1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 22 /* 23 * Copyright 2007 Sun Microsystems, Inc. All rights reserved. 24 * Use is subject to license terms. 25 * 26 * Portions Copyright (c) 2012 by Delphix. All rights reserved. 27 * Portions Copyright (c) 2016 by Joyent, Inc. 28 */ 29 30 #ifndef _SYS_DTRACE_IMPL_H 31 #define _SYS_DTRACE_IMPL_H 32 33 #ifdef __cplusplus 34 extern "C" { 35 #endif 36 37 /* 38 * DTrace Dynamic Tracing Software: Kernel Implementation Interfaces 39 * 40 * Note: The contents of this file are private to the implementation of the 41 * Solaris system and DTrace subsystem and are subject to change at any time 42 * without notice. Applications and drivers using these interfaces will fail 43 * to run on future releases. These interfaces should not be used for any 44 * purpose except those expressly outlined in dtrace(7D) and libdtrace(3LIB). 45 * Please refer to the "Solaris Dynamic Tracing Guide" for more information. 46 */ 47 48 #include <sys/dtrace.h> 49 50 /* 51 * DTrace Implementation Locks 52 */ 53 extern lck_attr_t dtrace_lck_attr; 54 extern lck_grp_t dtrace_lck_grp; 55 extern lck_mtx_t dtrace_procwaitfor_lock; 56 57 /* 58 * DTrace Implementation Constants and Typedefs 59 */ 60 #define DTRACE_MAXPROPLEN 128 61 #define DTRACE_DYNVAR_CHUNKSIZE 256 62 63 struct dtrace_probe; 64 struct dtrace_ecb; 65 struct dtrace_predicate; 66 struct dtrace_action; 67 struct dtrace_provider; 68 struct dtrace_state; 69 70 typedef struct dtrace_probe dtrace_probe_t; 71 typedef struct dtrace_ecb dtrace_ecb_t; 72 typedef struct dtrace_predicate dtrace_predicate_t; 73 typedef struct dtrace_action dtrace_action_t; 74 typedef struct dtrace_provider dtrace_provider_t; 75 typedef struct dtrace_meta dtrace_meta_t; 76 typedef struct dtrace_state dtrace_state_t; 77 typedef uint32_t dtrace_optid_t; 78 typedef uint32_t dtrace_specid_t; 79 typedef uint64_t dtrace_genid_t; 80 81 /* 82 * DTrace Probes 83 * 84 * The probe is the fundamental unit of the DTrace architecture. Probes are 85 * created by DTrace providers, and managed by the DTrace framework. A probe 86 * is identified by a unique <provider, module, function, name> tuple, and has 87 * a unique probe identifier assigned to it. (Some probes are not associated 88 * with a specific point in text; these are called _unanchored probes_ and have 89 * no module or function associated with them.) Probes are represented as a 90 * dtrace_probe structure. To allow quick lookups based on each element of the 91 * probe tuple, probes are hashed by each of provider, module, function and 92 * name. (If a lookup is performed based on a regular expression, a 93 * dtrace_probekey is prepared, and a linear search is performed.) Each probe 94 * is additionally pointed to by a linear array indexed by its identifier. The 95 * identifier is the provider's mechanism for indicating to the DTrace 96 * framework that a probe has fired: the identifier is passed as the first 97 * argument to dtrace_probe(), where it is then mapped into the corresponding 98 * dtrace_probe structure. From the dtrace_probe structure, dtrace_probe() can 99 * iterate over the probe's list of enabling control blocks; see "DTrace 100 * Enabling Control Blocks", below.) 101 */ 102 struct dtrace_probe { 103 dtrace_id_t dtpr_id; /* probe identifier */ 104 dtrace_ecb_t *dtpr_ecb; /* ECB list; see below */ 105 dtrace_ecb_t *dtpr_ecb_last; /* last ECB in list */ 106 void *dtpr_arg; /* provider argument */ 107 dtrace_cacheid_t dtpr_predcache; /* predicate cache ID */ 108 int dtpr_aframes; /* artificial frames */ 109 dtrace_provider_t *dtpr_provider; /* pointer to provider */ 110 char *dtpr_mod; /* probe's module name */ 111 char *dtpr_func; /* probe's function name */ 112 char *dtpr_name; /* probe's name */ 113 dtrace_probe_t *dtpr_nextprov; /* next in provider hash */ 114 dtrace_probe_t *dtpr_prevprov; /* previous in provider hash */ 115 dtrace_probe_t *dtpr_nextmod; /* next in module hash */ 116 dtrace_probe_t *dtpr_prevmod; /* previous in module hash */ 117 dtrace_probe_t *dtpr_nextfunc; /* next in function hash */ 118 dtrace_probe_t *dtpr_prevfunc; /* previous in function hash */ 119 dtrace_probe_t *dtpr_nextname; /* next in name hash */ 120 dtrace_probe_t *dtpr_prevname; /* previous in name hash */ 121 dtrace_genid_t dtpr_gen; /* probe generation ID */ 122 }; 123 124 typedef int dtrace_probekey_f(const char *, const char *, int); 125 126 typedef struct dtrace_probekey { 127 const char *dtpk_prov; /* provider name to match */ 128 dtrace_probekey_f *dtpk_pmatch; /* provider matching function */ 129 const char *dtpk_mod; /* module name to match */ 130 dtrace_probekey_f *dtpk_mmatch; /* module matching function */ 131 const char *dtpk_func; /* func name to match */ 132 dtrace_probekey_f *dtpk_fmatch; /* func matching function */ 133 const char *dtpk_name; /* name to match */ 134 dtrace_probekey_f *dtpk_nmatch; /* name matching function */ 135 dtrace_id_t dtpk_id; /* identifier to match */ 136 } dtrace_probekey_t; 137 138 typedef struct dtrace_hashbucket { 139 struct dtrace_hashbucket *dthb_next; /* next on hash chain */ 140 void *dthb_chain; /* chain of elements */ 141 int dthb_len; /* number of probes here */ 142 } dtrace_hashbucket_t; 143 144 typedef const char* dtrace_strkey_f(void*, uintptr_t); 145 146 typedef struct dtrace_hash { 147 dtrace_hashbucket_t **dth_tab; /* hash table */ 148 int dth_size; /* size of hash table */ 149 int dth_mask; /* mask to index into table */ 150 int dth_nbuckets; /* total number of buckets */ 151 uintptr_t dth_nextoffs; /* offset of next in element */ 152 uintptr_t dth_prevoffs; /* offset of prev in element */ 153 dtrace_strkey_f *dth_getstr; /* func to retrieve str in element */ 154 uintptr_t dth_stroffs; /* offset of str in element */ 155 } dtrace_hash_t; 156 157 /* 158 * DTrace Enabling Control Blocks 159 * 160 * When a provider wishes to fire a probe, it calls into dtrace_probe(), 161 * passing the probe identifier as the first argument. As described above, 162 * dtrace_probe() maps the identifier into a pointer to a dtrace_probe_t 163 * structure. This structure contains information about the probe, and a 164 * pointer to the list of Enabling Control Blocks (ECBs). Each ECB points to 165 * DTrace consumer state, and contains an optional predicate, and a list of 166 * actions. (Shown schematically below.) The ECB abstraction allows a single 167 * probe to be multiplexed across disjoint consumers, or across disjoint 168 * enablings of a single probe within one consumer. 169 * 170 * Enabling Control Block 171 * dtrace_ecb_t 172 * +------------------------+ 173 * | dtrace_epid_t ---------+--------------> Enabled Probe ID (EPID) 174 * | dtrace_state_t * ------+--------------> State associated with this ECB 175 * | dtrace_predicate_t * --+---------+ 176 * | dtrace_action_t * -----+----+ | 177 * | dtrace_ecb_t * ---+ | | | Predicate (if any) 178 * +-------------------+----+ | | dtrace_predicate_t 179 * | | +---> +--------------------+ 180 * | | | dtrace_difo_t * ---+----> DIFO 181 * | | +--------------------+ 182 * | | 183 * Next ECB | | Action 184 * (if any) | | dtrace_action_t 185 * : +--> +-------------------+ 186 * : | dtrace_actkind_t -+------> kind 187 * v | dtrace_difo_t * --+------> DIFO (if any) 188 * | dtrace_recdesc_t -+------> record descr. 189 * | dtrace_action_t * +------+ 190 * +-------------------+ | 191 * | Next action 192 * +-------------------------------+ (if any) 193 * | 194 * | Action 195 * | dtrace_action_t 196 * +--> +-------------------+ 197 * | dtrace_actkind_t -+------> kind 198 * | dtrace_difo_t * --+------> DIFO (if any) 199 * | dtrace_action_t * +------+ 200 * +-------------------+ | 201 * | Next action 202 * +-------------------------------+ (if any) 203 * | 204 * : 205 * v 206 * 207 * 208 * dtrace_probe() iterates over the ECB list. If the ECB needs less space 209 * than is available in the principal buffer, the ECB is processed: if the 210 * predicate is non-NULL, the DIF object is executed. If the result is 211 * non-zero, the action list is processed, with each action being executed 212 * accordingly. When the action list has been completely executed, processing 213 * advances to the next ECB. The ECB abstraction allows disjoint consumers 214 * to multiplex on single probes. 