buffers. Most of these functions 10743 * are called in the context of establishing or processing consumer state; 10744 * exceptions are explicitly noted. 10745 */ 10746 10747/* 10748 * Note: called from cross call context. This function switches the two 10749 * buffers on a given CPU. The atomicity of this operation is assured by 10750 * disabling interrupts while the actual switch takes place; the disabling of 10751 * interrupts serializes the execution with any execution of dtrace_probe() on 10752 * the same CPU. 10753 */ 10754static void 10755dtrace_buffer_switch(dtrace_buffer_t *buf) 10756{ 10757 caddr_t tomax = buf->dtb_tomax; 10758 caddr_t xamot = buf->dtb_xamot; 10759 dtrace_icookie_t cookie; 10760 hrtime_t now; 10761 10762 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH)); 10763 ASSERT(!(buf->dtb_flags & DTRACEBUF_RING)); 10764 10765 cookie = dtrace_interrupt_disable(); 10766 now = dtrace_gethrtime(); 10767 buf->dtb_tomax = xamot; 10768 buf->dtb_xamot = tomax; 10769 buf->dtb_xamot_drops = buf->dtb_drops; 10770 buf->dtb_xamot_offset = buf->dtb_offset; 10771 buf->dtb_xamot_errors = buf->dtb_errors; 10772 buf->dtb_xamot_flags = buf->dtb_flags; 10773 buf->dtb_offset = 0; 10774 buf->dtb_drops = 0; 10775 buf->dtb_errors = 0; 10776 buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED); 10777 buf->dtb_interval = now - buf->dtb_switched; 10778 buf->dtb_switched = now; 10779 dtrace_interrupt_enable(cookie); 10780} 10781 10782/* 10783 * Note: called from cross call context. This function activates a buffer 10784 * on a CPU. As with dtrace_buffer_switch(), the atomicity of the operation 10785 * is guaranteed by the disabling of interrupts. 10786 */ 10787static void 10788dtrace_buffer_activate(dtrace_state_t *state) 10789{ 10790 dtrace_buffer_t *buf; 10791 dtrace_icookie_t cookie = dtrace_interrupt_disable(); 10792 10793 buf = &state->dts_buffer[curcpu]; 10794 10795 if (buf->dtb_tomax != NULL) { 10796 /* 10797 * We might like to assert that the buffer is marked inactive, 10798 * but this isn't necessarily true: the buffer for the CPU 10799 * that processes the BEGIN probe has its buffer activated 10800 * manually. In this case, we take the (harmless) action 10801 * re-clearing the bit INACTIVE bit. 10802 */ 10803 buf->dtb_flags &= ~DTRACEBUF_INACTIVE; 10804 } 10805 10806 dtrace_interrupt_enable(cookie); 10807} 10808 10809static int 10810dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags, 10811 processorid_t cpu, int *factor) 10812{ 10813#if defined(sun) 10814 cpu_t *cp; 10815#endif 10816 dtrace_buffer_t *buf; 10817 int allocated = 0, desired = 0; 10818 10819#if defined(sun) 10820 ASSERT(MUTEX_HELD(&cpu_lock)); 10821 ASSERT(MUTEX_HELD(&dtrace_lock)); 10822 10823 *factor = 1; 10824 10825 if (size > dtrace_nonroot_maxsize && 10826 !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE)) 10827 return (EFBIG); 10828 10829 cp = cpu_list; 10830 10831 do { 10832 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id) 10833 continue; 10834 10835 buf = &bufs[cp->cpu_id]; 10836 10837 /* 10838 * If there is already a buffer allocated for this CPU, it 10839 * is only possible that this is a DR event. In this case, 10840 */ 10841 if (buf->dtb_tomax != NULL) { 10842 ASSERT(buf->dtb_size == size); 10843 continue; 10844 } 10845 10846 ASSERT(buf->dtb_xamot == NULL); 10847 10848 if ((buf->dtb_tomax = kmem_zalloc(size, 10849 KM_NOSLEEP | KM_NORMALPRI)) == NULL) 10850 goto err; 10851 10852 buf->dtb_size = size; 10853 buf->dtb_flags = flags; 10854 buf->dtb_offset = 0; 10855 buf->dtb_drops = 0; 10856 10857 if (flags & DTRACEBUF_NOSWITCH) 10858 continue; 10859 10860 if ((buf->dtb_xamot = kmem_zalloc(size, 10861 KM_NOSLEEP | KM_NORMALPRI)) == NULL) 10862 goto err; 10863 } while ((cp = cp->cpu_next) != cpu_list); 10864 10865 return (0); 10866 10867err: 10868 cp = cpu_list; 10869 10870 do { 10871 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id) 10872 continue; 10873 10874 buf = &bufs[cp->cpu_id]; 10875 desired += 2; 10876 10877 if (buf->dtb_xamot != NULL) { 10878 ASSERT(buf->dtb_tomax != NULL); 10879 ASSERT(buf->dtb_size == size); 10880 kmem_free(buf->dtb_xamot, size); 10881 allocated++; 10882 } 10883 10884 if (buf->dtb_tomax != NULL) { 10885 ASSERT(buf->dtb_size == size); 10886 kmem_free(buf->dtb_tomax, size); 10887 allocated++; 10888 } 10889 10890 buf->dtb_tomax = NULL; 10891 buf->dtb_xamot = NULL; 10892 buf->dtb_size = 0; 10893 } while ((cp = cp->cpu_next) != cpu_list); 10894#else 10895 int i; 10896 10897 *factor = 1; 10898#if defined(__amd64__) || defined(__mips__) || defined(__powerpc__) 10899 /* 10900 * FreeBSD isn't good at limiting the amount of memory we 10901 * ask to malloc, so let's place a limit here before trying 10902 * to do something that might well end in tears at bedtime. 10903 */ 10904 if (size > physmem * PAGE_SIZE / (128 * (mp_maxid + 1))) 10905 return (ENOMEM); 10906#endif 10907 10908 ASSERT(MUTEX_HELD(&dtrace_lock)); 10909 CPU_FOREACH(i) { 10910 if (cpu != DTRACE_CPUALL && cpu != i) 10911 continue; 10912 10913 buf = &bufs[i]; 10914 10915 /* 10916 * If there is already a buffer allocated for this CPU, it 10917 * is only possible that this is a DR event. In this case, 10918 * the buffer size must match our specified size. 10919 */ 10920 if (buf->dtb_tomax != NULL) { 10921 ASSERT(buf->dtb_size == size); 10922 continue; 10923 } 10924 10925 ASSERT(buf->dtb_xamot == NULL); 10926 10927 if ((buf->dtb_tomax = kmem_zalloc(size, 10928 KM_NOSLEEP | KM_NORMALPRI)) == NULL) 10929 goto err; 10930 10931 buf->dtb_size = size; 10932 buf->dtb_flags = flags; 10933 buf->dtb_offset = 0; 10934 buf->dtb_drops = 0; 10935 10936 if (flags & DTRACEBUF_NOSWITCH) 10937 continue; 10938 10939 if ((buf->dtb_xamot = kmem_zalloc(size, 10940 KM_NOSLEEP | KM_NORMALPRI)) == NULL) 10941 goto err; 10942 } 10943 10944 return (0); 10945 10946err: 10947 /* 10948 * Error allocating memory, so free the buffers that were 10949 * allocated before the failed allocation. 10950 */ 10951 CPU_FOREACH(i) { 10952 if (cpu != DTRACE_CPUALL && cpu != i) 10953 continue; 10954 10955 buf = &bufs[i]; 10956 desired += 2; 10957 10958 if (buf->dtb_xamot != NULL) { 10959 ASSERT(buf->dtb_tomax != NULL); 10960 ASSERT(buf->dtb_size == size); 10961 kmem_free(buf->dtb_xamot, size); 10962 allocated++; 10963 } 10964 10965 if (buf->dtb_tomax != NULL) { 10966 ASSERT(buf->dtb_size == size); 10967 kmem_free(buf->dtb_tomax, size); 10968 allocated++; 10969 } 10970 10971 buf->dtb_tomax = NULL; 10972 buf->dtb_xamot = NULL; 10973 buf->dtb_size = 0; 10974 10975 } 10976#endif 10977 *factor = desired / (allocated > 0 ? allocated : 1); 10978 10979 return (ENOMEM); 10980} 10981 10982/* 10983 * Note: called from probe context. This function just increments the drop 10984 * count on a buffer. It has been made a function to allow for the 10985 * possibility of understanding the source of mysterious drop counts. (A 10986 * problem for which one may be particularly disappointed that DTrace cannot 10987 * be used to understand DTrace.) 10988 */ 10989static void 10990dtrace_buffer_drop(dtrace_buffer_t *buf) 10991{ 10992 buf->dtb_drops++; 10993} 10994 10995/* 10996 * Note: called from probe context. This function is called to reserve space 10997 * in a buffer. If mstate is non-NULL, sets the scratch base and size in the 10998 * mstate. Returns the new offset in the buffer, or a negative value if an 10999 * error has occurred. 11000 */ 11001static intptr_t 11002dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align, 11003 dtrace_state_t *state, dtrace_mstate_t *mstate) 11004{ 11005 intptr_t offs = buf->dtb_offset, soffs; 11006 intptr_t woffs; 11007 caddr_t tomax; 11008 size_t total; 11009 11010 if (buf->dtb_flags & DTRACEBUF_INACTIVE) 11011 return (-1); 11012 11013 if ((tomax = buf->dtb_tomax) == NULL) { 11014 dtrace_buffer_drop(buf); 11015 return (-1); 11016 } 11017 11018 if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) { 11019 while (offs & (align - 1)) { 11020 /* 11021 * Assert that our alignment is off by a number which 11022 * is itself sizeof (uint32_t) aligned. 11023 */ 11024 ASSERT(!((align - (offs & (align - 1))) & 11025 (sizeof (uint32_t) - 1))); 11026 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE); 11027 offs += sizeof (uint32_t); 11028 } 11029 11030 if ((soffs = offs + needed) > buf->dtb_size) { 11031 dtrace_buffer_drop(buf); 11032 return (-1); 11033 } 11034 11035 if (mstate == NULL) 11036 return (offs); 11037 11038 mstate->dtms_scratch_base = (uintptr_t)tomax + soffs; 11039 mstate->dtms_scratch_size = buf->dtb_size - soffs; 11040 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base; 11041 11042 return (offs); 11043 } 11044 11045 if (buf->dtb_flags & DTRACEBUF_FILL) { 11046 if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN && 11047 (buf->dtb_flags & DTRACEBUF_FULL)) 11048 return (-1); 11049 goto out; 11050 } 11051 11052 total = needed + (offs & (align - 1)); 11053 11054 /* 11055 * For a ring buffer, life is quite a bit more complicated. Before 11056 * we can store any padding, we need to adjust our wrapping offset. 11057 * (If we've never before wrapped or we're not about to, no adjustment 11058 * is required.) 11059 */ 11060 if ((buf->dtb_flags & DTRACEBUF_WRAPPED) || 11061 offs + total > buf->dtb_size) { 11062 woffs = buf->dtb_xamot_offset; 11063 11064 if (offs + total > buf->dtb_size) { 11065 /* 11066 * We can't fit in the end of the buffer. First, a 11067 * sanity check that we can fit in the buffer at all. 11068 */ 11069 if (total > buf->dtb_size) { 11070 dtrace_buffer_drop(buf); 11071 return (-1); 11072 } 11073 11074 /* 11075 * We're going to be storing at the top of the buffer, 11076 * so now we need to deal with the wrapped offset. We 11077 * only reset our wrapped offset to 0 if it is 11078 * currently greater than the current offset. If it 11079 * is less than the current offset, it is because a 11080 * previous allocation induced a wrap -- but the 11081 * allocation didn't subsequently take the space due 11082 * to an error or false predicate evaluation. In this 11083 * case, we'll just leave the wrapped offset alone: if 11084 * the wrapped offset hasn't been advanced far enough 11085 * for this allocation, it will be adjusted in the 11086 * lower loop. 11087 */ 11088 if (buf->dtb_flags & DTRACEBUF_WRAPPED) { 11089 if (woffs >= offs) 11090 woffs = 0; 11091 } else { 11092 woffs = 0; 11093 } 11094 11095 /* 11096 * Now we know that we're going to be storing to the 11097 * top of the buffer and that there is room for us 11098 * there. We need to clear the buffer from the current 11099 * offset to the end (there may be old gunk there). 11100 */ 11101 while (offs < buf->dtb_size) 11102 tomax[offs++] = 0; 11103 11104 /* 11105 * We need to set our offset to zero. And because we 11106 * are wrapping, we need to set the bit indicating as 11107 * much. We can also adjust our needed space back 11108 * down to the space required by the ECB -- we know 11109 * that the top of the buffer is aligned. 11110 */ 11111 offs = 0; 11112 total = needed; 11113 buf->dtb_flags |= DTRACEBUF_WRAPPED; 11114 } else { 11115 /* 11116 * There is room for us in the buffer, so we simply 11117 * need to check the wrapped offset. 11118 */ 11119 if (woffs < offs) { 11120 /* 11121 * The wrapped offset is less than the offset. 11122 * This can happen if we allocated buffer space 11123 * that induced a wrap, but then we didn't 11124 * subsequently take the space due to an error 11125 * or false predicate evaluation. This is 11126 * okay; we know that _this_ allocation isn't 11127 * going to induce a wrap. We still can't 11128 * reset the wrapped offset to be zero, 11129 * however: the space may have been trashed in 11130 * the previous failed probe attempt. But at 11131 * least the wrapped offset doesn't need to 11132 * be adjusted at all... 11133 */ 11134 goto out; 11135 } 11136 } 11137 11138 while (offs + total > woffs) { 11139 dtrace_epid_t epid = *(uint32_t *)(tomax + woffs); 11140 size_t size; 11141 11142 if (epid == DTRACE_EPIDNONE) { 11143 size = sizeof (uint32_t); 11144 } else { 11145 ASSERT3U(epid, <=, state->dts_necbs); 11146 ASSERT(state->dts_ecbs[epid - 1] != NULL); 11147 11148 size = state->dts_ecbs[epid - 1]->dte_size; 11149 } 11150 11151 ASSERT(woffs + size <= buf->dtb_size); 11152 ASSERT(size != 0); 11153 11154 if (woffs + size == buf->dtb_size) { 11155 /* 11156 * We've reached the end of the buffer; we want 11157 * to set the wrapped offset to 0 and break 11158 * out. However, if the offs is 0, then we're 11159 * in a strange edge-condition: the amount of 11160 * space that we want to reserve plus the size 11161 * of the record that we're overwriting is 11162 * greater than the size of the buffer. This 11163 * is problematic because if we reserve the 11164 * space but subsequently don't consume it (due 11165 * to a failed predicate or error) the wrapped 11166 * offset will be 0 -- yet the EPID at offset 0 11167 * will not be committed. This situation is 11168 * relatively easy to deal with: if we're in 11169 * this case, the buffer is indistinguishable 11170 * from one that hasn't wrapped; we need only 11171 * finish the job by clearing the wrapped bit, 11172 * explicitly setting the offset to be 0, and 11173 * zero'ing out the old data in the buffer. 11174 */ 11175 if (offs == 0) { 11176 buf->dtb_flags &= ~DTRACEBUF_WRAPPED; 11177 buf->dtb_offset = 0; 11178 woffs = total; 11179 11180 while (woffs < buf->dtb_size) 11181 tomax[woffs++] = 0; 11182 } 11183 11184 woffs = 0; 11185 break; 11186 } 11187 11188 woffs += size; 11189 } 11190 11191 /* 11192 * We have a wrapped offset. It may be that the wrapped offset 11193 * has become zero -- that's okay. 11194 */ 11195 buf->dtb_xamot_offset = woffs; 11196 } 11197 11198out: 11199 /* 11200 * Now we can plow the buffer with any necessary padding. 11201 */ 11202 while (offs & (align - 1)) { 11203 /* 11204 * Assert that our alignment is off by a number which 11205 * is itself sizeof (uint32_t) aligned. 11206 */ 11207 ASSERT(!((align - (offs & (align - 1))) & 11208 (sizeof (uint32_t) - 1))); 11209 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE); 11210 offs += sizeof (uint32_t); 11211 } 11212 11213 if (buf->dtb_flags & DTRACEBUF_FILL) { 11214 if (offs + needed > buf->dtb_size - state->dts_reserve) { 11215 buf->dtb_flags |= DTRACEBUF_FULL; 11216 return (-1); 11217 } 11218 } 11219 11220 if (mstate == NULL) 11221 return (offs); 11222 11223 /* 11224 * For ring buffers and fill buffers, the scratch space is always 11225 * the inactive buffer. 11226 */ 11227 mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot; 11228 mstate->dtms_scratch_size = buf->dtb_size; 11229 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base; 11230 11231 return (offs); 11232} 11233 11234static void 11235dtrace_buffer_polish(dtrace_buffer_t *buf) 11236{ 11237 ASSERT(buf->dtb_flags & DTRACEBUF_RING); 11238 ASSERT(MUTEX_HELD(&dtrace_lock)); 11239 11240 if (!(buf->dtb_flags & DTRACEBUF_WRAPPED)) 11241 return; 11242 11243 /* 11244 * We need to polish the ring buffer. There are three cases: 11245 * 11246 * - The first (and presumably most common) is that there is no gap 11247 * between the buffer offset and the wrapped offset. In this case, 11248 * there is nothing in the buffer that isn't valid data; we can 11249 * mark the buffer as polished and return. 11250 * 11251 * - The second (less common than the first but still more common 11252 * than the third) is that there is a gap between the buffer offset 11253 * and the wrapped offset, and the wrapped offset is larger than the 11254 * buffer offset. This can happen because of an alignment issue, or 11255 * can happen because of a call to dtrace_buffer_reserve() that 11256 * didn't subsequently consume the buffer space. In this case, 11257 * we need to zero the data from the buffer offset to the wrapped 11258 * offset. 11259 * 11260 * - The third (and least common) is that there is a gap between the 11261 * buffer offset and the wrapped offset, but the wrapped offset is 11262 * _less_ than the buffer offset. This can only happen because a 11263 * call to dtrace_buffer_reserve() induced a wrap, but the space 11264 * was not subsequently consumed. In this case, we need to zero the 11265 * space from the offset to the end of the buffer _and_ from the 11266 * top of the buffer to the wrapped offset. 11267 */ 11268 if (buf->dtb_offset < buf->dtb_xamot_offset) { 11269 bzero(buf->dtb_tomax + buf->dtb_offset, 11270 buf->dtb_xamot_offset - buf->dtb_offset); 11271 } 11272 11273 if (buf->dtb_offset > buf->dtb_xamot_offset) { 11274 bzero(buf->dtb_tomax + buf->dtb_offset, 11275 buf->dtb_size - buf->dtb_offset); 11276 bzero(buf->dtb_tomax, buf->dtb_xamot_offset); 11277 } 11278} 11279 11280/* 11281 * This routine determines if data generated at the specified time has likely 11282 * been entirely consumed at user-level. This routine is called to determine 11283 * if an ECB on a defunct probe (but for an active enabling) can be safely 11284 * disabled and destroyed. 11285 */ 11286static int 11287dtrace_buffer_consumed(dtrace_buffer_t *bufs, hrtime_t when) 11288{ 11289 int i; 11290 11291 for (i = 0; i < NCPU; i++) { 11292 dtrace_buffer_t *buf = &bufs[i]; 11293 11294 if (buf->dtb_size == 0) 11295 continue; 11296 11297 if (buf->dtb_flags & DTRACEBUF_RING) 11298 return (0); 11299 11300 if (!buf->dtb_switched && buf->dtb_offset != 0) 11301 return (0); 11302 11303 if (buf->dtb_switched - buf->dtb_interval < when) 11304 return (0); 11305 } 11306 11307 return (1); 11308} 11309 11310static void 11311dtrace_buffer_free(dtrace_buffer_t *bufs) 11312{ 11313 int i; 11314 11315 for (i = 0; i < NCPU; i++) { 11316 dtrace_buffer_t *buf = &bufs[i]; 11317 11318 if (buf->dtb_tomax == NULL) { 11319 ASSERT(buf->dtb_xamot == NULL); 11320 ASSERT(buf->dtb_size == 0); 11321 continue; 11322 } 11323 11324 if (buf->dtb_xamot != NULL) { 11325 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH)); 11326 kmem_free(buf->dtb_xamot, buf->dtb_size); 11327 } 11328 11329 kmem_free(buf->dtb_tomax, buf->dtb_size); 11330 buf->dtb_size = 0; 11331 buf->dtb_tomax = NULL; 11332 buf->dtb_xamot = NULL; 11333 } 11334} 11335 11336/* 11337 * DTrace Enabling Functions 11338 */ 11339static dtrace_enabling_t * 11340dtrace_enabling_create(dtrace_vstate_t *vstate) 11341{ 11342 dtrace_enabling_t *enab; 11343 11344 enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP); 11345 enab->dten_vstate = vstate; 11346 11347 return (enab); 11348} 11349 11350static void 11351dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb) 11352{ 11353 dtrace_ecbdesc_t **ndesc; 11354 size_t osize, nsize; 11355 11356 /* 11357 * We can't add to enablings after we've enabled them, or after we've 11358 * retained them. 11359 */ 11360 ASSERT(enab->dten_probegen == 0); 11361 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL); 11362 11363 if (enab->dten_ndesc < enab->dten_maxdesc) { 11364 enab->dten_desc[enab->dten_ndesc++] = ecb; 11365 return; 11366 } 11367 11368 osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *); 11369 11370 if (enab->dten_maxdesc == 0) { 11371 enab->dten_maxdesc = 1; 11372 } else { 11373 enab->dten_maxdesc <<= 1; 11374 } 11375 11376 ASSERT(enab->dten_ndesc < enab->dten_maxdesc); 11377 11378 nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *); 11379 ndesc = kmem_zalloc(nsize, KM_SLEEP); 11380 bcopy(enab->dten_desc, ndesc, osize); 11381 if (enab->dten_desc != NULL) 11382 kmem_free(enab->dten_desc, osize); 11383 11384 enab->dten_desc = ndesc; 11385 enab->dten_desc[enab->dten_ndesc++] = ecb; 11386} 11387 11388static void 11389dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb, 11390 dtrace_probedesc_t *pd) 11391{ 11392 dtrace_ecbdesc_t *new; 11393 dtrace_predicate_t *pred; 11394 dtrace_actdesc_t *act; 11395 11396 /* 11397 * We're going to create a new ECB description that matches the 11398 * specified ECB in every way, but has the specified probe description. 11399 */ 11400 new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP); 11401 11402 if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL) 11403 dtrace_predicate_hold(pred); 11404 11405 for (act = ecb->dted_action; act != NULL; act = act->dtad_next) 11406 dtrace_actdesc_hold(act); 11407 11408 new->dted_action = ecb->dted_action; 11409 new->dted_pred = ecb->dted_pred; 11410 new->dted_probe = *pd; 11411 new->dted_uarg = ecb->dted_uarg; 11412 11413 dtrace_enabling_add(enab, new); 11414} 11415 11416static void 11417dtrace_enabling_dump(dtrace_enabling_t *enab) 11418{ 11419 int i; 11420 11421 for (i = 0; i < enab->dten_ndesc; i++) { 11422 dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe; 11423 11424 cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i, 11425 desc->dtpd_provider, desc->dtpd_mod, 11426 desc->dtpd_func, desc->dtpd_name); 11427 } 11428} 11429 11430static void 11431dtrace_enabling_destroy(dtrace_enabling_t *enab) 11432{ 11433 int i; 11434 dtrace_ecbdesc_t *ep; 11435 dtrace_vstate_t *vstate = enab->dten_vstate; 11436 11437 ASSERT(MUTEX_HELD(&dtrace_lock)); 11438 11439 for (i = 0; i < enab->dten_ndesc; i++) { 11440 dtrace_actdesc_t *act, *next; 11441 dtrace_predicate_t *pred; 11442 11443 ep = enab->dten_desc[i]; 11444 11445 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) 11446 dtrace_predicate_release(pred, vstate); 11447 11448 for (act = ep->dted_action; act != NULL; act = next) { 11449 next = act->dtad_next; 11450 dtrace_actdesc_release(act, vstate); 11451 } 11452 11453 kmem_free(ep, sizeof (dtrace_ecbdesc_t)); 11454 } 11455 11456 if (enab->dten_desc != NULL) 11457 kmem_free(enab->dten_desc, 11458 enab->dten_maxdesc * sizeof (dtrace_enabling_t *)); 11459 11460 /* 11461 * If this was a retained enabling, decrement the dts_nretained count 11462 * and take it off of the dtrace_retained list. 11463 */ 11464 if (enab->dten_prev != NULL || enab->dten_next != NULL || 11465 dtrace_retained == enab) { 11466 ASSERT(enab->dten_vstate->dtvs_state != NULL); 11467 ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0); 11468 enab->dten_vstate->dtvs_state->dts_nretained--; 11469 dtrace_retained_gen++; 11470 } 11471 11472 if (enab->dten_prev == NULL) { 11473 if (dtrace_retained == enab) { 11474 dtrace_retained = enab->dten_next; 11475 11476 if (dtrace_retained != NULL) 11477 dtrace_retained->dten_prev = NULL; 11478 } 11479 } else { 11480 ASSERT(enab != dtrace_retained); 11481 ASSERT(dtrace_retained != NULL); 11482 enab->dten_prev->dten_next = enab->dten_next; 11483 } 11484 11485 if (enab->dten_next != NULL) { 11486 ASSERT(dtrace_retained != NULL); 11487 enab->dten_next->dten_prev = enab->dten_prev; 11488 } 11489 11490 kmem_free(enab, sizeof (dtrace_enabling_t)); 11491} 11492 11493static int 11494dtrace_enabling_retain(dtrace_enabling_t *enab) 11495{ 11496 dtrace_state_t *state; 11497 11498 ASSERT(MUTEX_HELD(&dtrace_lock)); 11499 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL); 11500 ASSERT(enab->dten_vstate != NULL); 11501 11502 state = enab->dten_vstate->dtvs_state; 11503 ASSERT(state != NULL); 11504 11505 /* 11506 * We only allow each state to retain dtrace_retain_max enablings. 