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 #ifdef __sparcv9 55 typedef uint32_t pc_t; 56 #else 57 typedef uintptr_t pc_t; 58 #endif 59 typedef u_long greg_t; 60 #endif 61 62 /* 63 * DTrace Implementation Constants and Typedefs 64 */ 65 #define DTRACE_MAXPROPLEN 128 66 #define DTRACE_DYNVAR_CHUNKSIZE 256 67 68 #ifdef __FreeBSD__ 69 #define NCPU MAXCPU 70 #endif /* __FreeBSD__ */ 71 72 struct dtrace_probe; 73 struct dtrace_ecb; 74 struct dtrace_predicate; 75 struct dtrace_action; 76 struct dtrace_provider; 77 struct dtrace_state; 78 79 typedef struct dtrace_probe dtrace_probe_t; 80 typedef struct dtrace_ecb dtrace_ecb_t; 81 typedef struct dtrace_predicate dtrace_predicate_t; 82 typedef struct dtrace_action dtrace_action_t; 83 typedef struct dtrace_provider dtrace_provider_t; 84 typedef struct dtrace_meta dtrace_meta_t; 85 typedef struct dtrace_state dtrace_state_t; 86 typedef uint32_t dtrace_optid_t; 87 typedef uint32_t dtrace_specid_t; 88 typedef uint64_t dtrace_genid_t; 89 90 /* 91 * DTrace Probes 92 * 93 * The probe is the fundamental unit of the DTrace architecture. Probes are 94 * created by DTrace providers, and managed by the DTrace framework. A probe 95 * is identified by a unique <provider, module, function, name> tuple, and has 96 * a unique probe identifier assigned to it. (Some probes are not associated 97 * with a specific point in text; these are called _unanchored probes_ and have 98 * no module or function associated with them.) Probes are represented as a 99 * dtrace_probe structure. To allow quick lookups based on each element of the 100 * probe tuple, probes are hashed by each of provider, module, function and 101 * name. (If a lookup is performed based on a regular expression, a 102 * dtrace_probekey is prepared, and a linear search is performed.) Each probe 103 * is additionally pointed to by a linear array indexed by its identifier. The 104 * identifier is the provider's mechanism for indicating to the DTrace 105 * framework that a probe has fired: the identifier is passed as the first 106 * argument to dtrace_probe(), where it is then mapped into the corresponding 107 * dtrace_probe structure. From the dtrace_probe structure, dtrace_probe() can 108 * iterate over the probe's list of enabling control blocks; see "DTrace 109 * Enabling Control Blocks", below.) 110 */ 111 struct dtrace_probe { 112 dtrace_id_t dtpr_id; /* probe identifier */ 113 dtrace_ecb_t *dtpr_ecb; /* ECB list; see below */ 114 dtrace_ecb_t *dtpr_ecb_last; /* last ECB in list */ 115 void *dtpr_arg; /* provider argument */ 116 dtrace_cacheid_t dtpr_predcache; /* predicate cache ID */ 117 int dtpr_aframes; /* artificial frames */ 118 dtrace_provider_t *dtpr_provider; /* pointer to provider */ 119 char *dtpr_mod; /* probe's module name */ 120 char *dtpr_func; /* probe's function name */ 121 char *dtpr_name; /* probe's name */ 122 dtrace_probe_t *dtpr_nextmod; /* next in module hash */ 123 dtrace_probe_t *dtpr_prevmod; /* previous in module hash */ 124 dtrace_probe_t *dtpr_nextfunc; /* next in function hash */ 125 dtrace_probe_t *dtpr_prevfunc; /* previous in function hash */ 126 dtrace_probe_t *dtpr_nextname; /* next in name hash */ 127 dtrace_probe_t *dtpr_prevname; /* previous in name hash */ 128 dtrace_genid_t dtpr_gen; /* probe generation ID */ 129 }; 130 131 typedef int dtrace_probekey_f(const char *, const char *, int); 132 133 typedef struct dtrace_probekey { 134 char *dtpk_prov; /* provider name to match */ 135 dtrace_probekey_f *dtpk_pmatch; /* provider matching function */ 136 char *dtpk_mod; /* module name to match */ 137 dtrace_probekey_f *dtpk_mmatch; /* module matching function */ 138 char *dtpk_func; /* func name to match */ 139 dtrace_probekey_f *dtpk_fmatch; /* func matching function */ 140 char *dtpk_name; /* name to match */ 141 dtrace_probekey_f *dtpk_nmatch; /* name matching function */ 142 dtrace_id_t dtpk_id; /* identifier to match */ 143 } dtrace_probekey_t; 144 145 typedef struct dtrace_hashbucket { 146 struct dtrace_hashbucket *dthb_next; /* next on hash chain */ 147 dtrace_probe_t *dthb_chain; /* chain of probes */ 148 int dthb_len; /* number of probes here */ 149 } dtrace_hashbucket_t; 150 151 typedef struct dtrace_hash { 152 dtrace_hashbucket_t **dth_tab; /* hash table */ 153 int dth_size; /* size of hash table */ 154 int dth_mask; /* mask to index into table */ 155 int dth_nbuckets; /* total number of buckets */ 156 uintptr_t dth_nextoffs; /* offset of next in probe */ 157 uintptr_t dth_prevoffs; /* offset of prev in probe */ 158 uintptr_t dth_stroffs; /* offset of str in probe */ 159 } dtrace_hash_t; 160 161 /* 162 * DTrace Enabling Control Blocks 163 * 164 * When a provider wishes to fire a probe, it calls into dtrace_probe(), 165 * passing the probe identifier as the first argument. As described above, 166 * dtrace_probe() maps the identifier into a pointer to a dtrace_probe_t 167 * structure. This structure contains information about the probe, and a 168 * pointer to the list of Enabling Control Blocks (ECBs). Each ECB points to 169 * DTrace consumer state, and contains an optional predicate, and a list of 170 * actions. (Shown schematically below.) The ECB abstraction allows a single 171 * probe to be multiplexed across disjoint consumers, or across disjoint 172 * enablings of a single probe within one consumer. 173 * 174 * Enabling Control Block 175 * dtrace_ecb_t 176 * +------------------------+ 177 * | dtrace_epid_t ---------+--------------> Enabled Probe ID (EPID) 178 * | dtrace_state_t * ------+--------------> State associated with this ECB 179 * | dtrace_predicate_t * --+---------+ 180 * | dtrace_action_t * -----+----+ | 181 * | dtrace_ecb_t * ---+ | | | Predicate (if any) 182 * +-------------------+----+ | | dtrace_predicate_t 183 * | | +---> +--------------------+ 184 * | | | dtrace_difo_t * ---+----> DIFO 185 * | | +--------------------+ 186 * | | 187 * Next ECB | | Action 188 * (if any) | | dtrace_action_t 189 * : +--> +-------------------+ 190 * : | dtrace_actkind_t -+------> kind 191 * v | dtrace_difo_t * --+------> DIFO (if any) 192 * | dtrace_recdesc_t -+------> record descr. 193 * | dtrace_action_t * +------+ 194 * +-------------------+ | 195 * | Next action 196 * +-------------------------------+ (if any) 197 * | 198 * | Action 199 * | dtrace_action_t 200 * +--> +-------------------+ 201 * | dtrace_actkind_t -+------> kind 202 * | dtrace_difo_t * --+------> DIFO (if any) 203 * | dtrace_action_t * +------+ 204 * +-------------------+ | 205 * | Next action 206 * +-------------------------------+ (if any) 207 * | 208 * : 209 * v 210 * 211 * 212 * dtrace_probe() iterates over the ECB list. If the ECB needs less space 213 * than is available in the principal buffer, the ECB is processed: if the 214 * predicate is non-NULL, the DIF object is executed. If the result is 215 * non-zero, the action list is processed, with each action being executed 216 * accordingly. When the action list has been completely executed, processing 217 * advances to the next ECB. The ECB abstraction allows disjoint consumers 218 * to multiplex on single probes. 