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 27 /* 28 * Copyright 2016 Joyent, Inc. 29 * Copyright (c) 2012 by Delphix. All rights reserved. 30 */ 31 32 #ifndef _SYS_DTRACE_IMPL_H 33 #define _SYS_DTRACE_IMPL_H 34 35 #ifdef __cplusplus 36 extern "C" { 37 #endif 38 39 /* 40 * DTrace Dynamic Tracing Software: Kernel Implementation Interfaces 41 * 42 * Note: The contents of this file are private to the implementation of the 43 * Solaris system and DTrace subsystem and are subject to change at any time 44 * without notice. Applications and drivers using these interfaces will fail 45 * to run on future releases. These interfaces should not be used for any 46 * purpose except those expressly outlined in dtrace(7D) and libdtrace(3LIB). 47 * Please refer to the "Solaris Dynamic Tracing Guide" for more information. 48 */ 49 50 #include <sys/dtrace.h> 51 #include <sys/file.h> 52 53 #ifndef illumos 54 typedef uintptr_t pc_t; 55 typedef u_long greg_t; 56 #endif 57 58 /* 59 * DTrace Implementation Constants and Typedefs 60 */ 61 #define DTRACE_MAXPROPLEN 128 62 #define DTRACE_DYNVAR_CHUNKSIZE 256 63 64 #ifdef __FreeBSD__ 65 #define NCPU MAXCPU 66 #endif /* __FreeBSD__ */ 67 68 struct dtrace_probe; 69 struct dtrace_ecb; 70 struct dtrace_predicate; 71 struct dtrace_action; 72 struct dtrace_provider; 73 struct dtrace_state; 74 75 typedef struct dtrace_probe dtrace_probe_t; 76 typedef struct dtrace_ecb dtrace_ecb_t; 77 typedef struct dtrace_predicate dtrace_predicate_t; 78 typedef struct dtrace_action dtrace_action_t; 79 typedef struct dtrace_provider dtrace_provider_t; 80 typedef struct dtrace_meta dtrace_meta_t; 81 typedef struct dtrace_state dtrace_state_t; 82 typedef uint32_t dtrace_optid_t; 83 typedef uint32_t dtrace_specid_t; 84 typedef uint64_t dtrace_genid_t; 85 86 /* 87 * DTrace Probes 88 * 89 * The probe is the fundamental unit of the DTrace architecture. Probes are 90 * created by DTrace providers, and managed by the DTrace framework. A probe 91 * is identified by a unique <provider, module, function, name> tuple, and has 92 * a unique probe identifier assigned to it. (Some probes are not associated 93 * with a specific point in text; these are called _unanchored probes_ and have 94 * no module or function associated with them.) Probes are represented as a 95 * dtrace_probe structure. To allow quick lookups based on each element of the 96 * probe tuple, probes are hashed by each of provider, module, function and 97 * name. (If a lookup is performed based on a regular expression, a 98 * dtrace_probekey is prepared, and a linear search is performed.) Each probe 99 * is additionally pointed to by a linear array indexed by its identifier. The 100 * identifier is the provider's mechanism for indicating to the DTrace 101 * framework that a probe has fired: the identifier is passed as the first 102 * argument to dtrace_probe(), where it is then mapped into the corresponding 103 * dtrace_probe structure. From the dtrace_probe structure, dtrace_probe() can 104 * iterate over the probe's list of enabling control blocks; see "DTrace 105 * Enabling Control Blocks", below.) 106 */ 107 struct dtrace_probe { 108 dtrace_id_t dtpr_id; /* probe identifier */ 109 dtrace_ecb_t *dtpr_ecb; /* ECB list; see below */ 110 dtrace_ecb_t *dtpr_ecb_last; /* last ECB in list */ 111 void *dtpr_arg; /* provider argument */ 112 dtrace_cacheid_t dtpr_predcache; /* predicate cache ID */ 113 int dtpr_aframes; /* artificial frames */ 114 dtrace_provider_t *dtpr_provider; /* pointer to provider */ 115 char *dtpr_mod; /* probe's module name */ 116 char *dtpr_func; /* probe's function name */ 117 char *dtpr_name; /* probe's name */ 118 dtrace_probe_t *dtpr_nextmod; /* next in module hash */ 119 dtrace_probe_t *dtpr_prevmod; /* previous in module hash */ 120 dtrace_probe_t *dtpr_nextfunc; /* next in function hash */ 121 dtrace_probe_t *dtpr_prevfunc; /* previous in function hash */ 122 dtrace_probe_t *dtpr_nextname; /* next in name hash */ 123 dtrace_probe_t *dtpr_prevname; /* previous in name hash */ 124 dtrace_genid_t dtpr_gen; /* probe generation ID */ 125 }; 126 127 typedef int dtrace_probekey_f(const char *, const char *, int); 128 129 typedef struct dtrace_probekey { 130 char *dtpk_prov; /* provider name to match */ 131 dtrace_probekey_f *dtpk_pmatch; /* provider matching function */ 132 char *dtpk_mod; /* module name to match */ 133 dtrace_probekey_f *dtpk_mmatch; /* module matching function */ 134 char *dtpk_func; /* func name to match */ 135 dtrace_probekey_f *dtpk_fmatch; /* func matching function */ 136 char *dtpk_name; /* name to match */ 137 dtrace_probekey_f *dtpk_nmatch; /* name matching function */ 138 dtrace_id_t dtpk_id; /* identifier to match */ 139 } dtrace_probekey_t; 140 141 typedef struct dtrace_hashbucket { 142 struct dtrace_hashbucket *dthb_next; /* next on hash chain */ 143 dtrace_probe_t *dthb_chain; /* chain of probes */ 144 int dthb_len; /* number of probes here */ 145 } dtrace_hashbucket_t; 146 147 typedef struct dtrace_hash { 148 dtrace_hashbucket_t **dth_tab; /* hash table */ 149 int dth_size; /* size of hash table */ 150 int dth_mask; /* mask to index into table */ 151 int dth_nbuckets; /* total number of buckets */ 152 uintptr_t dth_nextoffs; /* offset of next in probe */ 153 uintptr_t dth_prevoffs; /* offset of prev in probe */ 154 uintptr_t dth_stroffs; /* offset of str in probe */ 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_size; /* size of buffer */ 431 uint32_t dtb_flags; /* flags */ 432 uint32_t dtb_drops; /* number of drops */ 433 caddr_t dtb_tomax; /* active buffer */ 434 caddr_t dtb_xamot; /* inactive buffer */ 435 uint32_t dtb_xamot_flags; /* inactive flags */ 436 uint32_t dtb_xamot_drops; /* drops in inactive buffer */ 437 uint64_t dtb_xamot_offset; /* offset in inactive buffer */ 438 uint32_t dtb_errors; /* number of errors */ 439 uint32_t dtb_xamot_errors; /* errors in inactive buffer */ 440 #ifndef _LP64 441 