11507 */ 11508 if (state->dts_nretained >= dtrace_retain_max) 11509 return (ENOSPC); 11510 11511 state->dts_nretained++; 11512 dtrace_retained_gen++; 11513 11514 if (dtrace_retained == NULL) { 11515 dtrace_retained = enab; 11516 return (0); 11517 } 11518 11519 enab->dten_next = dtrace_retained; 11520 dtrace_retained->dten_prev = enab; 11521 dtrace_retained = enab; 11522 11523 return (0); 11524} 11525 11526static int 11527dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match, 11528 dtrace_probedesc_t *create) 11529{ 11530 dtrace_enabling_t *new, *enab; 11531 int found = 0, err = ENOENT; 11532 11533 ASSERT(MUTEX_HELD(&dtrace_lock)); 11534 ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN); 11535 ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN); 11536 ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN); 11537 ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN); 11538 11539 new = dtrace_enabling_create(&state->dts_vstate); 11540 11541 /* 11542 * Iterate over all retained enablings, looking for enablings that 11543 * match the specified state. 11544 */ 11545 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) { 11546 int i; 11547 11548 /* 11549 * dtvs_state can only be NULL for helper enablings -- and 11550 * helper enablings can't be retained. 11551 */ 11552 ASSERT(enab->dten_vstate->dtvs_state != NULL); 11553 11554 if (enab->dten_vstate->dtvs_state != state) 11555 continue; 11556 11557 /* 11558 * Now iterate over each probe description; we're looking for 11559 * an exact match to the specified probe description. 11560 */ 11561 for (i = 0; i < enab->dten_ndesc; i++) { 11562 dtrace_ecbdesc_t *ep = enab->dten_desc[i]; 11563 dtrace_probedesc_t *pd = &ep->dted_probe; 11564 11565 if (strcmp(pd->dtpd_provider, match->dtpd_provider)) 11566 continue; 11567 11568 if (strcmp(pd->dtpd_mod, match->dtpd_mod)) 11569 continue; 11570 11571 if (strcmp(pd->dtpd_func, match->dtpd_func)) 11572 continue; 11573 11574 if (strcmp(pd->dtpd_name, match->dtpd_name)) 11575 continue; 11576 11577 /* 11578 * We have a winning probe! Add it to our growing 11579 * enabling. 11580 */ 11581 found = 1; 11582 dtrace_enabling_addlike(new, ep, create); 11583 } 11584 } 11585 11586 if (!found || (err = dtrace_enabling_retain(new)) != 0) { 11587 dtrace_enabling_destroy(new); 11588 return (err); 11589 } 11590 11591 return (0); 11592} 11593 11594static void 11595dtrace_enabling_retract(dtrace_state_t *state) 11596{ 11597 dtrace_enabling_t *enab, *next; 11598 11599 ASSERT(MUTEX_HELD(&dtrace_lock)); 11600 11601 /* 11602 * Iterate over all retained enablings, destroy the enablings retained 11603 * for the specified state. 11604 */ 11605 for (enab = dtrace_retained; enab != NULL; enab = next) { 11606 next = enab->dten_next; 11607 11608 /* 11609 * dtvs_state can only be NULL for helper enablings -- and 11610 * helper enablings can't be retained. 11611 */ 11612 ASSERT(enab->dten_vstate->dtvs_state != NULL); 11613 11614 if (enab->dten_vstate->dtvs_state == state) { 11615 ASSERT(state->dts_nretained > 0); 11616 dtrace_enabling_destroy(enab); 11617 } 11618 } 11619 11620 ASSERT(state->dts_nretained == 0); 11621} 11622 11623static int 11624dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched) 11625{ 11626 int i = 0; 11627 int matched = 0; 11628 11629 ASSERT(MUTEX_HELD(&cpu_lock)); 11630 ASSERT(MUTEX_HELD(&dtrace_lock)); 11631 11632 for (i = 0; i < enab->dten_ndesc; i++) { 11633 dtrace_ecbdesc_t *ep = enab->dten_desc[i]; 11634 11635 enab->dten_current = ep; 11636 enab->dten_error = 0; 11637 11638 matched += dtrace_probe_enable(&ep->dted_probe, enab); 11639 11640 if (enab->dten_error != 0) { 11641 /* 11642 * If we get an error half-way through enabling the 11643 * probes, we kick out -- perhaps with some number of 11644 * them enabled. Leaving enabled probes enabled may 11645 * be slightly confusing for user-level, but we expect 11646 * that no one will attempt to actually drive on in 11647 * the face of such errors. If this is an anonymous 11648 * enabling (indicated with a NULL nmatched pointer), 11649 * we cmn_err() a message. We aren't expecting to 11650 * get such an error -- such as it can exist at all, 11651 * it would be a result of corrupted DOF in the driver 11652 * properties. 11653 */ 11654 if (nmatched == NULL) { 11655 cmn_err(CE_WARN, "dtrace_enabling_match() " 11656 "error on %p: %d", (void *)ep, 11657 enab->dten_error); 11658 } 11659 11660 return (enab->dten_error); 11661 } 11662 } 11663 11664 enab->dten_probegen = dtrace_probegen; 11665 if (nmatched != NULL) 11666 *nmatched = matched; 11667 11668 return (0); 11669} 11670 11671static void 11672dtrace_enabling_matchall(void) 11673{ 11674 dtrace_enabling_t *enab; 11675 11676 mutex_enter(&cpu_lock); 11677 mutex_enter(&dtrace_lock); 11678 11679 /* 11680 * Iterate over all retained enablings to see if any probes match 11681 * against them. We only perform this operation on enablings for which 11682 * we have sufficient permissions by virtue of being in the global zone 11683 * or in the same zone as the DTrace client. Because we can be called 11684 * after dtrace_detach() has been called, we cannot assert that there 11685 * are retained enablings. We can safely load from dtrace_retained, 11686 * however: the taskq_destroy() at the end of dtrace_detach() will 11687 * block pending our completion. 11688 */ 11689 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) { 11690#if defined(sun) 11691 cred_t *cr = enab->dten_vstate->dtvs_state->dts_cred.dcr_cred; 11692 11693 if (INGLOBALZONE(curproc) || 11694 cr != NULL && getzoneid() == crgetzoneid(cr)) 11695#endif 11696 (void) dtrace_enabling_match(enab, NULL); 11697 } 11698 11699 mutex_exit(&dtrace_lock); 11700 mutex_exit(&cpu_lock); 11701} 11702 11703/* 11704 * If an enabling is to be enabled without having matched probes (that is, if 11705 * dtrace_state_go() is to be called on the underlying dtrace_state_t), the 11706 * enabling must be _primed_ by creating an ECB for every ECB description. 11707 * This must be done to assure that we know the number of speculations, the 11708 * number of aggregations, the minimum buffer size needed, etc. before we 11709 * transition out of DTRACE_ACTIVITY_INACTIVE. To do this without actually 11710 * enabling any probes, we create ECBs for every ECB decription, but with a 11711 * NULL probe -- which is exactly what this function does. 11712 */ 11713static void 11714dtrace_enabling_prime(dtrace_state_t *state) 11715{ 11716 dtrace_enabling_t *enab; 11717 int i; 11718 11719 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) { 11720 ASSERT(enab->dten_vstate->dtvs_state != NULL); 11721 11722 if (enab->dten_vstate->dtvs_state != state) 11723 continue; 11724 11725 /* 11726 * We don't want to prime an enabling more than once, lest 11727 * we allow a malicious user to induce resource exhaustion. 11728 * (The ECBs that result from priming an enabling aren't 11729 * leaked -- but they also aren't deallocated until the 11730 * consumer state is destroyed.) 11731 */ 11732 if (enab->dten_primed) 11733 continue; 11734 11735 for (i = 0; i < enab->dten_ndesc; i++) { 11736 enab->dten_current = enab->dten_desc[i]; 11737 (void) dtrace_probe_enable(NULL, enab); 11738 } 11739 11740 enab->dten_primed = 1; 11741 } 11742} 11743 11744/* 11745 * Called to indicate that probes should be provided due to retained 11746 * enablings. This is implemented in terms of dtrace_probe_provide(), but it 11747 * must take an initial lap through the enabling calling the dtps_provide() 11748 * entry point explicitly to allow for autocreated probes. 11749 */ 11750static void 11751dtrace_enabling_provide(dtrace_provider_t *prv) 11752{ 11753 int i, all = 0; 11754 dtrace_probedesc_t desc; 11755 dtrace_genid_t gen; 11756 11757 ASSERT(MUTEX_HELD(&dtrace_lock)); 11758 ASSERT(MUTEX_HELD(&dtrace_provider_lock)); 11759 11760 if (prv == NULL) { 11761 all = 1; 11762 prv = dtrace_provider; 11763 } 11764 11765 do { 11766 dtrace_enabling_t *enab; 11767 void *parg = prv->dtpv_arg; 11768 11769retry: 11770 gen = dtrace_retained_gen; 11771 for (enab = dtrace_retained; enab != NULL; 11772 enab = enab->dten_next) { 11773 for (i = 0; i < enab->dten_ndesc; i++) { 11774 desc = enab->dten_desc[i]->dted_probe; 11775 mutex_exit(&dtrace_lock); 11776 prv->dtpv_pops.dtps_provide(parg, &desc); 11777 mutex_enter(&dtrace_lock); 11778 /* 11779 * Process the retained enablings again if 11780 * they have changed while we weren't holding 11781 * dtrace_lock. 11782 */ 11783 if (gen != dtrace_retained_gen) 11784 goto retry; 11785 } 11786 } 11787 } while (all && (prv = prv->dtpv_next) != NULL); 11788 11789 mutex_exit(&dtrace_lock); 11790 dtrace_probe_provide(NULL, all ? NULL : prv); 11791 mutex_enter(&dtrace_lock); 11792} 11793 11794/* 11795 * Called to reap ECBs that are attached to probes from defunct providers. 11796 */ 11797static void 11798dtrace_enabling_reap(void) 11799{ 11800 dtrace_provider_t *prov; 11801 dtrace_probe_t *probe; 11802 dtrace_ecb_t *ecb; 11803 hrtime_t when; 11804 int i; 11805 11806 mutex_enter(&cpu_lock); 11807 mutex_enter(&dtrace_lock); 11808 11809 for (i = 0; i < dtrace_nprobes; i++) { 11810 if ((probe = dtrace_probes[i]) == NULL) 11811 continue; 11812 11813 if (probe->dtpr_ecb == NULL) 11814 continue; 11815 11816 prov = probe->dtpr_provider; 11817 11818 if ((when = prov->dtpv_defunct) == 0) 11819 continue; 11820 11821 /* 11822 * We have ECBs on a defunct provider: we want to reap these 11823 * ECBs to allow the provider to unregister. The destruction 11824 * of these ECBs must be done carefully: if we destroy the ECB 11825 * and the consumer later wishes to consume an EPID that 11826 * corresponds to the destroyed ECB (and if the EPID metadata 11827 * has not been previously consumed), the consumer will abort 11828 * processing on the unknown EPID. To reduce (but not, sadly, 11829 * eliminate) the possibility of this, we will only destroy an 11830 * ECB for a defunct provider if, for the state that 11831 * corresponds to the ECB: 11832 * 11833 * (a) There is no speculative tracing (which can effectively 11834 * cache an EPID for an arbitrary amount of time). 11835 * 11836 * (b) The principal buffers have been switched twice since the 11837 * provider became defunct. 11838 * 11839 * (c) The aggregation buffers are of zero size or have been 11840 * switched twice since the provider became defunct. 11841 * 11842 * We use dts_speculates to determine (a) and call a function 11843 * (dtrace_buffer_consumed()) to determine (b) and (c). Note 11844 * that as soon as we've been unable to destroy one of the ECBs 11845 * associated with the probe, we quit trying -- reaping is only 11846 * fruitful in as much as we can destroy all ECBs associated 11847 * with the defunct provider's probes. 11848 */ 11849 while ((ecb = probe->dtpr_ecb) != NULL) { 11850 dtrace_state_t *state = ecb->dte_state; 11851 dtrace_buffer_t *buf = state->dts_buffer; 11852 dtrace_buffer_t *aggbuf = state->dts_aggbuffer; 11853 11854 if (state->dts_speculates) 11855 break; 11856 11857 if (!dtrace_buffer_consumed(buf, when)) 11858 break; 11859 11860 if (!dtrace_buffer_consumed(aggbuf, when)) 11861 break; 11862 11863 dtrace_ecb_disable(ecb); 11864 ASSERT(probe->dtpr_ecb != ecb); 11865 dtrace_ecb_destroy(ecb); 11866 } 11867 } 11868 11869 mutex_exit(&dtrace_lock); 11870 mutex_exit(&cpu_lock); 11871} 11872 11873/* 11874 * DTrace DOF Functions 11875 */ 11876/*ARGSUSED*/ 11877static void 11878dtrace_dof_error(dof_hdr_t *dof, const char *str) 11879{ 11880 if (dtrace_err_verbose) 11881 cmn_err(CE_WARN, "failed to process DOF: %s", str); 11882 11883#ifdef DTRACE_ERRDEBUG 11884 dtrace_errdebug(str); 11885#endif 11886} 11887 11888/* 11889 * Create DOF out of a currently enabled state. Right now, we only create 11890 * DOF containing the run-time options -- but this could be expanded to create 11891 * complete DOF representing the enabled state. 11892 */ 11893static dof_hdr_t * 11894dtrace_dof_create(dtrace_state_t *state) 11895{ 11896 dof_hdr_t *dof; 11897 dof_sec_t *sec; 11898 dof_optdesc_t *opt; 11899 int i, len = sizeof (dof_hdr_t) + 11900 roundup(sizeof (dof_sec_t), sizeof (uint64_t)) + 11901 sizeof (dof_optdesc_t) * DTRACEOPT_MAX; 11902 11903 ASSERT(MUTEX_HELD(&dtrace_lock)); 11904 11905 dof = kmem_zalloc(len, KM_SLEEP); 11906 dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0; 11907 dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1; 11908 dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2; 11909 dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3; 11910 11911 dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE; 11912 dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE; 11913 dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION; 11914 dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION; 11915 dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS; 11916 dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS; 11917 11918 dof->dofh_flags = 0; 11919 dof->dofh_hdrsize = sizeof (dof_hdr_t); 11920 dof->dofh_secsize = sizeof (dof_sec_t); 11921 dof->dofh_secnum = 1; /* only DOF_SECT_OPTDESC */ 11922 dof->dofh_secoff = sizeof (dof_hdr_t); 11923 dof->dofh_loadsz = len; 11924 dof->dofh_filesz = len; 11925 dof->dofh_pad = 0; 11926 11927 /* 11928 * Fill in the option section header... 11929 */ 11930 sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t)); 11931 sec->dofs_type = DOF_SECT_OPTDESC; 11932 sec->dofs_align = sizeof (uint64_t); 11933 sec->dofs_flags = DOF_SECF_LOAD; 11934 sec->dofs_entsize = sizeof (dof_optdesc_t); 11935 11936 opt = (dof_optdesc_t *)((uintptr_t)sec + 11937 roundup(sizeof (dof_sec_t), sizeof (uint64_t))); 11938 11939 sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof; 11940 sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX; 11941 11942 for (i = 0; i < DTRACEOPT_MAX; i++) { 11943 opt[i].dofo_option = i; 11944 opt[i].dofo_strtab = DOF_SECIDX_NONE; 11945 opt[i].dofo_value = state->dts_options[i]; 11946 } 11947 11948 return (dof); 11949} 11950 11951static dof_hdr_t * 11952dtrace_dof_copyin(uintptr_t uarg, int *errp) 11953{ 11954 dof_hdr_t hdr, *dof; 11955 11956 ASSERT(!MUTEX_HELD(&dtrace_lock)); 11957 11958 /* 11959 * First, we're going to copyin() the sizeof (dof_hdr_t). 11960 */ 11961 if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) { 11962 dtrace_dof_error(NULL, "failed to copyin DOF header"); 11963 *errp = EFAULT; 11964 return (NULL); 11965 } 11966 11967 /* 11968 * Now we'll allocate the entire DOF and copy it in -- provided 11969 * that the length isn't outrageous. 11970 */ 11971 if (hdr.dofh_loadsz >= dtrace_dof_maxsize) { 11972 dtrace_dof_error(&hdr, "load size exceeds maximum"); 11973 *errp = E2BIG; 11974 return (NULL); 11975 } 11976 11977 if (hdr.dofh_loadsz < sizeof (hdr)) { 11978 dtrace_dof_error(&hdr, "invalid load size"); 11979 *errp = EINVAL; 11980 return (NULL); 11981 } 11982 11983 dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP); 11984 11985 if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0 || 11986 dof->dofh_loadsz != hdr.dofh_loadsz) { 11987 kmem_free(dof, hdr.dofh_loadsz); 11988 *errp = EFAULT; 11989 return (NULL); 11990 } 11991 11992 return (dof); 11993} 11994 11995#if !defined(sun) 11996static __inline uchar_t 11997dtrace_dof_char(char c) { 11998 switch (c) { 11999 case '0': 12000 case '1': 12001 case '2': 12002 case '3': 12003 case '4': 12004 case '5': 12005 case '6': 12006 case '7': 12007 case '8': 12008 case '9': 12009 return (c - '0'); 12010 case 'A': 12011 case 'B': 12012 case 'C': 12013 case 'D': 12014 case 'E': 12015 case 'F': 12016 return (c - 'A' + 10); 12017 case 'a': 12018 case 'b': 12019 case 'c': 12020 case 'd': 12021 case 'e': 12022 case 'f': 12023 return (c - 'a' + 10); 12024 } 12025 /* Should not reach here. */ 12026 return (0); 12027} 12028#endif 12029 12030static dof_hdr_t * 12031dtrace_dof_property(const char *name) 12032{ 12033 uchar_t *buf; 12034 uint64_t loadsz; 12035 unsigned int len, i; 12036 dof_hdr_t *dof; 12037 12038#if defined(sun) 12039 /* 12040 * Unfortunately, array of values in .conf files are always (and 12041 * only) interpreted to be integer arrays. We must read our DOF 12042 * as an integer array, and then squeeze it into a byte array. 12043 */ 12044 if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0, 12045 (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS) 12046 return (NULL); 12047 12048 for (i = 0; i < len; i++) 12049 buf[i] = (uchar_t)(((int *)buf)[i]); 12050 12051 if (len < sizeof (dof_hdr_t)) { 12052 ddi_prop_free(buf); 12053 dtrace_dof_error(NULL, "truncated header"); 12054 return (NULL); 12055 } 12056 12057 if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) { 12058 ddi_prop_free(buf); 12059 dtrace_dof_error(NULL, "truncated DOF"); 12060 return (NULL); 12061 } 12062 12063 if (loadsz >= dtrace_dof_maxsize) { 12064 ddi_prop_free(buf); 12065 dtrace_dof_error(NULL, "oversized DOF"); 12066 return (NULL); 12067 } 12068 12069 dof = kmem_alloc(loadsz, KM_SLEEP); 12070 bcopy(buf, dof, loadsz); 12071 ddi_prop_free(buf); 12072#else 12073 char *p; 12074 char *p_env; 12075 12076 if ((p_env = getenv(name)) == NULL) 12077 return (NULL); 12078 12079 len = strlen(p_env) / 2; 12080 12081 buf = kmem_alloc(len, KM_SLEEP); 12082 12083 dof = (dof_hdr_t *) buf; 12084 12085 p = p_env; 12086 12087 for (i = 0; i < len; i++) { 12088 buf[i] = (dtrace_dof_char(p[0]) << 4) | 12089 dtrace_dof_char(p[1]); 12090 p += 2; 12091 } 12092 12093 freeenv(p_env); 12094 12095 if (len < sizeof (dof_hdr_t)) { 12096 kmem_free(buf, 0); 12097 dtrace_dof_error(NULL, "truncated header"); 12098 return (NULL); 12099 } 12100 12101 if (len < (loadsz = dof->dofh_loadsz)) { 12102 kmem_free(buf, 0); 12103 dtrace_dof_error(NULL, "truncated DOF"); 12104 return (NULL); 12105 } 12106 12107 if (loadsz >= dtrace_dof_maxsize) { 12108 kmem_free(buf, 0); 12109 dtrace_dof_error(NULL, "oversized DOF"); 12110 return (NULL); 12111 } 12112#endif 12113 12114 return (dof); 12115} 12116 12117static void 12118dtrace_dof_destroy(dof_hdr_t *dof) 12119{ 12120 kmem_free(dof, dof->dofh_loadsz); 12121} 12122 12123/* 12124 * Return the dof_sec_t pointer corresponding to a given section index. If the 12125 * index is not valid, dtrace_dof_error() is called and NULL is returned. If 12126 * a type other than DOF_SECT_NONE is specified, the header is checked against 12127 * this type and NULL is returned if the types do not match. 12128 */ 12129static dof_sec_t * 12130dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i) 12131{ 12132 dof_sec_t *sec = (dof_sec_t *)(uintptr_t) 12133 ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize); 12134 12135 if (i >= dof->dofh_secnum) { 12136 dtrace_dof_error(dof, "referenced section index is invalid"); 12137 return (NULL); 12138 } 12139 12140 if (!(sec->dofs_flags & DOF_SECF_LOAD)) { 12141 dtrace_dof_error(dof, "referenced section is not loadable"); 12142 return (NULL); 12143 } 12144 12145 if (type != DOF_SECT_NONE && type != sec->dofs_type) { 12146 dtrace_dof_error(dof, "referenced section is the wrong type"); 12147 return (NULL); 12148 } 12149 12150 return (sec); 12151} 12152 12153static dtrace_probedesc_t * 12154dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc) 12155{ 12156 dof_probedesc_t *probe; 12157 dof_sec_t *strtab; 12158 uintptr_t daddr = (uintptr_t)dof; 12159 uintptr_t str; 12160 size_t size; 12161 12162 if (sec->dofs_type != DOF_SECT_PROBEDESC) { 12163 dtrace_dof_error(dof, "invalid probe section"); 12164 return (NULL); 12165 } 12166 12167 if (sec->dofs_align != sizeof (dof_secidx_t)) { 12168 dtrace_dof_error(dof, "bad alignment in probe description"); 12169 return (NULL); 12170 } 12171 12172 if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) { 12173 dtrace_dof_error(dof, "truncated probe description"); 12174 return (NULL); 12175 } 12176 12177 probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset); 12178 strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab); 12179 12180 if (strtab == NULL) 12181 return (NULL); 12182 12183 str = daddr + strtab->dofs_offset; 12184 size = strtab->dofs_size; 12185 12186 if (probe->dofp_provider >= strtab->dofs_size) { 12187 dtrace_dof_error(dof, "corrupt probe provider"); 12188 return (NULL); 12189 } 12190 12191 (void) strncpy(desc->dtpd_provider, 12192 (char *)(str + probe->dofp_provider), 12193 MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider)); 12194 12195 if (probe->dofp_mod >= strtab->dofs_size) { 12196 dtrace_dof_error(dof, "corrupt probe module"); 12197 return (NULL); 12198 } 12199 12200 (void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod), 12201 MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod)); 12202 12203 if (probe->dofp_func >= strtab->dofs_size) { 12204 dtrace_dof_error(dof, "corrupt probe function"); 12205 return (NULL); 12206 } 12207 12208 (void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func), 12209 MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func)); 12210 12211 if (probe->dofp_name >= strtab->dofs_size) { 12212 dtrace_dof_error(dof, "corrupt probe name"); 12213 return (NULL); 12214 } 12215 12216 (void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name), 12217 MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name)); 12218 12219 return (desc); 12220} 12221 12222static dtrace_difo_t * 12223dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate, 12224 cred_t *cr) 12225{ 12226 dtrace_difo_t *dp; 12227 size_t ttl = 0; 12228 dof_difohdr_t *dofd; 12229 uintptr_t daddr = (uintptr_t)dof; 12230 size_t max = dtrace_difo_maxsize; 12231 int i, l, n; 12232 12233 static const struct { 12234 int section; 12235 int bufoffs; 12236 int lenoffs; 12237 int entsize; 12238 int align; 12239 const char *msg; 12240 } difo[] = { 12241 { DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf), 12242 offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t), 12243 sizeof (dif_instr_t), "multiple DIF sections" }, 12244 12245 { DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab), 12246 offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t), 12247 sizeof (uint64_t), "multiple integer tables" }, 12248 12249 { DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab), 12250 offsetof(dtrace_difo_t, dtdo_strlen), 0, 12251 sizeof (char), "multiple string tables" }, 12252 12253 { DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab), 12254 offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t), 12255 sizeof (uint_t), "multiple variable tables" }, 12256 12257 { DOF_SECT_NONE, 0, 0, 0, 0, NULL } 12258 }; 12259 12260 if (sec->dofs_type != DOF_SECT_DIFOHDR) { 12261 dtrace_dof_error(dof, "invalid DIFO header section"); 12262 return (NULL); 12263 } 12264 12265 if (sec->dofs_align != sizeof (dof_secidx_t)) { 12266 dtrace_dof_error(dof, "bad alignment in DIFO header"); 12267 return (NULL); 12268 } 12269 12270 if (sec->dofs_size < sizeof (dof_difohdr_t) || 12271 sec->dofs_size % sizeof (dof_secidx_t)) { 12272 dtrace_dof_error(dof, "bad size in DIFO header"); 12273 return (NULL); 12274 } 12275 12276 dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset); 12277 n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1; 12278 12279 dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP); 12280 dp->dtdo_rtype = dofd->dofd_rtype; 12281 12282 for (l = 0; l < n; l++) { 12283 dof_sec_t *subsec; 12284 void **bufp; 12285 uint32_t *lenp; 12286 12287 if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE, 12288 dofd->dofd_links[l])) == NULL) 12289 goto err; /* invalid section link */ 12290 12291 if (ttl + subsec->dofs_size > max) { 12292 dtrace_dof_error(dof, "exceeds maximum size"); 12293 goto err; 12294 } 12295 12296 ttl += subsec->dofs_size; 12297 12298 for (i = 0; difo[i].section != DOF_SECT_NONE; i++) { 12299 if (subsec->dofs_type != difo[i].section) 12300 continue; 12301 12302 if (!