219 * 220 * Execution of the ECB results in consuming dte_size bytes in the buffer 221 * to record data. During execution, dte_needed bytes must be available in 222 * the buffer. This space is used for both recorded data and tuple data. 223 */ 224 struct dtrace_ecb { 225 dtrace_epid_t dte_epid; /* enabled probe ID */ 226 uint32_t dte_alignment; /* required alignment */ 227 size_t dte_needed; /* space needed for execution */ 228 size_t dte_size; /* size of recorded payload */ 229 dtrace_predicate_t *dte_predicate; /* predicate, if any */ 230 dtrace_action_t *dte_action; /* actions, if any */ 231 dtrace_ecb_t *dte_next; /* next ECB on probe */ 232 dtrace_state_t *dte_state; /* pointer to state */ 233 uint32_t dte_cond; /* security condition */ 234 dtrace_probe_t *dte_probe; /* pointer to probe */ 235 dtrace_action_t *dte_action_last; /* last action on ECB */ 236 uint64_t dte_uarg; /* library argument */ 237 }; 238 239 struct dtrace_predicate { 240 dtrace_difo_t *dtp_difo; /* DIF object */ 241 dtrace_cacheid_t dtp_cacheid; /* cache identifier */ 242 int dtp_refcnt; /* reference count */ 243 }; 244 245 struct dtrace_action { 246 dtrace_actkind_t dta_kind; /* kind of action */ 247 uint16_t dta_intuple; /* boolean: in aggregation */ 248 uint32_t dta_refcnt; /* reference count */ 249 dtrace_difo_t *dta_difo; /* pointer to DIFO */ 250 dtrace_recdesc_t dta_rec; /* record description */ 251 dtrace_action_t *dta_prev; /* previous action */ 252 dtrace_action_t *dta_next; /* next action */ 253 }; 254 255 typedef struct dtrace_aggregation { 256 dtrace_action_t dtag_action; /* action; must be first */ 257 dtrace_aggid_t dtag_id; /* identifier */ 258 dtrace_ecb_t *dtag_ecb; /* corresponding ECB */ 259 dtrace_action_t *dtag_first; /* first action in tuple */ 260 uint32_t dtag_base; /* base of aggregation */ 261 uint8_t dtag_hasarg; /* boolean: has argument */ 262 uint64_t dtag_initial; /* initial value */ 263 void (*dtag_aggregate)(uint64_t *, uint64_t, uint64_t); 264 } dtrace_aggregation_t; 265 266 /* 267 * DTrace Buffers 268 * 269 * Principal buffers, aggregation buffers, and speculative buffers are all 270 * managed with the dtrace_buffer structure. By default, this structure 271 * includes twin data buffers -- dtb_tomax and dtb_xamot -- that serve as the 272 * active and passive buffers, respectively. For speculative buffers, 273 * dtb_xamot will be NULL; for "ring" and "fill" buffers, dtb_xamot will point 274 * to a scratch buffer. For all buffer types, the dtrace_buffer structure is 275 * always allocated on a per-CPU basis; a single dtrace_buffer structure is 276 * never shared among CPUs. (That is, there is never true sharing of the 277 * dtrace_buffer structure; to prevent false sharing of the structure, it must 278 * always be aligned to the coherence granularity -- generally 64 bytes.) 279 * 280 * One of the critical design decisions of DTrace is that a given ECB always 281 * stores the same quantity and type of data. This is done to assure that the 282 * only metadata required for an ECB's traced data is the EPID. That is, from 283 * the EPID, the consumer can determine the data layout. (The data buffer 284 * layout is shown schematically below.) By assuring that one can determine 285 * data layout from the EPID, the metadata stream can be separated from the 286 * data stream -- simplifying the data stream enormously. The ECB always 287 * proceeds the recorded data as part of the dtrace_rechdr_t structure that 288 * includes the EPID and a high-resolution timestamp used for output ordering 289 * consistency. 290 * 291 * base of data buffer ---> +--------+--------------------+--------+ 292 * | rechdr | data | rechdr | 293 * +--------+------+--------+----+--------+ 294 * | data | rechdr | data | 295 * +---------------+--------+-------------+ 296 * | data, cont. | 297 * +--------+--------------------+--------+ 298 * | rechdr | data | | 299 * +--------+--------------------+ | 300 * | || | 301 * | || | 302 * | \/ | 303 * : : 304 * . . 305 * . . 306 * . . 307 * : : 308 * | | 309 * limit of data buffer ---> +--------------------------------------+ 310 * 311 * When evaluating an ECB, dtrace_probe() determines if the ECB's needs of the 312 * principal buffer (both scratch and payload) exceed the available space. If 313 * the ECB's needs exceed available space (and if the principal buffer policy 314 * is the default "switch" policy), the ECB is dropped, the buffer's drop count 315 * is incremented, and processing advances to the next ECB. If the ECB's needs 316 * can be met with the available space, the ECB is processed, but the offset in 317 * the principal buffer is only advanced if the ECB completes processing 318 * without error. 319 * 320 * When a buffer is to be switched (either because the buffer is the principal 321 * buffer with a "switch" policy or because it is an aggregation buffer), a 322 * cross call is issued to the CPU associated with the buffer. In the cross 323 * call context, interrupts are disabled, and the active and the inactive 324 * buffers are atomically switched. This involves switching the data pointers, 325 * copying the various state fields (offset, drops, errors, etc.) into their 326 * inactive equivalents, and clearing the state fields. Because interrupts are 327 * disabled during this procedure, the switch is guaranteed to appear atomic to 328 * dtrace_probe(). 329 * 330 * DTrace Ring Buffering 331 * 332 * To process a ring buffer correctly, one must know the oldest valid record. 333 * Processing starts at the oldest record in the buffer and continues until 334 * the end of the buffer is reached. Processing then resumes starting with 335 * the record stored at offset 0 in the buffer, and continues until the 336 * youngest record is processed. If trace records are of a fixed-length, 337 * determining the oldest record is trivial: 338 * 339 * - If the ring buffer has not wrapped, the oldest record is the record 340 * stored at offset 0. 341 * 342 * - If the ring buffer has wrapped, the oldest record is the record stored 343 * at the current offset. 344 * 345 * With variable length records, however, just knowing the current offset 346 * doesn't suffice for determining the oldest valid record: assuming that one 347 * allows for arbitrary data, one has no way of searching forward from the 348 * current offset to find the oldest valid record. (That is, one has no way 349 * of separating data from metadata.) It would be possible to simply refuse to 350 * process any data in the ring buffer between the current offset and the 351 * limit, but this leaves (potentially) an enormous amount of otherwise valid 352 * data unprocessed. 