uint64_t dtb_pad1; /* pad out to 64 bytes */ 442 #endif 443 uint64_t dtb_switched; /* time of last switch */ 444 uint64_t dtb_interval; /* observed switch interval */ 445 uint64_t dtb_pad2[6]; /* pad to avoid false sharing */ 446 } dtrace_buffer_t; 447 448 /* 449 * DTrace Aggregation Buffers 450 * 451 * Aggregation buffers use much of the same mechanism as described above 452 * ("DTrace Buffers"). However, because an aggregation is fundamentally a 453 * hash, there exists dynamic metadata associated with an aggregation buffer 454 * that is not associated with other kinds of buffers. This aggregation 455 * metadata is _only_ relevant for the in-kernel implementation of 456 * aggregations; it is not actually relevant to user-level consumers. To do 457 * this, we allocate dynamic aggregation data (hash keys and hash buckets) 458 * starting below the _limit_ of the buffer, and we allocate data from the 459 * _base_ of the buffer. When the aggregation buffer is copied out, _only_ the 460 * data is copied out; the metadata is simply discarded. Schematically, 461 * aggregation buffers look like: 462 * 463 * base of data buffer ---> +-------+------+-----------+-------+ 464 * | aggid | key | value | aggid | 465 * +-------+------+-----------+-------+ 466 * | key | 467 * +-------+-------+-----+------------+ 468 * | value | aggid | key | value | 469 * +-------+------++-----+------+-----+ 470 * | aggid | key | value | | 471 * +-------+------+-------------+ | 472 * | || | 473 * | || | 474 * | \/ | 475 * : : 476 * . . 477 * . . 478 * . . 479 * : : 480 * | /\ | 481 * | || +------------+ 482 * | || | | 483 * +---------------------+ | 484 * | hash keys | 485 * | (dtrace_aggkey structures) | 486 * | | 487 * +----------------------------------+ 488 * | hash buckets | 489 * | (dtrace_aggbuffer structure) | 490 * | | 491 * limit of data buffer ---> +----------------------------------+ 492 * 493 * 494 * As implied above, just as we assure that ECBs always store a constant 495 * amount of data, we assure that a given aggregation -- identified by its 496 * aggregation ID -- always stores data of a constant quantity and type. 497 * As with EPIDs, this allows the aggregation ID to serve as the metadata for a 498 * given record. 499 * 500 * Note that the size of the dtrace_aggkey structure must be sizeof (uintptr_t) 501 * aligned. (If this the structure changes such that this becomes false, an 502 * assertion will fail in dtrace_aggregate().) 503 */ 504 typedef struct dtrace_aggkey { 505 uint32_t dtak_hashval; /* hash value */ 506 uint32_t dtak_action:4; /* action -- 4 bits */ 507 uint32_t dtak_size:28; /* size -- 28 bits */ 508 caddr_t dtak_data; /* data pointer */ 509 struct dtrace_aggkey *dtak_next; /* next in hash chain */ 510 } dtrace_aggkey_t; 511 512 typedef struct dtrace_aggbuffer { 513 uintptr_t dtagb_hashsize; /* number of buckets */ 514 uintptr_t dtagb_free; /* free list of keys */ 515 dtrace_aggkey_t **dtagb_hash; /* hash table */ 516 } dtrace_aggbuffer_t; 517 518 /* 519 * DTrace Speculations 520 * 521 * Speculations have a per-CPU buffer and a global state. Once a speculation 522 * buffer has been comitted or discarded, it cannot be reused until all CPUs 523 * have taken the same action (commit or discard) on their respective 524 * speculative buffer. However, because DTrace probes may execute in arbitrary 525 * context, other CPUs cannot simply be cross-called at probe firing time to 526 * perform the necessary commit or discard. The speculation states thus 527 * optimize for the case that a speculative buffer is only active on one CPU at 528 * the time of a commit() or discard() -- for if this is the case, other CPUs 529 * need not take action, and the speculation is immediately available for 530 * reuse. If the speculation is active on multiple CPUs, it must be 531 * asynchronously cleaned -- potentially leading to a higher rate of dirty 532 * speculative drops. The speculation states are as follows: 533 * 534 * DTRACESPEC_INACTIVE <= Initial state; inactive speculation 535 * DTRACESPEC_ACTIVE <= Allocated, but not yet speculatively traced to 536 * DTRACESPEC_ACTIVEONE <= Speculatively traced to on one CPU 537 * DTRACESPEC_ACTIVEMANY <= Speculatively traced to on more than one CPU 538 * DTRACESPEC_COMMITTING <= Currently being commited on one CPU 539 * DTRACESPEC_COMMITTINGMANY <= Currently being commited on many CPUs 540 * DTRACESPEC_DISCARDING <= Currently being discarded on many CPUs 541 * 542 * The state transition diagram is as follows: 543 * 544 * +----------------------------------------------------------+ 545 * | | 546 * | +------------+ | 547 * | +-------------------| COMMITTING |<-----------------+ | 548 * | | +------------+ | | 549 * | | copied spec. ^ commit() on | | discard() on 550 * | | into principal | active CPU | | active CPU 551 * | | | commit() | | 552 * V V | | | 553 * +----------+ +--------+ +-----------+ 554 * | INACTIVE |---------------->| ACTIVE |--------------->| ACTIVEONE | 555 * +----------+ speculation() +--------+ speculate() +-----------+ 556 * ^ ^ | | | 557 * | | | discard() | | 558 * | | asynchronously | discard() on | | speculate() 559 * | | cleaned V inactive CPU | | on inactive 560 * | | +------------+ | | CPU 561 * | +-------------------| DISCARDING |<-----------------+ | 562 * | +------------+ | 563 * | asynchronously ^ | 564 * | copied spec. | discard() | 565 * | into principal +------------------------+ | 566 * | | V 567 * +----------------+ commit() +------------+ 568 * | COMMITTINGMANY |<----------------------------------| ACTIVEMANY | 569 * +----------------+ +------------+ 570 */ 571 typedef enum dtrace_speculation_state { 572 DTRACESPEC_INACTIVE = 0, 573 DTRACESPEC_ACTIVE, 574 DTRACESPEC_ACTIVEONE, 575 DTRACESPEC_ACTIVEMANY, 576 DTRACESPEC_COMMITTING, 577 DTRACESPEC_COMMITTINGMANY, 578 DTRACESPEC_DISCARDING 579 } dtrace_speculation_state_t; 580 581 typedef struct dtrace_speculation { 582 