215 * 216 * Execution of the ECB results in consuming dte_size bytes in the buffer 217 * to record data. During execution, dte_needed bytes must be available in 218 * the buffer. This space is used for both recorded data and tuple data. 219 */ 220 struct dtrace_ecb { 221 dtrace_epid_t dte_epid; /* enabled probe ID */ 222 uint32_t dte_alignment; /* required alignment */ 223 size_t dte_needed; /* space needed for execution */ 224 size_t dte_size; /* size of recorded payload */ 225 dtrace_predicate_t *dte_predicate; /* predicate, if any */ 226 dtrace_action_t *dte_action; /* actions, if any */ 227 dtrace_ecb_t *dte_next; /* next ECB on probe */ 228 dtrace_state_t *dte_state; /* pointer to state */ 229 uint32_t dte_cond; /* security condition */ 230 dtrace_probe_t *dte_probe; /* pointer to probe */ 231 dtrace_action_t *dte_action_last; /* last action on ECB */ 232 uint64_t dte_uarg; /* library argument */ 233 }; 234 235 struct dtrace_predicate { 236 dtrace_difo_t *dtp_difo; /* DIF object */ 237 dtrace_cacheid_t dtp_cacheid; /* cache identifier */ 238 int dtp_refcnt; /* reference count */ 239 }; 240 241 struct dtrace_action { 242 dtrace_actkind_t dta_kind; /* kind of action */ 243 uint16_t dta_intuple; /* boolean: in aggregation */ 244 uint32_t dta_refcnt; /* reference count */ 245 dtrace_difo_t *dta_difo; /* pointer to DIFO */ 246 dtrace_recdesc_t dta_rec; /* record description */ 247 dtrace_action_t *dta_prev; /* previous action */ 248 dtrace_action_t *dta_next; /* next action */ 249 }; 250 251 typedef struct dtrace_aggregation { 252 dtrace_action_t dtag_action; /* action; must be first */ 253 dtrace_aggid_t dtag_id; /* identifier */ 254 dtrace_ecb_t *dtag_ecb; /* corresponding ECB */ 255 dtrace_action_t *dtag_first; /* first action in tuple */ 256 uint32_t dtag_base; /* base of aggregation */ 257 uint8_t dtag_hasarg; /* boolean: has argument */ 258 uint64_t dtag_initial; /* initial value */ 259 void (*dtag_aggregate)(uint64_t *, uint64_t, uint64_t); 260 } dtrace_aggregation_t; 261 262 /* 263 * DTrace Buffers 264 * 265 * Principal buffers, aggregation buffers, and speculative buffers are all 266 * managed with the dtrace_buffer structure. By default, this structure 267 * includes twin data buffers -- dtb_tomax and dtb_xamot -- that serve as the 268 * active and passive buffers, respectively. For speculative buffers, 269 * dtb_xamot will be NULL; for "ring" and "fill" buffers, dtb_xamot will point 270 * to a scratch buffer. For all buffer types, the dtrace_buffer structure is 271 * always allocated on a per-CPU basis; a single dtrace_buffer structure is 272 * never shared among CPUs. (That is, there is never true sharing of the 273 * dtrace_buffer structure; to prevent false sharing of the structure, it must 274 * always be aligned to the coherence granularity -- generally 64 bytes.) 275 * 276 * One of the critical design decisions of DTrace is that a given ECB always 277 * stores the same quantity and type of data. This is done to assure that the 278 * only metadata required for an ECB's traced data is the EPID. That is, from 279 * the EPID, the consumer can determine the data layout. (The data buffer 280 * layout is shown schematically below.) By assuring that one can determine 281 * data layout from the EPID, the metadata stream can be separated from the 282 * data stream -- simplifying the data stream enormously. The ECB always 283 * proceeds the recorded data as part of the dtrace_rechdr_t structure that 284 * includes the EPID and a high-resolution timestamp used for output ordering 285 * consistency. 286 * 287 * base of data buffer ---> +--------+--------------------+--------+ 288 * | rechdr | data | rechdr | 289 * +--------+------+--------+----+--------+ 290 * | data | rechdr | data | 291 * +---------------+--------+-------------+ 292 * | data, cont. | 293 * +--------+--------------------+--------+ 294 * | rechdr | data | | 295 * +--------+--------------------+ | 296 * | || | 297 * | || | 298 * | \/ | 299 * : : 300 * . . 301 * . . 302 * . . 303 * : : 304 * | | 305 * limit of data buffer ---> +--------------------------------------+ 306 * 307 * When evaluating an ECB, dtrace_probe() determines if the ECB's needs of the 308 * principal buffer (both scratch and payload) exceed the available space. If 309 * the ECB's needs exceed available space (and if the principal buffer policy 310 * is the default "switch" policy), the ECB is dropped, the buffer's drop count 311 * is incremented, and processing advances to the next ECB. If the ECB's needs 312 * can be met with the available space, the ECB is processed, but the offset in 313 * the principal buffer is only advanced if the ECB completes processing 314 * without error. 315 * 316 * When a buffer is to be switched (either because the buffer is the principal 317 * buffer with a "switch" policy or because it is an aggregation buffer), a 318 * cross call is issued to the CPU associated with the buffer. In the cross 319 * call context, interrupts are disabled, and the active and the inactive 320 * buffers are atomically switched. This involves switching the data pointers, 321 * copying the various state fields (offset, drops, errors, etc.) into their 322 * inactive equivalents, and clearing the state fields. Because interrupts are 323 * disabled during this procedure, the switch is guaranteed to appear atomic to 324 * dtrace_probe(). 325 * 326 * DTrace Ring Buffering 327 * 328 * To process a ring buffer correctly, one must know the oldest valid record. 329 * Processing starts at the oldest record in the buffer and continues until 330 * the end of the buffer is reached. Processing then resumes starting with 331 * the record stored at offset 0 in the buffer, and continues until the 332 * youngest record is processed. If trace records are of a fixed-length, 333 * determining the oldest record is trivial: 334 * 335 * - If the ring buffer has not wrapped, the oldest record is the record 336 * stored at offset 0. 337 * 338 * - If the ring buffer has wrapped, the oldest record is the record stored 339 * at the current offset. 340 * 341 * With variable length records, however, just knowing the current offset 342 * doesn't suffice for determining the oldest valid record: assuming that one 343 * allows for arbitrary data, one has no way of searching forward from the 344 * current offset to find the oldest valid record. (That is, one has no way 345 * of separating data from metadata.) It would be possible to simply refuse to 346 * process any data in the ring buffer between the current offset and the 347 * limit, but this leaves (potentially) an enormous amount of otherwise valid 348 * data unprocessed. 349 * 350 * To effect ring buffering, we track two offsets in the buffer: the current 351 * offset and the _wrapped_ offset. If a request is made to reserve some 352 * amount of data, and the buffer has wrapped, the wrapped offset is 353 * incremented until the wrapped offset minus the current offset is greater 354 * than or equal to the reserve request. This is done by repeatedly looking 355 * up the ECB corresponding to the EPID at the current wrapped offset, and 356 * incrementing the wrapped offset by the size of the data payload 357 * corresponding to that ECB. If this offset is greater than or equal to the 358 * limit of the data buffer, the wrapped offset is set to 0. Thus, the 359 * current offset effectively "chases" the wrapped offset around the buffer. 360 * Schematically: 361 * 362 * base of data buffer ---> +------+--------------------+------+ 363 * | EPID | data | EPID | 364 * +------+--------+------+----+------+ 365 * | data | EPID | data | 366 * +---------------+------+-----------+ 367 * | data, cont. | 368 * +------+---------------------------+ 369 * | EPID | data | 370 * current offset ---> +------+---------------------------+ 371 * | invalid data | 372 * wrapped offset ---> +------+--------------------+------+ 373 * | EPID | data | EPID | 374 * +------+--------+------+----+------+ 375 * | data | EPID | data | 376 * +---------------+------+-----------+ 377 * : : 378 * . . 379 * . ... valid data ... . 380 * . . 