(subsec->dofs_flags & DOF_SECF_LOAD)) { 12303 dtrace_dof_error(dof, "section not loaded"); 12304 goto err; 12305 } 12306 12307 if (subsec->dofs_align != difo[i].align) { 12308 dtrace_dof_error(dof, "bad alignment"); 12309 goto err; 12310 } 12311 12312 bufp = (void **)((uintptr_t)dp + difo[i].bufoffs); 12313 lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs); 12314 12315 if (*bufp != NULL) { 12316 dtrace_dof_error(dof, difo[i].msg); 12317 goto err; 12318 } 12319 12320 if (difo[i].entsize != subsec->dofs_entsize) { 12321 dtrace_dof_error(dof, "entry size mismatch"); 12322 goto err; 12323 } 12324 12325 if (subsec->dofs_entsize != 0 && 12326 (subsec->dofs_size % subsec->dofs_entsize) != 0) { 12327 dtrace_dof_error(dof, "corrupt entry size"); 12328 goto err; 12329 } 12330 12331 *lenp = subsec->dofs_size; 12332 *bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP); 12333 bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset), 12334 *bufp, subsec->dofs_size); 12335 12336 if (subsec->dofs_entsize != 0) 12337 *lenp /= subsec->dofs_entsize; 12338 12339 break; 12340 } 12341 12342 /* 12343 * If we encounter a loadable DIFO sub-section that is not 12344 * known to us, assume this is a broken program and fail. 12345 */ 12346 if (difo[i].section == DOF_SECT_NONE && 12347 (subsec->dofs_flags & DOF_SECF_LOAD)) { 12348 dtrace_dof_error(dof, "unrecognized DIFO subsection"); 12349 goto err; 12350 } 12351 } 12352 12353 if (dp->dtdo_buf == NULL) { 12354 /* 12355 * We can't have a DIF object without DIF text. 12356 */ 12357 dtrace_dof_error(dof, "missing DIF text"); 12358 goto err; 12359 } 12360 12361 /* 12362 * Before we validate the DIF object, run through the variable table 12363 * looking for the strings -- if any of their size are under, we'll set 12364 * their size to be the system-wide default string size. Note that 12365 * this should _not_ happen if the "strsize" option has been set -- 12366 * in this case, the compiler should have set the size to reflect the 12367 * setting of the option. 12368 */ 12369 for (i = 0; i < dp->dtdo_varlen; i++) { 12370 dtrace_difv_t *v = &dp->dtdo_vartab[i]; 12371 dtrace_diftype_t *t = &v->dtdv_type; 12372 12373 if (v->dtdv_id < DIF_VAR_OTHER_UBASE) 12374 continue; 12375 12376 if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0) 12377 t->dtdt_size = dtrace_strsize_default; 12378 } 12379 12380 if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0) 12381 goto err; 12382 12383 dtrace_difo_init(dp, vstate); 12384 return (dp); 12385 12386err: 12387 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t)); 12388 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t)); 12389 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen); 12390 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t)); 12391 12392 kmem_free(dp, sizeof (dtrace_difo_t)); 12393 return (NULL); 12394} 12395 12396static dtrace_predicate_t * 12397dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate, 12398 cred_t *cr) 12399{ 12400 dtrace_difo_t *dp; 12401 12402 if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL) 12403 return (NULL); 12404 12405 return (dtrace_predicate_create(dp)); 12406} 12407 12408static dtrace_actdesc_t * 12409dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate, 12410 cred_t *cr) 12411{ 12412 dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next; 12413 dof_actdesc_t *desc; 12414 dof_sec_t *difosec; 12415 size_t offs; 12416 uintptr_t daddr = (uintptr_t)dof; 12417 uint64_t arg; 12418 dtrace_actkind_t kind; 12419 12420 if (sec->dofs_type != DOF_SECT_ACTDESC) { 12421 dtrace_dof_error(dof, "invalid action section"); 12422 return (NULL); 12423 } 12424 12425 if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) { 12426 dtrace_dof_error(dof, "truncated action description"); 12427 return (NULL); 12428 } 12429 12430 if (sec->dofs_align != sizeof (uint64_t)) { 12431 dtrace_dof_error(dof, "bad alignment in action description"); 12432 return (NULL); 12433 } 12434 12435 if (sec->dofs_size < sec->dofs_entsize) { 12436 dtrace_dof_error(dof, "section entry size exceeds total size"); 12437 return (NULL); 12438 } 12439 12440 if (sec->dofs_entsize != sizeof (dof_actdesc_t)) { 12441 dtrace_dof_error(dof, "bad entry size in action description"); 12442 return (NULL); 12443 } 12444 12445 if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) { 12446 dtrace_dof_error(dof, "actions exceed dtrace_actions_max"); 12447 return (NULL); 12448 } 12449 12450 for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) { 12451 desc = (dof_actdesc_t *)(daddr + 12452 (uintptr_t)sec->dofs_offset + offs); 12453 kind = (dtrace_actkind_t)desc->dofa_kind; 12454 12455 if ((DTRACEACT_ISPRINTFLIKE(kind) && 12456 (kind != DTRACEACT_PRINTA || 12457 desc->dofa_strtab != DOF_SECIDX_NONE)) || 12458 (kind == DTRACEACT_DIFEXPR && 12459 desc->dofa_strtab != DOF_SECIDX_NONE)) { 12460 dof_sec_t *strtab; 12461 char *str, *fmt; 12462 uint64_t i; 12463 12464 /* 12465 * The argument to these actions is an index into the 12466 * DOF string table. For printf()-like actions, this 12467 * is the format string. For print(), this is the 12468 * CTF type of the expression result. 12469 */ 12470 if ((strtab = dtrace_dof_sect(dof, 12471 DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL) 12472 goto err; 12473 12474 str = (char *)((uintptr_t)dof + 12475 (uintptr_t)strtab->dofs_offset); 12476 12477 for (i = desc->dofa_arg; i < strtab->dofs_size; i++) { 12478 if (str[i] == '\0') 12479 break; 12480 } 12481 12482 if (i >= strtab->dofs_size) { 12483 dtrace_dof_error(dof, "bogus format string"); 12484 goto err; 12485 } 12486 12487 if (i == desc->dofa_arg) { 12488 dtrace_dof_error(dof, "empty format string"); 12489 goto err; 12490 } 12491 12492 i -= desc->dofa_arg; 12493 fmt = kmem_alloc(i + 1, KM_SLEEP); 12494 bcopy(&str[desc->dofa_arg], fmt, i + 1); 12495 arg = (uint64_t)(uintptr_t)fmt; 12496 } else { 12497 if (kind == DTRACEACT_PRINTA) { 12498 ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE); 12499 arg = 0; 12500 } else { 12501 arg = desc->dofa_arg; 12502 } 12503 } 12504 12505 act = dtrace_actdesc_create(kind, desc->dofa_ntuple, 12506 desc->dofa_uarg, arg); 12507 12508 if (last != NULL) { 12509 last->dtad_next = act; 12510 } else { 12511 first = act; 12512 } 12513 12514 last = act; 12515 12516 if (desc->dofa_difo == DOF_SECIDX_NONE) 12517 continue; 12518 12519 if ((difosec = dtrace_dof_sect(dof, 12520 DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL) 12521 goto err; 12522 12523 act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr); 12524 12525 if (act->dtad_difo == NULL) 12526 goto err; 12527 } 12528 12529 ASSERT(first != NULL); 12530 return (first); 12531 12532err: 12533 for (act = first; act != NULL; act = next) { 12534 next = act->dtad_next; 12535 dtrace_actdesc_release(act, vstate); 12536 } 12537 12538 return (NULL); 12539} 12540 12541static dtrace_ecbdesc_t * 12542dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate, 12543 cred_t *cr) 12544{ 12545 dtrace_ecbdesc_t *ep; 12546 dof_ecbdesc_t *ecb; 12547 dtrace_probedesc_t *desc; 12548 dtrace_predicate_t *pred = NULL; 12549 12550 if (sec->dofs_size < sizeof (dof_ecbdesc_t)) { 12551 dtrace_dof_error(dof, "truncated ECB description"); 12552 return (NULL); 12553 } 12554 12555 if (sec->dofs_align != sizeof (uint64_t)) { 12556 dtrace_dof_error(dof, "bad alignment in ECB description"); 12557 return (NULL); 12558 } 12559 12560 ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset); 12561 sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes); 12562 12563 if (sec == NULL) 12564 return (NULL); 12565 12566 ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP); 12567 ep->dted_uarg = ecb->dofe_uarg; 12568 desc = &ep->dted_probe; 12569 12570 if (dtrace_dof_probedesc(dof, sec, desc) == NULL) 12571 goto err; 12572 12573 if (ecb->dofe_pred != DOF_SECIDX_NONE) { 12574 if ((sec = dtrace_dof_sect(dof, 12575 DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL) 12576 goto err; 12577 12578 if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL) 12579 goto err; 12580 12581 ep->dted_pred.dtpdd_predicate = pred; 12582 } 12583 12584 if (ecb->dofe_actions != DOF_SECIDX_NONE) { 12585 if ((sec = dtrace_dof_sect(dof, 12586 DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL) 12587 goto err; 12588 12589 ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr); 12590 12591 if (ep->dted_action == NULL) 12592 goto err; 12593 } 12594 12595 return (ep); 12596 12597err: 12598 if (pred != NULL) 12599 dtrace_predicate_release(pred, vstate); 12600 kmem_free(ep, sizeof (dtrace_ecbdesc_t)); 12601 return (NULL); 12602} 12603 12604/* 12605 * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the 12606 * specified DOF. At present, this amounts to simply adding 'ubase' to the 12607 * site of any user SETX relocations to account for load object base address. 12608 * In the future, if we need other relocations, this function can be extended. 12609 */ 12610static int 12611dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase) 12612{ 12613 uintptr_t daddr = (uintptr_t)dof; 12614 dof_relohdr_t *dofr = 12615 (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset); 12616 dof_sec_t *ss, *rs, *ts; 12617 dof_relodesc_t *r; 12618 uint_t i, n; 12619 12620 if (sec->dofs_size < sizeof (dof_relohdr_t) || 12621 sec->dofs_align != sizeof (dof_secidx_t)) { 12622 dtrace_dof_error(dof, "invalid relocation header"); 12623 return (-1); 12624 } 12625 12626 ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab); 12627 rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec); 12628 ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec); 12629 12630 if (ss == NULL || rs == NULL || ts == NULL) 12631 return (-1); /* dtrace_dof_error() has been called already */ 12632 12633 if (rs->dofs_entsize < sizeof (dof_relodesc_t) || 12634 rs->dofs_align != sizeof (uint64_t)) { 12635 dtrace_dof_error(dof, "invalid relocation section"); 12636 return (-1); 12637 } 12638 12639 r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset); 12640 n = rs->dofs_size / rs->dofs_entsize; 12641 12642 for (i = 0; i < n; i++) { 12643 uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset; 12644 12645 switch (r->dofr_type) { 12646 case DOF_RELO_NONE: 12647 break; 12648 case DOF_RELO_SETX: 12649 if (r->dofr_offset >= ts->dofs_size || r->dofr_offset + 12650 sizeof (uint64_t) > ts->dofs_size) { 12651 dtrace_dof_error(dof, "bad relocation offset"); 12652 return (-1); 12653 } 12654 12655 if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) { 12656 dtrace_dof_error(dof, "misaligned setx relo"); 12657 return (-1); 12658 } 12659 12660 *(uint64_t *)taddr += ubase; 12661 break; 12662 default: 12663 dtrace_dof_error(dof, "invalid relocation type"); 12664 return (-1); 12665 } 12666 12667 r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize); 12668 } 12669 12670 return (0); 12671} 12672 12673/* 12674 * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated 12675 * header: it should be at the front of a memory region that is at least 12676 * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in 12677 * size. It need not be validated in any other way. 12678 */ 12679static int 12680dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr, 12681 dtrace_enabling_t **enabp, uint64_t ubase, int noprobes) 12682{ 12683 uint64_t len = dof->dofh_loadsz, seclen; 12684 uintptr_t daddr = (uintptr_t)dof; 12685 dtrace_ecbdesc_t *ep; 12686 dtrace_enabling_t *enab; 12687 uint_t i; 12688 12689 ASSERT(MUTEX_HELD(&dtrace_lock)); 12690 ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t)); 12691 12692 /* 12693 * Check the DOF header identification bytes. In addition to checking 12694 * valid settings, we also verify that unused bits/bytes are zeroed so 12695 * we can use them later without fear of regressing existing binaries. 12696 */ 12697 if (bcmp(&dof->dofh_ident[DOF_ID_MAG0], 12698 DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) { 12699 dtrace_dof_error(dof, "DOF magic string mismatch"); 12700 return (-1); 12701 } 12702 12703 if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 && 12704 dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) { 12705 dtrace_dof_error(dof, "DOF has invalid data model"); 12706 return (-1); 12707 } 12708 12709 if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) { 12710 dtrace_dof_error(dof, "DOF encoding mismatch"); 12711 return (-1); 12712 } 12713 12714 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 && 12715 dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) { 12716 dtrace_dof_error(dof, "DOF version mismatch"); 12717 return (-1); 12718 } 12719 12720 if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) { 12721 dtrace_dof_error(dof, "DOF uses unsupported instruction set"); 12722 return (-1); 12723 } 12724 12725 if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) { 12726 dtrace_dof_error(dof, "DOF uses too many integer registers"); 12727 return (-1); 12728 } 12729 12730 if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) { 12731 dtrace_dof_error(dof, "DOF uses too many tuple registers"); 12732 return (-1); 12733 } 12734 12735 for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) { 12736 if (dof->dofh_ident[i] != 0) { 12737 dtrace_dof_error(dof, "DOF has invalid ident byte set"); 12738 return (-1); 12739 } 12740 } 12741 12742 if (dof->dofh_flags & ~DOF_FL_VALID) { 12743 dtrace_dof_error(dof, "DOF has invalid flag bits set"); 12744 return (-1); 12745 } 12746 12747 if (dof->dofh_secsize == 0) { 12748 dtrace_dof_error(dof, "zero section header size"); 12749 return (-1); 12750 } 12751 12752 /* 12753 * Check that the section headers don't exceed the amount of DOF 12754 * data. Note that we cast the section size and number of sections 12755 * to uint64_t's to prevent possible overflow in the multiplication. 12756 */ 12757 seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize; 12758 12759 if (dof->dofh_secoff > len || seclen > len || 12760 dof->dofh_secoff + seclen > len) { 12761 dtrace_dof_error(dof, "truncated section headers"); 12762 return (-1); 12763 } 12764 12765 if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) { 12766 dtrace_dof_error(dof, "misaligned section headers"); 12767 return (-1); 12768 } 12769 12770 if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) { 12771 dtrace_dof_error(dof, "misaligned section size"); 12772 return (-1); 12773 } 12774 12775 /* 12776 * Take an initial pass through the section headers to be sure that 12777 * the headers don't have stray offsets. If the 'noprobes' flag is 12778 * set, do not permit sections relating to providers, probes, or args. 12779 */ 12780 for (i = 0; i < dof->dofh_secnum; i++) { 12781 dof_sec_t *sec = (dof_sec_t *)(daddr + 12782 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize); 12783 12784 if (noprobes) { 12785 switch (sec->dofs_type) { 12786 case DOF_SECT_PROVIDER: 12787 case DOF_SECT_PROBES: 12788 case DOF_SECT_PRARGS: 12789 case DOF_SECT_PROFFS: 12790 dtrace_dof_error(dof, "illegal sections " 12791 "for enabling"); 12792 return (-1); 12793 } 12794 } 12795 12796 if (DOF_SEC_ISLOADABLE(sec->dofs_type) && 12797 !(sec->dofs_flags & DOF_SECF_LOAD)) { 12798 dtrace_dof_error(dof, "loadable section with load " 12799 "flag unset"); 12800 return (-1); 12801 } 12802 12803 if (!(sec->dofs_flags & DOF_SECF_LOAD)) 12804 continue; /* just ignore non-loadable sections */ 12805 12806 if (sec->dofs_align & (sec->dofs_align - 1)) { 12807 dtrace_dof_error(dof, "bad section alignment"); 12808 return (-1); 12809 } 12810 12811 if (sec->dofs_offset & (sec->dofs_align - 1)) { 12812 dtrace_dof_error(dof, "misaligned section"); 12813 return (-1); 12814 } 12815 12816 if (sec->dofs_offset > len || sec->dofs_size > len || 12817 sec->dofs_offset + sec->dofs_size > len) { 12818 dtrace_dof_error(dof, "corrupt section header"); 12819 return (-1); 12820 } 12821 12822 if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr + 12823 sec->dofs_offset + sec->dofs_size - 1) != '\0') { 12824 dtrace_dof_error(dof, "non-terminating string table"); 12825 return (-1); 12826 } 12827 } 12828 12829 /* 12830 * Take a second pass through the sections and locate and perform any 12831 * relocations that are present. We do this after the first pass to 12832 * be sure that all sections have had their headers validated. 12833 */ 12834 for (i = 0; i < dof->dofh_secnum; i++) { 12835 dof_sec_t *sec = (dof_sec_t *)(daddr + 12836 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize); 12837 12838 if (!(sec->dofs_flags & DOF_SECF_LOAD)) 12839 continue; /* skip sections that are not loadable */ 12840 12841 switch (sec->dofs_type) { 12842 case DOF_SECT_URELHDR: 12843 if (dtrace_dof_relocate(dof, sec, ubase) != 0) 12844 return (-1); 12845 break; 12846 } 12847 } 12848 12849 if ((enab = *enabp) == NULL) 12850 enab = *enabp = dtrace_enabling_create(vstate); 12851 12852 for (i = 0; i < dof->dofh_secnum; i++) { 12853 dof_sec_t *sec = (dof_sec_t *)(daddr + 12854 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize); 12855 12856 if (sec->dofs_type != DOF_SECT_ECBDESC) 12857 continue; 12858 12859 if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) { 12860 dtrace_enabling_destroy(enab); 12861 *enabp = NULL; 12862 return (-1); 12863 } 12864 12865 dtrace_enabling_add(enab, ep); 12866 } 12867 12868 return (0); 12869} 12870 12871/* 12872 * Process DOF for any options. This routine assumes that the DOF has been 12873 * at least processed by dtrace_dof_slurp(). 12874 */ 12875static int 12876dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state) 12877{ 12878 int i, rval; 12879 uint32_t entsize; 12880 size_t offs; 12881 dof_optdesc_t *desc; 12882 12883 for (i = 0; i < dof->dofh_secnum; i++) { 12884 dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof + 12885 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize); 12886 12887 if (sec->dofs_type != DOF_SECT_OPTDESC) 12888 continue; 12889 12890 if (sec->dofs_align != sizeof (uint64_t)) { 12891 dtrace_dof_error(dof, "bad alignment in " 12892 "option description"); 12893 return (EINVAL); 12894 } 12895 12896 if ((entsize = sec->dofs_entsize) == 0) { 12897 dtrace_dof_error(dof, "zeroed option entry size"); 12898 return (EINVAL); 12899 } 12900 12901 if (entsize < sizeof (dof_optdesc_t)) { 12902 dtrace_dof_error(dof, "bad option entry size"); 12903 return (EINVAL); 12904 } 12905 12906 for (offs = 0; offs < sec->dofs_size; offs += entsize) { 12907 desc = (dof_optdesc_t *)((uintptr_t)dof + 12908 (uintptr_t)sec->dofs_offset + offs); 12909 12910 if (desc->dofo_strtab != DOF_SECIDX_NONE) { 12911 dtrace_dof_error(dof, "non-zero option string"); 12912 return (EINVAL); 12913 } 12914 12915 if (desc->dofo_value == DTRACEOPT_UNSET) { 12916 dtrace_dof_error(dof, "unset option"); 12917 return (EINVAL); 12918 } 12919 12920 if ((rval = dtrace_state_option(state, 12921 desc->dofo_option, desc->dofo_value)) != 0) { 12922 dtrace_dof_error(dof, "rejected option"); 12923 return (rval); 12924 } 12925 } 12926 } 12927 12928 return (0); 12929} 12930 12931/* 12932 * DTrace Consumer State Functions 12933 */ 12934static int 12935dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size) 12936{ 12937 size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize; 12938 void *base; 12939 uintptr_t limit; 12940 dtrace_dynvar_t *dvar, *next, *start; 12941 int i; 12942 12943 ASSERT(MUTEX_HELD(&dtrace_lock)); 12944 ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL); 12945 12946 bzero(dstate, sizeof (dtrace_dstate_t)); 12947 12948 if ((dstate->dtds_chunksize = chunksize) == 0) 12949 dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE; 12950 12951 if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t))) 12952 size = min; 12953 12954 if ((base = kmem_zalloc(size, KM_NOSLEEP | KM_NORMALPRI)) == NULL) 12955 return (ENOMEM); 12956 12957 dstate->dtds_size = size; 12958 dstate->dtds_base = base; 12959 dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP); 12960 bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t)); 12961 12962 hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)); 12963 12964 if (hashsize != 1 && (hashsize & 1)) 12965 hashsize--; 12966 12967 dstate->dtds_hashsize = hashsize; 12968 dstate->dtds_hash = dstate->dtds_base; 12969 12970 /* 12971 * Set all of our hash buckets to point to the single sink, and (if 12972 * it hasn't already been set), set the sink's hash value to be the 12973 * sink sentinel value. The sink is needed for dynamic variable 12974 * lookups to know that they have iterated over an entire, valid hash 12975 * chain. 12976 */ 12977 for (i = 0; i < hashsize; i++) 12978 dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink; 12979 12980 if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK) 12981 dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK; 12982 12983 /* 12984 * Determine number of active CPUs. Divide free list evenly among 12985 * active CPUs. 12986 */ 12987 start = (dtrace_dynvar_t *) 12988 ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t)); 12989 limit = (uintptr_t)base + size; 12990 12991 maxper = (limit - (uintptr_t)start) / NCPU; 12992 maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize; 12993 12994#if !defined(sun) 12995 CPU_FOREACH(i) { 12996#else 12997 for (i = 0; i < NCPU; i++) { 12998#endif 12999 dstate->dtds_percpu[i].dtdsc_free = dvar = start; 13000 13001 /* 13002 * If we don't even have enough chunks to make it once through 13003 * NCPUs, we're just going to allocate everything to the first 13004 * CPU. And if we're on the last CPU, we're going to allocate 13005 * whatever is left over. In either case, we set the limit to 13006 * be the limit of the dynamic variable space. 