dtrace_speculation_state_t dtsp_state; /* current speculation state */ 583 int dtsp_cleaning; /* non-zero if being cleaned */ 584 dtrace_buffer_t *dtsp_buffer; /* speculative buffer */ 585 } dtrace_speculation_t; 586 587 /* 588 * DTrace Dynamic Variables 589 * 590 * The dynamic variable problem is obviously decomposed into two subproblems: 591 * allocating new dynamic storage, and freeing old dynamic storage. The 592 * presence of the second problem makes the first much more complicated -- or 593 * rather, the absence of the second renders the first trivial. This is the 594 * case with aggregations, for which there is effectively no deallocation of 595 * dynamic storage. (Or more accurately, all dynamic storage is deallocated 596 * when a snapshot is taken of the aggregation.) As DTrace dynamic variables 597 * allow for both dynamic allocation and dynamic deallocation, the 598 * implementation of dynamic variables is quite a bit more complicated than 599 * that of their aggregation kin. 600 * 601 * We observe that allocating new dynamic storage is tricky only because the 602 * size can vary -- the allocation problem is much easier if allocation sizes 603 * are uniform. We further observe that in D, the size of dynamic variables is 604 * actually _not_ dynamic -- dynamic variable sizes may be determined by static 605 * analysis of DIF text. (This is true even of putatively dynamically-sized 606 * objects like strings and stacks, the sizes of which are dictated by the 607 * "stringsize" and "stackframes" variables, respectively.) We exploit this by 608 * performing this analysis on all DIF before enabling any probes. For each 609 * dynamic load or store, we calculate the dynamically-allocated size plus the 610 * size of the dtrace_dynvar structure plus the storage required to key the 611 * data. For all DIF, we take the largest value and dub it the _chunksize_. 612 * We then divide dynamic memory into two parts: a hash table that is wide 613 * enough to have every chunk in its own bucket, and a larger region of equal 614 * chunksize units. Whenever we wish to dynamically allocate a variable, we 615 * always allocate a single chunk of memory. Depending on the uniformity of 616 * allocation, this will waste some amount of memory -- but it eliminates the 617 * non-determinism inherent in traditional heap fragmentation. 618 * 619 * Dynamic objects are allocated by storing a non-zero value to them; they are 620 * deallocated by storing a zero value to them. Dynamic variables are 621 * complicated enormously by being shared between CPUs. In particular, 622 * consider the following scenario: 623 * 624 * CPU A CPU B 625 * +---------------------------------+ +---------------------------------+ 626 * | | | | 627 * | allocates dynamic object a[123] | | | 628 * | by storing the value 345 to it | | | 629 * | ---------> | 630 * | | | wishing to load from object | 631 * | | | a[123], performs lookup in | 632 * | | | dynamic variable space | 633 * | <--------- | 634 * | deallocates object a[123] by | | | 635 * | storing 0 to it | | | 636 * | | | | 637 * | allocates dynamic object b[567] | | performs load from a[123] | 638 * | by storing the value 789 to it | | | 639 * : : : : 640 * . . . . 641 * 642 * This is obviously a race in the D program, but there are nonetheless only 643 * two valid values for CPU B's load from a[123]: 345 or 0. Most importantly, 644 * CPU B may _not_ see the value 789 for a[123]. 645 * 646 * There are essentially two ways to deal with this: 647 * 648 * (1) Explicitly spin-lock variables. That is, if CPU B wishes to load 649 * from a[123], it needs to lock a[123] and hold the lock for the 650 * duration that it wishes to manipulate it. 651 * 652 * (2) Avoid reusing freed chunks until it is known that no CPU is referring 653 * to them. 654 * 655 * The implementation of (1) is rife with complexity, because it requires the 656 * user of a dynamic variable to explicitly decree when they are done using it. 657 * Were all variables by value, this perhaps wouldn't be debilitating -- but 658 * dynamic variables of non-scalar types are tracked by reference. That is, if 659 * a dynamic variable is, say, a string, and that variable is to be traced to, 660 * say, the principal buffer, the DIF emulation code returns to the main 661 * dtrace_probe() loop a pointer to the underlying storage, not the contents of 662 * the storage. Further, code calling on DIF emulation would have to be aware 663 * that the DIF emulation has returned a reference to a dynamic variable that 664 * has been potentially locked. The variable would have to be unlocked after 665 * the main dtrace_probe() loop is finished with the variable, and the main 666 * dtrace_probe() loop would have to be careful to not call any further DIF 667 * emulation while the variable is locked to avoid deadlock. More generally, 668 * if one were to implement (1), DIF emulation code dealing with dynamic 669 * variables could only deal with one dynamic variable at a time (lest deadlock 670 * result). To sum, (1) exports too much subtlety to the users of dynamic 671 * variables -- increasing maintenance burden and imposing serious constraints 672 * on future DTrace development. 673 * 674 * The implementation of (2) is also complex, but the complexity is more 675 * manageable. We need to be sure that when a variable is deallocated, it is 676 * not placed on a traditional free list, but rather on a _dirty_ list. Once a 677 * variable is on a dirty list, it cannot be found by CPUs performing a 678 * subsequent lookup of the variable -- but it may still be in use by other 679 * CPUs. To assure that all CPUs that may be seeing the old variable have 680 * cleared out of probe context, a dtrace_sync() can be issued. Once the 681 * dtrace_sync() has completed, it can be known that all CPUs are done 682 * manipulating the dynamic variable -- the dirty list can be atomically 683 * appended to the free list. Unfortunately, there's a slight hiccup in this 684 * mechanism: dtrace_sync() may not be issued from probe context. The 685 * dtrace_sync() must be therefore issued asynchronously from non-probe 686 * context. For this we rely on the DTrace cleaner, a cyclic that runs at the 687 * "cleanrate" frequency. To ease this implementation, we define several chunk 688 * lists: 689 * 690 * - Dirty. Deallocated chunks, not yet cleaned. Not available. 