381 * : : 382 * +------+-------------+------+------+ 383 * | EPID | data | EPID | data | 384 * +------+------------++------+------+ 385 * | data, cont. | leftover | 386 * limit of data buffer ---> +-------------------+--------------+ 387 * 388 * If the amount of requested buffer space exceeds the amount of space 389 * available between the current offset and the end of the buffer: 390 * 391 * (1) all words in the data buffer between the current offset and the limit 392 * of the data buffer (marked "leftover", above) are set to 393 * DTRACE_EPIDNONE 394 * 395 * (2) the wrapped offset is set to zero 396 * 397 * (3) the iteration process described above occurs until the wrapped offset 398 * is greater than the amount of desired space. 399 * 400 * The wrapped offset is implemented by (re-)using the inactive offset. 401 * In a "switch" buffer policy, the inactive offset stores the offset in 402 * the inactive buffer; in a "ring" buffer policy, it stores the wrapped 403 * offset. 404 * 405 * DTrace Scratch Buffering 406 * 407 * Some ECBs may wish to allocate dynamically-sized temporary scratch memory. 408 * To accommodate such requests easily, scratch memory may be allocated in 409 * the buffer beyond the current offset plus the needed memory of the current 410 * ECB. If there isn't sufficient room in the buffer for the requested amount 411 * of scratch space, the allocation fails and an error is generated. Scratch 412 * memory is tracked in the dtrace_mstate_t and is automatically freed when 413 * the ECB ceases processing. Note that ring buffers cannot allocate their 414 * scratch from the principal buffer -- lest they needlessly overwrite older, 415 * valid data. Ring buffers therefore have their own dedicated scratch buffer 416 * from which scratch is allocated. 417 */ 418 #define DTRACEBUF_RING 0x0001 /* bufpolicy set to "ring" */ 419 #define DTRACEBUF_FILL 0x0002 /* bufpolicy set to "fill" */ 420 #define DTRACEBUF_NOSWITCH 0x0004 /* do not switch buffer */ 421 #define DTRACEBUF_WRAPPED 0x0008 /* ring buffer has wrapped */ 422 #define DTRACEBUF_DROPPED 0x0010 /* drops occurred */ 423 #define DTRACEBUF_ERROR 0x0020 /* errors occurred */ 424 #define DTRACEBUF_FULL 0x0040 /* "fill" buffer is full */ 425 #define DTRACEBUF_CONSUMED 0x0080 /* buffer has been consumed */ 426 #define DTRACEBUF_INACTIVE 0x0100 /* buffer is not yet active */ 427 428 typedef struct dtrace_buffer { 429 uint64_t dtb_offset; /* current offset in buffer */ 430 uint64_t dtb_cur_limit; /* current limit before signaling/dropping */ 431 uint64_t dtb_limit; /* limit before signaling */ 432 uint64_t dtb_size; /* size of buffer */ 433 uint32_t dtb_flags; /* flags */ 434 uint32_t dtb_drops; /* number of drops */ 435 caddr_t dtb_tomax; /* active buffer */ 436 caddr_t dtb_xamot; /* inactive buffer */ 437 uint32_t dtb_xamot_flags; /* inactive flags */ 438 uint32_t dtb_xamot_drops; /* drops in inactive buffer */ 439 uint64_t dtb_xamot_offset; /* offset in inactive buffer */ 440 uint32_t dtb_errors; /* number of errors */ 441 uint32_t dtb_xamot_errors; /* errors in inactive buffer */ 442 #ifndef _LP64 443 uint64_t dtb_pad1; 444 #endif 445 uint64_t dtb_switched; /* time of last switch */ 446 uint64_t dtb_interval; /* observed switch interval */ 447 uint64_t dtb_pad2[4]; /* pad to avoid false sharing */ 448 } dtrace_buffer_t; 449 450 /* 451 * DTrace Aggregation Buffers 452 * 453 * Aggregation buffers use much of the same mechanism as described above 454 * ("DTrace Buffers"). However, because an aggregation is fundamentally a 455 * hash, there exists dynamic metadata associated with an aggregation buffer 456 * that is not associated with other kinds of buffers. This aggregation 457 * metadata is _only_ relevant for the in-kernel implementation of 458 * aggregations; it is not actually relevant to user-level consumers. To do 459 * this, we allocate dynamic aggregation data (hash keys and hash buckets) 460 * starting below the _limit_ of the buffer, and we allocate data from the 461 * _base_ of the buffer. When the aggregation buffer is copied out, _only_ the 462 * data is copied out; the metadata is simply discarded. Schematically, 463 * aggregation buffers look like: 464 * 465 * base of data buffer ---> +-------+------+-----------+-------+ 466 * | aggid | key | value | aggid | 467 * +-------+------+-----------+-------+ 468 * | key | 469 * +-------+-------+-----+------------+ 470 * | value | aggid | key | value | 471 * +-------+------++-----+------+-----+ 472 * | aggid | key | value | | 473 * +-------+------+-------------+ | 474 * | || | 475 * | || | 476 * | \/ | 477 * : : 478 * . . 479 * . . 480 * . . 481 * : : 482 * | /\ | 483 * | || +------------+ 484 * | || | | 485 * +---------------------+ | 486 * | hash keys | 487 * | (dtrace_aggkey structures) | 488 * | | 489 * +----------------------------------+ 490 * | hash buckets | 491 * | (dtrace_aggbuffer structure) | 492 * | | 493 * limit of data buffer ---> +----------------------------------+ 494 * 495 * 496 * As implied above, just as we assure that ECBs always store a constant 497 * amount of data, we assure that a given aggregation -- identified by its 498 * aggregation ID -- always stores data of a constant quantity and type. 499 * As with EPIDs, this allows the aggregation ID to serve as the metadata for a 500 * given record. 501 * 502 * Note that the size of the dtrace_aggkey structure must be sizeof (uintptr_t) 503 * aligned. (If this the structure changes such that this becomes false, an 504 * assertion will fail in dtrace_aggregate().) 505 */ 506 typedef struct dtrace_aggkey { 507 uint32_t dtak_hashval; /* hash value */ 508 uint32_t dtak_action:4; /* action -- 4 bits */ 509 uint32_t dtak_size:28; /* size -- 28 bits */ 510 caddr_t dtak_data; /* data pointer */ 511 struct dtrace_aggkey *dtak_next; /* next in hash chain */ 512 } dtrace_aggkey_t; 513 514 typedef struct dtrace_aggbuffer { 515 uintptr_t dtagb_hashsize; /* number of buckets */ 516 uintptr_t dtagb_free; /* free list of keys */ 517 dtrace_aggkey_t **dtagb_hash; /* hash table */ 518 } dtrace_aggbuffer_t; 519 520 /* 521 * DTrace Speculations 522 * 523 * Speculations have a per-CPU buffer and a global state. Once a speculation 524 * buffer has been comitted or discarded, it cannot be reused until all CPUs 525 * have taken the same action (commit or discard) on their respective 526 * speculative buffer. However, because DTrace probes may execute in arbitrary 527 * context, other CPUs cannot simply be cross-called at probe firing time to 528 * perform the necessary commit or discard. The speculation states thus 529 * optimize for the case that a speculative buffer is only active on one CPU at 530 * the time of a commit() or discard() -- for if this is the case, other CPUs 531 * need not take action, and the speculation is immediately available for 532 * reuse. If the speculation is active on multiple CPUs, it must be 533 * asynchronously cleaned -- potentially leading to a higher rate of dirty 534 * speculative drops. The speculation states are as follows: 535 * 536 * DTRACESPEC_INACTIVE <= Initial state; inactive speculation 537 * DTRACESPEC_ACTIVE <= Allocated, but not yet speculatively traced to 538 * DTRACESPEC_ACTIVEONE <= Speculatively traced to on one CPU 539 * DTRACESPEC_ACTIVEMANY <= Speculatively traced to on more than one CPU 540 * DTRACESPEC_COMMITTING <= Currently being commited on one CPU 541 * DTRACESPEC_COMMITTINGMANY <= Currently being commited on many CPUs 542 * DTRACESPEC_DISCARDING <= Currently being discarded on many CPUs 543 * 544 * The state transition diagram is as follows: 545 * 546 * +----------------------------------------------------------+ 547 * | | 548 * | +------------+ | 549 * | +-------------------| COMMITTING |<-----------------+ | 550 * | | +------------+ | | 551 * | | copied spec. ^ commit() on | | discard() on 552 * | | into principal | active CPU | | active CPU 553 * | | | commit() | | 554 * V V | | | 555 * +----------+ +--------+ +-----------+ 556 * | INACTIVE |---------------->| ACTIVE |--------------->| ACTIVEONE | 557 * +----------+ speculation() +--------+ speculate() +-----------+ 558 * ^ ^ | | | 559 * | | | discard() | | 560 * | | asynchronously | discard() on | | speculate() 561 * | | cleaned V inactive CPU | | on inactive 562 * | | +------------+ | | CPU 563 * | +-------------------| DISCARDING |<-----------------+ | 564 * | +------------+ | 565 * | asynchronously ^ | 566 * | copied spec. | discard() | 