353 * 354 * To effect ring buffering, we track two offsets in the buffer: the current 355 * offset and the _wrapped_ offset. If a request is made to reserve some 356 * amount of data, and the buffer has wrapped, the wrapped offset is 357 * incremented until the wrapped offset minus the current offset is greater 358 * than or equal to the reserve request. This is done by repeatedly looking 359 * up the ECB corresponding to the EPID at the current wrapped offset, and 360 * incrementing the wrapped offset by the size of the data payload 361 * corresponding to that ECB. If this offset is greater than or equal to the 362 * limit of the data buffer, the wrapped offset is set to 0. Thus, the 363 * current offset effectively "chases" the wrapped offset around the buffer. 364 * Schematically: 365 * 366 * base of data buffer ---> +------+--------------------+------+ 367 * | EPID | data | EPID | 368 * +------+--------+------+----+------+ 369 * | data | EPID | data | 370 * +---------------+------+-----------+ 371 * | data, cont. | 372 * +------+---------------------------+ 373 * | EPID | data | 374 * current offset ---> +------+---------------------------+ 375 * | invalid data | 376 * wrapped offset ---> +------+--------------------+------+ 377 * | EPID | data | EPID | 378 * +------+--------+------+----+------+ 379 * | data | EPID | data | 380 * +---------------+------+-----------+ 381 * : : 382 * . . 383 * . ... valid data ... . 384 * . . 385 * : : 386 * +------+-------------+------+------+ 387 * | EPID | data | EPID | data | 388 * +------+------------++------+------+ 389 * | data, cont. | leftover | 390 * limit of data buffer ---> +-------------------+--------------+ 391 * 392 * If the amount of requested buffer space exceeds the amount of space 393 * available between the current offset and the end of the buffer: 394 * 395 * (1) all words in the data buffer between the current offset and the limit 396 * of the data buffer (marked "leftover", above) are set to 397 * DTRACE_EPIDNONE 398 * 399 * (2) the wrapped offset is set to zero 400 * 401 * (3) the iteration process described above occurs until the wrapped offset 402 * is greater than the amount of desired space. 403 * 404 * The wrapped offset is implemented by (re-)using the inactive offset. 405 * In a "switch" buffer policy, the inactive offset stores the offset in 406 * the inactive buffer; in a "ring" buffer policy, it stores the wrapped 407 * offset. 408 * 409 * DTrace Scratch Buffering 410 * 411 * Some ECBs may wish to allocate dynamically-sized temporary scratch memory. 412 * To accommodate such requests easily, scratch memory may be allocated in 413 * the buffer beyond the current offset plus the needed memory of the current 414 * ECB. If there isn't sufficient room in the buffer for the requested amount 415 * of scratch space, the allocation fails and an error is generated. Scratch 416 * memory is tracked in the dtrace_mstate_t and is automatically freed when 417 * the ECB ceases processing. Note that ring buffers cannot allocate their 418 * scratch from the principal buffer -- lest they needlessly overwrite older, 419 * valid data. Ring buffers therefore have their own dedicated scratch buffer 420 * from which scratch is allocated. 421 */ 422 #define DTRACEBUF_RING 0x0001 /* bufpolicy set to "ring" */ 423 #define DTRACEBUF_FILL 0x0002 /* bufpolicy set to "fill" */ 424 #define DTRACEBUF_NOSWITCH 0x0004 /* do not switch buffer */ 425 #define DTRACEBUF_WRAPPED 0x0008 /* ring buffer has wrapped */ 426 #define DTRACEBUF_DROPPED 0x0010 /* drops occurred */ 427 #define DTRACEBUF_ERROR 0x0020 /* errors occurred */ 428 #define DTRACEBUF_FULL 0x0040 /* "fill" buffer is full */ 429 #define DTRACEBUF_CONSUMED 0x0080 /* buffer has been consumed */ 430 #define DTRACEBUF_INACTIVE 0x0100 /* buffer is not yet active */ 431 432 typedef struct dtrace_buffer { 433 uint64_t dtb_offset; /* current offset in buffer */ 434 uint64_t dtb_size; /* size of buffer */ 435 uint32_t dtb_flags; /* flags */ 436 uint32_t dtb_drops; /* number of drops */ 437 caddr_t dtb_tomax; /* active buffer */ 438 caddr_t dtb_xamot; /* inactive buffer */ 439 uint32_t dtb_xamot_flags; /* inactive flags */ 440 uint32_t dtb_xamot_drops; /* drops in inactive buffer */ 441 uint64_t dtb_xamot_offset; /* offset in inactive buffer */ 442 uint32_t dtb_errors; /* number of errors */ 443 uint32_t dtb_xamot_errors; /* errors in inactive buffer */ 444 #ifndef _LP64 445 uint64_t dtb_pad1; /* pad out to 64 bytes */ 446 #endif 447 uint64_t dtb_switched; /* time of last switch */ 448 uint64_t dtb_interval; /* observed switch interval */ 449 uint64_t dtb_pad2[6]; /* pad to avoid false sharing */ 450 } dtrace_buffer_t; 451 452 /* 453 * DTrace Aggregation Buffers 454 * 455 * Aggregation buffers use much of the same mechanism as described above 456 * ("DTrace Buffers"). However, because an aggregation is fundamentally a 457 * hash, there exists dynamic metadata associated with an aggregation buffer 458 * that is not associated with other kinds of buffers. This aggregation 459 * metadata is _only_ relevant for the in-kernel implementation of 460 * aggregations; it is not actually relevant to user-level consumers. To do 461 * this, we allocate dynamic aggregation data (hash keys and hash buckets) 462 * starting below the _limit_ of the buffer, and we allocate data from the 463 * _base_ of the buffer. When the aggregation buffer is copied out, _only_ the 464 * data is copied out; the metadata is simply discarded. Schematically, 465 * aggregation buffers look like: 466 * 467 * base of data buffer ---> +-------+------+-----------+-------+ 468 * | aggid | key | value | aggid | 469 * +-------+------+-----------+-------+ 470 * | key | 471 * +-------+-------+-----+------------+ 472 * | value | aggid | key | value | 473 * +-------+------++-----+------+-----+ 474 * | aggid | key | value | | 475 * +-------+------+-------------+ | 476 * | || | 477 * | || | 478 * | \/ | 479 * : : 480 * . . 481 * . . 482 * . . 483 * : : 484 * | /\ | 485 * | || +------------+ 486 * | || | | 487 * +---------------------+ | 488 * | hash keys | 489 * | (dtrace_aggkey structures) | 490 * | | 491 * +----------------------------------+ 492 * | hash buckets | 493 * | (dtrace_aggbuffer structure) | 494 * | | 495 * limit of data buffer ---> +----------------------------------+ 496 * 497 * 498 * As implied above, just as we assure that ECBs always store a constant 499 * amount of data, we assure that a given aggregation -- identified by its 500 * aggregation ID -- always stores data of a constant quantity and type. 501 * As with EPIDs, this allows the aggregation ID to serve as the metadata for a 502 * given record. 503 * 504 * Note that the size of the dtrace_aggkey structure must be sizeof (uintptr_t) 505 * aligned. (If this the structure changes such that this becomes false, an 506 * assertion will fail in dtrace_aggregate().) 