691 * 692 * - Rinsing. Formerly dirty chunks that are currently being asynchronously 693 * cleaned. Not available, but will be shortly. Dynamic variable 694 * allocation may not spin or block for availability, however. 695 * 696 * - Clean. Clean chunks, ready for allocation -- but not on the free list. 697 * 698 * - Free. Available for allocation. 699 * 700 * Moreover, to avoid absurd contention, _each_ of these lists is implemented 701 * on a per-CPU basis. This is only for performance, not correctness; chunks 702 * may be allocated from another CPU's free list. The algorithm for allocation 703 * then is this: 704 * 705 * (1) Attempt to atomically allocate from current CPU's free list. If list 706 * is non-empty and allocation is successful, allocation is complete. 707 * 708 * (2) If the clean list is non-empty, atomically move it to the free list, 709 * and reattempt (1). 710 * 711 * (3) If the dynamic variable space is in the CLEAN state, look for free 712 * and clean lists on other CPUs by setting the current CPU to the next 713 * CPU, and reattempting (1). If the next CPU is the current CPU (that 714 * is, if all CPUs have been checked), atomically switch the state of 715 * the dynamic variable space based on the following: 716 * 717 * - If no free chunks were found and no dirty chunks were found, 718 * atomically set the state to EMPTY. 719 * 720 * - If dirty chunks were found, atomically set the state to DIRTY. 721 * 722 * - If rinsing chunks were found, atomically set the state to RINSING. 723 * 724 * (4) Based on state of dynamic variable space state, increment appropriate 725 * counter to indicate dynamic drops (if in EMPTY state) vs. dynamic 726 * dirty drops (if in DIRTY state) vs. dynamic rinsing drops (if in 727 * RINSING state). Fail the allocation. 728 * 729 * The cleaning cyclic operates with the following algorithm: for all CPUs 730 * with a non-empty dirty list, atomically move the dirty list to the rinsing 731 * list. Perform a dtrace_sync(). For all CPUs with a non-empty rinsing list, 732 * atomically move the rinsing list to the clean list. Perform another 733 * dtrace_sync(). By this point, all CPUs have seen the new clean list; the 734 * state of the dynamic variable space can be restored to CLEAN. 735 * 736 * There exist two final races that merit explanation. The first is a simple 737 * allocation race: 738 * 739 * CPU A CPU B 740 * +---------------------------------+ +---------------------------------+ 741 * | | | | 742 * | allocates dynamic object a[123] | | allocates dynamic object a[123] | 743 * | by storing the value 345 to it | | by storing the value 567 to it | 744 * | | | | 745 * : : : : 746 * . . . . 747 * 748 * Again, this is a race in the D program. It can be resolved by having a[123] 749 * hold the value 345 or a[123] hold the value 567 -- but it must be true that 750 * a[123] have only _one_ of these values. (That is, the racing CPUs may not 751 * put the same element twice on the same hash chain.) This is resolved 752 * simply: before the allocation is undertaken, the start of the new chunk's 753 * hash chain is noted. Later, after the allocation is complete, the hash 754 * chain is atomically switched to point to the new element. If this fails 755 * (because of either concurrent allocations or an allocation concurrent with a 756 * deletion), the newly allocated chunk is deallocated to the dirty list, and 757 * the whole process of looking up (and potentially allocating) the dynamic 758 * variable is reattempted. 759 * 760 * The final race is a simple deallocation race: 761 * 762 * CPU A CPU B 763 * +---------------------------------+ +---------------------------------+ 764 * | | | | 765 * | deallocates dynamic object | | deallocates dynamic object | 766 * | a[123] by storing the value 0 | | a[123] by storing the value 0 | 767 * | to it | | to it | 768 * | | | | 769 * : : : : 770 * . . . . 771 * 772 * Once again, this is a race in the D program, but it is one that we must 773 * handle without corrupting the underlying data structures. Because 774 * deallocations require the deletion of a chunk from the middle of a hash 775 * chain, we cannot use a single-word atomic operation to remove it. For this, 776 * we add a spin lock to the hash buckets that is _only_ used for deallocations 777 * (allocation races are handled as above). Further, this spin lock is _only_ 778 * held for the duration of the delete; before control is returned to the DIF 779 * emulation code, the hash bucket is unlocked. 780 */ 781 typedef struct dtrace_key { 782 uint64_t dttk_value; /* data value or data pointer */ 783 uint64_t dttk_size; /* 0 if by-val, >0 if by-ref */ 784 } dtrace_key_t; 785 786 typedef struct dtrace_tuple { 787 uint32_t dtt_nkeys; /* number of keys in tuple */ 788 uint32_t dtt_pad; /* padding */ 789 dtrace_key_t dtt_key[1]; /* array of tuple keys */ 790 } dtrace_tuple_t; 791 792 typedef struct dtrace_dynvar { 793 uint64_t dtdv_hashval; /* hash value -- 0 if free */ 794 struct dtrace_dynvar *dtdv_next; /* next on list or hash chain */ 795 void *dtdv_data; /* pointer to data */ 796 dtrace_tuple_t dtdv_tuple; /* tuple key */ 797 } dtrace_dynvar_t; 798 799 typedef enum dtrace_dynvar_op { 800 DTRACE_DYNVAR_ALLOC, 801 DTRACE_DYNVAR_NOALLOC, 802 DTRACE_DYNVAR_DEALLOC 803 } dtrace_dynvar_op_t; 804 805 typedef struct dtrace_dynhash { 806 dtrace_dynvar_t *dtdh_chain; /* hash chain for this bucket */ 807 uintptr_t dtdh_lock; /* deallocation lock */ 808 #ifdef _LP64 809 