567 * | into principal +------------------------+ | 568 * | | V 569 * +----------------+ commit() +------------+ 570 * | COMMITTINGMANY |<----------------------------------| ACTIVEMANY | 571 * +----------------+ +------------+ 572 */ 573 typedef enum dtrace_speculation_state { 574 DTRACESPEC_INACTIVE = 0, 575 DTRACESPEC_ACTIVE, 576 DTRACESPEC_ACTIVEONE, 577 DTRACESPEC_ACTIVEMANY, 578 DTRACESPEC_COMMITTING, 579 DTRACESPEC_COMMITTINGMANY, 580 DTRACESPEC_DISCARDING 581 } dtrace_speculation_state_t; 582 583 typedef struct dtrace_speculation { 584 dtrace_speculation_state_t dtsp_state; /* current speculation state */ 585 int dtsp_cleaning; /* non-zero if being cleaned */ 586 dtrace_buffer_t *dtsp_buffer; /* speculative buffer */ 587 } dtrace_speculation_t; 588 589 /* 590 * DTrace Dynamic Variables 591 * 592 * The dynamic variable problem is obviously decomposed into two subproblems: 593 * allocating new dynamic storage, and freeing old dynamic storage. The 594 * presence of the second problem makes the first much more complicated -- or 595 * rather, the absence of the second renders the first trivial. This is the 596 * case with aggregations, for which there is effectively no deallocation of 597 * dynamic storage. (Or more accurately, all dynamic storage is deallocated 598 * when a snapshot is taken of the aggregation.) As DTrace dynamic variables 599 * allow for both dynamic allocation and dynamic deallocation, the 600 * implementation of dynamic variables is quite a bit more complicated than 601 * that of their aggregation kin. 602 * 603 * We observe that allocating new dynamic storage is tricky only because the 604 * size can vary -- the allocation problem is much easier if allocation sizes 605 * are uniform. We further observe that in D, the size of dynamic variables is 606 * actually _not_ dynamic -- dynamic variable sizes may be determined by static 607 * analysis of DIF text. (This is true even of putatively dynamically-sized 608 * objects like strings and stacks, the sizes of which are dictated by the 609 * "stringsize" and "stackframes" variables, respectively.) We exploit this by 610 * performing this analysis on all DIF before enabling any probes. For each 611 * dynamic load or store, we calculate the dynamically-allocated size plus the 612 * size of the dtrace_dynvar structure plus the storage required to key the 613 * data. For all DIF, we take the largest value and dub it the _chunksize_. 614 * We then divide dynamic memory into two parts: a hash table that is wide 615 * enough to have every chunk in its own bucket, and a larger region of equal 616 * chunksize units. Whenever we wish to dynamically allocate a variable, we 617 * always allocate a single chunk of memory. Depending on the uniformity of 618 * allocation, this will waste some amount of memory -- but it eliminates the 619 * non-determinism inherent in traditional heap fragmentation. 620 * 621 * Dynamic objects are allocated by storing a non-zero value to them; they are 622 * deallocated by storing a zero value to them. Dynamic variables are 623 * complicated enormously by being shared between CPUs. In particular, 624 * consider the following scenario: 625 * 626 * CPU A CPU B 627 * +---------------------------------+ +---------------------------------+ 628 * | | | | 629 * | allocates dynamic object a[123] | | | 630 * | by storing the value 345 to it | | | 631 * | ---------> | 632 * | | | wishing to load from object | 633 * | | | a[123], performs lookup in | 634 * | | | dynamic variable space | 635 * | <--------- | 636 * | deallocates object a[123] by | | | 637 * | storing 0 to it | | | 638 * | | | | 639 * | allocates dynamic object b[567] | | performs load from a[123] | 640 * | by storing the value 789 to it | | | 641 * : : : : 642 * . . . . 643 * 644 * This is obviously a race in the D program, but there are nonetheless only 645 * two valid values for CPU B's load from a[123]: 345 or 0. Most importantly, 646 * CPU B may _not_ see the value 789 for a[123]. 647 * 648 * There are essentially two ways to deal with this: 649 * 650 * (1) Explicitly spin-lock variables. That is, if CPU B wishes to load 651 * from a[123], it needs to lock a[123] and hold the lock for the 652 * duration that it wishes to manipulate it. 653 * 654 * (2) Avoid reusing freed chunks until it is known that no CPU is referring 655 * to them. 656 * 657 * The implementation of (1) is rife with complexity, because it requires the 658 * user of a dynamic variable to explicitly decree when they are done using it. 659 * Were all variables by value, this perhaps wouldn't be debilitating -- but 660 * dynamic variables of non-scalar types are tracked by reference. That is, if 661 * a dynamic variable is, say, a string, and that variable is to be traced to, 662 * say, the principal buffer, the DIF emulation code returns to the main 663 * dtrace_probe() loop a pointer to the underlying storage, not the contents of 664 * the storage. Further, code calling on DIF emulation would have to be aware 665 * that the DIF emulation has returned a reference to a dynamic variable that 666 * has been potentially locked. The variable would have to be unlocked after 667 * the main dtrace_probe() loop is finished with the variable, and the main 668 * dtrace_probe() loop would have to be careful to not call any further DIF 669 * emulation while the variable is locked to avoid deadlock. More generally, 670 * if one were to implement (1), DIF emulation code dealing with dynamic 671 * variables could only deal with one dynamic variable at a time (lest deadlock 672 * result). To sum, (1) exports too much subtlety to the users of dynamic 673 * variables -- increasing maintenance burden and imposing serious constraints 674 * on future DTrace development. 675 * 676 * The implementation of (2) is also complex, but the complexity is more 677 * manageable. We need to be sure that when a variable is deallocated, it is 678 * not placed on a traditional free list, but rather on a _dirty_ list. Once a 679 * variable is on a dirty list, it cannot be found by CPUs performing a 680 * subsequent lookup of the variable -- but it may still be in use by other 681 * CPUs. To assure that all CPUs that may be seeing the old variable have 682 * cleared out of probe context, a dtrace_sync() can be issued. Once the 683 * dtrace_sync() has completed, it can be known that all CPUs are done 684 * manipulating the dynamic variable -- the dirty list can be atomically 685 * appended to the free list. Unfortunately, there's a slight hiccup in this 686 * mechanism: dtrace_sync() may not be issued from probe context. The 687 * dtrace_sync() must be therefore issued asynchronously from non-probe 688 * context. For this we rely on the DTrace cleaner, a cyclic that runs at the 689 * "cleanrate" frequency. To ease this implementation, we define several chunk 690 * lists: 691 * 692 * - Dirty. Deallocated chunks, not yet cleaned. Not available. 693 * 694 * - Rinsing. Formerly dirty chunks that are currently being asynchronously 695 * cleaned. Not available, but will be shortly. Dynamic variable 696 * allocation may not spin or block for availability, however. 697 * 698 * - Clean. Clean chunks, ready for allocation -- but not on the free list. 699 * 700 * - Free. Available for allocation. 701 * 702 * Moreover, to avoid absurd contention, _each_ of these lists is implemented 703 * on a per-CPU basis. This is only for performance, not correctness; chunks 704 * may be allocated from another CPU's free list. The algorithm for allocation 705 * then is this: 706 * 707 * (1) Attempt to atomically allocate from current CPU's free list. If list 708 * is non-empty and allocation is successful, allocation is complete. 709 * 710 * (2) If the clean list is non-empty, atomically move it to the free list, 711 * and reattempt (1). 712 * 713 * (3) If the dynamic variable space is in the CLEAN state, look for free 714 * and clean lists on other CPUs by setting the current CPU to the next 715 * CPU, and reattempting (1). If the next CPU is the current CPU (that 716 * is, if all CPUs have been checked), atomically switch the state of 717 * the dynamic variable space based on the following: 718 * 719 * - If no free chunks were found and no dirty chunks were found, 720 * atomically set the state to EMPTY. 