13007 */ 13008 if (maxper == 0 || i == NCPU - 1) { 13009 limit = (uintptr_t)base + size; 13010 start = NULL; 13011 } else { 13012 limit = (uintptr_t)start + maxper; 13013 start = (dtrace_dynvar_t *)limit; 13014 } 13015 13016 ASSERT(limit <= (uintptr_t)base + size); 13017 13018 for (;;) { 13019 next = (dtrace_dynvar_t *)((uintptr_t)dvar + 13020 dstate->dtds_chunksize); 13021 13022 if ((uintptr_t)next + dstate->dtds_chunksize >= limit) 13023 break; 13024 13025 dvar->dtdv_next = next; 13026 dvar = next; 13027 } 13028 13029 if (maxper == 0) 13030 break; 13031 } 13032 13033 return (0); 13034} 13035 13036static void 13037dtrace_dstate_fini(dtrace_dstate_t *dstate) 13038{ 13039 ASSERT(MUTEX_HELD(&cpu_lock)); 13040 13041 if (dstate->dtds_base == NULL) 13042 return; 13043 13044 kmem_free(dstate->dtds_base, dstate->dtds_size); 13045 kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu); 13046} 13047 13048static void 13049dtrace_vstate_fini(dtrace_vstate_t *vstate) 13050{ 13051 /* 13052 * Logical XOR, where are you? 13053 */ 13054 ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL)); 13055 13056 if (vstate->dtvs_nglobals > 0) { 13057 kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals * 13058 sizeof (dtrace_statvar_t *)); 13059 } 13060 13061 if (vstate->dtvs_ntlocals > 0) { 13062 kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals * 13063 sizeof (dtrace_difv_t)); 13064 } 13065 13066 ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL)); 13067 13068 if (vstate->dtvs_nlocals > 0) { 13069 kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals * 13070 sizeof (dtrace_statvar_t *)); 13071 } 13072} 13073 13074#if defined(sun) 13075static void 13076dtrace_state_clean(dtrace_state_t *state) 13077{ 13078 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) 13079 return; 13080 13081 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars); 13082 dtrace_speculation_clean(state); 13083} 13084 13085static void 13086dtrace_state_deadman(dtrace_state_t *state) 13087{ 13088 hrtime_t now; 13089 13090 dtrace_sync(); 13091 13092 now = dtrace_gethrtime(); 13093 13094 if (state != dtrace_anon.dta_state && 13095 now - state->dts_laststatus >= dtrace_deadman_user) 13096 return; 13097 13098 /* 13099 * We must be sure that dts_alive never appears to be less than the 13100 * value upon entry to dtrace_state_deadman(), and because we lack a 13101 * dtrace_cas64(), we cannot store to it atomically. We thus instead 13102 * store INT64_MAX to it, followed by a memory barrier, followed by 13103 * the new value. This assures that dts_alive never appears to be 13104 * less than its true value, regardless of the order in which the 13105 * stores to the underlying storage are issued. 13106 */ 13107 state->dts_alive = INT64_MAX; 13108 dtrace_membar_producer(); 13109 state->dts_alive = now; 13110} 13111#else 13112static void 13113dtrace_state_clean(void *arg) 13114{ 13115 dtrace_state_t *state = arg; 13116 dtrace_optval_t *opt = state->dts_options; 13117 13118 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) 13119 return; 13120 13121 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars); 13122 dtrace_speculation_clean(state); 13123 13124 callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC, 13125 dtrace_state_clean, state); 13126} 13127 13128static void 13129dtrace_state_deadman(void *arg) 13130{ 13131 dtrace_state_t *state = arg; 13132 hrtime_t now; 13133 13134 dtrace_sync(); 13135 13136 dtrace_debug_output(); 13137 13138 now = dtrace_gethrtime(); 13139 13140 if (state != dtrace_anon.dta_state && 13141 now - state->dts_laststatus >= dtrace_deadman_user) 13142 return; 13143 13144 /* 13145 * We must be sure that dts_alive never appears to be less than the 13146 * value upon entry to dtrace_state_deadman(), and because we lack a 13147 * dtrace_cas64(), we cannot store to it atomically. We thus instead 13148 * store INT64_MAX to it, followed by a memory barrier, followed by 13149 * the new value. This assures that dts_alive never appears to be 13150 * less than its true value, regardless of the order in which the 13151 * stores to the underlying storage are issued. 13152 */ 13153 state->dts_alive = INT64_MAX; 13154 dtrace_membar_producer(); 13155 state->dts_alive = now; 13156 13157 callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC, 13158 dtrace_state_deadman, state); 13159} 13160#endif 13161 13162static dtrace_state_t * 13163#if defined(sun) 13164dtrace_state_create(dev_t *devp, cred_t *cr) 13165#else 13166dtrace_state_create(struct cdev *dev) 13167#endif 13168{ 13169#if defined(sun) 13170 minor_t minor; 13171 major_t major; 13172#else 13173 cred_t *cr = NULL; 13174 int m = 0; 13175#endif 13176 char c[30]; 13177 dtrace_state_t *state; 13178 dtrace_optval_t *opt; 13179 int bufsize = NCPU * sizeof (dtrace_buffer_t), i; 13180 13181 ASSERT(MUTEX_HELD(&dtrace_lock)); 13182 ASSERT(MUTEX_HELD(&cpu_lock)); 13183 13184#if defined(sun) 13185 minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1, 13186 VM_BESTFIT | VM_SLEEP); 13187 13188 if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) { 13189 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1); 13190 return (NULL); 13191 } 13192 13193 state = ddi_get_soft_state(dtrace_softstate, minor); 13194#else 13195 if (dev != NULL) { 13196 cr = dev->si_cred; 13197 m = dev2unit(dev); 13198 } 13199 13200 /* Allocate memory for the state. */ 13201 state = kmem_zalloc(sizeof(dtrace_state_t), KM_SLEEP); 13202#endif 13203 13204 state->dts_epid = DTRACE_EPIDNONE + 1; 13205 13206 (void) snprintf(c, sizeof (c), "dtrace_aggid_%d", m); 13207#if defined(sun) 13208 state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1, 13209 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER); 13210 13211 if (devp != NULL) { 13212 major = getemajor(*devp); 13213 } else { 13214 major = ddi_driver_major(dtrace_devi); 13215 } 13216 13217 state->dts_dev = makedevice(major, minor); 13218 13219 if (devp != NULL) 13220 *devp = state->dts_dev; 13221#else 13222 state->dts_aggid_arena = new_unrhdr(1, INT_MAX, &dtrace_unr_mtx); 13223 state->dts_dev = dev; 13224#endif 13225 13226 /* 13227 * We allocate NCPU buffers. On the one hand, this can be quite 13228 * a bit of memory per instance (nearly 36K on a Starcat). On the 13229 * other hand, it saves an additional memory reference in the probe 13230 * path. 13231 */ 13232 state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP); 13233 state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP); 13234 13235#if defined(sun) 13236 state->dts_cleaner = CYCLIC_NONE; 13237 state->dts_deadman = CYCLIC_NONE; 13238#else 13239 callout_init(&state->dts_cleaner, CALLOUT_MPSAFE); 13240 callout_init(&state->dts_deadman, CALLOUT_MPSAFE); 13241#endif 13242 state->dts_vstate.dtvs_state = state; 13243 13244 for (i = 0; i < DTRACEOPT_MAX; i++) 13245 state->dts_options[i] = DTRACEOPT_UNSET; 13246 13247 /* 13248 * Set the default options. 13249 */ 13250 opt = state->dts_options; 13251 opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH; 13252 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO; 13253 opt[DTRACEOPT_NSPEC] = dtrace_nspec_default; 13254 opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default; 13255 opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL; 13256 opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default; 13257 opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default; 13258 opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default; 13259 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default; 13260 opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default; 13261 opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default; 13262 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default; 13263 opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default; 13264 opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default; 13265 13266 state->dts_activity = DTRACE_ACTIVITY_INACTIVE; 13267 13268 /* 13269 * Depending on the user credentials, we set flag bits which alter probe 13270 * visibility or the amount of destructiveness allowed. In the case of 13271 * actual anonymous tracing, or the possession of all privileges, all of 13272 * the normal checks are bypassed. 13273 */ 13274 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) { 13275 state->dts_cred.dcr_visible = DTRACE_CRV_ALL; 13276 state->dts_cred.dcr_action = DTRACE_CRA_ALL; 13277 } else { 13278 /* 13279 * Set up the credentials for this instantiation. We take a 13280 * hold on the credential to prevent it from disappearing on 13281 * us; this in turn prevents the zone_t referenced by this 13282 * credential from disappearing. This means that we can 13283 * examine the credential and the zone from probe context. 13284 */ 13285 crhold(cr); 13286 state->dts_cred.dcr_cred = cr; 13287 13288 /* 13289 * CRA_PROC means "we have *some* privilege for dtrace" and 13290 * unlocks the use of variables like pid, zonename, etc. 13291 */ 13292 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) || 13293 PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) { 13294 state->dts_cred.dcr_action |= DTRACE_CRA_PROC; 13295 } 13296 13297 /* 13298 * dtrace_user allows use of syscall and profile providers. 13299 * If the user also has proc_owner and/or proc_zone, we 13300 * extend the scope to include additional visibility and 13301 * destructive power. 13302 */ 13303 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) { 13304 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) { 13305 state->dts_cred.dcr_visible |= 13306 DTRACE_CRV_ALLPROC; 13307 13308 state->dts_cred.dcr_action |= 13309 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER; 13310 } 13311 13312 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) { 13313 state->dts_cred.dcr_visible |= 13314 DTRACE_CRV_ALLZONE; 13315 13316 state->dts_cred.dcr_action |= 13317 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE; 13318 } 13319 13320 /* 13321 * If we have all privs in whatever zone this is, 13322 * we can do destructive things to processes which 13323 * have altered credentials. 13324 */ 13325#if defined(sun) 13326 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE), 13327 cr->cr_zone->zone_privset)) { 13328 state->dts_cred.dcr_action |= 13329 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG; 13330 } 13331#endif 13332 } 13333 13334 /* 13335 * Holding the dtrace_kernel privilege also implies that 13336 * the user has the dtrace_user privilege from a visibility 13337 * perspective. But without further privileges, some 13338 * destructive actions are not available. 13339 */ 13340 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) { 13341 /* 13342 * Make all probes in all zones visible. However, 13343 * this doesn't mean that all actions become available 13344 * to all zones. 13345 */ 13346 state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL | 13347 DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE; 13348 13349 state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL | 13350 DTRACE_CRA_PROC; 13351 /* 13352 * Holding proc_owner means that destructive actions 13353 * for *this* zone are allowed. 13354 */ 13355 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) 13356 state->dts_cred.dcr_action |= 13357 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER; 13358 13359 /* 13360 * Holding proc_zone means that destructive actions 13361 * for this user/group ID in all zones is allowed. 13362 */ 13363 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) 13364 state->dts_cred.dcr_action |= 13365 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE; 13366 13367#if defined(sun) 13368 /* 13369 * If we have all privs in whatever zone this is, 13370 * we can do destructive things to processes which 13371 * have altered credentials. 13372 */ 13373 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE), 13374 cr->cr_zone->zone_privset)) { 13375 state->dts_cred.dcr_action |= 13376 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG; 13377 } 13378#endif 13379 } 13380 13381 /* 13382 * Holding the dtrace_proc privilege gives control over fasttrap 13383 * and pid providers. We need to grant wider destructive 13384 * privileges in the event that the user has proc_owner and/or 13385 * proc_zone. 13386 */ 13387 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) { 13388 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) 13389 state->dts_cred.dcr_action |= 13390 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER; 13391 13392 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) 13393 state->dts_cred.dcr_action |= 13394 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE; 13395 } 13396 } 13397 13398 return (state); 13399} 13400 13401static int 13402dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which) 13403{ 13404 dtrace_optval_t *opt = state->dts_options, size; 13405 processorid_t cpu = 0;; 13406 int flags = 0, rval, factor, divisor = 1; 13407 13408 ASSERT(MUTEX_HELD(&dtrace_lock)); 13409 ASSERT(MUTEX_HELD(&cpu_lock)); 13410 ASSERT(which < DTRACEOPT_MAX); 13411 ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE || 13412 (state == dtrace_anon.dta_state && 13413 state->dts_activity == DTRACE_ACTIVITY_ACTIVE)); 13414 13415 if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0) 13416 return (0); 13417 13418 if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET) 13419 cpu = opt[DTRACEOPT_CPU]; 13420 13421 if (which == DTRACEOPT_SPECSIZE) 13422 flags |= DTRACEBUF_NOSWITCH; 13423 13424 if (which == DTRACEOPT_BUFSIZE) { 13425 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING) 13426 flags |= DTRACEBUF_RING; 13427 13428 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL) 13429 flags |= DTRACEBUF_FILL; 13430 13431 if (state != dtrace_anon.dta_state || 13432 state->dts_activity != DTRACE_ACTIVITY_ACTIVE) 13433 flags |= DTRACEBUF_INACTIVE; 13434 } 13435 13436 for (size = opt[which]; size >= sizeof (uint64_t); size /= divisor) { 13437 /* 13438 * The size must be 8-byte aligned. If the size is not 8-byte 13439 * aligned, drop it down by the difference. 13440 */ 13441 if (size & (sizeof (uint64_t) - 1)) 13442 size -= size & (sizeof (uint64_t) - 1); 13443 13444 if (size < state->dts_reserve) { 13445 /* 13446 * Buffers always must be large enough to accommodate 13447 * their prereserved space. We return E2BIG instead 13448 * of ENOMEM in this case to allow for user-level 13449 * software to differentiate the cases. 13450 */ 13451 return (E2BIG); 13452 } 13453 13454 rval = dtrace_buffer_alloc(buf, size, flags, cpu, &factor); 13455 13456 if (rval != ENOMEM) { 13457 opt[which] = size; 13458 return (rval); 13459 } 13460 13461 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL) 13462 return (rval); 13463 13464 for (divisor = 2; divisor < factor; divisor <<= 1) 13465 continue; 13466 } 13467 13468 return (ENOMEM); 13469} 13470 13471static int 13472dtrace_state_buffers(dtrace_state_t *state) 13473{ 13474 dtrace_speculation_t *spec = state->dts_speculations; 13475 int rval, i; 13476 13477 if ((rval = dtrace_state_buffer(state, state->dts_buffer, 13478 DTRACEOPT_BUFSIZE)) != 0) 13479 return (rval); 13480 13481 if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer, 13482 DTRACEOPT_AGGSIZE)) != 0) 13483 return (rval); 13484 13485 for (i = 0; i < state->dts_nspeculations; i++) { 13486 if ((rval = dtrace_state_buffer(state, 13487 spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0) 13488 return (rval); 13489 } 13490 13491 return (0); 13492} 13493 13494static void 13495dtrace_state_prereserve(dtrace_state_t *state) 13496{ 13497 dtrace_ecb_t *ecb; 13498 dtrace_probe_t *probe; 13499 13500 state->dts_reserve = 0; 13501 13502 if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL) 13503 return; 13504 13505 /* 13506 * If our buffer policy is a "fill" buffer policy, we need to set the 13507 * prereserved space to be the space required by the END probes. 13508 */ 13509 probe = dtrace_probes[dtrace_probeid_end - 1]; 13510 ASSERT(probe != NULL); 13511 13512 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) { 13513 if (ecb->dte_state != state) 13514 continue; 13515 13516 state->dts_reserve += ecb->dte_needed + ecb->dte_alignment; 13517 } 13518} 13519 13520static int 13521dtrace_state_go(dtrace_state_t *state, processorid_t *cpu) 13522{ 13523 dtrace_optval_t *opt = state->dts_options, sz, nspec; 13524 dtrace_speculation_t *spec; 13525 dtrace_buffer_t *buf; 13526#if defined(sun) 13527 cyc_handler_t hdlr; 13528 cyc_time_t when; 13529#endif 13530 int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t); 13531 dtrace_icookie_t cookie; 13532 13533 mutex_enter(&cpu_lock); 13534 mutex_enter(&dtrace_lock); 13535 13536 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) { 13537 rval = EBUSY; 13538 goto out; 13539 } 13540 13541 /* 13542 * Before we can perform any checks, we must prime all of the 13543 * retained enablings that correspond to this state. 13544 */ 13545 dtrace_enabling_prime(state); 13546 13547 if (state->dts_destructive && !state->dts_cred.dcr_destructive) { 13548 rval = EACCES; 13549 goto out; 13550 } 13551 13552 dtrace_state_prereserve(state); 13553 13554 /* 13555 * Now we want to do is try to allocate our speculations. 13556 * We do not automatically resize the number of speculations; if 13557 * this fails, we will fail the operation. 13558 */ 13559 nspec = opt[DTRACEOPT_NSPEC]; 13560 ASSERT(nspec != DTRACEOPT_UNSET); 13561 13562 if (nspec > INT_MAX) { 13563 rval = ENOMEM; 13564 goto out; 13565 } 13566 13567 spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t), 13568 KM_NOSLEEP | KM_NORMALPRI); 13569 13570 if (spec == NULL) { 13571 rval = ENOMEM; 13572 goto out; 13573 } 13574 13575 state->dts_speculations = spec; 13576 state->dts_nspeculations = (int)nspec; 13577 13578 for (i = 0; i < nspec; i++) { 13579 if ((buf = kmem_zalloc(bufsize, 13580 KM_NOSLEEP | KM_NORMALPRI)) == NULL) { 13581 rval = ENOMEM; 13582 goto err; 13583 } 13584 13585 spec[i].dtsp_buffer = buf; 13586 } 13587 13588 if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) { 13589 if (dtrace_anon.dta_state == NULL) { 13590 rval = ENOENT; 13591 goto out; 13592 } 13593 13594 if (state->dts_necbs != 0) { 13595 rval = EALREADY; 13596 goto out; 13597 } 13598 13599 state->dts_anon = dtrace_anon_grab(); 13600 ASSERT(state->dts_anon != NULL); 13601 state = state->dts_anon; 13602 13603 /* 13604 * We want "grabanon" to be set in the grabbed state, so we'll 13605 * copy that option value from the grabbing state into the 13606 * grabbed state. 13607 */ 13608 state->dts_options[DTRACEOPT_GRABANON] = 13609 opt[DTRACEOPT_GRABANON]; 13610 13611 *cpu = dtrace_anon.dta_beganon; 13612 13613 /* 13614 * If the anonymous state is active (as it almost certainly 13615 * is if the anonymous enabling ultimately matched anything), 13616 * we don't allow any further option processing -- but we 13617 * don't return failure. 13618 */ 13619 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) 13620 goto out; 13621 } 13622 13623 if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET && 13624 opt[DTRACEOPT_AGGSIZE] != 0) { 13625 if (state->dts_aggregations == NULL) { 13626 /* 13627 * We're not going to create an aggregation buffer 13628 * because we don't have any ECBs that contain 13629 * aggregations -- set this option to 0. 13630 */ 13631 opt[DTRACEOPT_AGGSIZE] = 0; 13632 } else { 13633 /* 13634 * If we have an aggregation buffer, we must also have 13635 * a buffer to use as scratch. 13636 */ 13637 if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET || 13638 opt[DTRACEOPT_BUFSIZE] < state->dts_needed) { 13639 opt[DTRACEOPT_BUFSIZE] = state->dts_needed; 13640 } 13641 } 13642 } 13643 13644 if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET && 13645 opt[DTRACEOPT_SPECSIZE] != 0) { 13646 if (!state->dts_speculates) { 13647 /* 13648 * We're not going to create speculation buffers 13649 * because we don't have any ECBs that actually 13650 * speculate -- set the speculation size to 0. 13651 */ 13652 opt[DTRACEOPT_SPECSIZE] = 0; 13653 } 13654 } 13655 13656 /* 13657 * The bare minimum size for any buffer that we're actually going to 13658 * do anything to is sizeof (uint64_t). 13659 */ 13660 sz = sizeof (uint64_t); 13661 13662 if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) || 13663 (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) || 13664 (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) { 13665 /* 13666 * A buffer size has been explicitly set to 0 (or to a size 13667 * that will be adjusted to 0) and we need the space -- we 13668 * need to return failure. We return ENOSPC to differentiate 13669 * it from failing to allocate a buffer due to failure to meet 13670 * the reserve (for which we return E2BIG). 13671 */ 13672 rval = ENOSPC; 13673 goto out; 13674 } 13675 13676 if ((rval = dtrace_state_buffers(state)) != 0) 13677 goto err; 13678 13679 if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET) 13680 sz = dtrace_dstate_defsize; 13681 13682 do { 13683 rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz); 13684 13685 if (rval == 0) 13686 break; 13687 13688 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL) 13689 goto err; 13690 } while (sz >>= 1); 13691 13692 opt[DTRACEOPT_DYNVARSIZE] = sz; 13693 13694 if (rval != 0) 13695 goto err; 13696 13697 if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max) 13698 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max; 13699 13700 if (opt[DTRACEOPT_CLEANRATE] == 0) 13701 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max; 13702 13703 if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min) 13704 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min; 13705 13706 if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max) 13707 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max; 13708 13709 state->dts_alive = state->dts_laststatus = dtrace_gethrtime(); 13710#if defined(sun) 13711 hdlr.cyh_func = (cyc_func_t)dtrace_state_clean; 13712 hdlr.cyh_arg = state; 13713 hdlr.cyh_level = CY_LOW_LEVEL; 13714 13715 when.cyt_when = 0; 13716 when.cyt_interval = opt[DTRACEOPT_CLEANRATE]; 13717 13718 state->dts_cleaner = cyclic_add(&hdlr, &when); 13719 13720 hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman; 13721 hdlr.cyh_arg = state; 13722 hdlr.cyh_level = CY_LOW_LEVEL; 13723 13724 when.cyt_when = 0; 13725 when.cyt_interval = dtrace_deadman_interval; 13726 13727 state->dts_deadman = cyclic_add(&hdlr, &when); 13728#else 13729 callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC, 13730 dtrace_state_clean, state); 13731 callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC, 13732 dtrace_state_deadman, state); 13733#endif 13734 13735 state->dts_activity = DTRACE_ACTIVITY_WARMUP; 13736 13737 /* 13738 * Now it's time to actually fire the BEGIN probe. We need to disable 13739 * interrupts here both to record the CPU on which we fired the BEGIN 13740 * probe (the data from this CPU will be processed first at user 13741 * level) and to manually activate the buffer for this CPU. 13742 */ 13743 cookie = dtrace_interrupt_disable(); 13744 *cpu = curcpu; 13745 ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE); 13746 state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE; 13747 13748 dtrace_probe(dtrace_probeid_begin, 13749 (uint64_t)(uintptr_t)state, 0, 0, 0, 0); 13750 dtrace_interrupt_enable(cookie); 13751 /* 13752 * We may have had an exit action from a BEGIN probe; only change our 13753 * state to ACTIVE if we're still in WARMUP. 