uintptr_t dtdh_pad[6]; /* pad to avoid false sharing */ 810 #else 811 uintptr_t dtdh_pad[14]; /* pad to avoid false sharing */ 812 #endif 813 } dtrace_dynhash_t; 814 815 typedef struct dtrace_dstate_percpu { 816 dtrace_dynvar_t *dtdsc_free; /* free list for this CPU */ 817 dtrace_dynvar_t *dtdsc_dirty; /* dirty list for this CPU */ 818 dtrace_dynvar_t *dtdsc_rinsing; /* rinsing list for this CPU */ 819 dtrace_dynvar_t *dtdsc_clean; /* clean list for this CPU */ 820 uint64_t dtdsc_drops; /* number of capacity drops */ 821 uint64_t dtdsc_dirty_drops; /* number of dirty drops */ 822 uint64_t dtdsc_rinsing_drops; /* number of rinsing drops */ 823 #ifdef _LP64 824 uint64_t dtdsc_pad; /* pad to avoid false sharing */ 825 #else 826 uint64_t dtdsc_pad[2]; /* pad to avoid false sharing */ 827 #endif 828 } dtrace_dstate_percpu_t; 829 830 typedef enum dtrace_dstate_state { 831 DTRACE_DSTATE_CLEAN = 0, 832 DTRACE_DSTATE_EMPTY, 833 DTRACE_DSTATE_DIRTY, 834 DTRACE_DSTATE_RINSING 835 } dtrace_dstate_state_t; 836 837 typedef struct dtrace_dstate { 838 void *dtds_base; /* base of dynamic var. space */ 839 size_t dtds_size; /* size of dynamic var. space */ 840 size_t dtds_hashsize; /* number of buckets in hash */ 841 size_t dtds_chunksize; /* size of each chunk */ 842 dtrace_dynhash_t *dtds_hash; /* pointer to hash table */ 843 dtrace_dstate_state_t dtds_state; /* current dynamic var. state */ 844 dtrace_dstate_percpu_t *dtds_percpu; /* per-CPU dyn. var. state */ 845 } dtrace_dstate_t; 846 847 /* 848 * DTrace Variable State 849 * 850 * The DTrace variable state tracks user-defined variables in its dtrace_vstate 851 * structure. Each DTrace consumer has exactly one dtrace_vstate structure, 852 * but some dtrace_vstate structures may exist without a corresponding DTrace 853 * consumer (see "DTrace Helpers", below). As described in <sys/dtrace.h>, 854 * user-defined variables can have one of three scopes: 855 * 856 * DIFV_SCOPE_GLOBAL => global scope 857 * DIFV_SCOPE_THREAD => thread-local scope (i.e. "self->" variables) 858 * DIFV_SCOPE_LOCAL => clause-local scope (i.e. "this->" variables) 859 * 860 * The variable state tracks variables by both their scope and their allocation 861 * type: 862 * 863 * - The dtvs_globals and dtvs_locals members each point to an array of 864 * dtrace_statvar structures. These structures contain both the variable 865 * metadata (dtrace_difv structures) and the underlying storage for all 866 * statically allocated variables, including statically allocated 867 * DIFV_SCOPE_GLOBAL variables and all DIFV_SCOPE_LOCAL variables. 868 * 869 * - The dtvs_tlocals member points to an array of dtrace_difv structures for 870 * DIFV_SCOPE_THREAD variables. As such, this array tracks _only_ the 871 * variable metadata for DIFV_SCOPE_THREAD variables; the underlying storage 872 * is allocated out of the dynamic variable space. 873 * 874 * - The dtvs_dynvars member is the dynamic variable state associated with the 875 * variable state. The dynamic variable state (described in "DTrace Dynamic 876 * Variables", above) tracks all DIFV_SCOPE_THREAD variables and all 877 * dynamically-allocated DIFV_SCOPE_GLOBAL variables. 878 */ 879 typedef struct dtrace_statvar { 880 uint64_t dtsv_data; /* data or pointer to it */ 881 size_t dtsv_size; /* size of pointed-to data */ 882 int dtsv_refcnt; /* reference count */ 883 dtrace_difv_t dtsv_var; /* variable metadata */ 884 } dtrace_statvar_t; 885 886 typedef struct dtrace_vstate { 887 dtrace_state_t *dtvs_state; /* back pointer to state */ 888 dtrace_statvar_t **dtvs_globals; /* statically-allocated glbls */ 889 int dtvs_nglobals; /* number of globals */ 890 dtrace_difv_t *dtvs_tlocals; /* thread-local metadata */ 891 int dtvs_ntlocals; /* number of thread-locals */ 892 dtrace_statvar_t **dtvs_locals; /* clause-local data */ 893 int dtvs_nlocals; /* number of clause-locals */ 894 dtrace_dstate_t dtvs_dynvars; /* dynamic variable state */ 895 } dtrace_vstate_t; 896 897 /* 898 * DTrace Machine State 899 * 900 * In the process of processing a fired probe, DTrace needs to track and/or 901 * cache some per-CPU state associated with that particular firing. This is 902 * state that is always discarded after the probe firing has completed, and 903 * much of it is not specific to any DTrace consumer, remaining valid across 904 * all ECBs. This state is tracked in the dtrace_mstate structure. 905 */ 906 #define DTRACE_MSTATE_ARGS 0x00000001 907 #define DTRACE_MSTATE_PROBE 0x00000002 908 #define DTRACE_MSTATE_EPID 0x00000004 909 #define DTRACE_MSTATE_TIMESTAMP 0x00000008 910 #define DTRACE_MSTATE_STACKDEPTH 0x00000010 911 #define DTRACE_MSTATE_CALLER 0x00000020 912 #define DTRACE_MSTATE_IPL 0x00000040 913 #define DTRACE_MSTATE_FLTOFFS 0x00000080 914 #define DTRACE_MSTATE_WALLTIMESTAMP 0x00000100 915 #define DTRACE_MSTATE_USTACKDEPTH 0x00000200 916 #define DTRACE_MSTATE_UCALLER 0x00000400 917 918 typedef struct dtrace_mstate { 919 uintptr_t dtms_scratch_base; /* base of scratch space */ 920 uintptr_t dtms_scratch_ptr; /* current scratch pointer */ 921 size_t dtms_scratch_size; /* scratch size */ 922 uint32_t dtms_present; /* variables that are present */ 923 uint64_t dtms_arg[5]; /* cached arguments */ 924 dtrace_epid_t dtms_epid; /* current EPID */ 925 uint64_t dtms_timestamp; /* cached timestamp */ 926 hrtime_t dtms_walltimestamp; /* cached wall timestamp */ 927 int dtms_stackdepth; /* cached stackdepth */ 928 int dtms_ustackdepth; /* cached ustackdepth */ 929 struct dtrace_probe *dtms_probe; /* current probe */ 930 uintptr_t dtms_caller; /* cached caller */ 931 uint64_t dtms_ucaller; /* cached user-level caller */ 932 int dtms_ipl; /* cached interrupt pri lev */ 933 int dtms_fltoffs; /* faulting DIFO offset */ 934 uintptr_t dtms_strtok; /* saved strtok() pointer */ 935 uintptr_t dtms_strtok_limit; /* upper bound of strtok ptr */ 936 uint32_t dtms_access; /* memory access rights */ 937 dtrace_difo_t *dtms_difo; /* current dif object */ 938 file_t *dtms_getf; /* cached rval of getf() */ 939 } dtrace_mstate_t; 940 941 #define DTRACE_COND_OWNER 0x1 942 #define DTRACE_COND_USERMODE 0x2 943 #define DTRACE_COND_ZONEOWNER 0x4 944 945 #define DTRACE_PROBEKEY_MAXDEPTH 8 /* max glob recursion depth */ 946 947 /* 948 * Access flag used by dtrace_mstate.dtms_access. 