507 */ 508 typedef struct dtrace_aggkey { 509 uint32_t dtak_hashval; /* hash value */ 510 uint32_t dtak_action:4; /* action -- 4 bits */ 511 uint32_t dtak_size:28; /* size -- 28 bits */ 512 caddr_t dtak_data; /* data pointer */ 513 struct dtrace_aggkey *dtak_next; /* next in hash chain */ 514 } dtrace_aggkey_t; 515 516 typedef struct dtrace_aggbuffer { 517 uintptr_t dtagb_hashsize; /* number of buckets */ 518 uintptr_t dtagb_free; /* free list of keys */ 519 dtrace_aggkey_t **dtagb_hash; /* hash table */ 520 } dtrace_aggbuffer_t; 521 522 /* 523 * DTrace Speculations 524 * 525 * Speculations have a per-CPU buffer and a global state. Once a speculation 526 * buffer has been comitted or discarded, it cannot be reused until all CPUs 527 * have taken the same action (commit or discard) on their respective 528 * speculative buffer. However, because DTrace probes may execute in arbitrary 529 * context, other CPUs cannot simply be cross-called at probe firing time to 530 * perform the necessary commit or discard. The speculation states thus 531 * optimize for the case that a speculative buffer is only active on one CPU at 532 * the time of a commit() or discard() -- for if this is the case, other CPUs 533 * need not take action, and the speculation is immediately available for 534 * reuse. If the speculation is active on multiple CPUs, it must be 535 * asynchronously cleaned -- potentially leading to a higher rate of dirty 536 * speculative drops. The speculation states are as follows: 537 * 538 * DTRACESPEC_INACTIVE <= Initial state; inactive speculation 539 * DTRACESPEC_ACTIVE <= Allocated, but not yet speculatively traced to 540 * DTRACESPEC_ACTIVEONE <= Speculatively traced to on one CPU 541 * DTRACESPEC_ACTIVEMANY <= Speculatively traced to on more than one CPU 542 * DTRACESPEC_COMMITTING <= Currently being commited on one CPU 543 * DTRACESPEC_COMMITTINGMANY <= Currently being commited on many CPUs 544 * DTRACESPEC_DISCARDING <= Currently being discarded on many CPUs 545 * 546 * The state transition diagram is as follows: 547 * 548 * +----------------------------------------------------------+ 549 * | | 550 * | +------------+ | 551 * | +-------------------| COMMITTING |<-----------------+ | 552 * | | +------------+ | | 553 * | | copied spec. ^ commit() on | | discard() on 554 * | | into principal | active CPU | | active CPU 555 * | | | commit() | | 556 * V V | | | 557 * +----------+ +--------+ +-----------+ 558 * | INACTIVE |---------------->| ACTIVE |--------------->| ACTIVEONE | 559 * +----------+ speculation() +--------+ speculate() +-----------+ 560 * ^ ^ | | | 561 * | | | discard() | | 562 * | | asynchronously | discard() on | | speculate() 563 * | | cleaned V inactive CPU | | on inactive 564 * | | +------------+ | | CPU 565 * | +-------------------| DISCARDING |<-----------------+ | 566 * | +------------+ | 567 * | asynchronously ^ | 568 * | copied spec. | discard() | 569 * | into principal +------------------------+ | 570 * | | V 571 * +----------------+ commit() +------------+ 572 * | COMMITTINGMANY |<----------------------------------| ACTIVEMANY | 573 * +----------------+ +------------+ 574 */ 575 typedef enum dtrace_speculation_state { 576 DTRACESPEC_INACTIVE = 0, 577 DTRACESPEC_ACTIVE, 578 DTRACESPEC_ACTIVEONE, 579 DTRACESPEC_ACTIVEMANY, 580 DTRACESPEC_COMMITTING, 581 DTRACESPEC_COMMITTINGMANY, 582 DTRACESPEC_DISCARDING 583 } dtrace_speculation_state_t; 584 585 typedef struct dtrace_speculation { 586 dtrace_speculation_state_t dtsp_state; /* current speculation state */ 587 int dtsp_cleaning; /* non-zero if being cleaned */ 588 dtrace_buffer_t *dtsp_buffer; /* speculative buffer */ 589 } dtrace_speculation_t; 590 591 /* 592 * DTrace Dynamic Variables 593 * 594 * The dynamic variable problem is obviously decomposed into two subproblems: 595 * allocating new dynamic storage, and freeing old dynamic storage. The 596 * presence of the second problem makes the first much more complicated -- or 597 * rather, the absence of the second renders the first trivial. This is the 598 * case with aggregations, for which there is effectively no deallocation of 599 * dynamic storage. (Or more accurately, all dynamic storage is deallocated 600 * when a snapshot is taken of the aggregation.) As DTrace dynamic variables 601 * allow for both dynamic allocation and dynamic deallocation, the 602 * implementation of dynamic variables is quite a bit more complicated than 603 * that of their aggregation kin. 604 * 605 * We observe that allocating new dynamic storage is tricky only because the 606 * size can vary -- the allocation problem is much easier if allocation sizes 607 * are uniform. We further observe that in D, the size of dynamic variables is 608 * actually _not_ dynamic -- dynamic variable sizes may be determined by static 609 * analysis of DIF text. (This is true even of putatively dynamically-sized 610 * objects like strings and stacks, the sizes of which are dictated by the 611 * "stringsize" and "stackframes" variables, respectively.) We exploit this by 612 * performing this analysis on all DIF before enabling any probes. For each 613 * dynamic load or store, we calculate the dynamically-allocated size plus the 614 * size of the dtrace_dynvar structure plus the storage required to key the 615 * data. For all DIF, we take the largest value and dub it the _chunksize_. 616 * We then divide dynamic memory into two parts: a hash table that is wide 617 * enough to have every chunk in its own bucket, and a larger region of equal 618 * chunksize units. Whenever we wish to dynamically allocate a variable, we 619 * always allocate a single chunk of memory. Depending on the uniformity of 620 * allocation, this will waste some amount of memory -- but it eliminates the 621 * non-determinism inherent in traditional heap fragmentation. 622 * 623 * Dynamic objects are allocated by storing a non-zero value to them; they are 624 * deallocated by storing a zero value to them. Dynamic variables are 625 * complicated enormously by being shared between CPUs. In particular, 626 * consider the following scenario: 627 * 628 * CPU A CPU B 629 * +---------------------------------+ +---------------------------------+ 630 * | | | | 631 * | allocates dynamic object a[123] | | | 632 * | by storing the value 345 to it | | | 633 * | ---------> | 634 * | | | wishing to load from object | 635 * | | | a[123], performs lookup in | 636 * | | | dynamic variable space | 637 * | <--------- | 638 * | deallocates object a[123] by | | | 639 * | storing 0 to it | | | 640 * | | | | 641 * | allocates dynamic object b[567] | | performs load from a[123] | 642 * | by storing the value 789 to it | | | 643 * : : : : 644 * . . . . 645 * 646 * This is obviously a race in the D program, but there are nonetheless only 647 * two valid values for CPU B's load from a[123]: 345 or 0. Most importantly, 648 * CPU B may _not_ see the value 789 for a[123]. 