13754 */ 13755 ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP || 13756 state->dts_activity == DTRACE_ACTIVITY_DRAINING); 13757 13758 if (state->dts_activity == DTRACE_ACTIVITY_WARMUP) 13759 state->dts_activity = DTRACE_ACTIVITY_ACTIVE; 13760 13761 /* 13762 * Regardless of whether or not now we're in ACTIVE or DRAINING, we 13763 * want each CPU to transition its principal buffer out of the 13764 * INACTIVE state. Doing this assures that no CPU will suddenly begin 13765 * processing an ECB halfway down a probe's ECB chain; all CPUs will 13766 * atomically transition from processing none of a state's ECBs to 13767 * processing all of them. 13768 */ 13769 dtrace_xcall(DTRACE_CPUALL, 13770 (dtrace_xcall_t)dtrace_buffer_activate, state); 13771 goto out; 13772 13773err: 13774 dtrace_buffer_free(state->dts_buffer); 13775 dtrace_buffer_free(state->dts_aggbuffer); 13776 13777 if ((nspec = state->dts_nspeculations) == 0) { 13778 ASSERT(state->dts_speculations == NULL); 13779 goto out; 13780 } 13781 13782 spec = state->dts_speculations; 13783 ASSERT(spec != NULL); 13784 13785 for (i = 0; i < state->dts_nspeculations; i++) { 13786 if ((buf = spec[i].dtsp_buffer) == NULL) 13787 break; 13788 13789 dtrace_buffer_free(buf); 13790 kmem_free(buf, bufsize); 13791 } 13792 13793 kmem_free(spec, nspec * sizeof (dtrace_speculation_t)); 13794 state->dts_nspeculations = 0; 13795 state->dts_speculations = NULL; 13796 13797out: 13798 mutex_exit(&dtrace_lock); 13799 mutex_exit(&cpu_lock); 13800 13801 return (rval); 13802} 13803 13804static int 13805dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu) 13806{ 13807 dtrace_icookie_t cookie; 13808 13809 ASSERT(MUTEX_HELD(&dtrace_lock)); 13810 13811 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE && 13812 state->dts_activity != DTRACE_ACTIVITY_DRAINING) 13813 return (EINVAL); 13814 13815 /* 13816 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync 13817 * to be sure that every CPU has seen it. See below for the details 13818 * on why this is done. 13819 */ 13820 state->dts_activity = DTRACE_ACTIVITY_DRAINING; 13821 dtrace_sync(); 13822 13823 /* 13824 * By this point, it is impossible for any CPU to be still processing 13825 * with DTRACE_ACTIVITY_ACTIVE. We can thus set our activity to 13826 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any 13827 * other CPU in dtrace_buffer_reserve(). This allows dtrace_probe() 13828 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN 13829 * iff we're in the END probe. 13830 */ 13831 state->dts_activity = DTRACE_ACTIVITY_COOLDOWN; 13832 dtrace_sync(); 13833 ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN); 13834 13835 /* 13836 * Finally, we can release the reserve and call the END probe. We 13837 * disable interrupts across calling the END probe to allow us to 13838 * return the CPU on which we actually called the END probe. This 13839 * allows user-land to be sure that this CPU's principal buffer is 13840 * processed last. 13841 */ 13842 state->dts_reserve = 0; 13843 13844 cookie = dtrace_interrupt_disable(); 13845 *cpu = curcpu; 13846 dtrace_probe(dtrace_probeid_end, 13847 (uint64_t)(uintptr_t)state, 0, 0, 0, 0); 13848 dtrace_interrupt_enable(cookie); 13849 13850 state->dts_activity = DTRACE_ACTIVITY_STOPPED; 13851 dtrace_sync(); 13852 13853 return (0); 13854} 13855 13856static int 13857dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option, 13858 dtrace_optval_t val) 13859{ 13860 ASSERT(MUTEX_HELD(&dtrace_lock)); 13861 13862 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) 13863 return (EBUSY); 13864 13865 if (option >= DTRACEOPT_MAX) 13866 return (EINVAL); 13867 13868 if (option != DTRACEOPT_CPU && val < 0) 13869 return (EINVAL); 13870 13871 switch (option) { 13872 case DTRACEOPT_DESTRUCTIVE: 13873 if (dtrace_destructive_disallow) 13874 return (EACCES); 13875 13876 state->dts_cred.dcr_destructive = 1; 13877 break; 13878 13879 case DTRACEOPT_BUFSIZE: 13880 case DTRACEOPT_DYNVARSIZE: 13881 case DTRACEOPT_AGGSIZE: 13882 case DTRACEOPT_SPECSIZE: 13883 case DTRACEOPT_STRSIZE: 13884 if (val < 0) 13885 return (EINVAL); 13886 13887 if (val >= LONG_MAX) { 13888 /* 13889 * If this is an otherwise negative value, set it to 13890 * the highest multiple of 128m less than LONG_MAX. 13891 * Technically, we're adjusting the size without 13892 * regard to the buffer resizing policy, but in fact, 13893 * this has no effect -- if we set the buffer size to 13894 * ~LONG_MAX and the buffer policy is ultimately set to 13895 * be "manual", the buffer allocation is guaranteed to 13896 * fail, if only because the allocation requires two 13897 * buffers. (We set the the size to the highest 13898 * multiple of 128m because it ensures that the size 13899 * will remain a multiple of a megabyte when 13900 * repeatedly halved -- all the way down to 15m.) 13901 */ 13902 val = LONG_MAX - (1 << 27) + 1; 13903 } 13904 } 13905 13906 state->dts_options[option] = val; 13907 13908 return (0); 13909} 13910 13911static void 13912dtrace_state_destroy(dtrace_state_t *state) 13913{ 13914 dtrace_ecb_t *ecb; 13915 dtrace_vstate_t *vstate = &state->dts_vstate; 13916#if defined(sun) 13917 minor_t minor = getminor(state->dts_dev); 13918#endif 13919 int i, bufsize = NCPU * sizeof (dtrace_buffer_t); 13920 dtrace_speculation_t *spec = state->dts_speculations; 13921 int nspec = state->dts_nspeculations; 13922 uint32_t match; 13923 13924 ASSERT(MUTEX_HELD(&dtrace_lock)); 13925 ASSERT(MUTEX_HELD(&cpu_lock)); 13926 13927 /* 13928 * First, retract any retained enablings for this state. 13929 */ 13930 dtrace_enabling_retract(state); 13931 ASSERT(state->dts_nretained == 0); 13932 13933 if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE || 13934 state->dts_activity == DTRACE_ACTIVITY_DRAINING) { 13935 /* 13936 * We have managed to come into dtrace_state_destroy() on a 13937 * hot enabling -- almost certainly because of a disorderly 13938 * shutdown of a consumer. (That is, a consumer that is 13939 * exiting without having called dtrace_stop().) In this case, 13940 * we're going to set our activity to be KILLED, and then 13941 * issue a sync to be sure that everyone is out of probe 13942 * context before we start blowing away ECBs. 13943 */ 13944 state->dts_activity = DTRACE_ACTIVITY_KILLED; 13945 dtrace_sync(); 13946 } 13947 13948 /* 13949 * Release the credential hold we took in dtrace_state_create(). 13950 */ 13951 if (state->dts_cred.dcr_cred != NULL) 13952 crfree(state->dts_cred.dcr_cred); 13953 13954 /* 13955 * Now we can safely disable and destroy any enabled probes. Because 13956 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress 13957 * (especially if they're all enabled), we take two passes through the 13958 * ECBs: in the first, we disable just DTRACE_PRIV_KERNEL probes, and 13959 * in the second we disable whatever is left over. 13960 */ 13961 for (match = DTRACE_PRIV_KERNEL; ; match = 0) { 13962 for (i = 0; i < state->dts_necbs; i++) { 13963 if ((ecb = state->dts_ecbs[i]) == NULL) 13964 continue; 13965 13966 if (match && ecb->dte_probe != NULL) { 13967 dtrace_probe_t *probe = ecb->dte_probe; 13968 dtrace_provider_t *prov = probe->dtpr_provider; 13969 13970 if (!(prov->dtpv_priv.dtpp_flags & match)) 13971 continue; 13972 } 13973 13974 dtrace_ecb_disable(ecb); 13975 dtrace_ecb_destroy(ecb); 13976 } 13977 13978 if (!match) 13979 break; 13980 } 13981 13982 /* 13983 * Before we free the buffers, perform one more sync to assure that 13984 * every CPU is out of probe context. 13985 */ 13986 dtrace_sync(); 13987 13988 dtrace_buffer_free(state->dts_buffer); 13989 dtrace_buffer_free(state->dts_aggbuffer); 13990 13991 for (i = 0; i < nspec; i++) 13992 dtrace_buffer_free(spec[i].dtsp_buffer); 13993 13994#if defined(sun) 13995 if (state->dts_cleaner != CYCLIC_NONE) 13996 cyclic_remove(state->dts_cleaner); 13997 13998 if (state->dts_deadman != CYCLIC_NONE) 13999 cyclic_remove(state->dts_deadman); 14000#else 14001 callout_stop(&state->dts_cleaner); 14002 callout_drain(&state->dts_cleaner); 14003 callout_stop(&state->dts_deadman); 14004 callout_drain(&state->dts_deadman); 14005#endif 14006 14007 dtrace_dstate_fini(&vstate->dtvs_dynvars); 14008 dtrace_vstate_fini(vstate); 14009 if (state->dts_ecbs != NULL) 14010 kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *)); 14011 14012 if (state->dts_aggregations != NULL) { 14013#ifdef DEBUG 14014 for (i = 0; i < state->dts_naggregations; i++) 14015 ASSERT(state->dts_aggregations[i] == NULL); 14016#endif 14017 ASSERT(state->dts_naggregations > 0); 14018 kmem_free(state->dts_aggregations, 14019 state->dts_naggregations * sizeof (dtrace_aggregation_t *)); 14020 } 14021 14022 kmem_free(state->dts_buffer, bufsize); 14023 kmem_free(state->dts_aggbuffer, bufsize); 14024 14025 for (i = 0; i < nspec; i++) 14026 kmem_free(spec[i].dtsp_buffer, bufsize); 14027 14028 if (spec != NULL) 14029 kmem_free(spec, nspec * sizeof (dtrace_speculation_t)); 14030 14031 dtrace_format_destroy(state); 14032 14033 if (state->dts_aggid_arena != NULL) { 14034#if defined(sun) 14035 vmem_destroy(state->dts_aggid_arena); 14036#else 14037 delete_unrhdr(state->dts_aggid_arena); 14038#endif 14039 state->dts_aggid_arena = NULL; 14040 } 14041#if defined(sun) 14042 ddi_soft_state_free(dtrace_softstate, minor); 14043 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1); 14044#endif 14045} 14046 14047/* 14048 * DTrace Anonymous Enabling Functions 14049 */ 14050static dtrace_state_t * 14051dtrace_anon_grab(void) 14052{ 14053 dtrace_state_t *state; 14054 14055 ASSERT(MUTEX_HELD(&dtrace_lock)); 14056 14057 if ((state = dtrace_anon.dta_state) == NULL) { 14058 ASSERT(dtrace_anon.dta_enabling == NULL); 14059 return (NULL); 14060 } 14061 14062 ASSERT(dtrace_anon.dta_enabling != NULL); 14063 ASSERT(dtrace_retained != NULL); 14064 14065 dtrace_enabling_destroy(dtrace_anon.dta_enabling); 14066 dtrace_anon.dta_enabling = NULL; 14067 dtrace_anon.dta_state = NULL; 14068 14069 return (state); 14070} 14071 14072static void 14073dtrace_anon_property(void) 14074{ 14075 int i, rv; 14076 dtrace_state_t *state; 14077 dof_hdr_t *dof; 14078 char c[32]; /* enough for "dof-data-" + digits */ 14079 14080 ASSERT(MUTEX_HELD(&dtrace_lock)); 14081 ASSERT(MUTEX_HELD(&cpu_lock)); 14082 14083 for (i = 0; ; i++) { 14084 (void) snprintf(c, sizeof (c), "dof-data-%d", i); 14085 14086 dtrace_err_verbose = 1; 14087 14088 if ((dof = dtrace_dof_property(c)) == NULL) { 14089 dtrace_err_verbose = 0; 14090 break; 14091 } 14092 14093#if defined(sun) 14094 /* 14095 * We want to create anonymous state, so we need to transition 14096 * the kernel debugger to indicate that DTrace is active. If 14097 * this fails (e.g. because the debugger has modified text in 14098 * some way), we won't continue with the processing. 14099 */ 14100 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) { 14101 cmn_err(CE_NOTE, "kernel debugger active; anonymous " 14102 "enabling ignored."); 14103 dtrace_dof_destroy(dof); 14104 break; 14105 } 14106#endif 14107 14108 /* 14109 * If we haven't allocated an anonymous state, we'll do so now. 14110 */ 14111 if ((state = dtrace_anon.dta_state) == NULL) { 14112#if defined(sun) 14113 state = dtrace_state_create(NULL, NULL); 14114#else 14115 state = dtrace_state_create(NULL); 14116#endif 14117 dtrace_anon.dta_state = state; 14118 14119 if (state == NULL) { 14120 /* 14121 * This basically shouldn't happen: the only 14122 * failure mode from dtrace_state_create() is a 14123 * failure of ddi_soft_state_zalloc() that 14124 * itself should never happen. Still, the 14125 * interface allows for a failure mode, and 14126 * we want to fail as gracefully as possible: 14127 * we'll emit an error message and cease 14128 * processing anonymous state in this case. 14129 */ 14130 cmn_err(CE_WARN, "failed to create " 14131 "anonymous state"); 14132 dtrace_dof_destroy(dof); 14133 break; 14134 } 14135 } 14136 14137 rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(), 14138 &dtrace_anon.dta_enabling, 0, B_TRUE); 14139 14140 if (rv == 0) 14141 rv = dtrace_dof_options(dof, state); 14142 14143 dtrace_err_verbose = 0; 14144 dtrace_dof_destroy(dof); 14145 14146 if (rv != 0) { 14147 /* 14148 * This is malformed DOF; chuck any anonymous state 14149 * that we created. 14150 */ 14151 ASSERT(dtrace_anon.dta_enabling == NULL); 14152 dtrace_state_destroy(state); 14153 dtrace_anon.dta_state = NULL; 14154 break; 14155 } 14156 14157 ASSERT(dtrace_anon.dta_enabling != NULL); 14158 } 14159 14160 if (dtrace_anon.dta_enabling != NULL) { 14161 int rval; 14162 14163 /* 14164 * dtrace_enabling_retain() can only fail because we are 14165 * trying to retain more enablings than are allowed -- but 14166 * we only have one anonymous enabling, and we are guaranteed 14167 * to be allowed at least one retained enabling; we assert 14168 * that dtrace_enabling_retain() returns success. 14169 */ 14170 rval = dtrace_enabling_retain(dtrace_anon.dta_enabling); 14171 ASSERT(rval == 0); 14172 14173 dtrace_enabling_dump(dtrace_anon.dta_enabling); 14174 } 14175} 14176 14177/* 14178 * DTrace Helper Functions 14179 */ 14180static void 14181dtrace_helper_trace(dtrace_helper_action_t *helper, 14182 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where) 14183{ 14184 uint32_t size, next, nnext, i; 14185 dtrace_helptrace_t *ent; 14186 uint16_t flags = cpu_core[curcpu].cpuc_dtrace_flags; 14187 14188 if (!dtrace_helptrace_enabled) 14189 return; 14190 14191 ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals); 14192 14193 /* 14194 * What would a tracing framework be without its own tracing 14195 * framework? (Well, a hell of a lot simpler, for starters...) 14196 */ 14197 size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals * 14198 sizeof (uint64_t) - sizeof (uint64_t); 14199 14200 /* 14201 * Iterate until we can allocate a slot in the trace buffer. 14202 */ 14203 do { 14204 next = dtrace_helptrace_next; 14205 14206 if (next + size < dtrace_helptrace_bufsize) { 14207 nnext = next + size; 14208 } else { 14209 nnext = size; 14210 } 14211 } while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next); 14212 14213 /* 14214 * We have our slot; fill it in. 14215 */ 14216 if (nnext == size) 14217 next = 0; 14218 14219 ent = (dtrace_helptrace_t *)&dtrace_helptrace_buffer[next]; 14220 ent->dtht_helper = helper; 14221 ent->dtht_where = where; 14222 ent->dtht_nlocals = vstate->dtvs_nlocals; 14223 14224 ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ? 14225 mstate->dtms_fltoffs : -1; 14226 ent->dtht_fault = DTRACE_FLAGS2FLT(flags); 14227 ent->dtht_illval = cpu_core[curcpu].cpuc_dtrace_illval; 14228 14229 for (i = 0; i < vstate->dtvs_nlocals; i++) { 14230 dtrace_statvar_t *svar; 14231 14232 if ((svar = vstate->dtvs_locals[i]) == NULL) 14233 continue; 14234 14235 ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t)); 14236 ent->dtht_locals[i] = 14237 ((uint64_t *)(uintptr_t)svar->dtsv_data)[curcpu]; 14238 } 14239} 14240 14241static uint64_t 14242dtrace_helper(int which, dtrace_mstate_t *mstate, 14243 dtrace_state_t *state, uint64_t arg0, uint64_t arg1) 14244{ 14245 uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags; 14246 uint64_t sarg0 = mstate->dtms_arg[0]; 14247 uint64_t sarg1 = mstate->dtms_arg[1]; 14248 uint64_t rval = 0; 14249 dtrace_helpers_t *helpers = curproc->p_dtrace_helpers; 14250 dtrace_helper_action_t *helper; 14251 dtrace_vstate_t *vstate; 14252 dtrace_difo_t *pred; 14253 int i, trace = dtrace_helptrace_enabled; 14254 14255 ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS); 14256 14257 if (helpers == NULL) 14258 return (0); 14259 14260 if ((helper = helpers->dthps_actions[which]) == NULL) 14261 return (0); 14262 14263 vstate = &helpers->dthps_vstate; 14264 mstate->dtms_arg[0] = arg0; 14265 mstate->dtms_arg[1] = arg1; 14266 14267 /* 14268 * Now iterate over each helper. If its predicate evaluates to 'true', 14269 * we'll call the corresponding actions. Note that the below calls 14270 * to dtrace_dif_emulate() may set faults in machine state. This is 14271 * okay: our caller (the outer dtrace_dif_emulate()) will simply plow 14272 * the stored DIF offset with its own (which is the desired behavior). 14273 * Also, note the calls to dtrace_dif_emulate() may allocate scratch 14274 * from machine state; this is okay, too. 14275 */ 14276 for (; helper != NULL; helper = helper->dtha_next) { 14277 if ((pred = helper->dtha_predicate) != NULL) { 14278 if (trace) 14279 dtrace_helper_trace(helper, mstate, vstate, 0); 14280 14281 if (!dtrace_dif_emulate(pred, mstate, vstate, state)) 14282 goto next; 14283 14284 if (*flags & CPU_DTRACE_FAULT) 14285 goto err; 14286 } 14287 14288 for (i = 0; i < helper->dtha_nactions; i++) { 14289 if (trace) 14290 dtrace_helper_trace(helper, 14291 mstate, vstate, i + 1); 14292 14293 rval = dtrace_dif_emulate(helper->dtha_actions[i], 14294 mstate, vstate, state); 14295 14296 if (*flags & CPU_DTRACE_FAULT) 14297 goto err; 14298 } 14299 14300next: 14301 if (trace) 14302 dtrace_helper_trace(helper, mstate, vstate, 14303 DTRACE_HELPTRACE_NEXT); 14304 } 14305 14306 if (trace) 14307 dtrace_helper_trace(helper, mstate, vstate, 14308 DTRACE_HELPTRACE_DONE); 14309 14310 /* 14311 * Restore the arg0 that we saved upon entry. 14312 */ 14313 mstate->dtms_arg[0] = sarg0; 14314 mstate->dtms_arg[1] = sarg1; 14315 14316 return (rval); 14317 14318err: 14319 if (trace) 14320 dtrace_helper_trace(helper, mstate, vstate, 14321 DTRACE_HELPTRACE_ERR); 14322 14323 /* 14324 * Restore the arg0 that we saved upon entry. 14325 */ 14326 mstate->dtms_arg[0] = sarg0; 14327 mstate->dtms_arg[1] = sarg1; 14328 14329 return (0); 14330} 14331 14332static void 14333dtrace_helper_action_destroy(dtrace_helper_action_t *helper, 14334 dtrace_vstate_t *vstate) 14335{ 14336 int i; 14337 14338 if (helper->dtha_predicate != NULL) 14339 dtrace_difo_release(helper->dtha_predicate, vstate); 14340 14341 for (i = 0; i < helper->dtha_nactions; i++) { 14342 ASSERT(helper->dtha_actions[i] != NULL); 14343 dtrace_difo_release(helper->dtha_actions[i], vstate); 14344 } 14345 14346 kmem_free(helper->dtha_actions, 14347 helper->dtha_nactions * sizeof (dtrace_difo_t *)); 14348 kmem_free(helper, sizeof (dtrace_helper_action_t)); 14349} 14350 14351static int 14352dtrace_helper_destroygen(int gen) 14353{ 14354 proc_t *p = curproc; 14355 dtrace_helpers_t *help = p->p_dtrace_helpers; 14356 dtrace_vstate_t *vstate; 14357 int i; 14358 14359 ASSERT(MUTEX_HELD(&dtrace_lock)); 14360 14361 if (help == NULL || gen > help->dthps_generation) 14362 return (EINVAL); 14363 14364 vstate = &help->dthps_vstate; 14365 14366 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) { 14367 dtrace_helper_action_t *last = NULL, *h, *next; 14368 14369 for (h = help->dthps_actions[i]; h != NULL; h = next) { 14370 next = h->dtha_next; 14371 14372 if (h->dtha_generation == gen) { 14373 if (last != NULL) { 14374 last->dtha_next = next; 14375 } else { 14376 help->dthps_actions[i] = next; 14377 } 14378 14379 dtrace_helper_action_destroy(h, vstate); 14380 } else { 14381 last = h; 14382 } 14383 } 14384 } 14385 14386 /* 14387 * Interate until we've cleared out all helper providers with the 14388 * given generation number. 14389 */ 14390 for (;;) { 14391 dtrace_helper_provider_t *prov; 14392 14393 /* 14394 * Look for a helper provider with the right generation. We 14395 * have to start back at the beginning of the list each time 14396 * because we drop dtrace_lock. It's unlikely that we'll make 14397 * more than two passes. 14398 */ 14399 for (i = 0; i < help->dthps_nprovs; i++) { 14400 prov = help->dthps_provs[i]; 14401 14402 if (prov->dthp_generation == gen) 14403 break; 14404 } 14405 14406 /* 14407 * If there were no matches, we're done. 14408 */ 14409 if (i == help->dthps_nprovs) 14410 break; 14411 14412 /* 14413 * Move the last helper provider into this slot. 14414 */ 14415 help->dthps_nprovs--; 14416 help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs]; 14417 help->dthps_provs[help->dthps_nprovs] = NULL; 14418 14419 mutex_exit(&dtrace_lock); 14420 14421 /* 14422 * If we have a meta provider, remove this helper provider. 14423 */ 14424 mutex_enter(&dtrace_meta_lock); 14425 if (dtrace_meta_pid != NULL) { 14426 ASSERT(dtrace_deferred_pid == NULL); 14427 dtrace_helper_provider_remove(&prov->dthp_prov, 14428 p->p_pid); 14429 } 14430 mutex_exit(&dtrace_meta_lock); 14431 14432 dtrace_helper_provider_destroy(prov); 14433 14434 mutex_enter(&dtrace_lock); 14435 } 14436 14437 return (0); 14438} 14439 14440static int 14441dtrace_helper_validate(dtrace_helper_action_t *helper) 14442{ 14443 int err = 0, i; 14444 dtrace_difo_t *dp; 14445 14446 if ((dp = helper->dtha_predicate) != NULL) 14447 err += dtrace_difo_validate_helper(dp); 14448 14449 for (i = 0; i < helper->dtha_nactions; i++) 14450 err += dtrace_difo_validate_helper(helper->dtha_actions[i]); 14451 14452 return (err == 0); 14453} 14454 14455static int 14456dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep) 14457{ 14458 dtrace_helpers_t *help; 14459 dtrace_helper_action_t *helper, *last; 14460 dtrace_actdesc_t *act; 14461 dtrace_vstate_t *vstate; 14462 dtrace_predicate_t *pred; 14463 int count = 0, nactions = 0, i; 14464 14465 if (which < 0 || which >= DTRACE_NHELPER_ACTIONS) 14466 return (EINVAL); 14467 14468 help = curproc->p_dtrace_helpers; 14469 last = help->dthps_actions[which]; 14470 vstate = &help->dthps_vstate; 14471 14472 for (count = 0; last != NULL; last = last->dtha_next) { 14473 count++; 14474 if (last->dtha_next == NULL) 14475 break; 14476 } 14477 14478 /* 14479 * If we already have dtrace_helper_actions_max helper actions for this 14480 * helper action type, we'll refuse to add a new one. 14481 */ 14482 if (count >= dtrace_helper_actions_max) 14483 return (ENOSPC); 14484 14485 helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP); 14486 helper->dtha_generation = help->dthps_generation; 14487 14488 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) { 14489 ASSERT(pred->dtp_difo != NULL); 14490 dtrace_difo_hold(pred->dtp_difo); 14491 helper->dtha_predicate = pred->dtp_difo; 14492 } 14493 14494 for (act = ep->dted_action; act != NULL; act = act->dtad_next) { 14495 if (act->dtad_kind != DTRACEACT_DIFEXPR) 14496 goto err; 14497 14498 if (act->dtad_difo == NULL) 14499 goto err; 14500 14501 nactions++; 14502 } 14503 14504 helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) * 14505 (helper->dtha_nactions = nactions), KM_SLEEP); 14506 14507 for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) { 14508 dtrace_difo_hold(act->dtad_difo); 14509 helper->dtha_actions[i++] = act->dtad_difo; 14510 } 14511 14512 if (!dtrace_helper_validate(helper)) 14513 goto err; 14514 14515 if (last == NULL) { 14516 help->dthps_actions[which] = helper; 14517 } else { 14518 last->dtha_next = helper; 14519 } 14520 14521 if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) { 14522 dtrace_helptrace_nlocals = vstate->dtvs_nlocals; 14523 dtrace_helptrace_next = 0; 14524 } 14525 14526 return (0); 14527err: 14528 dtrace_helper_action_destroy(helper, vstate); 14529 return (EINVAL); 14530} 14531 14532static void 14533dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help, 14534 dof_helper_t *dofhp) 14535{ 14536 ASSERT(MUTEX_NOT_HELD(&dtrace_lock)); 14537 14538 mutex_enter(&dtrace_meta_lock); 14539 mutex_enter(&dtrace_lock); 14540 14541 if (!dtrace_attached() || dtrace_meta_pid == NULL) { 14542 /* 14543 * If the dtrace module is loaded but not attached, or if 14544 * there aren't isn't a meta provider registered to deal with 14545 * these provider descriptions, we need to postpone creating 14546 * the actual providers until later. 