949 */ 950 #define DTRACE_ACCESS_KERNEL 0x1 /* the priv to read kmem */ 951 952 953 /* 954 * DTrace Activity 955 * 956 * Each DTrace consumer is in one of several states, which (for purposes of 957 * avoiding yet-another overloading of the noun "state") we call the current 958 * _activity_. The activity transitions on dtrace_go() (from DTRACIOCGO), on 959 * dtrace_stop() (from DTRACIOCSTOP) and on the exit() action. Activities may 960 * only transition in one direction; the activity transition diagram is a 961 * directed acyclic graph. The activity transition diagram is as follows: 962 * 963 * 964 * +----------+ +--------+ +--------+ 965 * | INACTIVE |------------------>| WARMUP |------------------>| ACTIVE | 966 * +----------+ dtrace_go(), +--------+ dtrace_go(), +--------+ 967 * before BEGIN | after BEGIN | | | 968 * | | | | 969 * exit() action | | | | 970 * from BEGIN ECB | | | | 971 * | | | | 972 * v | | | 973 * +----------+ exit() action | | | 974 * +-----------------------------| DRAINING |<-------------------+ | | 975 * | +----------+ | | 976 * | | | | 977 * | dtrace_stop(), | | | 978 * | before END | | | 979 * | | | | 980 * | v | | 981 * | +---------+ +----------+ | | 982 * | | STOPPED |<----------------| COOLDOWN |<----------------------+ | 983 * | +---------+ dtrace_stop(), +----------+ dtrace_stop(), | 984 * | after END before END | 985 * | | 986 * | +--------+ | 987 * +----------------------------->| KILLED |<--------------------------+ 988 * deadman timeout or +--------+ deadman timeout or 989 * killed consumer killed consumer 990 * 991 * Note that once a DTrace consumer has stopped tracing, there is no way to 992 * restart it; if a DTrace consumer wishes to restart tracing, it must reopen 993 * the DTrace pseudodevice. 994 */ 995 typedef enum dtrace_activity { 996 DTRACE_ACTIVITY_INACTIVE = 0, /* not yet running */ 997 DTRACE_ACTIVITY_WARMUP, /* while starting */ 998 DTRACE_ACTIVITY_ACTIVE, /* running */ 999 DTRACE_ACTIVITY_DRAINING, /* before stopping */ 1000 DTRACE_ACTIVITY_COOLDOWN, /* while stopping */ 1001 DTRACE_ACTIVITY_STOPPED, /* after stopping */ 1002 DTRACE_ACTIVITY_KILLED /* killed */ 1003 } dtrace_activity_t; 1004 1005 /* 1006 * DTrace Helper Implementation 1007 * 1008 * A description of the helper architecture may be found in <sys/dtrace.h>. 1009 * Each process contains a pointer to its helpers in its p_dtrace_helpers 1010 * member. This is a pointer to a dtrace_helpers structure, which contains an 1011 * array of pointers to dtrace_helper structures, helper variable state (shared 1012 * among a process's helpers) and a generation count. (The generation count is 1013 * used to provide an identifier when a helper is added so that it may be 1014 * subsequently removed.) The dtrace_helper structure is self-explanatory, 1015 * containing pointers to the objects needed to execute the helper. Note that 1016 * helpers are _duplicated_ across fork(2), and destroyed on exec(2). No more 1017 * than dtrace_helpers_max are allowed per-process. 1018 */ 1019 #define DTRACE_HELPER_ACTION_USTACK 0 1020 #define DTRACE_NHELPER_ACTIONS 1 1021 1022 typedef struct dtrace_helper_action { 1023 int dtha_generation; /* helper action generation */ 1024 int dtha_nactions; /* number of actions */ 1025 dtrace_difo_t *dtha_predicate; /* helper action predicate */ 1026 dtrace_difo_t **dtha_actions; /* array of actions */ 1027 struct dtrace_helper_action *dtha_next; /* next helper action */ 1028 } dtrace_helper_action_t; 1029 1030 typedef struct dtrace_helper_provider { 1031 int dthp_generation; /* helper provider generation */ 1032 uint32_t dthp_ref; /* reference count */ 1033 dof_helper_t dthp_prov; /* DOF w/ provider and probes */ 1034 } dtrace_helper_provider_t; 1035 1036 typedef struct dtrace_helpers { 1037 dtrace_helper_action_t **dthps_actions; /* array of helper actions */ 1038 dtrace_vstate_t dthps_vstate; /* helper action var. state */ 1039 dtrace_helper_provider_t **dthps_provs; /* array of providers */ 1040 uint_t dthps_nprovs; /* count of providers */ 1041 uint_t dthps_maxprovs; /* provider array size */ 1042 int dthps_generation; /* current generation */ 1043 pid_t dthps_pid; /* pid of associated proc */ 1044 int dthps_deferred; /* helper in deferred list */ 1045 struct dtrace_helpers *dthps_next; /* next pointer */ 1046 struct dtrace_helpers *dthps_prev; /* prev pointer */ 1047 } dtrace_helpers_t; 1048 1049 /* 1050 * DTrace Helper Action Tracing 1051 * 1052 * Debugging helper actions can be arduous. To ease the development and 1053 * debugging of helpers, DTrace contains a tracing-framework-within-a-tracing- 1054 * framework: helper tracing. If dtrace_helptrace_enabled is non-zero (which 1055 * it is by default on DEBUG kernels), all helper activity will be traced to a 1056 * global, in-kernel ring buffer. Each entry includes a pointer to the specific 1057 * helper, the location within the helper, and a trace of all local variables. 