649 * 650 * There are essentially two ways to deal with this: 651 * 652 * (1) Explicitly spin-lock variables. That is, if CPU B wishes to load 653 * from a[123], it needs to lock a[123] and hold the lock for the 654 * duration that it wishes to manipulate it. 655 * 656 * (2) Avoid reusing freed chunks until it is known that no CPU is referring 657 * to them. 658 * 659 * The implementation of (1) is rife with complexity, because it requires the 660 * user of a dynamic variable to explicitly decree when they are done using it. 661 * Were all variables by value, this perhaps wouldn't be debilitating -- but 662 * dynamic variables of non-scalar types are tracked by reference. That is, if 663 * a dynamic variable is, say, a string, and that variable is to be traced to, 664 * say, the principal buffer, the DIF emulation code returns to the main 665 * dtrace_probe() loop a pointer to the underlying storage, not the contents of 666 * the storage. Further, code calling on DIF emulation would have to be aware 667 * that the DIF emulation has returned a reference to a dynamic variable that 668 * has been potentially locked. The variable would have to be unlocked after 669 * the main dtrace_probe() loop is finished with the variable, and the main 670 * dtrace_probe() loop would have to be careful to not call any further DIF 671 * emulation while the variable is locked to avoid deadlock. More generally, 672 * if one were to implement (1), DIF emulation code dealing with dynamic 673 * variables could only deal with one dynamic variable at a time (lest deadlock 674 * result). To sum, (1) exports too much subtlety to the users of dynamic 675 * variables -- increasing maintenance burden and imposing serious constraints 676 * on future DTrace development. 677 * 678 * The implementation of (2) is also complex, but the complexity is more 679 * manageable. We need to be sure that when a variable is deallocated, it is 680 * not placed on a traditional free list, but rather on a _dirty_ list. Once a 681 * variable is on a dirty list, it cannot be found by CPUs performing a 682 * subsequent lookup of the variable -- but it may still be in use by other 683 * CPUs. To assure that all CPUs that may be seeing the old variable have 684 * cleared out of probe context, a dtrace_sync() can be issued. Once the 685 * dtrace_sync() has completed, it can be known that all CPUs are done 686 * manipulating the dynamic variable -- the dirty list can be atomically 687 * appended to the free list. Unfortunately, there's a slight hiccup in this 688 * mechanism: dtrace_sync() may not be issued from probe context. The 689 * dtrace_sync() must be therefore issued asynchronously from non-probe 690 * context. For this we rely on the DTrace cleaner, a cyclic that runs at the 691 * "cleanrate" frequency. To ease this implementation, we define several chunk 692 * lists: 693 * 694 * - Dirty. Deallocated chunks, not yet cleaned. Not available. 695 * 696 * - Rinsing. Formerly dirty chunks that are currently being asynchronously 697 * cleaned. Not available, but will be shortly. Dynamic variable 698 * allocation may not spin or block for availability, however. 699 * 700 * - Clean. Clean chunks, ready for allocation -- but not on the free list. 701 * 702 * - Free. Available for allocation. 703 * 704 * Moreover, to avoid absurd contention, _each_ of these lists is implemented 705 * on a per-CPU basis. This is only for performance, not correctness; chunks 706 * may be allocated from another CPU's free list. The algorithm for allocation 707 * then is this: 708 * 709 * (1) Attempt to atomically allocate from current CPU's free list. If list 710 * is non-empty and allocation is successful, allocation is complete. 711 * 712 * (2) If the clean list is non-empty, atomically move it to the free list, 713 * and reattempt (1). 714 * 715 * (3) If the dynamic variable space is in the CLEAN state, look for free 716 * and clean lists on other CPUs by setting the current CPU to the next 717 * CPU, and reattempting (1). If the next CPU is the current CPU (that 718 * is, if all CPUs have been checked), atomically switch the state of 719 * the dynamic variable space based on the following: 720 * 721 * - If no free chunks were found and no dirty chunks were found, 722 * atomically set the state to EMPTY. 723 * 724 * - If dirty chunks were found, atomically set the state to DIRTY. 725 * 726 * - If rinsing chunks were found, atomically set the state to RINSING. 727 * 728 * (4) Based on state of dynamic variable space state, increment appropriate 729 * counter to indicate dynamic drops (if in EMPTY state) vs. dynamic 730 * dirty drops (if in DIRTY state) vs. dynamic rinsing drops (if in 731 * RINSING state). Fail the allocation. 732 * 733 * The cleaning cyclic operates with the following algorithm: for all CPUs 734 * with a non-empty dirty list, atomically move the dirty list to the rinsing 735 * list. Perform a dtrace_sync(). For all CPUs with a non-empty rinsing list, 736 * atomically move the rinsing list to the clean list. Perform another 737 * dtrace_sync(). By this point, all CPUs have seen the new clean list; the 738 * state of the dynamic variable space can be restored to CLEAN. 739 * 740 * There exist two final races that merit explanation. The first is a simple 741 * allocation race: 742 * 743 * CPU A CPU B 744 * +---------------------------------+ +---------------------------------+ 745 * | | | | 746 * | allocates dynamic object a[123] | | allocates dynamic object a[123] | 747 * | by storing the value 345 to it | | by storing the value 567 to it | 748 * | | | | 749 * : : : : 750 * . . . . 751 * 752 * Again, this is a race in the D program. It can be resolved by having a[123] 753 * hold the value 345 or a[123] hold the value 567 -- but it must be true that 754 * a[123] have only _one_ of these values. (That is, the racing CPUs may not 755 * put the same element twice on the same hash chain.) This is resolved 756 * simply: before the allocation is undertaken, the start of the new chunk's 757 * hash chain is noted. Later, after the allocation is complete, the hash 758 * chain is atomically switched to point to the new element. If this fails 759 * (because of either concurrent allocations or an allocation concurrent with a 760 * deletion), the newly allocated chunk is deallocated to the dirty list, and 761 * the whole process of looking up (and potentially allocating) the dynamic 762 * variable is reattempted. 763 * 764 * The final race is a simple deallocation race: 765 * 766 * CPU A CPU B 767 * +---------------------------------+ +---------------------------------+ 768 * | | | | 769 * | deallocates dynamic object | | deallocates dynamic object | 770 * | a[123] by storing the value 0 | | a[123] by storing the value 0 | 771 * | to it | | to it | 772 * | | | | 773 * : : : : 774 * . . . . 775 * 776 * Once again, this is a race in the D program, but it is one that we must 777 * handle without corrupting the underlying data structures. Because 778 * deallocations require the deletion of a chunk from the middle of a hash 779 * chain, we cannot use a single-word atomic operation to remove it. For this, 780 * we add a spin lock to the hash buckets that is _only_ used for deallocations 781 * (allocation races are handled as above). Further, this spin lock is _only_ 782 * held for the duration of the delete; before control is returned to the DIF 783 * emulation code, the hash bucket is unlocked. 