14547 */ 14548 14549 if (help->dthps_next == NULL && help->dthps_prev == NULL && 14550 dtrace_deferred_pid != help) { 14551 help->dthps_deferred = 1; 14552 help->dthps_pid = p->p_pid; 14553 help->dthps_next = dtrace_deferred_pid; 14554 help->dthps_prev = NULL; 14555 if (dtrace_deferred_pid != NULL) 14556 dtrace_deferred_pid->dthps_prev = help; 14557 dtrace_deferred_pid = help; 14558 } 14559 14560 mutex_exit(&dtrace_lock); 14561 14562 } else if (dofhp != NULL) { 14563 /* 14564 * If the dtrace module is loaded and we have a particular 14565 * helper provider description, pass that off to the 14566 * meta provider. 14567 */ 14568 14569 mutex_exit(&dtrace_lock); 14570 14571 dtrace_helper_provide(dofhp, p->p_pid); 14572 14573 } else { 14574 /* 14575 * Otherwise, just pass all the helper provider descriptions 14576 * off to the meta provider. 14577 */ 14578 14579 int i; 14580 mutex_exit(&dtrace_lock); 14581 14582 for (i = 0; i < help->dthps_nprovs; i++) { 14583 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov, 14584 p->p_pid); 14585 } 14586 } 14587 14588 mutex_exit(&dtrace_meta_lock); 14589} 14590 14591static int 14592dtrace_helper_provider_add(dof_helper_t *dofhp, int gen) 14593{ 14594 dtrace_helpers_t *help; 14595 dtrace_helper_provider_t *hprov, **tmp_provs; 14596 uint_t tmp_maxprovs, i; 14597 14598 ASSERT(MUTEX_HELD(&dtrace_lock)); 14599 14600 help = curproc->p_dtrace_helpers; 14601 ASSERT(help != NULL); 14602 14603 /* 14604 * If we already have dtrace_helper_providers_max helper providers, 14605 * we're refuse to add a new one. 14606 */ 14607 if (help->dthps_nprovs >= dtrace_helper_providers_max) 14608 return (ENOSPC); 14609 14610 /* 14611 * Check to make sure this isn't a duplicate. 14612 */ 14613 for (i = 0; i < help->dthps_nprovs; i++) { 14614 if (dofhp->dofhp_dof == 14615 help->dthps_provs[i]->dthp_prov.dofhp_dof) 14616 return (EALREADY); 14617 } 14618 14619 hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP); 14620 hprov->dthp_prov = *dofhp; 14621 hprov->dthp_ref = 1; 14622 hprov->dthp_generation = gen; 14623 14624 /* 14625 * Allocate a bigger table for helper providers if it's already full. 14626 */ 14627 if (help->dthps_maxprovs == help->dthps_nprovs) { 14628 tmp_maxprovs = help->dthps_maxprovs; 14629 tmp_provs = help->dthps_provs; 14630 14631 if (help->dthps_maxprovs == 0) 14632 help->dthps_maxprovs = 2; 14633 else 14634 help->dthps_maxprovs *= 2; 14635 if (help->dthps_maxprovs > dtrace_helper_providers_max) 14636 help->dthps_maxprovs = dtrace_helper_providers_max; 14637 14638 ASSERT(tmp_maxprovs < help->dthps_maxprovs); 14639 14640 help->dthps_provs = kmem_zalloc(help->dthps_maxprovs * 14641 sizeof (dtrace_helper_provider_t *), KM_SLEEP); 14642 14643 if (tmp_provs != NULL) { 14644 bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs * 14645 sizeof (dtrace_helper_provider_t *)); 14646 kmem_free(tmp_provs, tmp_maxprovs * 14647 sizeof (dtrace_helper_provider_t *)); 14648 } 14649 } 14650 14651 help->dthps_provs[help->dthps_nprovs] = hprov; 14652 help->dthps_nprovs++; 14653 14654 return (0); 14655} 14656 14657static void 14658dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov) 14659{ 14660 mutex_enter(&dtrace_lock); 14661 14662 if (--hprov->dthp_ref == 0) { 14663 dof_hdr_t *dof; 14664 mutex_exit(&dtrace_lock); 14665 dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof; 14666 dtrace_dof_destroy(dof); 14667 kmem_free(hprov, sizeof (dtrace_helper_provider_t)); 14668 } else { 14669 mutex_exit(&dtrace_lock); 14670 } 14671} 14672 14673static int 14674dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec) 14675{ 14676 uintptr_t daddr = (uintptr_t)dof; 14677 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec; 14678 dof_provider_t *provider; 14679 dof_probe_t *probe; 14680 uint8_t *arg; 14681 char *strtab, *typestr; 14682 dof_stridx_t typeidx; 14683 size_t typesz; 14684 uint_t nprobes, j, k; 14685 14686 ASSERT(sec->dofs_type == DOF_SECT_PROVIDER); 14687 14688 if (sec->dofs_offset & (sizeof (uint_t) - 1)) { 14689 dtrace_dof_error(dof, "misaligned section offset"); 14690 return (-1); 14691 } 14692 14693 /* 14694 * The section needs to be large enough to contain the DOF provider 14695 * structure appropriate for the given version. 14696 */ 14697 if (sec->dofs_size < 14698 ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ? 14699 offsetof(dof_provider_t, dofpv_prenoffs) : 14700 sizeof (dof_provider_t))) { 14701 dtrace_dof_error(dof, "provider section too small"); 14702 return (-1); 14703 } 14704 14705 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset); 14706 str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab); 14707 prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes); 14708 arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs); 14709 off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs); 14710 14711 if (str_sec == NULL || prb_sec == NULL || 14712 arg_sec == NULL || off_sec == NULL) 14713 return (-1); 14714 14715 enoff_sec = NULL; 14716 14717 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 && 14718 provider->dofpv_prenoffs != DOF_SECT_NONE && 14719 (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS, 14720 provider->dofpv_prenoffs)) == NULL) 14721 return (-1); 14722 14723 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset); 14724 14725 if (provider->dofpv_name >= str_sec->dofs_size || 14726 strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) { 14727 dtrace_dof_error(dof, "invalid provider name"); 14728 return (-1); 14729 } 14730 14731 if (prb_sec->dofs_entsize == 0 || 14732 prb_sec->dofs_entsize > prb_sec->dofs_size) { 14733 dtrace_dof_error(dof, "invalid entry size"); 14734 return (-1); 14735 } 14736 14737 if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) { 14738 dtrace_dof_error(dof, "misaligned entry size"); 14739 return (-1); 14740 } 14741 14742 if (off_sec->dofs_entsize != sizeof (uint32_t)) { 14743 dtrace_dof_error(dof, "invalid entry size"); 14744 return (-1); 14745 } 14746 14747 if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) { 14748 dtrace_dof_error(dof, "misaligned section offset"); 14749 return (-1); 14750 } 14751 14752 if (arg_sec->dofs_entsize != sizeof (uint8_t)) { 14753 dtrace_dof_error(dof, "invalid entry size"); 14754 return (-1); 14755 } 14756 14757 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset); 14758 14759 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize; 14760 14761 /* 14762 * Take a pass through the probes to check for errors. 14763 */ 14764 for (j = 0; j < nprobes; j++) { 14765 probe = (dof_probe_t *)(uintptr_t)(daddr + 14766 prb_sec->dofs_offset + j * prb_sec->dofs_entsize); 14767 14768 if (probe->dofpr_func >= str_sec->dofs_size) { 14769 dtrace_dof_error(dof, "invalid function name"); 14770 return (-1); 14771 } 14772 14773 if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) { 14774 dtrace_dof_error(dof, "function name too long"); 14775 return (-1); 14776 } 14777 14778 if (probe->dofpr_name >= str_sec->dofs_size || 14779 strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) { 14780 dtrace_dof_error(dof, "invalid probe name"); 14781 return (-1); 14782 } 14783 14784 /* 14785 * The offset count must not wrap the index, and the offsets 14786 * must also not overflow the section's data. 14787 */ 14788 if (probe->dofpr_offidx + probe->dofpr_noffs < 14789 probe->dofpr_offidx || 14790 (probe->dofpr_offidx + probe->dofpr_noffs) * 14791 off_sec->dofs_entsize > off_sec->dofs_size) { 14792 dtrace_dof_error(dof, "invalid probe offset"); 14793 return (-1); 14794 } 14795 14796 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) { 14797 /* 14798 * If there's no is-enabled offset section, make sure 14799 * there aren't any is-enabled offsets. Otherwise 14800 * perform the same checks as for probe offsets 14801 * (immediately above). 14802 */ 14803 if (enoff_sec == NULL) { 14804 if (probe->dofpr_enoffidx != 0 || 14805 probe->dofpr_nenoffs != 0) { 14806 dtrace_dof_error(dof, "is-enabled " 14807 "offsets with null section"); 14808 return (-1); 14809 } 14810 } else if (probe->dofpr_enoffidx + 14811 probe->dofpr_nenoffs < probe->dofpr_enoffidx || 14812 (probe->dofpr_enoffidx + probe->dofpr_nenoffs) * 14813 enoff_sec->dofs_entsize > enoff_sec->dofs_size) { 14814 dtrace_dof_error(dof, "invalid is-enabled " 14815 "offset"); 14816 return (-1); 14817 } 14818 14819 if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) { 14820 dtrace_dof_error(dof, "zero probe and " 14821 "is-enabled offsets"); 14822 return (-1); 14823 } 14824 } else if (probe->dofpr_noffs == 0) { 14825 dtrace_dof_error(dof, "zero probe offsets"); 14826 return (-1); 14827 } 14828 14829 if (probe->dofpr_argidx + probe->dofpr_xargc < 14830 probe->dofpr_argidx || 14831 (probe->dofpr_argidx + probe->dofpr_xargc) * 14832 arg_sec->dofs_entsize > arg_sec->dofs_size) { 14833 dtrace_dof_error(dof, "invalid args"); 14834 return (-1); 14835 } 14836 14837 typeidx = probe->dofpr_nargv; 14838 typestr = strtab + probe->dofpr_nargv; 14839 for (k = 0; k < probe->dofpr_nargc; k++) { 14840 if (typeidx >= str_sec->dofs_size) { 14841 dtrace_dof_error(dof, "bad " 14842 "native argument type"); 14843 return (-1); 14844 } 14845 14846 typesz = strlen(typestr) + 1; 14847 if (typesz > DTRACE_ARGTYPELEN) { 14848 dtrace_dof_error(dof, "native " 14849 "argument type too long"); 14850 return (-1); 14851 } 14852 typeidx += typesz; 14853 typestr += typesz; 14854 } 14855 14856 typeidx = probe->dofpr_xargv; 14857 typestr = strtab + probe->dofpr_xargv; 14858 for (k = 0; k < probe->dofpr_xargc; k++) { 14859 if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) { 14860 dtrace_dof_error(dof, "bad " 14861 "native argument index"); 14862 return (-1); 14863 } 14864 14865 if (typeidx >= str_sec->dofs_size) { 14866 dtrace_dof_error(dof, "bad " 14867 "translated argument type"); 14868 return (-1); 14869 } 14870 14871 typesz = strlen(typestr) + 1; 14872 if (typesz > DTRACE_ARGTYPELEN) { 14873 dtrace_dof_error(dof, "translated argument " 14874 "type too long"); 14875 return (-1); 14876 } 14877 14878 typeidx += typesz; 14879 typestr += typesz; 14880 } 14881 } 14882 14883 return (0); 14884} 14885 14886static int 14887dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp) 14888{ 14889 dtrace_helpers_t *help; 14890 dtrace_vstate_t *vstate; 14891 dtrace_enabling_t *enab = NULL; 14892 int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1; 14893 uintptr_t daddr = (uintptr_t)dof; 14894 14895 ASSERT(MUTEX_HELD(&dtrace_lock)); 14896 14897 if ((help = curproc->p_dtrace_helpers) == NULL) 14898 help = dtrace_helpers_create(curproc); 14899 14900 vstate = &help->dthps_vstate; 14901 14902 if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab, 14903 dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) { 14904 dtrace_dof_destroy(dof); 14905 return (rv); 14906 } 14907 14908 /* 14909 * Look for helper providers and validate their descriptions. 14910 */ 14911 if (dhp != NULL) { 14912 for (i = 0; i < dof->dofh_secnum; i++) { 14913 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr + 14914 dof->dofh_secoff + i * dof->dofh_secsize); 14915 14916 if (sec->dofs_type != DOF_SECT_PROVIDER) 14917 continue; 14918 14919 if (dtrace_helper_provider_validate(dof, sec) != 0) { 14920 dtrace_enabling_destroy(enab); 14921 dtrace_dof_destroy(dof); 14922 return (-1); 14923 } 14924 14925 nprovs++; 14926 } 14927 } 14928 14929 /* 14930 * Now we need to walk through the ECB descriptions in the enabling. 14931 */ 14932 for (i = 0; i < enab->dten_ndesc; i++) { 14933 dtrace_ecbdesc_t *ep = enab->dten_desc[i]; 14934 dtrace_probedesc_t *desc = &ep->dted_probe; 14935 14936 if (strcmp(desc->dtpd_provider, "dtrace") != 0) 14937 continue; 14938 14939 if (strcmp(desc->dtpd_mod, "helper") != 0) 14940 continue; 14941 14942 if (strcmp(desc->dtpd_func, "ustack") != 0) 14943 continue; 14944 14945 if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK, 14946 ep)) != 0) { 14947 /* 14948 * Adding this helper action failed -- we are now going 14949 * to rip out the entire generation and return failure. 14950 */ 14951 (void) dtrace_helper_destroygen(help->dthps_generation); 14952 dtrace_enabling_destroy(enab); 14953 dtrace_dof_destroy(dof); 14954 return (-1); 14955 } 14956 14957 nhelpers++; 14958 } 14959 14960 if (nhelpers < enab->dten_ndesc) 14961 dtrace_dof_error(dof, "unmatched helpers"); 14962 14963 gen = help->dthps_generation++; 14964 dtrace_enabling_destroy(enab); 14965 14966 if (dhp != NULL && nprovs > 0) { 14967 dhp->dofhp_dof = (uint64_t)(uintptr_t)dof; 14968 if (dtrace_helper_provider_add(dhp, gen) == 0) { 14969 mutex_exit(&dtrace_lock); 14970 dtrace_helper_provider_register(curproc, help, dhp); 14971 mutex_enter(&dtrace_lock); 14972 14973 destroy = 0; 14974 } 14975 } 14976 14977 if (destroy) 14978 dtrace_dof_destroy(dof); 14979 14980 return (gen); 14981} 14982 14983static dtrace_helpers_t * 14984dtrace_helpers_create(proc_t *p) 14985{ 14986 dtrace_helpers_t *help; 14987 14988 ASSERT(MUTEX_HELD(&dtrace_lock)); 14989 ASSERT(p->p_dtrace_helpers == NULL); 14990 14991 help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP); 14992 help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) * 14993 DTRACE_NHELPER_ACTIONS, KM_SLEEP); 14994 14995 p->p_dtrace_helpers = help; 14996 dtrace_helpers++; 14997 14998 return (help); 14999} 15000 15001#if defined(sun) 15002static 15003#endif 15004void 15005dtrace_helpers_destroy(proc_t *p) 15006{ 15007 dtrace_helpers_t *help; 15008 dtrace_vstate_t *vstate; 15009#if defined(sun) 15010 proc_t *p = curproc; 15011#endif 15012 int i; 15013 15014 mutex_enter(&dtrace_lock); 15015 15016 ASSERT(p->p_dtrace_helpers != NULL); 15017 ASSERT(dtrace_helpers > 0); 15018 15019 help = p->p_dtrace_helpers; 15020 vstate = &help->dthps_vstate; 15021 15022 /* 15023 * We're now going to lose the help from this process. 15024 */ 15025 p->p_dtrace_helpers = NULL; 15026 dtrace_sync(); 15027 15028 /* 15029 * Destory the helper actions. 15030 */ 15031 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) { 15032 dtrace_helper_action_t *h, *next; 15033 15034 for (h = help->dthps_actions[i]; h != NULL; h = next) { 15035 next = h->dtha_next; 15036 dtrace_helper_action_destroy(h, vstate); 15037 h = next; 15038 } 15039 } 15040 15041 mutex_exit(&dtrace_lock); 15042 15043 /* 15044 * Destroy the helper providers. 15045 */ 15046 if (help->dthps_maxprovs > 0) { 15047 mutex_enter(&dtrace_meta_lock); 15048 if (dtrace_meta_pid != NULL) { 15049 ASSERT(dtrace_deferred_pid == NULL); 15050 15051 for (i = 0; i < help->dthps_nprovs; i++) { 15052 dtrace_helper_provider_remove( 15053 &help->dthps_provs[i]->dthp_prov, p->p_pid); 15054 } 15055 } else { 15056 mutex_enter(&dtrace_lock); 15057 ASSERT(help->dthps_deferred == 0 || 15058 help->dthps_next != NULL || 15059 help->dthps_prev != NULL || 15060 help == dtrace_deferred_pid); 15061 15062 /* 15063 * Remove the helper from the deferred list. 15064 */ 15065 if (help->dthps_next != NULL) 15066 help->dthps_next->dthps_prev = help->dthps_prev; 15067 if (help->dthps_prev != NULL) 15068 help->dthps_prev->dthps_next = help->dthps_next; 15069 if (dtrace_deferred_pid == help) { 15070 dtrace_deferred_pid = help->dthps_next; 15071 ASSERT(help->dthps_prev == NULL); 15072 } 15073 15074 mutex_exit(&dtrace_lock); 15075 } 15076 15077 mutex_exit(&dtrace_meta_lock); 15078 15079 for (i = 0; i < help->dthps_nprovs; i++) { 15080 dtrace_helper_provider_destroy(help->dthps_provs[i]); 15081 } 15082 15083 kmem_free(help->dthps_provs, help->dthps_maxprovs * 15084 sizeof (dtrace_helper_provider_t *)); 15085 } 15086 15087 mutex_enter(&dtrace_lock); 15088 15089 dtrace_vstate_fini(&help->dthps_vstate); 15090 kmem_free(help->dthps_actions, 15091 sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS); 15092 kmem_free(help, sizeof (dtrace_helpers_t)); 15093 15094 --dtrace_helpers; 15095 mutex_exit(&dtrace_lock); 15096} 15097 15098#if defined(sun) 15099static 15100#endif 15101void 15102dtrace_helpers_duplicate(proc_t *from, proc_t *to) 15103{ 15104 dtrace_helpers_t *help, *newhelp; 15105 dtrace_helper_action_t *helper, *new, *last; 15106 dtrace_difo_t *dp; 15107 dtrace_vstate_t *vstate; 15108 int i, j, sz, hasprovs = 0; 15109 15110 mutex_enter(&dtrace_lock); 15111 ASSERT(from->p_dtrace_helpers != NULL); 15112 ASSERT(dtrace_helpers > 0); 15113 15114 help = from->p_dtrace_helpers; 15115 newhelp = dtrace_helpers_create(to); 15116 ASSERT(to->p_dtrace_helpers != NULL); 15117 15118 newhelp->dthps_generation = help->dthps_generation; 15119 vstate = &newhelp->dthps_vstate; 15120 15121 /* 15122 * Duplicate the helper actions. 15123 */ 15124 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) { 15125 if ((helper = help->dthps_actions[i]) == NULL) 15126 continue; 15127 15128 for (last = NULL; helper != NULL; helper = helper->dtha_next) { 15129 new = kmem_zalloc(sizeof (dtrace_helper_action_t), 15130 KM_SLEEP); 15131 new->dtha_generation = helper->dtha_generation; 15132 15133 if ((dp = helper->dtha_predicate) != NULL) { 15134 dp = dtrace_difo_duplicate(dp, vstate); 15135 new->dtha_predicate = dp; 15136 } 15137 15138 new->dtha_nactions = helper->dtha_nactions; 15139 sz = sizeof (dtrace_difo_t *) * new->dtha_nactions; 15140 new->dtha_actions = kmem_alloc(sz, KM_SLEEP); 15141 15142 for (j = 0; j < new->dtha_nactions; j++) { 15143 dtrace_difo_t *dp = helper->dtha_actions[j]; 15144 15145 ASSERT(dp != NULL); 15146 dp = dtrace_difo_duplicate(dp, vstate); 15147 new->dtha_actions[j] = dp; 15148 } 15149 15150 if (last != NULL) { 15151 last->dtha_next = new; 15152 } else { 15153 newhelp->dthps_actions[i] = new; 15154 } 15155 15156 last = new; 15157 } 15158 } 15159 15160 /* 15161 * Duplicate the helper providers and register them with the 15162 * DTrace framework. 15163 */ 15164 if (help->dthps_nprovs > 0) { 15165 newhelp->dthps_nprovs = help->dthps_nprovs; 15166 newhelp->dthps_maxprovs = help->dthps_nprovs; 15167 newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs * 15168 sizeof (dtrace_helper_provider_t *), KM_SLEEP); 15169 for (i = 0; i < newhelp->dthps_nprovs; i++) { 15170 newhelp->dthps_provs[i] = help->dthps_provs[i]; 15171 newhelp->dthps_provs[i]->dthp_ref++; 15172 } 15173 15174 hasprovs = 1; 15175 } 15176 15177 mutex_exit(&dtrace_lock); 15178 15179 if (hasprovs) 15180 dtrace_helper_provider_register(to, newhelp, NULL); 15181} 15182 15183/* 15184 * DTrace Hook Functions 15185 */ 15186static void 15187dtrace_module_loaded(modctl_t *ctl) 15188{ 15189 dtrace_provider_t *prv; 15190 15191 mutex_enter(&dtrace_provider_lock); 15192#if defined(sun) 15193 mutex_enter(&mod_lock); 15194#endif 15195 15196#if defined(sun) 15197 ASSERT(ctl->mod_busy); 15198#endif 15199 15200 /* 15201 * We're going to call each providers per-module provide operation 15202 * specifying only this module. 15203 */ 15204 for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next) 15205 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl); 15206 15207#if defined(sun) 15208 mutex_exit(&mod_lock); 15209#endif 15210 mutex_exit(&dtrace_provider_lock); 15211 15212 /* 15213 * If we have any retained enablings, we need to match against them. 15214 * Enabling probes requires that cpu_lock be held, and we cannot hold 15215 * cpu_lock here -- it is legal for cpu_lock to be held when loading a 15216 * module. (In particular, this happens when loading scheduling 15217 * classes.) So if we have any retained enablings, we need to dispatch 15218 * our task queue to do the match for us. 15219 */ 15220 mutex_enter(&dtrace_lock); 15221 15222 if (dtrace_retained == NULL) { 15223 mutex_exit(&dtrace_lock); 15224 return; 15225 } 15226 15227 (void) taskq_dispatch(dtrace_taskq, 15228 (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP); 15229 15230 mutex_exit(&dtrace_lock); 15231 15232 /* 15233 * And now, for a little heuristic sleaze: in general, we want to 15234 * match modules as soon as they load. However, we cannot guarantee 15235 * this, because it would lead us to the lock ordering violation 15236 * outlined above. The common case, of course, is that cpu_lock is 15237 * _not_ held -- so we delay here for a clock tick, hoping that that's 15238 * long enough for the task queue to do its work. If it's not, it's 15239 * not a serious problem -- it just means that the module that we 15240 * just loaded may not be immediately instrumentable. 15241 */ 15242 delay(1); 15243} 15244 15245static void 15246#if defined(sun) 15247dtrace_module_unloaded(modctl_t *ctl) 15248#else 15249dtrace_module_unloaded(modctl_t *ctl, int *error) 15250#endif 15251{ 15252 dtrace_probe_t template, *probe, *first, *next; 15253 dtrace_provider_t *prov; 15254#if !defined(sun) 15255 char modname[DTRACE_MODNAMELEN]; 15256 size_t len; 15257#endif 15258 15259#if defined(sun) 15260 template.dtpr_mod = ctl->mod_modname; 15261#else 15262 /* Handle the fact that ctl->filename may end in ".ko". */ 15263 strlcpy(modname, ctl->filename, sizeof(modname)); 15264 len = strlen(ctl->filename); 15265 if (len > 3 && strcmp(modname + len - 3, ".ko") == 0) 15266 modname[len - 3] = '\0'; 15267 template.dtpr_mod = modname; 15268#endif 15269 15270 mutex_enter(&dtrace_provider_lock); 15271#if defined(sun) 15272 mutex_enter(&mod_lock); 15273#endif 15274 mutex_enter(&dtrace_lock); 15275 15276#if !defined(sun) 15277 if (ctl->nenabled > 0) { 15278 /* Don't allow unloads if a probe is enabled. */ 15279 mutex_exit(&dtrace_provider_lock); 15280 mutex_exit(&dtrace_lock); 15281 *error = -1; 15282 printf( 15283 "kldunload: attempt to unload module that has DTrace probes enabled\n"); 15284 return; 15285 } 15286#endif 15287 15288 if (dtrace_bymod == NULL) { 15289 /* 15290 * The DTrace module is loaded (obviously) but not attached; 15291 * we don't have any work to do. 15292 */ 15293 mutex_exit(&dtrace_provider_lock); 15294#if defined(sun) 15295 mutex_exit(&mod_lock); 15296#endif 15297 mutex_exit(&dtrace_lock); 15298 return; 15299 } 15300 15301 for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template); 15302 probe != NULL; probe = probe->dtpr_nextmod) { 15303 if (probe->dtpr_ecb != NULL) { 15304 mutex_exit(&dtrace_provider_lock); 15305#if defined(sun) 15306 mutex_exit(&mod_lock); 15307#endif 15308 mutex_exit(&dtrace_lock); 15309 15310 /* 15311 * This shouldn't _actually_ be possible -- we're 15312 * unloading a module that has an enabled probe in it. 15313 * (It's normally up to the provider to make sure that 15314 * this can't happen.) However, because dtps_enable() 15315 * doesn't have a failure mode, there can be an 15316 * enable/unload race. Upshot: we don't want to 15317 * assert, but we're not going to disable the 15318 * probe, either. 