721 * 722 * - If dirty chunks were found, atomically set the state to DIRTY. 723 * 724 * - If rinsing chunks were found, atomically set the state to RINSING. 725 * 726 * (4) Based on state of dynamic variable space state, increment appropriate 727 * counter to indicate dynamic drops (if in EMPTY state) vs. dynamic 728 * dirty drops (if in DIRTY state) vs. dynamic rinsing drops (if in 729 * RINSING state). Fail the allocation. 730 * 731 * The cleaning cyclic operates with the following algorithm: for all CPUs 732 * with a non-empty dirty list, atomically move the dirty list to the rinsing 733 * list. Perform a dtrace_sync(). For all CPUs with a non-empty rinsing list, 734 * atomically move the rinsing list to the clean list. Perform another 735 * dtrace_sync(). By this point, all CPUs have seen the new clean list; the 736 * state of the dynamic variable space can be restored to CLEAN. 737 * 738 * There exist two final races that merit explanation. The first is a simple 739 * allocation race: 740 * 741 * CPU A CPU B 742 * +---------------------------------+ +---------------------------------+ 743 * | | | | 744 * | allocates dynamic object a[123] | | allocates dynamic object a[123] | 745 * | by storing the value 345 to it | | by storing the value 567 to it | 746 * | | | | 747 * : : : : 748 * . . . . 749 * 750 * Again, this is a race in the D program. It can be resolved by having a[123] 751 * hold the value 345 or a[123] hold the value 567 -- but it must be true that 752 * a[123] have only _one_ of these values. (That is, the racing CPUs may not 753 * put the same element twice on the same hash chain.) This is resolved 754 * simply: before the allocation is undertaken, the start of the new chunk's 755 * hash chain is noted. Later, after the allocation is complete, the hash 756 * chain is atomically switched to point to the new element. If this fails 757 * (because of either concurrent allocations or an allocation concurrent with a 758 * deletion), the newly allocated chunk is deallocated to the dirty list, and 759 * the whole process of looking up (and potentially allocating) the dynamic 760 * variable is reattempted. 761 * 762 * The final race is a simple deallocation race: 763 * 764 * CPU A CPU B 765 * +---------------------------------+ +---------------------------------+ 766 * | | | | 767 * | deallocates dynamic object | | deallocates dynamic object | 768 * | a[123] by storing the value 0 | | a[123] by storing the value 0 | 769 * | to it | | to it | 770 * | | | | 771 * : : : : 772 * . . . . 773 * 774 * Once again, this is a race in the D program, but it is one that we must 775 * handle without corrupting the underlying data structures. Because 776 * deallocations require the deletion of a chunk from the middle of a hash 777 * chain, we cannot use a single-word atomic operation to remove it. For this, 778 * we add a spin lock to the hash buckets that is _only_ used for deallocations 779 * (allocation races are handled as above). Further, this spin lock is _only_ 780 * held for the duration of the delete; before control is returned to the DIF 781 * emulation code, the hash bucket is unlocked. 782 */ 783 typedef struct dtrace_key { 784 uint64_t dttk_value; /* data value or data pointer */ 785 uint64_t dttk_size; /* 0 if by-val, >0 if by-ref */ 786 } dtrace_key_t; 787 788 typedef struct dtrace_tuple { 789 uint32_t dtt_nkeys; /* number of keys in tuple */ 790 uint32_t dtt_pad; /* padding */ 791 dtrace_key_t dtt_key[1]; /* array of tuple keys */ 792 } dtrace_tuple_t; 793 794 typedef struct dtrace_dynvar { 795 uint64_t dtdv_hashval; /* hash value -- 0 if free */ 796 struct dtrace_dynvar *dtdv_next; /* next on list or hash chain */ 797 void *dtdv_data; /* pointer to data */ 798 dtrace_tuple_t dtdv_tuple; /* tuple key */ 799 } dtrace_dynvar_t; 800 801 typedef enum dtrace_dynvar_op { 802 DTRACE_DYNVAR_ALLOC, 803 DTRACE_DYNVAR_NOALLOC, 804 DTRACE_DYNVAR_DEALLOC 805 } dtrace_dynvar_op_t; 806 807 typedef struct dtrace_dynhash { 808 dtrace_dynvar_t *dtdh_chain; /* hash chain for this bucket */ 809 uintptr_t dtdh_lock; /* deallocation lock */ 810 #ifdef _LP64 811 uintptr_t dtdh_pad[6]; /* pad to avoid false sharing */ 812 #else 813 uintptr_t dtdh_pad[14]; /* pad to avoid false sharing */ 814 #endif 815 } dtrace_dynhash_t; 816 817 typedef struct dtrace_dstate_percpu { 818 dtrace_dynvar_t *dtdsc_free; /* free list for this CPU */ 819 dtrace_dynvar_t *dtdsc_dirty; /* dirty list for this CPU */ 820 dtrace_dynvar_t *dtdsc_rinsing; /* rinsing list for this CPU */ 821 dtrace_dynvar_t *dtdsc_clean; /* clean list for this CPU */ 822 uint64_t dtdsc_drops; /* number of capacity drops */ 823 uint64_t dtdsc_dirty_drops; /* number of dirty drops */ 824 uint64_t dtdsc_rinsing_drops; /* number of rinsing drops */ 825 #ifdef _LP64 826 uint64_t dtdsc_pad; /* pad to avoid false sharing */ 827 #else 828 uint64_t dtdsc_pad[2]; /* pad to avoid false sharing */ 829 #endif 830 } dtrace_dstate_percpu_t; 831 832 typedef enum dtrace_dstate_state { 833 DTRACE_DSTATE_CLEAN = 0, 834 DTRACE_DSTATE_EMPTY, 835 DTRACE_DSTATE_DIRTY, 836 DTRACE_DSTATE_RINSING 837 } dtrace_dstate_state_t; 838 839 typedef struct dtrace_dstate { 840 void *dtds_base; /* base of dynamic var. space */ 841 size_t dtds_size; /* size of dynamic var. space */ 842 size_t dtds_hashsize; /* number of buckets in hash */ 843 size_t dtds_chunksize; /* size of each chunk */ 844 dtrace_dynhash_t *dtds_hash; /* pointer to hash table */ 845 dtrace_dstate_state_t dtds_state; /* current dynamic var. state */ 846 dtrace_dstate_percpu_t *dtds_percpu; /* per-CPU dyn. var. state */ 847 } dtrace_dstate_t; 848 849 /* 850 * DTrace Variable State 851 * 852 * The DTrace variable state tracks user-defined variables in its dtrace_vstate 853 * structure. Each DTrace consumer has exactly one dtrace_vstate structure, 854 * but some dtrace_vstate structures may exist without a corresponding DTrace 855 * consumer (see "DTrace Helpers", below). As described in <sys/dtrace.h>, 856 * user-defined variables can have one of three scopes: 857 * 858 * DIFV_SCOPE_GLOBAL => global scope 859 * DIFV_SCOPE_THREAD => thread-local scope (i.e. "self->" variables) 860 * DIFV_SCOPE_LOCAL => clause-local scope (i.e. "this->" variables) 861 * 862 * The variable state tracks variables by both their scope and their allocation 863 * type: 864 * 865 * - The dtvs_globals and dtvs_locals members each point to an array of 866 * dtrace_statvar structures. These structures contain both the variable 867 * metadata (dtrace_difv structures) and the underlying storage for all 868 * statically allocated variables, including statically allocated 869 * DIFV_SCOPE_GLOBAL variables and all DIFV_SCOPE_LOCAL variables. 870 * 871 * - The dtvs_tlocals member points to an array of dtrace_difv structures for 872 * DIFV_SCOPE_THREAD variables. As such, this array tracks _only_ the 873 * variable metadata for DIFV_SCOPE_THREAD variables; the underlying storage 874 * is allocated out of the dynamic variable space. 875 * 876 * - The dtvs_dynvars member is the dynamic variable state associated with the 877 * variable state. The dynamic variable state (described in "DTrace Dynamic 878 * Variables", above) tracks all DIFV_SCOPE_THREAD variables and all 879 * dynamically-allocated DIFV_SCOPE_GLOBAL variables. 880 */ 881 typedef struct dtrace_statvar { 882 uint64_t dtsv_data; /* data or pointer to it */ 883 size_t dtsv_size; /* size of pointed-to data */ 884 int dtsv_refcnt; /* reference count */ 885 dtrace_difv_t dtsv_var; /* variable metadata */ 886 } dtrace_statvar_t; 887 888 typedef struct dtrace_vstate { 889 dtrace_state_t *dtvs_state; /* back pointer to state */ 890 dtrace_statvar_t **dtvs_globals; /* statically-allocated glbls */ 891 int dtvs_nglobals; /* number of globals */ 892 dtrace_difv_t *dtvs_tlocals; /* thread-local metadata */ 893 int dtvs_ntlocals; /* number of thread-locals */ 894 dtrace_statvar_t **dtvs_locals; /* clause-local data */ 895 int dtvs_nlocals; /* number of clause-locals */ 896 dtrace_dstate_t dtvs_dynvars; /* dynamic variable state */ 897 } dtrace_vstate_t; 898 899 /* 900 * DTrace Machine State 901 * 902 * In the process of processing a fired probe, DTrace needs to track and/or 903 * cache some per-CPU state associated with that particular firing. This is 904 * state that is always discarded after the probe firing has completed, and 905 * much of it is not specific to any DTrace consumer, remaining valid across 906 * all ECBs. This state is tracked in the dtrace_mstate structure. 