784 */ 785 typedef struct dtrace_key { 786 uint64_t dttk_value; /* data value or data pointer */ 787 uint64_t dttk_size; /* 0 if by-val, >0 if by-ref */ 788 } dtrace_key_t; 789 790 typedef struct dtrace_tuple { 791 uint32_t dtt_nkeys; /* number of keys in tuple */ 792 uint32_t dtt_pad; /* padding */ 793 dtrace_key_t dtt_key[1]; /* array of tuple keys */ 794 } dtrace_tuple_t; 795 796 typedef struct dtrace_dynvar { 797 uint64_t dtdv_hashval; /* hash value -- 0 if free */ 798 struct dtrace_dynvar *dtdv_next; /* next on list or hash chain */ 799 void *dtdv_data; /* pointer to data */ 800 dtrace_tuple_t dtdv_tuple; /* tuple key */ 801 } dtrace_dynvar_t; 802 803 typedef enum dtrace_dynvar_op { 804 DTRACE_DYNVAR_ALLOC, 805 DTRACE_DYNVAR_NOALLOC, 806 DTRACE_DYNVAR_DEALLOC 807 } dtrace_dynvar_op_t; 808 809 typedef struct dtrace_dynhash { 810 dtrace_dynvar_t *dtdh_chain; /* hash chain for this bucket */ 811 uintptr_t dtdh_lock; /* deallocation lock */ 812 #ifdef _LP64 813 uintptr_t dtdh_pad[6]; /* pad to avoid false sharing */ 814 #else 815 uintptr_t dtdh_pad[14]; /* pad to avoid false sharing */ 816 #endif 817 } dtrace_dynhash_t; 818 819 typedef struct dtrace_dstate_percpu { 820 dtrace_dynvar_t *dtdsc_free; /* free list for this CPU */ 821 dtrace_dynvar_t *dtdsc_dirty; /* dirty list for this CPU */ 822 dtrace_dynvar_t *dtdsc_rinsing; /* rinsing list for this CPU */ 823 dtrace_dynvar_t *dtdsc_clean; /* clean list for this CPU */ 824 uint64_t dtdsc_drops; /* number of capacity drops */ 825 uint64_t dtdsc_dirty_drops; /* number of dirty drops */ 826 uint64_t dtdsc_rinsing_drops; /* number of rinsing drops */ 827 #ifdef _LP64 828 uint64_t dtdsc_pad; /* pad to avoid false sharing */ 829 #else 830 uint64_t dtdsc_pad[2]; /* pad to avoid false sharing */ 831 #endif 832 } dtrace_dstate_percpu_t; 833 834 typedef enum dtrace_dstate_state { 835 DTRACE_DSTATE_CLEAN = 0, 836 DTRACE_DSTATE_EMPTY, 837 DTRACE_DSTATE_DIRTY, 838 DTRACE_DSTATE_RINSING 839 } dtrace_dstate_state_t; 840 841 typedef struct dtrace_dstate { 842 void *dtds_base; /* base of dynamic var. space */ 843 size_t dtds_size; /* size of dynamic var. space */ 844 size_t dtds_hashsize; /* number of buckets in hash */ 845 size_t dtds_chunksize; /* size of each chunk */ 846 dtrace_dynhash_t *dtds_hash; /* pointer to hash table */ 847 dtrace_dstate_state_t dtds_state; /* current dynamic var. state */ 848 dtrace_dstate_percpu_t *dtds_percpu; /* per-CPU dyn. var. state */ 849 } dtrace_dstate_t; 850 851 /* 852 * DTrace Variable State 853 * 854 * The DTrace variable state tracks user-defined variables in its dtrace_vstate 855 * structure. Each DTrace consumer has exactly one dtrace_vstate structure, 856 * but some dtrace_vstate structures may exist without a corresponding DTrace 857 * consumer (see "DTrace Helpers", below). As described in <sys/dtrace.h>, 858 * user-defined variables can have one of three scopes: 859 * 860 * DIFV_SCOPE_GLOBAL => global scope 861 * DIFV_SCOPE_THREAD => thread-local scope (i.e. "self->" variables) 862 * DIFV_SCOPE_LOCAL => clause-local scope (i.e. "this->" variables) 863 * 864 * The variable state tracks variables by both their scope and their allocation 865 * type: 866 * 867 * - The dtvs_globals and dtvs_locals members each point to an array of 868 * dtrace_statvar structures. These structures contain both the variable 869 * metadata (dtrace_difv structures) and the underlying storage for all 870 * statically allocated variables, including statically allocated 871 * DIFV_SCOPE_GLOBAL variables and all DIFV_SCOPE_LOCAL variables. 872 * 873 * - The dtvs_tlocals member points to an array of dtrace_difv structures for 874 * DIFV_SCOPE_THREAD variables. As such, this array tracks _only_ the 875 * variable metadata for DIFV_SCOPE_THREAD variables; the underlying storage 876 * is allocated out of the dynamic variable space. 877 * 878 * - The dtvs_dynvars member is the dynamic variable state associated with the 879 * variable state. The dynamic variable state (described in "DTrace Dynamic 880 * Variables", above) tracks all DIFV_SCOPE_THREAD variables and all 881 * dynamically-allocated DIFV_SCOPE_GLOBAL variables. 882 */ 883 typedef struct dtrace_statvar { 884 uint64_t dtsv_data; /* data or pointer to it */ 885 size_t dtsv_size; /* size of pointed-to data */ 886 int dtsv_refcnt; /* reference count */ 887 dtrace_difv_t dtsv_var; /* variable metadata */ 888 } dtrace_statvar_t; 889 890 typedef struct dtrace_vstate { 891 dtrace_state_t *dtvs_state; /* back pointer to state */ 892 dtrace_statvar_t **dtvs_globals; /* statically-allocated glbls */ 893 int dtvs_nglobals; /* number of globals */ 894 dtrace_difv_t *dtvs_tlocals; /* thread-local metadata */ 895 int dtvs_ntlocals; /* number of thread-locals */ 896 dtrace_statvar_t **dtvs_locals; /* clause-local data */ 897 int dtvs_nlocals; /* number of clause-locals */ 898 dtrace_dstate_t dtvs_dynvars; /* dynamic variable state */ 899 } dtrace_vstate_t; 900 901 /* 902 * DTrace Machine State 903 * 904 * In the process of processing a fired probe, DTrace needs to track and/or 905 * cache some per-CPU state associated with that particular firing. This is 906 * state that is always discarded after the probe firing has completed, and 907 * much of it is not specific to any DTrace consumer, remaining valid across 908 * all ECBs. This state is tracked in the dtrace_mstate structure. 909 */ 910 #define DTRACE_MSTATE_ARGS 0x00000001 911 #define DTRACE_MSTATE_PROBE 0x00000002 912 #define DTRACE_MSTATE_EPID 0x00000004 913 #define DTRACE_MSTATE_TIMESTAMP 0x00000008 914 #define DTRACE_MSTATE_STACKDEPTH 0x00000010 915 #define DTRACE_MSTATE_CALLER 0x00000020 916 #define DTRACE_MSTATE_IPL 0x00000040 917 #define DTRACE_MSTATE_FLTOFFS 0x00000080 918 #define DTRACE_MSTATE_WALLTIMESTAMP 0x00000100 919 #define DTRACE_MSTATE_USTACKDEPTH 0x00000200 920 #define DTRACE_MSTATE_UCALLER 0x00000400 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 int dtms_stackdepth; /* cached stackdepth */ 932 int dtms_ustackdepth; /* cached ustackdepth */ 933 struct dtrace_probe *dtms_probe; /* current probe */ 934 uintptr_t dtms_caller; /* cached caller */ 935 uint64_t dtms_ucaller; /* cached user-level caller */ 936 int dtms_ipl; /* cached interrupt pri lev */ 937 int dtms_fltoffs; /* faulting DIFO offset */ 938 uintptr_t dtms_strtok; /* saved strtok() pointer */ 939 uintptr_t dtms_strtok_limit; /* upper bound of strtok ptr */ 940 uint32_t dtms_access; /* memory access rights */ 941 dtrace_difo_t *dtms_difo; /* current dif object */ 942 file_t *dtms_getf; /* cached rval of getf() */ 943 } dtrace_mstate_t; 944 945 #define DTRACE_COND_OWNER 0x1 946 #define DTRACE_COND_USERMODE 0x2 947 #define DTRACE_COND_ZONEOWNER 0x4 948 949 #define DTRACE_PROBEKEY_MAXDEPTH 8 /* max glob recursion depth */ 950 951 /* 952 * Access flag used by dtrace_mstate.dtms_access. 