1058 * The ring buffer may be displayed in a human-readable format with the 1059 * ::dtrace_helptrace mdb(1) dcmd. 1060 */ 1061 #define DTRACE_HELPTRACE_NEXT (-1) 1062 #define DTRACE_HELPTRACE_DONE (-2) 1063 #define DTRACE_HELPTRACE_ERR (-3) 1064 1065 typedef struct dtrace_helptrace { 1066 dtrace_helper_action_t *dtht_helper; /* helper action */ 1067 int dtht_where; /* where in helper action */ 1068 int dtht_nlocals; /* number of locals */ 1069 int dtht_fault; /* type of fault (if any) */ 1070 int dtht_fltoffs; /* DIF offset */ 1071 uint64_t dtht_illval; /* faulting value */ 1072 uint64_t dtht_locals[1]; /* local variables */ 1073 } dtrace_helptrace_t; 1074 1075 /* 1076 * DTrace Credentials 1077 * 1078 * In probe context, we have limited flexibility to examine the credentials 1079 * of the DTrace consumer that created a particular enabling. We use 1080 * the Least Privilege interfaces to cache the consumer's cred pointer and 1081 * some facts about that credential in a dtrace_cred_t structure. These 1082 * can limit the consumer's breadth of visibility and what actions the 1083 * consumer may take. 1084 */ 1085 #define DTRACE_CRV_ALLPROC 0x01 1086 #define DTRACE_CRV_KERNEL 0x02 1087 #define DTRACE_CRV_ALLZONE 0x04 1088 1089 #define DTRACE_CRV_ALL (DTRACE_CRV_ALLPROC | DTRACE_CRV_KERNEL | \ 1090 DTRACE_CRV_ALLZONE) 1091 1092 #define DTRACE_CRA_PROC 0x0001 1093 #define DTRACE_CRA_PROC_CONTROL 0x0002 1094 #define DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER 0x0004 1095 #define DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE 0x0008 1096 #define DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG 0x0010 1097 #define DTRACE_CRA_KERNEL 0x0020 1098 #define DTRACE_CRA_KERNEL_DESTRUCTIVE 0x0040 1099 1100 #define DTRACE_CRA_ALL (DTRACE_CRA_PROC | \ 1101 DTRACE_CRA_PROC_CONTROL | \ 1102 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER | \ 1103 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE | \ 1104 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG | \ 1105 DTRACE_CRA_KERNEL | \ 1106 DTRACE_CRA_KERNEL_DESTRUCTIVE) 1107 1108 typedef struct dtrace_cred { 1109 cred_t *dcr_cred; 1110 uint8_t dcr_destructive; 1111 uint8_t dcr_visible; 1112 uint16_t dcr_action; 1113 } dtrace_cred_t; 1114 1115 /* 1116 * DTrace Consumer State 1117 * 1118 * Each DTrace consumer has an associated dtrace_state structure that contains 1119 * its in-kernel DTrace state -- including options, credentials, statistics and 1120 * pointers to ECBs, buffers, speculations and formats. A dtrace_state 1121 * structure is also allocated for anonymous enablings. When anonymous state 1122 * is grabbed, the grabbing consumers dts_anon pointer is set to the grabbed 1123 * dtrace_state structure. 1124 */ 1125 struct dtrace_state { 1126 #ifdef illumos 1127 dev_t dts_dev; /* device */ 1128 #else 1129 struct cdev *dts_dev; /* device */ 1130 #endif 1131 int dts_necbs; /* total number of ECBs */ 1132 dtrace_ecb_t **dts_ecbs; /* array of ECBs */ 1133 dtrace_epid_t dts_epid; /* next EPID to allocate */ 1134 size_t dts_needed; /* greatest needed space */ 1135 struct dtrace_state *dts_anon; /* anon. state, if grabbed */ 1136 dtrace_activity_t dts_activity; /* current activity */ 1137 dtrace_vstate_t dts_vstate; /* variable state */ 1138 dtrace_buffer_t *dts_buffer; /* principal buffer */ 1139 dtrace_buffer_t *dts_aggbuffer; /* aggregation buffer */ 1140 dtrace_speculation_t *dts_speculations; /* speculation array */ 1141 int dts_nspeculations; /* number of speculations */ 1142 int dts_naggregations; /* number of aggregations */ 1143 dtrace_aggregation_t **dts_aggregations; /* aggregation array */ 1144 #ifdef illumos 1145 vmem_t *dts_aggid_arena; /* arena for aggregation IDs */ 1146 #else 1147 struct unrhdr *dts_aggid_arena; /* arena for aggregation IDs */ 1148 #endif 1149 uint64_t dts_errors; /* total number of errors */ 1150 uint32_t dts_speculations_busy; /* number of spec. busy */ 1151 uint32_t dts_speculations_unavail; /* number of spec unavail */ 1152 uint32_t dts_stkstroverflows; /* stack string tab overflows */ 1153 uint32_t dts_dblerrors; /* errors in ERROR probes */ 1154 uint32_t dts_reserve; /* space reserved for END */ 1155 hrtime_t dts_laststatus; /* time of last status */ 1156 #ifdef illumos 1157 cyclic_id_t dts_cleaner; /* cleaning cyclic */ 1158 cyclic_id_t dts_deadman; /* deadman cyclic */ 1159 #else 1160 struct callout dts_cleaner; /* Cleaning callout. */ 1161 struct callout dts_deadman; /* Deadman callout. */ 1162 #endif 1163 hrtime_t dts_alive; /* time last alive */ 1164 char dts_speculates; /* boolean: has speculations */ 1165 char dts_destructive; /* boolean: has dest. actions */ 1166 int dts_nformats; /* number of formats */ 1167 char **dts_formats; /* format string array */ 1168 dtrace_optval_t dts_options[DTRACEOPT_MAX]; /* options */ 1169 dtrace_cred_t dts_cred; /* credentials */ 1170 size_t dts_nretained; /* number of retained enabs */ 1171 int dts_getf; /* number of getf() calls */ 1172 uint64_t dts_rstate[NCPU][2]; /* per-CPU random state */ 1173 }; 1174 1175 struct dtrace_provider { 1176 dtrace_pattr_t dtpv_attr; /* provider attributes */ 1177 dtrace_ppriv_t dtpv_priv; /* provider privileges */ 1178 dtrace_pops_t dtpv_pops; /* provider operations */ 1179 char *dtpv_name; /* provider name */ 1180 void *dtpv_arg; /* provider argument */ 1181 hrtime_t dtpv_defunct; /* when made defunct */ 1182 struct dtrace_provider *dtpv_next; /* next provider */ 1183 }; 1184 1185 struct dtrace_meta { 1186 dtrace_mops_t dtm_mops; /* meta provider operations */ 1187 char *dtm_name; /* meta provider name */ 1188 void *dtm_arg; /* meta provider user arg */ 1189 uint64_t dtm_count; /* no. of associated provs. */ 1190 }; 1191 1192 /* 1193 * DTrace Enablings 1194 * 1195 * A dtrace_enabling structure is used to track a collection of ECB 1196 * descriptions -- before they have been turned into actual ECBs. This is 1197 * created as a result of DOF processing, and is generally used to generate 1198 * ECBs immediately thereafter. However, enablings are also generally 1199 * retained should the probes they describe be created at a later time; as 1200 * each new module or provider registers with the framework, the retained 1201 * enablings are reevaluated, with any new match resulting in new ECBs. To 1202 * prevent probes from being matched more than once, the enabling tracks the 1203 * last probe generation matched, and only matches probes from subsequent 1204 * generations. 