907 */ 908 #define DTRACE_MSTATE_ARGS 0x00000001 909 #define DTRACE_MSTATE_PROBE 0x00000002 910 #define DTRACE_MSTATE_EPID 0x00000004 911 #define DTRACE_MSTATE_TIMESTAMP 0x00000008 912 #define DTRACE_MSTATE_STACKDEPTH 0x00000010 913 #define DTRACE_MSTATE_CALLER 0x00000020 914 #define DTRACE_MSTATE_IPL 0x00000040 915 #define DTRACE_MSTATE_FLTOFFS 0x00000080 916 #define DTRACE_MSTATE_WALLTIMESTAMP 0x00000100 917 #define DTRACE_MSTATE_USTACKDEPTH 0x00000200 918 #define DTRACE_MSTATE_UCALLER 0x00000400 919 #define DTRACE_MSTATE_MACHTIMESTAMP 0x00000800 920 #define DTRACE_MSTATE_MACHCTIMESTAMP 0x00001000 921 922 typedef struct dtrace_mstate { 923 uintptr_t dtms_scratch_base; /* base of scratch space */ 924 uintptr_t dtms_scratch_ptr; /* current scratch pointer */ 925 size_t dtms_scratch_size; /* scratch size */ 926 uint32_t dtms_present; /* variables that are present */ 927 uint64_t dtms_arg[5]; /* cached arguments */ 928 dtrace_epid_t dtms_epid; /* current EPID */ 929 uint64_t dtms_timestamp; /* cached timestamp */ 930 hrtime_t dtms_walltimestamp; /* cached wall timestamp */ 931 uint64_t dtms_machtimestamp; /* cached mach absolute timestamp */ 932 uint64_t dtms_machctimestamp; /* cached mach continuous timestamp */ 933 int dtms_stackdepth; /* cached stackdepth */ 934 int dtms_ustackdepth; /* cached ustackdepth */ 935 struct dtrace_probe *dtms_probe; /* current probe */ 936 uintptr_t dtms_caller; /* cached caller */ 937 uint64_t dtms_ucaller; /* cached user-level caller */ 938 int dtms_ipl; /* cached interrupt pri lev */ 939 int dtms_fltoffs; /* faulting DIFO offset */ 940 uintptr_t dtms_strtok; /* saved strtok() pointer */ 941 uintptr_t dtms_strtok_limit; /* upper bound of strtok ptr */ 942 uint32_t dtms_access; /* memory access rights */ 943 dtrace_difo_t *dtms_difo; /* current dif object */ 944 } dtrace_mstate_t; 945 946 #define DTRACE_COND_OWNER 0x1 947 #define DTRACE_COND_USERMODE 0x2 948 #define DTRACE_COND_ZONEOWNER 0x4 949 950 #define DTRACE_PROBEKEY_MAXDEPTH 8 /* max glob recursion depth */ 951 952 /* 953 * Access flag used by dtrace_mstate.dtms_access. 954 */ 955 #define DTRACE_ACCESS_KERNEL 0x1 /* the priv to read kmem */ 956 957 958 /* 959 * DTrace Activity 960 * 961 * Each DTrace consumer is in one of several states, which (for purposes of 962 * avoiding yet-another overloading of the noun "state") we call the current 963 * _activity_. The activity transitions on dtrace_go() (from DTRACIOCGO), on 964 * dtrace_stop() (from DTRACIOCSTOP) and on the exit() action. Activities may 965 * only transition in one direction; the activity transition diagram is a 966 * directed acyclic graph. The activity transition diagram is as follows: 967 * 968 * 969 * 970 * +----------+ +--------+ +--------+ 971 * | INACTIVE |------------------>| WARMUP |------------------>| ACTIVE | 972 * +----------+ dtrace_go(), +--------+ dtrace_go(), +--------+ 973 * before BEGIN | after BEGIN | | | 974 * | | | | 975 * exit() action | | | | 976 * from BEGIN ECB | | | | 977 * | | | | 978 * v | | | 979 * +----------+ exit() action | | | 980 * +-----------------------------| DRAINING |<-------------------+ | | 981 * | +----------+ | | 982 * | | | | 983 * | dtrace_stop(), | | | 984 * | before END | | | 985 * | | | | 986 * | v | | 987 * | +---------+ +----------+ | | 988 * | | STOPPED |<----------------| COOLDOWN |<----------------------+ | 989 * | +---------+ dtrace_stop(), +----------+ dtrace_stop(), | 990 * | after END before END | 991 * | | 992 * | +--------+ | 993 * +----------------------------->| KILLED |<--------------------------+ 994 * deadman timeout or +--------+ deadman timeout or 995 * killed consumer killed consumer 996 * 997 * Note that once a DTrace consumer has stopped tracing, there is no way to 998 * restart it; if a DTrace consumer wishes to restart tracing, it must reopen 999 * the DTrace pseudodevice. 1000 */ 1001 typedef enum dtrace_activity { 1002 DTRACE_ACTIVITY_INACTIVE = 0, /* not yet running */ 1003 DTRACE_ACTIVITY_WARMUP, /* while starting */ 1004 DTRACE_ACTIVITY_ACTIVE, /* running */ 1005 DTRACE_ACTIVITY_DRAINING, /* before stopping */ 1006 DTRACE_ACTIVITY_COOLDOWN, /* while stopping */ 1007 DTRACE_ACTIVITY_STOPPED, /* after stopping */ 1008 DTRACE_ACTIVITY_KILLED /* killed */ 1009 } dtrace_activity_t; 1010 1011 1012 /* 1013 * APPLE NOTE: DTrace dof modes implementation 1014 * 1015 * DTrace has four "dof modes". They are: 1016 * 1017 * DTRACE_DOF_MODE_NEVER Never load any dof, period. 1018 * DTRACE_DOF_MODE_LAZY_ON Defer loading dof until later 1019 * DTRACE_DOF_MODE_LAZY_OFF Load all deferred dof now, and any new dof 1020 * DTRACE_DOF_MODE_NON_LAZY Load all dof immediately. 1021 * 1022 * It is legal to transition between the two lazy modes. The NEVER and 1023 * NON_LAZY modes are permanent, and must not change once set. 1024 * 1025 * The current dof mode is kept in dtrace_dof_mode, which is protected by the 1026 * dtrace_dof_mode_lock. This is a RW lock, reads require shared access, writes 1027 * require exclusive access. Because NEVER and NON_LAZY are permanent states, 1028 * it is legal to test for those modes without holding the dof mode lock. 1029 * 1030 * Lock ordering is dof mode lock before any dtrace lock, and before the 1031 * process p_dtrace_sprlock. In general, other locks should not be held when 1032 * taking the dof mode lock. Acquiring the dof mode lock in exclusive mode 1033 * will block process fork, exec, and exit, so it should be held exclusive 1034 * for as short a time as possible. 1035 */ 1036 1037 #define DTRACE_DOF_MODE_NEVER 0 1038 #define DTRACE_DOF_MODE_LAZY_ON 1 1039 #define DTRACE_DOF_MODE_LAZY_OFF 2 1040 #define DTRACE_DOF_MODE_NON_LAZY 3 1041 1042 /* 1043 * dtrace kernel symbol modes are used to control when the kernel may dispose of 1044 * symbol information used by the fbt/sdt provider. The kernel itself, as well as 1045 * every kext, has symbol table/nlist info that has historically been preserved 1046 * for dtrace's use. This allowed dtrace to be lazy about allocating fbt/sdt probes, 1047 * at the expense of keeping the symbol info in the kernel permanently. 1048 * 1049 * Starting in 10.7+, fbt probes may be created from userspace, in the same 1050 * fashion as pid probes. The kernel allows dtrace "first right of refusal" 1051 * whenever symbol data becomes available (such as a kext load). If dtrace is 1052 * active, it will immediately read/copy the needed data, and then the kernel 1053 * may free it. If dtrace is not active, it returns immediately, having done 1054 * no work or allocations, and the symbol data is freed. Should dtrace need 1055 * this data later, it is expected that the userspace client will push the 1056 * data into the kernel via ioctl calls. 1057 * 1058 * The kernel symbol modes are used to control what dtrace does with symbol data: 1059 * 1060 * DTRACE_KERNEL_SYMBOLS_NEVER Effectively disables fbt/sdt 1061 * DTRACE_KERNEL_SYMBOLS_FROM_KERNEL Immediately read/copy symbol data 1062 * DTRACE_KERNEL_SYMBOLS_FROM_USERSPACE Wait for symbols from userspace 1063 * DTRACE_KERNEL_SYMBOLS_ALWAYS_FROM_KERNEL Immediately read/copy symbol data 1064 * 1065 * It is legal to transition between DTRACE_KERNEL_SYMBOLS_FROM_KERNEL and 1066 * DTRACE_KERNEL_SYMBOLS_FROM_USERSPACE. The DTRACE_KERNEL_SYMBOLS_NEVER and 1067 * DTRACE_KERNEL_SYMBOLS_ALWAYS_FROM_KERNEL are permanent modes, intended to 1068 * disable fbt probes entirely, or prevent any symbols being loaded from 1069 * userspace. 1070 * 1071 * The kernel symbol mode is kept in dtrace_kernel_symbol_mode, which is protected 1072 * by the dtrace_lock. 1073 */ 1074 1075 #define DTRACE_KERNEL_SYMBOLS_NEVER 0 1076 #define DTRACE_KERNEL_SYMBOLS_FROM_KERNEL 1 1077 #define DTRACE_KERNEL_SYMBOLS_FROM_USERSPACE 2 1078 #define DTRACE_KERNEL_SYMBOLS_ALWAYS_FROM_KERNEL 3 1079 1080 1081 /* 1082 * DTrace Helper Implementation 1083 * 1084 * A description of the helper architecture may be found in <sys/dtrace.h>. 