15319 */ 15320 if (dtrace_err_verbose) { 15321#if defined(sun) 15322 cmn_err(CE_WARN, "unloaded module '%s' had " 15323 "enabled probes", ctl->mod_modname); 15324#else 15325 cmn_err(CE_WARN, "unloaded module '%s' had " 15326 "enabled probes", modname); 15327#endif 15328 } 15329 15330 return; 15331 } 15332 } 15333 15334 probe = first; 15335 15336 for (first = NULL; probe != NULL; probe = next) { 15337 ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe); 15338 15339 dtrace_probes[probe->dtpr_id - 1] = NULL; 15340 15341 next = probe->dtpr_nextmod; 15342 dtrace_hash_remove(dtrace_bymod, probe); 15343 dtrace_hash_remove(dtrace_byfunc, probe); 15344 dtrace_hash_remove(dtrace_byname, probe); 15345 15346 if (first == NULL) { 15347 first = probe; 15348 probe->dtpr_nextmod = NULL; 15349 } else { 15350 probe->dtpr_nextmod = first; 15351 first = probe; 15352 } 15353 } 15354 15355 /* 15356 * We've removed all of the module's probes from the hash chains and 15357 * from the probe array. Now issue a dtrace_sync() to be sure that 15358 * everyone has cleared out from any probe array processing. 15359 */ 15360 dtrace_sync(); 15361 15362 for (probe = first; probe != NULL; probe = first) { 15363 first = probe->dtpr_nextmod; 15364 prov = probe->dtpr_provider; 15365 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id, 15366 probe->dtpr_arg); 15367 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1); 15368 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1); 15369 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1); 15370#if defined(sun) 15371 vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1); 15372#else 15373 free_unr(dtrace_arena, probe->dtpr_id); 15374#endif 15375 kmem_free(probe, sizeof (dtrace_probe_t)); 15376 } 15377 15378 mutex_exit(&dtrace_lock); 15379#if defined(sun) 15380 mutex_exit(&mod_lock); 15381#endif 15382 mutex_exit(&dtrace_provider_lock); 15383} 15384 15385#if !defined(sun) 15386static void 15387dtrace_kld_load(void *arg __unused, linker_file_t lf) 15388{ 15389 15390 dtrace_module_loaded(lf); 15391} 15392 15393static void 15394dtrace_kld_unload_try(void *arg __unused, linker_file_t lf, int *error) 15395{ 15396 15397 if (*error != 0) 15398 /* We already have an error, so don't do anything. */ 15399 return; 15400 dtrace_module_unloaded(lf, error); 15401} 15402#endif 15403 15404#if defined(sun) 15405static void 15406dtrace_suspend(void) 15407{ 15408 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend)); 15409} 15410 15411static void 15412dtrace_resume(void) 15413{ 15414 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume)); 15415} 15416#endif 15417 15418static int 15419dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu) 15420{ 15421 ASSERT(MUTEX_HELD(&cpu_lock)); 15422 mutex_enter(&dtrace_lock); 15423 15424 switch (what) { 15425 case CPU_CONFIG: { 15426 dtrace_state_t *state; 15427 dtrace_optval_t *opt, rs, c; 15428 15429 /* 15430 * For now, we only allocate a new buffer for anonymous state. 15431 */ 15432 if ((state = dtrace_anon.dta_state) == NULL) 15433 break; 15434 15435 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) 15436 break; 15437 15438 opt = state->dts_options; 15439 c = opt[DTRACEOPT_CPU]; 15440 15441 if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu) 15442 break; 15443 15444 /* 15445 * Regardless of what the actual policy is, we're going to 15446 * temporarily set our resize policy to be manual. We're 15447 * also going to temporarily set our CPU option to denote 15448 * the newly configured CPU. 15449 */ 15450 rs = opt[DTRACEOPT_BUFRESIZE]; 15451 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL; 15452 opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu; 15453 15454 (void) dtrace_state_buffers(state); 15455 15456 opt[DTRACEOPT_BUFRESIZE] = rs; 15457 opt[DTRACEOPT_CPU] = c; 15458 15459 break; 15460 } 15461 15462 case CPU_UNCONFIG: 15463 /* 15464 * We don't free the buffer in the CPU_UNCONFIG case. (The 15465 * buffer will be freed when the consumer exits.) 15466 */ 15467 break; 15468 15469 default: 15470 break; 15471 } 15472 15473 mutex_exit(&dtrace_lock); 15474 return (0); 15475} 15476 15477#if defined(sun) 15478static void 15479dtrace_cpu_setup_initial(processorid_t cpu) 15480{ 15481 (void) dtrace_cpu_setup(CPU_CONFIG, cpu); 15482} 15483#endif 15484 15485static void 15486dtrace_toxrange_add(uintptr_t base, uintptr_t limit) 15487{ 15488 if (dtrace_toxranges >= dtrace_toxranges_max) { 15489 int osize, nsize; 15490 dtrace_toxrange_t *range; 15491 15492 osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t); 15493 15494 if (osize == 0) { 15495 ASSERT(dtrace_toxrange == NULL); 15496 ASSERT(dtrace_toxranges_max == 0); 15497 dtrace_toxranges_max = 1; 15498 } else { 15499 dtrace_toxranges_max <<= 1; 15500 } 15501 15502 nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t); 15503 range = kmem_zalloc(nsize, KM_SLEEP); 15504 15505 if (dtrace_toxrange != NULL) { 15506 ASSERT(osize != 0); 15507 bcopy(dtrace_toxrange, range, osize); 15508 kmem_free(dtrace_toxrange, osize); 15509 } 15510 15511 dtrace_toxrange = range; 15512 } 15513 15514 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == 0); 15515 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == 0); 15516 15517 dtrace_toxrange[dtrace_toxranges].dtt_base = base; 15518 dtrace_toxrange[dtrace_toxranges].dtt_limit = limit; 15519 dtrace_toxranges++; 15520} 15521 15522/* 15523 * DTrace Driver Cookbook Functions 15524 */ 15525#if defined(sun) 15526/*ARGSUSED*/ 15527static int 15528dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd) 15529{ 15530 dtrace_provider_id_t id; 15531 dtrace_state_t *state = NULL; 15532 dtrace_enabling_t *enab; 15533 15534 mutex_enter(&cpu_lock); 15535 mutex_enter(&dtrace_provider_lock); 15536 mutex_enter(&dtrace_lock); 15537 15538 if (ddi_soft_state_init(&dtrace_softstate, 15539 sizeof (dtrace_state_t), 0) != 0) { 15540 cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state"); 15541 mutex_exit(&cpu_lock); 15542 mutex_exit(&dtrace_provider_lock); 15543 mutex_exit(&dtrace_lock); 15544 return (DDI_FAILURE); 15545 } 15546 15547 if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR, 15548 DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE || 15549 ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR, 15550 DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) { 15551 cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes"); 15552 ddi_remove_minor_node(devi, NULL); 15553 ddi_soft_state_fini(&dtrace_softstate); 15554 mutex_exit(&cpu_lock); 15555 mutex_exit(&dtrace_provider_lock); 15556 mutex_exit(&dtrace_lock); 15557 return (DDI_FAILURE); 15558 } 15559 15560 ddi_report_dev(devi); 15561 dtrace_devi = devi; 15562 15563 dtrace_modload = dtrace_module_loaded; 15564 dtrace_modunload = dtrace_module_unloaded; 15565 dtrace_cpu_init = dtrace_cpu_setup_initial; 15566 dtrace_helpers_cleanup = dtrace_helpers_destroy; 15567 dtrace_helpers_fork = dtrace_helpers_duplicate; 15568 dtrace_cpustart_init = dtrace_suspend; 15569 dtrace_cpustart_fini = dtrace_resume; 15570 dtrace_debugger_init = dtrace_suspend; 15571 dtrace_debugger_fini = dtrace_resume; 15572 15573 register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL); 15574 15575 ASSERT(MUTEX_HELD(&cpu_lock)); 15576 15577 dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1, 15578 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER); 15579 dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE, 15580 UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0, 15581 VM_SLEEP | VMC_IDENTIFIER); 15582 dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri, 15583 1, INT_MAX, 0); 15584 15585 dtrace_state_cache = kmem_cache_create("dtrace_state_cache", 15586 sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN, 15587 NULL, NULL, NULL, NULL, NULL, 0); 15588 15589 ASSERT(MUTEX_HELD(&cpu_lock)); 15590 dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod), 15591 offsetof(dtrace_probe_t, dtpr_nextmod), 15592 offsetof(dtrace_probe_t, dtpr_prevmod)); 15593 15594 dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func), 15595 offsetof(dtrace_probe_t, dtpr_nextfunc), 15596 offsetof(dtrace_probe_t, dtpr_prevfunc)); 15597 15598 dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name), 15599 offsetof(dtrace_probe_t, dtpr_nextname), 15600 offsetof(dtrace_probe_t, dtpr_prevname)); 15601 15602 if (dtrace_retain_max < 1) { 15603 cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; " 15604 "setting to 1", dtrace_retain_max); 15605 dtrace_retain_max = 1; 15606 } 15607 15608 /* 15609 * Now discover our toxic ranges. 15610 */ 15611 dtrace_toxic_ranges(dtrace_toxrange_add); 15612 15613 /* 15614 * Before we register ourselves as a provider to our own framework, 15615 * we would like to assert that dtrace_provider is NULL -- but that's 15616 * not true if we were loaded as a dependency of a DTrace provider. 15617 * Once we've registered, we can assert that dtrace_provider is our 15618 * pseudo provider. 15619 */ 15620 (void) dtrace_register("dtrace", &dtrace_provider_attr, 15621 DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id); 15622 15623 ASSERT(dtrace_provider != NULL); 15624 ASSERT((dtrace_provider_id_t)dtrace_provider == id); 15625 15626 dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t) 15627 dtrace_provider, NULL, NULL, "BEGIN", 0, NULL); 15628 dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t) 15629 dtrace_provider, NULL, NULL, "END", 0, NULL); 15630 dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t) 15631 dtrace_provider, NULL, NULL, "ERROR", 1, NULL); 15632 15633 dtrace_anon_property(); 15634 mutex_exit(&cpu_lock); 15635 15636 /* 15637 * If DTrace helper tracing is enabled, we need to allocate the 15638 * trace buffer and initialize the values. 15639 */ 15640 if (dtrace_helptrace_enabled) { 15641 ASSERT(dtrace_helptrace_buffer == NULL); 15642 dtrace_helptrace_buffer = 15643 kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP); 15644 dtrace_helptrace_next = 0; 15645 } 15646 15647 /* 15648 * If there are already providers, we must ask them to provide their 15649 * probes, and then match any anonymous enabling against them. Note 15650 * that there should be no other retained enablings at this time: 15651 * the only retained enablings at this time should be the anonymous 15652 * enabling. 15653 */ 15654 if (dtrace_anon.dta_enabling != NULL) { 15655 ASSERT(dtrace_retained == dtrace_anon.dta_enabling); 15656 15657 dtrace_enabling_provide(NULL); 15658 state = dtrace_anon.dta_state; 15659 15660 /* 15661 * We couldn't hold cpu_lock across the above call to 15662 * dtrace_enabling_provide(), but we must hold it to actually 15663 * enable the probes. We have to drop all of our locks, pick 15664 * up cpu_lock, and regain our locks before matching the 15665 * retained anonymous enabling. 15666 */ 15667 mutex_exit(&dtrace_lock); 15668 mutex_exit(&dtrace_provider_lock); 15669 15670 mutex_enter(&cpu_lock); 15671 mutex_enter(&dtrace_provider_lock); 15672 mutex_enter(&dtrace_lock); 15673 15674 if ((enab = dtrace_anon.dta_enabling) != NULL) 15675 (void) dtrace_enabling_match(enab, NULL); 15676 15677 mutex_exit(&cpu_lock); 15678 } 15679 15680 mutex_exit(&dtrace_lock); 15681 mutex_exit(&dtrace_provider_lock); 15682 15683 if (state != NULL) { 15684 /* 15685 * If we created any anonymous state, set it going now. 15686 */ 15687 (void) dtrace_state_go(state, &dtrace_anon.dta_beganon); 15688 } 15689 15690 return (DDI_SUCCESS); 15691} 15692#endif 15693 15694#if !defined(sun) 15695#if __FreeBSD_version >= 800039 15696static void dtrace_dtr(void *); 15697#endif 15698#endif 15699 15700/*ARGSUSED*/ 15701static int 15702#if defined(sun) 15703dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p) 15704#else 15705dtrace_open(struct cdev *dev, int oflags, int devtype, struct thread *td) 15706#endif 15707{ 15708 dtrace_state_t *state; 15709 uint32_t priv; 15710 uid_t uid; 15711 zoneid_t zoneid; 15712 15713#if defined(sun) 15714 if (getminor(*devp) == DTRACEMNRN_HELPER) 15715 return (0); 15716 15717 /* 15718 * If this wasn't an open with the "helper" minor, then it must be 15719 * the "dtrace" minor. 15720 */ 15721 if (getminor(*devp) == DTRACEMNRN_DTRACE) 15722 return (ENXIO); 15723#else 15724 cred_t *cred_p = NULL; 15725 15726#if __FreeBSD_version < 800039 15727 /* 15728 * The first minor device is the one that is cloned so there is 15729 * nothing more to do here. 15730 */ 15731 if (dev2unit(dev) == 0) 15732 return 0; 15733 15734 /* 15735 * Devices are cloned, so if the DTrace state has already 15736 * been allocated, that means this device belongs to a 15737 * different client. Each client should open '/dev/dtrace' 15738 * to get a cloned device. 15739 */ 15740 if (dev->si_drv1 != NULL) 15741 return (EBUSY); 15742#endif 15743 15744 cred_p = dev->si_cred; 15745#endif 15746 15747 /* 15748 * If no DTRACE_PRIV_* bits are set in the credential, then the 15749 * caller lacks sufficient permission to do anything with DTrace. 15750 */ 15751 dtrace_cred2priv(cred_p, &priv, &uid, &zoneid); 15752 if (priv == DTRACE_PRIV_NONE) { 15753#if !defined(sun) 15754#if __FreeBSD_version < 800039 15755 /* Destroy the cloned device. */ 15756 destroy_dev(dev); 15757#endif 15758#endif 15759 15760 return (EACCES); 15761 } 15762 15763 /* 15764 * Ask all providers to provide all their probes. 15765 */ 15766 mutex_enter(&dtrace_provider_lock); 15767 dtrace_probe_provide(NULL, NULL); 15768 mutex_exit(&dtrace_provider_lock); 15769 15770 mutex_enter(&cpu_lock); 15771 mutex_enter(&dtrace_lock); 15772 dtrace_opens++; 15773 dtrace_membar_producer(); 15774 15775#if defined(sun) 15776 /* 15777 * If the kernel debugger is active (that is, if the kernel debugger 15778 * modified text in some way), we won't allow the open. 15779 */ 15780 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) { 15781 dtrace_opens--; 15782 mutex_exit(&cpu_lock); 15783 mutex_exit(&dtrace_lock); 15784 return (EBUSY); 15785 } 15786 15787 state = dtrace_state_create(devp, cred_p); 15788#else 15789 state = dtrace_state_create(dev); 15790#if __FreeBSD_version < 800039 15791 dev->si_drv1 = state; 15792#else 15793 devfs_set_cdevpriv(state, dtrace_dtr); 15794#endif 15795#endif 15796 15797 mutex_exit(&cpu_lock); 15798 15799 if (state == NULL) { 15800#if defined(sun) 15801 if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL) 15802 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE); 15803#else 15804 --dtrace_opens; 15805#endif 15806 mutex_exit(&dtrace_lock); 15807#if !defined(sun) 15808#if __FreeBSD_version < 800039 15809 /* Destroy the cloned device. */ 15810 destroy_dev(dev); 15811#endif 15812#endif 15813 return (EAGAIN); 15814 } 15815 15816 mutex_exit(&dtrace_lock); 15817 15818 return (0); 15819} 15820 15821/*ARGSUSED*/ 15822#if defined(sun) 15823static int 15824dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p) 15825#elif __FreeBSD_version < 800039 15826static int 15827dtrace_close(struct cdev *dev, int flags, int fmt __unused, struct thread *td) 15828#else 15829static void 15830dtrace_dtr(void *data) 15831#endif 15832{ 15833#if defined(sun) 15834 minor_t minor = getminor(dev); 15835 dtrace_state_t *state; 15836 15837 if (minor == DTRACEMNRN_HELPER) 15838 return (0); 15839 15840 state = ddi_get_soft_state(dtrace_softstate, minor); 15841#else 15842#if __FreeBSD_version < 800039 15843 dtrace_state_t *state = dev->si_drv1; 15844 15845 /* Check if this is not a cloned device. */ 15846 if (dev2unit(dev) == 0) 15847 return (0); 15848#else 15849 dtrace_state_t *state = data; 15850#endif 15851 15852#endif 15853 15854 mutex_enter(&cpu_lock); 15855 mutex_enter(&dtrace_lock); 15856 15857 if (state != NULL) { 15858 if (state->dts_anon) { 15859 /* 15860 * There is anonymous state. Destroy that first. 15861 */ 15862 ASSERT(dtrace_anon.dta_state == NULL); 15863 dtrace_state_destroy(state->dts_anon); 15864 } 15865 15866 dtrace_state_destroy(state); 15867 15868#if !defined(sun) 15869 kmem_free(state, 0); 15870#if __FreeBSD_version < 800039 15871 dev->si_drv1 = NULL; 15872#endif 15873#endif 15874 } 15875 15876 ASSERT(dtrace_opens > 0); 15877#if defined(sun) 15878 /* 15879 * Only relinquish control of the kernel debugger interface when there 15880 * are no consumers and no anonymous enablings. 15881 */ 15882 if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL) 15883 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE); 15884#else 15885 --dtrace_opens; 15886#endif 15887 15888 mutex_exit(&dtrace_lock); 15889 mutex_exit(&cpu_lock); 15890 15891#if __FreeBSD_version < 800039 15892 /* Schedule this cloned device to be destroyed. */ 15893 destroy_dev_sched(dev); 15894#endif 15895 15896#if defined(sun) || __FreeBSD_version < 800039 15897 return (0); 15898#endif 15899} 15900 15901#if defined(sun) 15902/*ARGSUSED*/ 15903static int 15904dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv) 15905{ 15906 int rval; 15907 dof_helper_t help, *dhp = NULL; 15908 15909 switch (cmd) { 15910 case DTRACEHIOC_ADDDOF: 15911 if (copyin((void *)arg, &help, sizeof (help)) != 0) { 15912 dtrace_dof_error(NULL, "failed to copyin DOF helper"); 15913 return (EFAULT); 15914 } 15915 15916 dhp = &help; 15917 arg = (intptr_t)help.dofhp_dof; 15918 /*FALLTHROUGH*/ 15919 15920 case DTRACEHIOC_ADD: { 15921 dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval); 15922 15923 if (dof == NULL) 15924 return (rval); 15925 15926 mutex_enter(&dtrace_lock); 15927 15928 /* 15929 * dtrace_helper_slurp() takes responsibility for the dof -- 15930 * it may free it now or it may save it and free it later. 15931 */ 15932 if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) { 15933 *rv = rval; 15934 rval = 0; 15935 } else { 15936 rval = EINVAL; 15937 } 15938 15939 mutex_exit(&dtrace_lock); 15940 return (rval); 15941 } 15942 15943 case DTRACEHIOC_REMOVE: { 15944 mutex_enter(&dtrace_lock); 15945 rval = dtrace_helper_destroygen(arg); 15946 mutex_exit(&dtrace_lock); 15947 15948 return (rval); 15949 } 15950 15951 default: 15952 break; 15953 } 15954 15955 return (ENOTTY); 15956} 15957 15958/*ARGSUSED*/ 15959static int 15960dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv) 15961{ 15962 minor_t minor = getminor(dev); 15963 dtrace_state_t *state; 15964 int rval; 15965 15966 if (minor == DTRACEMNRN_HELPER) 15967 return (dtrace_ioctl_helper(cmd, arg, rv)); 15968 15969 state = ddi_get_soft_state(dtrace_softstate, minor); 15970 15971 if (state->dts_anon) { 15972 ASSERT(dtrace_anon.dta_state == NULL); 15973 state = state->dts_anon; 15974 } 15975 15976 switch (cmd) { 15977 case DTRACEIOC_PROVIDER: { 15978 dtrace_providerdesc_t pvd; 15979 dtrace_provider_t *pvp; 15980 15981 if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0) 15982 return (EFAULT); 15983 15984 pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0'; 15985 mutex_enter(&dtrace_provider_lock); 15986 15987 for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) { 15988 if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0) 15989 break; 15990 } 15991 15992 mutex_exit(&dtrace_provider_lock); 15993 15994 if (pvp == NULL) 15995 return (ESRCH); 15996 15997 bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t)); 15998 bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t)); 15999 16000 if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0) 16001 return (EFAULT); 16002 16003 return (0); 16004 } 16005 16006 case DTRACEIOC_EPROBE: { 16007 dtrace_eprobedesc_t epdesc; 16008 dtrace_ecb_t *ecb; 16009 dtrace_action_t *act; 16010 void *buf; 16011 size_t size; 16012 uintptr_t dest; 16013 int nrecs; 16014 16015 if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0) 16016 return (EFAULT); 16017 16018 mutex_enter(&dtrace_lock); 16019 16020 if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) { 16021 mutex_exit(&dtrace_lock); 16022 return (EINVAL); 16023 } 16024 16025 if (ecb->dte_probe == NULL) { 16026 mutex_exit(&dtrace_lock); 16027 return (EINVAL); 16028 } 16029 16030 epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id; 16031 epdesc.dtepd_uarg = ecb->dte_uarg; 16032 epdesc.dtepd_size = ecb->dte_size; 16033 16034 nrecs = epdesc.dtepd_nrecs; 16035 epdesc.dtepd_nrecs = 0; 16036 for (act = ecb->dte_action; act != NULL; act = act->dta_next) { 16037 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple) 16038 continue; 16039 16040 epdesc.dtepd_nrecs++; 16041 } 16042 16043 /* 16044 * Now that we have the size, we need to allocate a temporary 16045 * buffer in which to store the complete description. We need 16046 * the temporary buffer to be able to drop dtrace_lock() 16047 * across the copyout(), below. 16048 */ 16049 size = sizeof (dtrace_eprobedesc_t) + 16050 (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t)); 16051 16052 buf = kmem_alloc(size, KM_SLEEP); 16053 dest = (uintptr_t)buf; 16054 16055 bcopy(&epdesc, (void *)dest, sizeof (epdesc)); 16056 dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]); 16057 16058 for (act = ecb->dte_action; act != NULL; act = act->dta_next) { 16059 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple) 16060 continue; 16061 16062 if (nrecs-- == 0) 16063 break; 16064 16065 bcopy(&act->dta_rec, (void *)dest, 16066 sizeof (dtrace_recdesc_t)); 16067 dest += sizeof (dtrace_recdesc_t); 16068 } 16069 16070 mutex_exit(&dtrace_lock); 16071 16072 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) { 16073 kmem_free(buf, size); 16074 return (EFAULT); 16075 } 16076 16077 kmem_free(buf, size); 16078 return (0); 16079 } 16080 16081 case DTRACEIOC_AGGDESC: { 16082 dtrace_aggdesc_t aggdesc; 16083 dtrace_action_t *act; 16084 dtrace_aggregation_t *agg; 16085 int nrecs; 16086 uint32_t offs; 16087 dtrace_recdesc_t *lrec; 16088 void *buf; 16089 size_t size; 16090 uintptr_t dest; 16091 16092 if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0) 16093 return (EFAULT); 16094 16095 mutex_enter(&dtrace_lock); 16096 16097 if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) { 16098 mutex_exit(&dtrace_lock); 16099 return (EINVAL); 16100 } 16101 16102 aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid; 16103 16104 nrecs = aggdesc.dtagd_nrecs; 16105 aggdesc.dtagd_nrecs = 0; 16106 16107 offs = agg->dtag_base; 16108 lrec = &agg->dtag_action.dta_rec; 16109 aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs; 16110 16111 for (act = agg->dtag_first; ; act = act->dta_next) { 16112 ASSERT(act->dta_intuple || 16113 DTRACEACT_ISAGG(act->dta_kind)); 16114 16115 /* 16116 * If this action has a record size of zero, it 16117 * denotes an argument to the aggregating action. 16118 * Because the presence of this record doesn't (or 16119 * shouldn't) affect the way the data is interpreted, 16120 * we don't copy it out to save user-level the 16121 * confusion of dealing with a zero-length record. 16122 */ 16123 if (act->dta_rec.dtrd_size == 0) { 16124 ASSERT(agg->dtag_hasarg); 16125 continue; 16126 } 16127 16128 aggdesc.dtagd_nrecs++; 16129 16130 if (act == &agg->dtag_action) 16131 break; 16132 } 16133 16134 /* 16135 * Now that we have the size, we need to allocate a temporary 16136 * buffer in which to store the complete description. We need 16137 * the temporary buffer to be able to drop dtrace_lock() 16138 * across the copyout(), below. 