953 */ 954 #define DTRACE_ACCESS_KERNEL 0x1 /* the priv to read kmem */ 955 956 957 /* 958 * DTrace Activity 959 * 960 * Each DTrace consumer is in one of several states, which (for purposes of 961 * avoiding yet-another overloading of the noun "state") we call the current 962 * _activity_. The activity transitions on dtrace_go() (from DTRACIOCGO), on 963 * dtrace_stop() (from DTRACIOCSTOP) and on the exit() action. Activities may 964 * only transition in one direction; the activity transition diagram is a 965 * directed acyclic graph. The activity transition diagram is as follows: 966 * 967 * 968 * +----------+ +--------+ +--------+ 969 * | INACTIVE |------------------>| WARMUP |------------------>| ACTIVE | 970 * +----------+ dtrace_go(), +--------+ dtrace_go(), +--------+ 971 * before BEGIN | after BEGIN | | | 972 * | | | | 973 * exit() action | | | | 974 * from BEGIN ECB | | | | 975 * | | | | 976 * v | | | 977 * +----------+ exit() action | | | 978 * +-----------------------------| DRAINING |<-------------------+ | | 979 * | +----------+ | | 980 * | | | | 981 * | dtrace_stop(), | | | 982 * | before END | | | 983 * | | | | 984 * | v | | 985 * | +---------+ +----------+ | | 986 * | | STOPPED |<----------------| COOLDOWN |<----------------------+ | 987 * | +---------+ dtrace_stop(), +----------+ dtrace_stop(), | 988 * | after END before END | 989 * | | 990 * | +--------+ | 991 * +----------------------------->| KILLED |<--------------------------+ 992 * deadman timeout or +--------+ deadman timeout or 993 * killed consumer killed consumer 994 * 995 * Note that once a DTrace consumer has stopped tracing, there is no way to 996 * restart it; if a DTrace consumer wishes to restart tracing, it must reopen 997 * the DTrace pseudodevice. 998 */ 999 typedef enum dtrace_activity { 1000 DTRACE_ACTIVITY_INACTIVE = 0, /* not yet running */ 1001 DTRACE_ACTIVITY_WARMUP, /* while starting */ 1002 DTRACE_ACTIVITY_ACTIVE, /* running */ 1003 DTRACE_ACTIVITY_DRAINING, /* before stopping */ 1004 DTRACE_ACTIVITY_COOLDOWN, /* while stopping */ 1005 DTRACE_ACTIVITY_STOPPED, /* after stopping */ 1006 DTRACE_ACTIVITY_KILLED /* killed */ 1007 } dtrace_activity_t; 1008 1009 /* 1010 * DTrace Helper Implementation 1011 * 1012 * A description of the helper architecture may be found in <sys/dtrace.h>. 1013 * Each process contains a pointer to its helpers in its p_dtrace_helpers 1014 * member. This is a pointer to a dtrace_helpers structure, which contains an 1015 * array of pointers to dtrace_helper structures, helper variable state (shared 1016 * among a process's helpers) and a generation count. (The generation count is 1017 * used to provide an identifier when a helper is added so that it may be 1018 * subsequently removed.) The dtrace_helper structure is self-explanatory, 1019 * containing pointers to the objects needed to execute the helper. Note that 1020 * helpers are _duplicated_ across fork(2), and destroyed on exec(2). No more 1021 * than dtrace_helpers_max are allowed per-process. 1022 */ 1023 #define DTRACE_HELPER_ACTION_USTACK 0 1024 #define DTRACE_NHELPER_ACTIONS 1 1025 1026 typedef struct dtrace_helper_action { 1027 int dtha_generation; /* helper action generation */ 1028 int dtha_nactions; /* number of actions */ 1029 dtrace_difo_t *dtha_predicate; /* helper action predicate */ 1030 dtrace_difo_t **dtha_actions; /* array of actions */ 1031 struct dtrace_helper_action *dtha_next; /* next helper action */ 1032 } dtrace_helper_action_t; 1033 1034 typedef struct dtrace_helper_provider { 1035 int dthp_generation; /* helper provider generation */ 1036 uint32_t dthp_ref; /* reference count */ 1037 dof_helper_t dthp_prov; /* DOF w/ provider and probes */ 1038 } dtrace_helper_provider_t; 1039 1040 typedef struct dtrace_helpers { 1041 dtrace_helper_action_t **dthps_actions; /* array of helper actions */ 1042 dtrace_vstate_t dthps_vstate; /* helper action var. state */ 1043 dtrace_helper_provider_t **dthps_provs; /* array of providers */ 1044 uint_t dthps_nprovs; /* count of providers */ 1045 uint_t dthps_maxprovs; /* provider array size */ 1046 int dthps_generation; /* current generation */ 1047 pid_t dthps_pid; /* pid of associated proc */ 1048 int dthps_deferred; /* helper in deferred list */ 1049 struct dtrace_helpers *dthps_next; /* next pointer */ 1050 struct dtrace_helpers *dthps_prev; /* prev pointer */ 1051 } dtrace_helpers_t; 1052 1053 /* 1054 * DTrace Helper Action Tracing 1055 * 1056 * Debugging helper actions can be arduous. To ease the development and 1057 * debugging of helpers, DTrace contains a tracing-framework-within-a-tracing- 1058 * framework: helper tracing. If dtrace_helptrace_enabled is non-zero (which 1059 * it is by default on DEBUG kernels), all helper activity will be traced to a 1060 * global, in-kernel ring buffer. Each entry includes a pointer to the specific 1061 * helper, the location within the helper, and a trace of all local variables. 1062 * The ring buffer may be displayed in a human-readable format with the 1063 * ::dtrace_helptrace mdb(1) dcmd. 1064 */ 1065 #define DTRACE_HELPTRACE_NEXT (-1) 1066 #define DTRACE_HELPTRACE_DONE (-2) 1067 #define DTRACE_HELPTRACE_ERR (-3) 1068 1069 typedef struct dtrace_helptrace { 1070 dtrace_helper_action_t *dtht_helper; /* helper action */ 1071 int dtht_where; /* where in helper action */ 1072 int dtht_nlocals; /* number of locals */ 1073 int dtht_fault; /* type of fault (if any) */ 1074 int dtht_fltoffs; /* DIF offset */ 1075 uint64_t dtht_illval; /* faulting value */ 1076 uint64_t dtht_locals[1]; /* local variables */ 1077 } dtrace_helptrace_t; 1078 1079 /* 1080 * DTrace Credentials 1081 * 1082 * In probe context, we have limited flexibility to examine the credentials 1083 * of the DTrace consumer that created a particular enabling. We use 1084 * the Least Privilege interfaces to cache the consumer's cred pointer and 1085 * some facts about that credential in a dtrace_cred_t structure. These 1086 * can limit the consumer's breadth of visibility and what actions the 1087 * consumer may take. 1088 */ 1089 #define DTRACE_CRV_ALLPROC 0x01 1090 #define DTRACE_CRV_KERNEL 0x02 1091 #define DTRACE_CRV_ALLZONE 0x04 1092 1093 #define DTRACE_CRV_ALL (DTRACE_CRV_ALLPROC | DTRACE_CRV_KERNEL | \ 1094 DTRACE_CRV_ALLZONE) 1095 1096 #define DTRACE_CRA_PROC 0x0001 1097 #define DTRACE_CRA_PROC_CONTROL 0x0002 1098 #define DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER 0x0004 1099 #define DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE 0x0008 1100 #define DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG 0x0010 1101 #define DTRACE_CRA_KERNEL 0x0020 1102 #define DTRACE_CRA_KERNEL_DESTRUCTIVE 0x0040 1103 1104 #define DTRACE_CRA_ALL (DTRACE_CRA_PROC | \ 1105 DTRACE_CRA_PROC_CONTROL | \ 1106 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER | \ 1107 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE | \ 1108 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG | \ 1109 DTRACE_CRA_KERNEL | \ 1110 DTRACE_CRA_KERNEL_DESTRUCTIVE) 1111 1112 typedef struct dtrace_cred { 1113 cred_t *dcr_cred; 1114 uint8_t dcr_destructive; 1115 uint8_t dcr_visible; 1116 uint16_t dcr_action; 1117 } dtrace_cred_t; 1118 1119 /* 1120 * DTrace Consumer State 1121 * 1122 * Each DTrace consumer has an associated dtrace_state structure that contains 1123 * its in-kernel DTrace state -- including options, credentials, statistics and 1124 * pointers to ECBs, buffers, speculations and formats. A dtrace_state 1125 * structure is also allocated for anonymous enablings. When anonymous state 1126 * is grabbed, the grabbing consumers dts_anon pointer is set to the grabbed 1127 * dtrace_state structure. 