1085 * Each process contains a pointer to its helpers in its p_dtrace_helpers 1086 * member. This is a pointer to a dtrace_helpers structure, which contains an 1087 * array of pointers to dtrace_helper structures, helper variable state (shared 1088 * among a process's helpers) and a generation count. (The generation count is 1089 * used to provide an identifier when a helper is added so that it may be 1090 * subsequently removed.) The dtrace_helper structure is self-explanatory, 1091 * containing pointers to the objects needed to execute the helper. Note that 1092 * helpers are _duplicated_ across fork(2), and destroyed on exec(2). No more 1093 * than dtrace_helpers_max are allowed per-process. 1094 */ 1095 #define DTRACE_HELPER_ACTION_USTACK 0 1096 #define DTRACE_NHELPER_ACTIONS 1 1097 1098 typedef struct dtrace_helper_action { 1099 int dtha_generation; /* helper action generation */ 1100 int dtha_nactions; /* number of actions */ 1101 dtrace_difo_t *dtha_predicate; /* helper action predicate */ 1102 dtrace_difo_t **dtha_actions; /* array of actions */ 1103 struct dtrace_helper_action *dtha_next; /* next helper action */ 1104 } dtrace_helper_action_t; 1105 1106 typedef struct dtrace_helper_provider { 1107 int dthp_generation; /* helper provider generation */ 1108 uint32_t dthp_ref; /* reference count */ 1109 dof_helper_t dthp_prov; /* DOF w/ provider and probes */ 1110 } dtrace_helper_provider_t; 1111 1112 typedef struct dtrace_helpers { 1113 dtrace_helper_action_t **dthps_actions; /* array of helper actions */ 1114 dtrace_vstate_t dthps_vstate; /* helper action var. state */ 1115 dtrace_helper_provider_t **dthps_provs; /* array of providers */ 1116 uint_t dthps_nprovs; /* count of providers */ 1117 uint_t dthps_maxprovs; /* provider array size */ 1118 int dthps_generation; /* current generation */ 1119 pid_t dthps_pid; /* pid of associated proc */ 1120 int dthps_deferred; /* helper in deferred list */ 1121 struct dtrace_helpers *dthps_next; /* next pointer */ 1122 struct dtrace_helpers *dthps_prev; /* prev pointer */ 1123 } dtrace_helpers_t; 1124 1125 /* 1126 * DTrace Helper Action Tracing 1127 * 1128 * Debugging helper actions can be arduous. To ease the development and 1129 * debugging of helpers, DTrace contains a tracing-framework-within-a-tracing- 1130 * framework: helper tracing. If dtrace_helptrace_enabled is non-zero (which 1131 * it is by default on DEBUG kernels), all helper activity will be traced to a 1132 * global, in-kernel ring buffer. Each entry includes a pointer to the specific 1133 * helper, the location within the helper, and a trace of all local variables. 1134 * The ring buffer may be displayed in a human-readable format with the 1135 * ::dtrace_helptrace mdb(1) dcmd. 1136 */ 1137 #define DTRACE_HELPTRACE_NEXT (-1) 1138 #define DTRACE_HELPTRACE_DONE (-2) 1139 #define DTRACE_HELPTRACE_ERR (-3) 1140 1141 1142 typedef struct dtrace_helptrace { 1143 dtrace_helper_action_t *dtht_helper; /* helper action */ 1144 int dtht_where; /* where in helper action */ 1145 int dtht_nlocals; /* number of locals */ 1146 int dtht_fault; /* type of fault (if any) */ 1147 int dtht_fltoffs; /* DIF offset */ 1148 uint64_t dtht_illval; /* faulting value */ 1149 uint64_t dtht_locals[1]; /* local variables */ 1150 } dtrace_helptrace_t; 1151 1152 /* 1153 * DTrace Credentials 1154 * 1155 * In probe context, we have limited flexibility to examine the credentials 1156 * of the DTrace consumer that created a particular enabling. We use 1157 * the Least Privilege interfaces to cache the consumer's cred pointer and 1158 * some facts about that credential in a dtrace_cred_t structure. These 1159 * can limit the consumer's breadth of visibility and what actions the 1160 * consumer may take. 1161 */ 1162 #define DTRACE_CRV_ALLPROC 0x01 1163 #define DTRACE_CRV_KERNEL 0x02 1164 #define DTRACE_CRV_ALLZONE 0x04 1165 1166 #define DTRACE_CRV_ALL (DTRACE_CRV_ALLPROC | DTRACE_CRV_KERNEL | \ 1167 DTRACE_CRV_ALLZONE) 1168 1169 #define DTRACE_CRA_PROC 0x0001 1170 #define DTRACE_CRA_PROC_CONTROL 0x0002 1171 #define DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER 0x0004 1172 #define DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE 0x0008 1173 #define DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG 0x0010 1174 #define DTRACE_CRA_KERNEL 0x0020 1175 #define DTRACE_CRA_KERNEL_DESTRUCTIVE 0x0040 1176 1177 #define DTRACE_CRA_ALL (DTRACE_CRA_PROC | \ 1178 DTRACE_CRA_PROC_CONTROL | \ 1179 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER | \ 1180 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE | \ 1181 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG | \ 1182 DTRACE_CRA_KERNEL | \ 1183 DTRACE_CRA_KERNEL_DESTRUCTIVE) 1184 1185 typedef struct dtrace_cred { 1186 cred_t *dcr_cred; 1187 uint8_t dcr_destructive; 1188 uint8_t dcr_visible; 1189 uint16_t dcr_action; 1190 } dtrace_cred_t; 1191 1192 typedef struct dtrace_format { 1193 uint64_t dtf_refcount; 1194 char dtf_str[]; 1195 } dtrace_format_t; 1196 1197 #define DTRACE_FORMAT_SIZE(fmt) (strlen(fmt->dtf_str) + 1 + sizeof(dtrace_format_t)) 1198 1199 /* 1200 * DTrace Consumer State 1201 * 1202 * Each DTrace consumer has an associated dtrace_state structure that contains 1203 * its in-kernel DTrace state -- including options, credentials, statistics and 1204 * pointers to ECBs, buffers, speculations and formats. A dtrace_state 1205 * structure is also allocated for anonymous enablings. When anonymous state 1206 * is grabbed, the grabbing consumers dts_anon pointer is set to the grabbed 1207 * dtrace_state structure. 1208 */ 1209 struct dtrace_state { 1210 dev_t dts_dev; /* device */ 1211 int dts_necbs; /* total number of ECBs */ 1212 dtrace_ecb_t **dts_ecbs; /* array of ECBs */ 1213 dtrace_epid_t dts_epid; /* next EPID to allocate */ 1214 size_t dts_needed; /* greatest needed space */ 1215 struct dtrace_state *dts_anon; /* anon. state, if grabbed */ 1216 dtrace_activity_t dts_activity; /* current activity */ 1217 dtrace_vstate_t dts_vstate; /* variable state */ 1218 dtrace_buffer_t *dts_buffer; /* principal buffer */ 1219 dtrace_buffer_t *dts_aggbuffer; /* aggregation buffer */ 1220 dtrace_speculation_t *dts_speculations; /* speculation array */ 1221 int dts_nspeculations; /* number of speculations */ 1222 int dts_naggregations; /* number of aggregations */ 1223 dtrace_aggregation_t **dts_aggregations; /* aggregation array */ 1224 vmem_t *dts_aggid_arena; /* arena for aggregation IDs */ 1225 uint64_t dts_errors; /* total number of errors */ 1226 uint32_t dts_speculations_busy; /* number of spec. busy */ 1227 uint32_t dts_speculations_unavail; /* number of spec unavail */ 1228 uint32_t dts_stkstroverflows; /* stack string tab overflows */ 1229 uint32_t dts_dblerrors; /* errors in ERROR probes */ 1230 uint32_t dts_reserve; /* space reserved for END */ 1231 hrtime_t dts_laststatus; /* time of last status */ 1232 cyclic_id_t dts_cleaner; /* cleaning cyclic */ 1233 cyclic_id_t dts_deadman; /* deadman cyclic */ 1234 hrtime_t dts_alive; /* time last alive */ 1235 char dts_speculates; /* boolean: has speculations */ 1236 char dts_destructive; /* boolean: has dest. actions */ 1237 int dts_nformats; /* number of formats */ 1238 dtrace_format_t **dts_formats; /* format string array */ 1239 dtrace_optval_t dts_options[DTRACEOPT_MAX]; /* options */ 1240 dtrace_cred_t dts_cred; /* credentials */ 1241 size_t dts_nretained; /* number of retained enabs */ 1242 uint64_t dts_arg_error_illval; 1243 uint32_t dts_buf_over_limit; /* number of bufs over dtb_limit */ 1244 uint64_t **dts_rstate; /* per-CPU random state */ 1245 }; 1246 1247 struct dtrace_provider { 1248 dtrace_pattr_t dtpv_attr; /* provider attributes */ 1249 dtrace_ppriv_t dtpv_priv; /* provider privileges */ 1250 dtrace_pops_t dtpv_pops; /* provider operations */ 1251 char *dtpv_name; /* provider name */ 1252 void *dtpv_arg; /* provider argument */ 1253 uint_t dtpv_defunct; /* boolean: defunct provider */ 1254 struct dtrace_provider *dtpv_next; /* next provider */ 1255 uint64_t dtpv_probe_count; /* number of associated probes */ 1256 uint64_t dtpv_ecb_count; /* number of associated enabled ECBs */ 1257 }; 1258 1259 struct dtrace_meta { 1260 dtrace_mops_t dtm_mops; /* meta provider operations */ 1261 char *dtm_name; /* meta provider name */ 1262 void *dtm_arg; /* meta provider user arg */ 1263 uint64_t dtm_count; /* number of associated providers */ 1264 }; 1265 1266 /* 1267 * DTrace Enablings 1268 * 1269 * A dtrace_enabling structure is used to track a collection of ECB 1270 * descriptions -- before they have been turned into actual ECBs. This is 1271 * created as a result of DOF processing, and is generally used to generate 1272 * ECBs immediately thereafter. However, enablings are also generally 1273 * retained should the probes they describe be created at a later time; as 1274 * each new module or provider registers with the framework, the retained 1275 * enablings are reevaluated, with any new match resulting in new ECBs. To 1276 * prevent probes from being matched more than once, the enabling tracks the 1277 * last probe generation matched, and only matches probes from subsequent 1278 * generations. 