16139 */ 16140 size = sizeof (dtrace_aggdesc_t) + 16141 (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t)); 16142 16143 buf = kmem_alloc(size, KM_SLEEP); 16144 dest = (uintptr_t)buf; 16145 16146 bcopy(&aggdesc, (void *)dest, sizeof (aggdesc)); 16147 dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]); 16148 16149 for (act = agg->dtag_first; ; act = act->dta_next) { 16150 dtrace_recdesc_t rec = act->dta_rec; 16151 16152 /* 16153 * See the comment in the above loop for why we pass 16154 * over zero-length records. 16155 */ 16156 if (rec.dtrd_size == 0) { 16157 ASSERT(agg->dtag_hasarg); 16158 continue; 16159 } 16160 16161 if (nrecs-- == 0) 16162 break; 16163 16164 rec.dtrd_offset -= offs; 16165 bcopy(&rec, (void *)dest, sizeof (rec)); 16166 dest += sizeof (dtrace_recdesc_t); 16167 16168 if (act == &agg->dtag_action) 16169 break; 16170 } 16171 16172 mutex_exit(&dtrace_lock); 16173 16174 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) { 16175 kmem_free(buf, size); 16176 return (EFAULT); 16177 } 16178 16179 kmem_free(buf, size); 16180 return (0); 16181 } 16182 16183 case DTRACEIOC_ENABLE: { 16184 dof_hdr_t *dof; 16185 dtrace_enabling_t *enab = NULL; 16186 dtrace_vstate_t *vstate; 16187 int err = 0; 16188 16189 *rv = 0; 16190 16191 /* 16192 * If a NULL argument has been passed, we take this as our 16193 * cue to reevaluate our enablings. 16194 */ 16195 if (arg == NULL) { 16196 dtrace_enabling_matchall(); 16197 16198 return (0); 16199 } 16200 16201 if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL) 16202 return (rval); 16203 16204 mutex_enter(&cpu_lock); 16205 mutex_enter(&dtrace_lock); 16206 vstate = &state->dts_vstate; 16207 16208 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) { 16209 mutex_exit(&dtrace_lock); 16210 mutex_exit(&cpu_lock); 16211 dtrace_dof_destroy(dof); 16212 return (EBUSY); 16213 } 16214 16215 if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) { 16216 mutex_exit(&dtrace_lock); 16217 mutex_exit(&cpu_lock); 16218 dtrace_dof_destroy(dof); 16219 return (EINVAL); 16220 } 16221 16222 if ((rval = dtrace_dof_options(dof, state)) != 0) { 16223 dtrace_enabling_destroy(enab); 16224 mutex_exit(&dtrace_lock); 16225 mutex_exit(&cpu_lock); 16226 dtrace_dof_destroy(dof); 16227 return (rval); 16228 } 16229 16230 if ((err = dtrace_enabling_match(enab, rv)) == 0) { 16231 err = dtrace_enabling_retain(enab); 16232 } else { 16233 dtrace_enabling_destroy(enab); 16234 } 16235 16236 mutex_exit(&cpu_lock); 16237 mutex_exit(&dtrace_lock); 16238 dtrace_dof_destroy(dof); 16239 16240 return (err); 16241 } 16242 16243 case DTRACEIOC_REPLICATE: { 16244 dtrace_repldesc_t desc; 16245 dtrace_probedesc_t *match = &desc.dtrpd_match; 16246 dtrace_probedesc_t *create = &desc.dtrpd_create; 16247 int err; 16248 16249 if (copyin((void *)arg, &desc, sizeof (desc)) != 0) 16250 return (EFAULT); 16251 16252 match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0'; 16253 match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0'; 16254 match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0'; 16255 match->dtpd_name[DTRACE_NAMELEN - 1] = '\0'; 16256 16257 create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0'; 16258 create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0'; 16259 create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0'; 16260 create->dtpd_name[DTRACE_NAMELEN - 1] = '\0'; 16261 16262 mutex_enter(&dtrace_lock); 16263 err = dtrace_enabling_replicate(state, match, create); 16264 mutex_exit(&dtrace_lock); 16265 16266 return (err); 16267 } 16268 16269 case DTRACEIOC_PROBEMATCH: 16270 case DTRACEIOC_PROBES: { 16271 dtrace_probe_t *probe = NULL; 16272 dtrace_probedesc_t desc; 16273 dtrace_probekey_t pkey; 16274 dtrace_id_t i; 16275 int m = 0; 16276 uint32_t priv; 16277 uid_t uid; 16278 zoneid_t zoneid; 16279 16280 if (copyin((void *)arg, &desc, sizeof (desc)) != 0) 16281 return (EFAULT); 16282 16283 desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0'; 16284 desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0'; 16285 desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0'; 16286 desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0'; 16287 16288 /* 16289 * Before we attempt to match this probe, we want to give 16290 * all providers the opportunity to provide it. 16291 */ 16292 if (desc.dtpd_id == DTRACE_IDNONE) { 16293 mutex_enter(&dtrace_provider_lock); 16294 dtrace_probe_provide(&desc, NULL); 16295 mutex_exit(&dtrace_provider_lock); 16296 desc.dtpd_id++; 16297 } 16298 16299 if (cmd == DTRACEIOC_PROBEMATCH) { 16300 dtrace_probekey(&desc, &pkey); 16301 pkey.dtpk_id = DTRACE_IDNONE; 16302 } 16303 16304 dtrace_cred2priv(cr, &priv, &uid, &zoneid); 16305 16306 mutex_enter(&dtrace_lock); 16307 16308 if (cmd == DTRACEIOC_PROBEMATCH) { 16309 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) { 16310 if ((probe = dtrace_probes[i - 1]) != NULL && 16311 (m = dtrace_match_probe(probe, &pkey, 16312 priv, uid, zoneid)) != 0) 16313 break; 16314 } 16315 16316 if (m < 0) { 16317 mutex_exit(&dtrace_lock); 16318 return (EINVAL); 16319 } 16320 16321 } else { 16322 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) { 16323 if ((probe = dtrace_probes[i - 1]) != NULL && 16324 dtrace_match_priv(probe, priv, uid, zoneid)) 16325 break; 16326 } 16327 } 16328 16329 if (probe == NULL) { 16330 mutex_exit(&dtrace_lock); 16331 return (ESRCH); 16332 } 16333 16334 dtrace_probe_description(probe, &desc); 16335 mutex_exit(&dtrace_lock); 16336 16337 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0) 16338 return (EFAULT); 16339 16340 return (0); 16341 } 16342 16343 case DTRACEIOC_PROBEARG: { 16344 dtrace_argdesc_t desc; 16345 dtrace_probe_t *probe; 16346 dtrace_provider_t *prov; 16347 16348 if (copyin((void *)arg, &desc, sizeof (desc)) != 0) 16349 return (EFAULT); 16350 16351 if (desc.dtargd_id == DTRACE_IDNONE) 16352 return (EINVAL); 16353 16354 if (desc.dtargd_ndx == DTRACE_ARGNONE) 16355 return (EINVAL); 16356 16357 mutex_enter(&dtrace_provider_lock); 16358 mutex_enter(&mod_lock); 16359 mutex_enter(&dtrace_lock); 16360 16361 if (desc.dtargd_id > dtrace_nprobes) { 16362 mutex_exit(&dtrace_lock); 16363 mutex_exit(&mod_lock); 16364 mutex_exit(&dtrace_provider_lock); 16365 return (EINVAL); 16366 } 16367 16368 if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) { 16369 mutex_exit(&dtrace_lock); 16370 mutex_exit(&mod_lock); 16371 mutex_exit(&dtrace_provider_lock); 16372 return (EINVAL); 16373 } 16374 16375 mutex_exit(&dtrace_lock); 16376 16377 prov = probe->dtpr_provider; 16378 16379 if (prov->dtpv_pops.dtps_getargdesc == NULL) { 16380 /* 16381 * There isn't any typed information for this probe. 16382 * Set the argument number to DTRACE_ARGNONE. 16383 */ 16384 desc.dtargd_ndx = DTRACE_ARGNONE; 16385 } else { 16386 desc.dtargd_native[0] = '\0'; 16387 desc.dtargd_xlate[0] = '\0'; 16388 desc.dtargd_mapping = desc.dtargd_ndx; 16389 16390 prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg, 16391 probe->dtpr_id, probe->dtpr_arg, &desc); 16392 } 16393 16394 mutex_exit(&mod_lock); 16395 mutex_exit(&dtrace_provider_lock); 16396 16397 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0) 16398 return (EFAULT); 16399 16400 return (0); 16401 } 16402 16403 case DTRACEIOC_GO: { 16404 processorid_t cpuid; 16405 rval = dtrace_state_go(state, &cpuid); 16406 16407 if (rval != 0) 16408 return (rval); 16409 16410 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0) 16411 return (EFAULT); 16412 16413 return (0); 16414 } 16415 16416 case DTRACEIOC_STOP: { 16417 processorid_t cpuid; 16418 16419 mutex_enter(&dtrace_lock); 16420 rval = dtrace_state_stop(state, &cpuid); 16421 mutex_exit(&dtrace_lock); 16422 16423 if (rval != 0) 16424 return (rval); 16425 16426 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0) 16427 return (EFAULT); 16428 16429 return (0); 16430 } 16431 16432 case DTRACEIOC_DOFGET: { 16433 dof_hdr_t hdr, *dof; 16434 uint64_t len; 16435 16436 if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0) 16437 return (EFAULT); 16438 16439 mutex_enter(&dtrace_lock); 16440 dof = dtrace_dof_create(state); 16441 mutex_exit(&dtrace_lock); 16442 16443 len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz); 16444 rval = copyout(dof, (void *)arg, len); 16445 dtrace_dof_destroy(dof); 16446 16447 return (rval == 0 ? 0 : EFAULT); 16448 } 16449 16450 case DTRACEIOC_AGGSNAP: 16451 case DTRACEIOC_BUFSNAP: { 16452 dtrace_bufdesc_t desc; 16453 caddr_t cached; 16454 dtrace_buffer_t *buf; 16455 16456 if (copyin((void *)arg, &desc, sizeof (desc)) != 0) 16457 return (EFAULT); 16458 16459 if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU) 16460 return (EINVAL); 16461 16462 mutex_enter(&dtrace_lock); 16463 16464 if (cmd == DTRACEIOC_BUFSNAP) { 16465 buf = &state->dts_buffer[desc.dtbd_cpu]; 16466 } else { 16467 buf = &state->dts_aggbuffer[desc.dtbd_cpu]; 16468 } 16469 16470 if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) { 16471 size_t sz = buf->dtb_offset; 16472 16473 if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) { 16474 mutex_exit(&dtrace_lock); 16475 return (EBUSY); 16476 } 16477 16478 /* 16479 * If this buffer has already been consumed, we're 16480 * going to indicate that there's nothing left here 16481 * to consume. 16482 */ 16483 if (buf->dtb_flags & DTRACEBUF_CONSUMED) { 16484 mutex_exit(&dtrace_lock); 16485 16486 desc.dtbd_size = 0; 16487 desc.dtbd_drops = 0; 16488 desc.dtbd_errors = 0; 16489 desc.dtbd_oldest = 0; 16490 sz = sizeof (desc); 16491 16492 if (copyout(&desc, (void *)arg, sz) != 0) 16493 return (EFAULT); 16494 16495 return (0); 16496 } 16497 16498 /* 16499 * If this is a ring buffer that has wrapped, we want 16500 * to copy the whole thing out. 16501 */ 16502 if (buf->dtb_flags & DTRACEBUF_WRAPPED) { 16503 dtrace_buffer_polish(buf); 16504 sz = buf->dtb_size; 16505 } 16506 16507 if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) { 16508 mutex_exit(&dtrace_lock); 16509 return (EFAULT); 16510 } 16511 16512 desc.dtbd_size = sz; 16513 desc.dtbd_drops = buf->dtb_drops; 16514 desc.dtbd_errors = buf->dtb_errors; 16515 desc.dtbd_oldest = buf->dtb_xamot_offset; 16516 desc.dtbd_timestamp = dtrace_gethrtime(); 16517 16518 mutex_exit(&dtrace_lock); 16519 16520 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0) 16521 return (EFAULT); 16522 16523 buf->dtb_flags |= DTRACEBUF_CONSUMED; 16524 16525 return (0); 16526 } 16527 16528 if (buf->dtb_tomax == NULL) { 16529 ASSERT(buf->dtb_xamot == NULL); 16530 mutex_exit(&dtrace_lock); 16531 return (ENOENT); 16532 } 16533 16534 cached = buf->dtb_tomax; 16535 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH)); 16536 16537 dtrace_xcall(desc.dtbd_cpu, 16538 (dtrace_xcall_t)dtrace_buffer_switch, buf); 16539 16540 state->dts_errors += buf->dtb_xamot_errors; 16541 16542 /* 16543 * If the buffers did not actually switch, then the cross call 16544 * did not take place -- presumably because the given CPU is 16545 * not in the ready set. If this is the case, we'll return 16546 * ENOENT. 16547 */ 16548 if (buf->dtb_tomax == cached) { 16549 ASSERT(buf->dtb_xamot != cached); 16550 mutex_exit(&dtrace_lock); 16551 return (ENOENT); 16552 } 16553 16554 ASSERT(cached == buf->dtb_xamot); 16555 16556 /* 16557 * We have our snapshot; now copy it out. 16558 */ 16559 if (copyout(buf->dtb_xamot, desc.dtbd_data, 16560 buf->dtb_xamot_offset) != 0) { 16561 mutex_exit(&dtrace_lock); 16562 return (EFAULT); 16563 } 16564 16565 desc.dtbd_size = buf->dtb_xamot_offset; 16566 desc.dtbd_drops = buf->dtb_xamot_drops; 16567 desc.dtbd_errors = buf->dtb_xamot_errors; 16568 desc.dtbd_oldest = 0; 16569 desc.dtbd_timestamp = buf->dtb_switched; 16570 16571 mutex_exit(&dtrace_lock); 16572 16573 /* 16574 * Finally, copy out the buffer description. 16575 */ 16576 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0) 16577 return (EFAULT); 16578 16579 return (0); 16580 } 16581 16582 case DTRACEIOC_CONF: { 16583 dtrace_conf_t conf; 16584 16585 bzero(&conf, sizeof (conf)); 16586 conf.dtc_difversion = DIF_VERSION; 16587 conf.dtc_difintregs = DIF_DIR_NREGS; 16588 conf.dtc_diftupregs = DIF_DTR_NREGS; 16589 conf.dtc_ctfmodel = CTF_MODEL_NATIVE; 16590 16591 if (copyout(&conf, (void *)arg, sizeof (conf)) != 0) 16592 return (EFAULT); 16593 16594 return (0); 16595 } 16596 16597 case DTRACEIOC_STATUS: { 16598 dtrace_status_t stat; 16599 dtrace_dstate_t *dstate; 16600 int i, j; 16601 uint64_t nerrs; 16602 16603 /* 16604 * See the comment in dtrace_state_deadman() for the reason 16605 * for setting dts_laststatus to INT64_MAX before setting 16606 * it to the correct value. 16607 */ 16608 state->dts_laststatus = INT64_MAX; 16609 dtrace_membar_producer(); 16610 state->dts_laststatus = dtrace_gethrtime(); 16611 16612 bzero(&stat, sizeof (stat)); 16613 16614 mutex_enter(&dtrace_lock); 16615 16616 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) { 16617 mutex_exit(&dtrace_lock); 16618 return (ENOENT); 16619 } 16620 16621 if (state->dts_activity == DTRACE_ACTIVITY_DRAINING) 16622 stat.dtst_exiting = 1; 16623 16624 nerrs = state->dts_errors; 16625 dstate = &state->dts_vstate.dtvs_dynvars; 16626 16627 for (i = 0; i < NCPU; i++) { 16628 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i]; 16629 16630 stat.dtst_dyndrops += dcpu->dtdsc_drops; 16631 stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops; 16632 stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops; 16633 16634 if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL) 16635 stat.dtst_filled++; 16636 16637 nerrs += state->dts_buffer[i].dtb_errors; 16638 16639 for (j = 0; j < state->dts_nspeculations; j++) { 16640 dtrace_speculation_t *spec; 16641 dtrace_buffer_t *buf; 16642 16643 spec = &state->dts_speculations[j]; 16644 buf = &spec->dtsp_buffer[i]; 16645 stat.dtst_specdrops += buf->dtb_xamot_drops; 16646 } 16647 } 16648 16649 stat.dtst_specdrops_busy = state->dts_speculations_busy; 16650 stat.dtst_specdrops_unavail = state->dts_speculations_unavail; 16651 stat.dtst_stkstroverflows = state->dts_stkstroverflows; 16652 stat.dtst_dblerrors = state->dts_dblerrors; 16653 stat.dtst_killed = 16654 (state->dts_activity == DTRACE_ACTIVITY_KILLED); 16655 stat.dtst_errors = nerrs; 16656 16657 mutex_exit(&dtrace_lock); 16658 16659 if (copyout(&stat, (void *)arg, sizeof (stat)) != 0) 16660 return (EFAULT); 16661 16662 return (0); 16663 } 16664 16665 case DTRACEIOC_FORMAT: { 16666 dtrace_fmtdesc_t fmt; 16667 char *str; 16668 int len; 16669 16670 if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0) 16671 return (EFAULT); 16672 16673 mutex_enter(&dtrace_lock); 16674 16675 if (fmt.dtfd_format == 0 || 16676 fmt.dtfd_format > state->dts_nformats) { 16677 mutex_exit(&dtrace_lock); 16678 return (EINVAL); 16679 } 16680 16681 /* 16682 * Format strings are allocated contiguously and they are 16683 * never freed; if a format index is less than the number 16684 * of formats, we can assert that the format map is non-NULL 16685 * and that the format for the specified index is non-NULL. 16686 */ 16687 ASSERT(state->dts_formats != NULL); 16688 str = state->dts_formats[fmt.dtfd_format - 1]; 16689 ASSERT(str != NULL); 16690 16691 len = strlen(str) + 1; 16692 16693 if (len > fmt.dtfd_length) { 16694 fmt.dtfd_length = len; 16695 16696 if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) { 16697 mutex_exit(&dtrace_lock); 16698 return (EINVAL); 16699 } 16700 } else { 16701 if (copyout(str, fmt.dtfd_string, len) != 0) { 16702 mutex_exit(&dtrace_lock); 16703 return (EINVAL); 16704 } 16705 } 16706 16707 mutex_exit(&dtrace_lock); 16708 return (0); 16709 } 16710 16711 default: 16712 break; 16713 } 16714 16715 return (ENOTTY); 16716} 16717 16718/*ARGSUSED*/ 16719static int 16720dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd) 16721{ 16722 dtrace_state_t *state; 16723 16724 switch (cmd) { 16725 case DDI_DETACH: 16726 break; 16727 16728 case DDI_SUSPEND: 16729 return (DDI_SUCCESS); 16730 16731 default: 16732 return (DDI_FAILURE); 16733 } 16734 16735 mutex_enter(&cpu_lock); 16736 mutex_enter(&dtrace_provider_lock); 16737 mutex_enter(&dtrace_lock); 16738 16739 ASSERT(dtrace_opens == 0); 16740 16741 if (dtrace_helpers > 0) { 16742 mutex_exit(&dtrace_provider_lock); 16743 mutex_exit(&dtrace_lock); 16744 mutex_exit(&cpu_lock); 16745 return (DDI_FAILURE); 16746 } 16747 16748 if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) { 16749 mutex_exit(&dtrace_provider_lock); 16750 mutex_exit(&dtrace_lock); 16751 mutex_exit(&cpu_lock); 16752 return (DDI_FAILURE); 16753 } 16754 16755 dtrace_provider = NULL; 16756 16757 if ((state = dtrace_anon_grab()) != NULL) { 16758 /* 16759 * If there were ECBs on this state, the provider should 16760 * have not been allowed to detach; assert that there is 16761 * none. 16762 */ 16763 ASSERT(state->dts_necbs == 0); 16764 dtrace_state_destroy(state); 16765 16766 /* 16767 * If we're being detached with anonymous state, we need to 16768 * indicate to the kernel debugger that DTrace is now inactive. 16769 */ 16770 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE); 16771 } 16772 16773 bzero(&dtrace_anon, sizeof (dtrace_anon_t)); 16774 unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL); 16775 dtrace_cpu_init = NULL; 16776 dtrace_helpers_cleanup = NULL; 16777 dtrace_helpers_fork = NULL; 16778 dtrace_cpustart_init = NULL; 16779 dtrace_cpustart_fini = NULL; 16780 dtrace_debugger_init = NULL; 16781 dtrace_debugger_fini = NULL; 16782 dtrace_modload = NULL; 16783 dtrace_modunload = NULL; 16784 16785 mutex_exit(&cpu_lock); 16786 16787 if (dtrace_helptrace_enabled) { 16788 kmem_free(dtrace_helptrace_buffer, dtrace_helptrace_bufsize); 16789 dtrace_helptrace_buffer = NULL; 16790 } 16791 16792 kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *)); 16793 dtrace_probes = NULL; 16794 dtrace_nprobes = 0; 16795 16796 dtrace_hash_destroy(dtrace_bymod); 16797 dtrace_hash_destroy(dtrace_byfunc); 16798 dtrace_hash_destroy(dtrace_byname); 16799 dtrace_bymod = NULL; 16800 dtrace_byfunc = NULL; 16801 dtrace_byname = NULL; 16802 16803 kmem_cache_destroy(dtrace_state_cache); 16804 vmem_destroy(dtrace_minor); 16805 vmem_destroy(dtrace_arena); 16806 16807 if (dtrace_toxrange != NULL) { 16808 kmem_free(dtrace_toxrange, 16809 dtrace_toxranges_max * sizeof (dtrace_toxrange_t)); 16810 dtrace_toxrange = NULL; 16811 dtrace_toxranges = 0; 16812 dtrace_toxranges_max = 0; 16813 } 16814 16815 ddi_remove_minor_node(dtrace_devi, NULL); 16816 dtrace_devi = NULL; 16817 16818 ddi_soft_state_fini(&dtrace_softstate); 16819 16820 ASSERT(dtrace_vtime_references == 0); 16821 ASSERT(dtrace_opens == 0); 16822 ASSERT(dtrace_retained == NULL); 16823 16824 mutex_exit(&dtrace_lock); 16825 mutex_exit(&dtrace_provider_lock); 16826 16827 /* 16828 * We don't destroy the task queue until after we have dropped our 16829 * locks (taskq_destroy() may block on running tasks). To prevent 16830 * attempting to do work after we have effectively detached but before 16831 * the task queue has been destroyed, all tasks dispatched via the 16832 * task queue must check that DTrace is still attached before 16833 * performing any operation. 16834 */ 16835 taskq_destroy(dtrace_taskq); 16836 dtrace_taskq = NULL; 16837 16838 return (DDI_SUCCESS); 16839} 16840#endif 16841 16842#if defined(sun) 16843/*ARGSUSED*/ 16844static int 16845dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result) 16846{ 16847 int error; 16848 16849 switch (infocmd) { 16850 case DDI_INFO_DEVT2DEVINFO: 16851 *result = (void *)dtrace_devi; 16852 error = DDI_SUCCESS; 16853 break; 16854 case DDI_INFO_DEVT2INSTANCE: 16855 *result = (void *)0; 16856 error = DDI_SUCCESS; 16857 break; 16858 default: 16859 error = DDI_FAILURE; 16860 } 16861 return (error); 16862} 16863#endif 16864 16865#if defined(sun) 16866static struct cb_ops dtrace_cb_ops = { 16867 dtrace_open, /* open */ 16868 dtrace_close, /* close */ 16869 nulldev, /* strategy */ 16870 nulldev, /* print */ 16871 nodev, /* dump */ 16872 nodev, /* read */ 16873 nodev, /* write */ 16874 dtrace_ioctl, /* ioctl */ 16875 nodev, /* devmap */ 16876 nodev, /* mmap */ 16877 nodev, /* segmap */ 16878 nochpoll, /* poll */ 16879 ddi_prop_op, /* cb_prop_op */ 16880 0, /* streamtab */ 16881 D_NEW | D_MP /* Driver compatibility flag */ 16882}; 16883 16884static struct dev_ops dtrace_ops = { 16885 DEVO_REV, /* devo_rev */ 16886 0, /* refcnt */ 16887 dtrace_info, /* get_dev_info */ 16888 nulldev, /* identify */ 16889 nulldev, /* probe */ 16890 dtrace_attach, /* attach */ 16891 dtrace_detach, /* detach */ 16892 nodev, /* reset */ 16893 &dtrace_cb_ops, /* driver operations */ 16894 NULL, /* bus operations */ 16895 nodev /* dev power */ 16896}; 16897 16898static struct modldrv modldrv = { 16899 &mod_driverops, /* module type (this is a pseudo driver) */ 16900 "Dynamic Tracing", /* name of module */ 16901 &dtrace_ops, /* driver ops */ 16902}; 16903 16904static struct modlinkage modlinkage = { 16905 MODREV_1, 16906 (void *)&modldrv, 16907 NULL 16908}; 16909 16910int 16911_init(void) 16912{ 16913 return (mod_install(&modlinkage)); 16914} 16915 16916int 16917_info(struct modinfo *modinfop) 16918{ 16919 return (mod_info(&modlinkage, modinfop)); 16920} 16921 16922int 16923_fini(void) 16924{ 16925 return (mod_remove(&modlinkage)); 16926} 16927#else 16928 16929static d_ioctl_t dtrace_ioctl; 16930static d_ioctl_t dtrace_ioctl_helper; 16931static void dtrace_load(void *); 16932static int dtrace_unload(void); 16933#if __FreeBSD_version < 800039 16934static void dtrace_clone(void *, struct ucred *, char *, int , struct cdev **); 16935static struct clonedevs *dtrace_clones; /* Ptr to the array of cloned devices. */ 16936static eventhandler_tag eh_tag; /* Event handler tag. */ 16937#else 16938static struct cdev *dtrace_dev; 16939static struct cdev *helper_dev; 16940#endif 16941 16942void dtrace_invop_init(void); 16943void dtrace_invop_uninit(void); 16944 16945static struct cdevsw dtrace_cdevsw = { 16946 .d_version = D_VERSION, 16947#if __FreeBSD_version < 800039 16948 .d_flags = D_TRACKCLOSE | D_NEEDMINOR, 16949 .d_close = dtrace_close, 16950#endif 16951 .d_ioctl = dtrace_ioctl, 16952 .d_open = dtrace_open, 16953 .d_name = "dtrace", 16954}; 16955 16956static struct cdevsw helper_cdevsw = { 16957 .d_version = D_VERSION, 16958 .d_ioctl = dtrace_ioctl_helper, 16959 .d_name = "helper", 16960}; 16961 16962#include <dtrace_anon.c> 16963#if __FreeBSD_version < 800039 16964#include <dtrace_clone.c> 16965#endif 16966#include <dtrace_ioctl.c> 16967#include <dtrace_load.c> 16968#include <dtrace_modevent.c> 16969#include <dtrace_sysctl.c> 16970#include <dtrace_unload.c> 16971#include <dtrace_vtime.c> 16972#include <dtrace_hacks.c> 16973#include <dtrace_isa.c> 16974 16975SYSINIT(dtrace_load, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_load, NULL); 16976SYSUNINIT(dtrace_unload, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_unload, NULL); 16977SYSINIT(dtrace_anon_init, SI_SUB_DTRACE_ANON, SI_ORDER_FIRST, dtrace_anon_init, NULL); 16978 16979DEV_MODULE(dtrace, dtrace_modevent, NULL); 16980MODULE_VERSION(dtrace, 1); 16981MODULE_DEPEND(dtrace, cyclic, 1, 1, 1); 16982MODULE_DEPEND(dtrace, opensolaris, 1, 1, 1); 16983#endif 16984