1205 */ 1206 typedef struct dtrace_enabling { 1207 dtrace_ecbdesc_t **dten_desc; /* all ECB descriptions */ 1208 int dten_ndesc; /* number of ECB descriptions */ 1209 int dten_maxdesc; /* size of ECB array */ 1210 dtrace_vstate_t *dten_vstate; /* associated variable state */ 1211 dtrace_genid_t dten_probegen; /* matched probe generation */ 1212 dtrace_ecbdesc_t *dten_current; /* current ECB description */ 1213 int dten_error; /* current error value */ 1214 int dten_primed; /* boolean: set if primed */ 1215 struct dtrace_enabling *dten_prev; /* previous enabling */ 1216 struct dtrace_enabling *dten_next; /* next enabling */ 1217 } dtrace_enabling_t; 1218 1219 /* 1220 * DTrace Anonymous Enablings 1221 * 1222 * Anonymous enablings are DTrace enablings that are not associated with a 1223 * controlling process, but rather derive their enabling from DOF stored as 1224 * properties in the dtrace.conf file. If there is an anonymous enabling, a 1225 * DTrace consumer state and enabling are created on attach. The state may be 1226 * subsequently grabbed by the first consumer specifying the "grabanon" 1227 * option. As long as an anonymous DTrace enabling exists, dtrace(7D) will 1228 * refuse to unload. 1229 */ 1230 typedef struct dtrace_anon { 1231 dtrace_state_t *dta_state; /* DTrace consumer state */ 1232 dtrace_enabling_t *dta_enabling; /* pointer to enabling */ 1233 processorid_t dta_beganon; /* which CPU BEGIN ran on */ 1234 } dtrace_anon_t; 1235 1236 /* 1237 * DTrace Error Debugging 1238 */ 1239 #ifdef DEBUG 1240 #define DTRACE_ERRDEBUG 1241 #endif 1242 1243 #ifdef DTRACE_ERRDEBUG 1244 1245 typedef struct dtrace_errhash { 1246 const char *dter_msg; /* error message */ 1247 int dter_count; /* number of times seen */ 1248 } dtrace_errhash_t; 1249 1250 #define DTRACE_ERRHASHSZ 256 /* must be > number of err msgs */ 1251 1252 #endif /* DTRACE_ERRDEBUG */ 1253 1254 /* 1255 * DTrace Toxic Ranges 1256 * 1257 * DTrace supports safe loads from probe context; if the address turns out to 1258 * be invalid, a bit will be set by the kernel indicating that DTrace 1259 * encountered a memory error, and DTrace will propagate the error to the user 1260 * accordingly. However, there may exist some regions of memory in which an 1261 * arbitrary load can change system state, and from which it is impossible to 1262 * recover from such a load after it has been attempted. Examples of this may 1263 * include memory in which programmable I/O registers are mapped (for which a 1264 * read may have some implications for the device) or (in the specific case of 1265 * UltraSPARC-I and -II) the virtual address hole. The platform is required 1266 * to make DTrace aware of these toxic ranges; DTrace will then check that 1267 * target addresses are not in a toxic range before attempting to issue a 1268 * safe load. 1269 */ 1270 typedef struct dtrace_toxrange { 1271 uintptr_t dtt_base; /* base of toxic range */ 1272 uintptr_t dtt_limit; /* limit of toxic range */ 1273 } dtrace_toxrange_t; 1274 1275 #ifdef illumos 1276 extern uint64_t dtrace_getarg(int, int); 1277 #else 1278 extern uint64_t __noinline dtrace_getarg(int, int); 1279 #endif 1280 extern greg_t dtrace_getfp(void); 1281 extern int dtrace_getipl(void); 1282 extern uintptr_t dtrace_caller(int); 1283 extern uint32_t dtrace_cas32(uint32_t *, uint32_t, uint32_t); 1284 extern void *dtrace_casptr(volatile void *, volatile void *, volatile void *); 1285 extern void dtrace_copyin(uintptr_t, uintptr_t, size_t, volatile uint16_t *); 1286 extern void dtrace_copyinstr(uintptr_t, uintptr_t, size_t, volatile uint16_t *); 1287 extern void dtrace_copyout(uintptr_t, uintptr_t, size_t, volatile uint16_t *); 1288 extern void dtrace_copyoutstr(uintptr_t, uintptr_t, size_t, 1289 volatile uint16_t *); 1290 extern void dtrace_getpcstack(pc_t *, int, int, uint32_t *); 1291 extern ulong_t dtrace_getreg(struct trapframe *, uint_t); 1292 extern int dtrace_getstackdepth(int); 1293 extern void dtrace_getupcstack(uint64_t *, int); 1294 extern void dtrace_getufpstack(uint64_t *, uint64_t *, int); 1295 extern int dtrace_getustackdepth(void); 1296 extern uintptr_t dtrace_fulword(void *); 1297 extern uint8_t dtrace_fuword8(void *); 1298 extern uint16_t dtrace_fuword16(void *); 1299 extern uint32_t dtrace_fuword32(void *); 1300 extern uint64_t dtrace_fuword64(void *); 1301 extern void dtrace_probe_error(dtrace_state_t *, dtrace_epid_t, int, int, 1302 int, uintptr_t); 1303 extern int dtrace_assfail(const char *, const char *, int); 1304 extern int dtrace_attached(void); 1305 #ifdef illumos 1306 extern hrtime_t dtrace_gethrestime(void); 1307 #endif 1308 1309 #ifdef __sparc 1310 extern void dtrace_flush_windows(void); 1311 extern void dtrace_flush_user_windows(void); 1312 extern uint_t dtrace_getotherwin(void); 1313 extern uint_t dtrace_getfprs(void); 1314 #else 1315 extern void dtrace_copy(uintptr_t, uintptr_t, size_t); 1316 extern void dtrace_copystr(uintptr_t, uintptr_t, size_t, volatile uint16_t *); 1317 #endif 1318 1319 /* 1320 * DTrace Assertions 1321 * 1322 * DTrace calls ASSERT and VERIFY from probe context. To assure that a failed 1323 * ASSERT or VERIFY does not induce a markedly more catastrophic failure (e.g., 1324 * one from which a dump cannot be gleaned), DTrace must define its own ASSERT 1325 * and VERIFY macros to be ones that may safely be called from probe context. 1326 * This header file must thus be included by any DTrace component that calls 1327 * ASSERT and/or VERIFY from probe context, and _only_ by those components. 1328 * (The only exception to this is kernel debugging infrastructure at user-level 1329 * that doesn't depend on calling ASSERT.) 1330 */ 1331 #undef ASSERT 1332 #undef VERIFY 1333 #define VERIFY(EX) ((void)((EX) || \ 1334 dtrace_assfail(#EX, __FILE__, __LINE__))) 1335 #ifdef DEBUG 1336 #define ASSERT(EX) ((void)((EX) || \ 1337 dtrace_assfail(#EX, __FILE__, __LINE__))) 1338 #else 1339 #define ASSERT(X) ((void)0) 1340 #endif 1341 1342 #ifdef __cplusplus 1343 } 1344 #endif 1345 1346 #endif /* _SYS_DTRACE_IMPL_H */ 1347