1279 */ 1280 typedef struct dtrace_enabling { 1281 dtrace_ecbdesc_t **dten_desc; /* all ECB descriptions */ 1282 int dten_ndesc; /* number of ECB descriptions */ 1283 int dten_maxdesc; /* size of ECB array */ 1284 dtrace_vstate_t *dten_vstate; /* associated variable state */ 1285 dtrace_genid_t dten_probegen; /* matched probe generation */ 1286 dtrace_ecbdesc_t *dten_current; /* current ECB description */ 1287 int dten_error; /* current error value */ 1288 int dten_primed; /* boolean: set if primed */ 1289 struct dtrace_enabling *dten_prev; /* previous enabling */ 1290 struct dtrace_enabling *dten_next; /* next enabling */ 1291 } dtrace_enabling_t; 1292 1293 /* 1294 * DTrace Anonymous Enablings 1295 * 1296 * Anonymous enablings are DTrace enablings that are not associated with a 1297 * controlling process, but rather derive their enabling from DOF stored as 1298 * properties in the dtrace.conf file. If there is an anonymous enabling, a 1299 * DTrace consumer state and enabling are created on attach. The state may be 1300 * subsequently grabbed by the first consumer specifying the "grabanon" 1301 * option. As long as an anonymous DTrace enabling exists, dtrace(7D) will 1302 * refuse to unload. 1303 */ 1304 typedef struct dtrace_anon { 1305 dtrace_state_t *dta_state; /* DTrace consumer state */ 1306 dtrace_enabling_t *dta_enabling; /* pointer to enabling */ 1307 processorid_t dta_beganon; /* which CPU BEGIN ran on */ 1308 } dtrace_anon_t; 1309 1310 /* 1311 * DTrace Error Debugging 1312 */ 1313 #if DEBUG 1314 #define DTRACE_ERRDEBUG 1315 #endif 1316 1317 #ifdef DTRACE_ERRDEBUG 1318 1319 typedef struct dtrace_errhash { 1320 const char *dter_msg; /* error message */ 1321 int dter_count; /* number of times seen */ 1322 } dtrace_errhash_t; 1323 1324 #define DTRACE_ERRHASHSZ 256 /* must be > number of err msgs */ 1325 1326 #endif /* DTRACE_ERRDEBUG */ 1327 1328 typedef struct dtrace_string dtrace_string_t; 1329 1330 typedef struct dtrace_string { 1331 dtrace_string_t *dtst_next; 1332 dtrace_string_t *dtst_prev; 1333 uint32_t dtst_refcount; 1334 char dtst_str[]; 1335 } dtrace_string_t; 1336 1337 /** 1338 * DTrace Matching pre-conditions 1339 * 1340 * Used when matching new probes to discard matching of enablings that 1341 * doesn't match the condition tested by dmc_func 1342 */ 1343 typedef struct dtrace_match_cond { 1344 int (*dmc_func)(dtrace_probedesc_t*, void*); 1345 void *dmc_data; 1346 } dtrace_match_cond_t; 1347 1348 1349 /* 1350 * DTrace Toxic Ranges 1351 * 1352 * DTrace supports safe loads from probe context; if the address turns out to 1353 * be invalid, a bit will be set by the kernel indicating that DTrace 1354 * encountered a memory error, and DTrace will propagate the error to the user 1355 * accordingly. However, there may exist some regions of memory in which an 1356 * arbitrary load can change system state, and from which it is impossible to 1357 * recover from such a load after it has been attempted. Examples of this may 1358 * include memory in which programmable I/O registers are mapped (for which a 1359 * read may have some implications for the device) or (in the specific case of 1360 * UltraSPARC-I and -II) the virtual address hole. The platform is required 1361 * to make DTrace aware of these toxic ranges; DTrace will then check that 1362 * target addresses are not in a toxic range before attempting to issue a 1363 * safe load. 1364 */ 1365 typedef struct dtrace_toxrange { 1366 uintptr_t dtt_base; /* base of toxic range */ 1367 uintptr_t dtt_limit; /* limit of toxic range */ 1368 } dtrace_toxrange_t; 1369 1370 extern uint64_t dtrace_getarg(int, int, dtrace_mstate_t*, dtrace_vstate_t*); 1371 extern int dtrace_getipl(void); 1372 extern uintptr_t dtrace_caller(int); 1373 extern uint32_t dtrace_cas32(uint32_t *, uint32_t, uint32_t); 1374 extern void *dtrace_casptr(void *, void *, void *); 1375 extern void dtrace_copyin(user_addr_t, uintptr_t, size_t, volatile uint16_t *); 1376 extern void dtrace_copyinstr(user_addr_t, uintptr_t, size_t, volatile uint16_t *); 1377 extern void dtrace_copyout(uintptr_t, user_addr_t, size_t, volatile uint16_t *); 1378 extern void dtrace_copyoutstr(uintptr_t, user_addr_t, size_t, volatile uint16_t *); 1379 extern void dtrace_getpcstack(pc_t *, int, int, uint32_t *); 1380 extern uint64_t dtrace_load64(uintptr_t); 1381 extern int dtrace_canload(uint64_t, size_t, dtrace_mstate_t*, dtrace_vstate_t*); 1382 1383 extern uint64_t dtrace_getreg(struct regs *, uint_t); 1384 extern uint64_t dtrace_getvmreg(uint_t); 1385 extern int dtrace_getstackdepth(int); 1386 extern void dtrace_getupcstack(uint64_t *, int); 1387 extern void dtrace_getufpstack(uint64_t *, uint64_t *, int); 1388 extern int dtrace_getustackdepth(void); 1389 extern uintptr_t dtrace_fulword(void *); 1390 extern uint8_t dtrace_fuword8(user_addr_t); 1391 extern uint16_t dtrace_fuword16(user_addr_t); 1392 extern uint32_t dtrace_fuword32(user_addr_t); 1393 extern uint64_t dtrace_fuword64(user_addr_t); 1394 extern int dtrace_proc_waitfor(dtrace_procdesc_t*); 1395 extern void dtrace_probe_error(dtrace_state_t *, dtrace_epid_t, int, int, 1396 int, uint64_t); 1397 extern int dtrace_assfail(const char *, const char *, int); 1398 extern int dtrace_attached(void); 1399 extern hrtime_t dtrace_gethrestime(void); 1400 1401 extern void dtrace_flush_caches(void); 1402 1403 extern void dtrace_copy(uintptr_t, uintptr_t, size_t); 1404 extern void dtrace_copystr(uintptr_t, uintptr_t, size_t, volatile uint16_t *); 1405 1406 extern void* dtrace_ptrauth_strip(void*, uint64_t); 1407 extern int dtrace_is_valid_ptrauth_key(uint64_t); 1408 1409 extern uint64_t dtrace_physmem_read(uint64_t, size_t); 1410 extern void dtrace_physmem_write(uint64_t, uint64_t, size_t); 1411 1412 extern void dtrace_livedump(char *, size_t); 1413 1414 /* 1415 * DTrace state handling 1416 */ 1417 extern minor_t dtrace_state_reserve(void); 1418 extern dtrace_state_t* dtrace_state_allocate(minor_t minor); 1419 extern dtrace_state_t* dtrace_state_get(minor_t minor); 1420 extern void dtrace_state_free(minor_t minor); 1421 1422 /* 1423 * DTrace restriction checks 1424 */ 1425 extern void dtrace_restriction_policy_load(void); 1426 extern boolean_t dtrace_is_restricted(void); 1427 extern boolean_t dtrace_are_restrictions_relaxed(void); 1428 extern boolean_t dtrace_fbt_probes_restricted(void); 1429 extern boolean_t dtrace_sdt_probes_restricted(void); 1430 extern boolean_t dtrace_can_attach_to_proc(proc_t); 1431 1432 /* 1433 * DTrace Assertions 1434 * 1435 * DTrace calls ASSERT and VERIFY from probe context. To assure that a failed 1436 * ASSERT or VERIFYdoes not induce a markedly more catastrophic failure (e.g., 1437 * one from which a dump cannot be gleaned), DTrace must define its own ASSERT 1438 * and VERIFY macros to be ones that may safely be called from probe context. 1439 * This header file must thus be included by any DTrace component that calls 1440 * ASSERT and/or VERIFY from probe context, and _only_ by those components. 1441 * (The only exception to this is kernel debugging infrastructure at user-level 1442 * that doesn't depend on calling ASSERT.) 1443 */ 1444 #undef ASSERT 1445 #undef VERIFY 1446 1447 #define VERIFY(EX) ((void)((EX) || \ 1448 dtrace_assfail(#EX, __FILE__, __LINE__))) 1449 1450 #if DEBUG 1451 #define ASSERT(EX) ((void)((EX) || \ 1452 dtrace_assfail(#EX, __FILE__, __LINE__))) 1453 #else 1454 #define ASSERT(X) ((void)0) 1455 #endif 1456 1457 #ifdef __cplusplus 1458 } 1459 #endif 1460 1461 #endif /* _SYS_DTRACE_IMPL_H */ 1462 1463