1128 */ 1129 struct dtrace_state { 1130 #ifdef illumos 1131 dev_t dts_dev; /* device */ 1132 #else 1133 struct cdev *dts_dev; /* device */ 1134 #endif 1135 int dts_necbs; /* total number of ECBs */ 1136 dtrace_ecb_t **dts_ecbs; /* array of ECBs */ 1137 dtrace_epid_t dts_epid; /* next EPID to allocate */ 1138 size_t dts_needed; /* greatest needed space */ 1139 struct dtrace_state *dts_anon; /* anon. state, if grabbed */ 1140 dtrace_activity_t dts_activity; /* current activity */ 1141 dtrace_vstate_t dts_vstate; /* variable state */ 1142 dtrace_buffer_t *dts_buffer; /* principal buffer */ 1143 dtrace_buffer_t *dts_aggbuffer; /* aggregation buffer */ 1144 dtrace_speculation_t *dts_speculations; /* speculation array */ 1145 int dts_nspeculations; /* number of speculations */ 1146 int dts_naggregations; /* number of aggregations */ 1147 dtrace_aggregation_t **dts_aggregations; /* aggregation array */ 1148 #ifdef illumos 1149 vmem_t *dts_aggid_arena; /* arena for aggregation IDs */ 1150 #else 1151 struct unrhdr *dts_aggid_arena; /* arena for aggregation IDs */ 1152 #endif 1153 uint64_t dts_errors; /* total number of errors */ 1154 uint32_t dts_speculations_busy; /* number of spec. busy */ 1155 uint32_t dts_speculations_unavail; /* number of spec unavail */ 1156 uint32_t dts_stkstroverflows; /* stack string tab overflows */ 1157 uint32_t dts_dblerrors; /* errors in ERROR probes */ 1158 uint32_t dts_reserve; /* space reserved for END */ 1159 hrtime_t dts_laststatus; /* time of last status */ 1160 #ifdef illumos 1161 cyclic_id_t dts_cleaner; /* cleaning cyclic */ 1162 cyclic_id_t dts_deadman; /* deadman cyclic */ 1163 #else 1164 struct callout dts_cleaner; /* Cleaning callout. */ 1165 struct callout dts_deadman; /* Deadman callout. */ 1166 #endif 1167 hrtime_t dts_alive; /* time last alive */ 1168 char dts_speculates; /* boolean: has speculations */ 1169 char dts_destructive; /* boolean: has dest. actions */ 1170 int dts_nformats; /* number of formats */ 1171 char **dts_formats; /* format string array */ 1172 dtrace_optval_t dts_options[DTRACEOPT_MAX]; /* options */ 1173 dtrace_cred_t dts_cred; /* credentials */ 1174 size_t dts_nretained; /* number of retained enabs */ 1175 int dts_getf; /* number of getf() calls */ 1176 uint64_t dts_rstate[NCPU][2]; /* per-CPU random state */ 1177 }; 1178 1179 struct dtrace_provider { 1180 dtrace_pattr_t dtpv_attr; /* provider attributes */ 1181 dtrace_ppriv_t dtpv_priv; /* provider privileges */ 1182 dtrace_pops_t dtpv_pops; /* provider operations */ 1183 char *dtpv_name; /* provider name */ 1184 void *dtpv_arg; /* provider argument */ 1185 hrtime_t dtpv_defunct; /* when made defunct */ 1186 struct dtrace_provider *dtpv_next; /* next provider */ 1187 }; 1188 1189 struct dtrace_meta { 1190 dtrace_mops_t dtm_mops; /* meta provider operations */ 1191 char *dtm_name; /* meta provider name */ 1192 void *dtm_arg; /* meta provider user arg */ 1193 uint64_t dtm_count; /* no. of associated provs. */ 1194 }; 1195 1196 /* 1197 * DTrace Enablings 1198 * 1199 * A dtrace_enabling structure is used to track a collection of ECB 1200 * descriptions -- before they have been turned into actual ECBs. This is 1201 * created as a result of DOF processing, and is generally used to generate 1202 * ECBs immediately thereafter. However, enablings are also generally 1203 * retained should the probes they describe be created at a later time; as 1204 * each new module or provider registers with the framework, the retained 1205 * enablings are reevaluated, with any new match resulting in new ECBs. To 1206 * prevent probes from being matched more than once, the enabling tracks the 1207 * last probe generation matched, and only matches probes from subsequent 1208 * generations. 1209 */ 1210 typedef struct dtrace_enabling { 1211 dtrace_ecbdesc_t **dten_desc; /* all ECB descriptions */ 1212 int dten_ndesc; /* number of ECB descriptions */ 1213 int dten_maxdesc; /* size of ECB array */ 1214 dtrace_vstate_t *dten_vstate; /* associated variable state */ 1215 dtrace_genid_t dten_probegen; /* matched probe generation */ 1216 dtrace_ecbdesc_t *dten_current; /* current ECB description */ 1217 int dten_error; /* current error value */ 1218 int dten_primed; /* boolean: set if primed */ 1219 struct dtrace_enabling *dten_prev; /* previous enabling */ 1220 struct dtrace_enabling *dten_next; /* next enabling */ 1221 } dtrace_enabling_t; 1222 1223 /* 1224 * DTrace Anonymous Enablings 1225 * 1226 * Anonymous enablings are DTrace enablings that are not associated with a 1227 * controlling process, but rather derive their enabling from DOF stored as 1228 * properties in the dtrace.conf file. If there is an anonymous enabling, a 1229 * DTrace consumer state and enabling are created on attach. The state may be 1230 * subsequently grabbed by the first consumer specifying the "grabanon" 1231 * option. As long as an anonymous DTrace enabling exists, dtrace(7D) will 1232 * refuse to unload. 1233 */ 1234 typedef struct dtrace_anon { 1235 dtrace_state_t *dta_state; /* DTrace consumer state */ 1236 dtrace_enabling_t *dta_enabling; /* pointer to enabling */ 1237 processorid_t dta_beganon; /* which CPU BEGIN ran on */ 1238 } dtrace_anon_t; 1239 1240 /* 1241 * DTrace Error Debugging 1242 */ 1243 #ifdef DEBUG 1244 #define DTRACE_ERRDEBUG 1245 #endif 1246 1247 #ifdef DTRACE_ERRDEBUG 1248 1249 typedef struct dtrace_errhash { 1250 const char *dter_msg; /* error message */ 1251 int dter_count; /* number of times seen */ 1252 } dtrace_errhash_t; 1253 1254 #define DTRACE_ERRHASHSZ 256 /* must be > number of err msgs */ 1255 1256 #endif /* DTRACE_ERRDEBUG */ 1257 1258 /* 1259 * DTrace Toxic Ranges 1260 * 1261 * DTrace supports safe loads from probe context; if the address turns out to 1262 * be invalid, a bit will be set by the kernel indicating that DTrace 1263 * encountered a memory error, and DTrace will propagate the error to the user 1264 * accordingly. However, there may exist some regions of memory in which an 1265 * arbitrary load can change system state, and from which it is impossible to 1266 * recover from such a load after it has been attempted. Examples of this may 1267 * include memory in which programmable I/O registers are mapped (for which a 1268 * read may have some implications for the device) or (in the specific case of 1269 * UltraSPARC-I and -II) the virtual address hole. The platform is required 1270 * to make DTrace aware of these toxic ranges; DTrace will then check that 1271 * target addresses are not in a toxic range before attempting to issue a 1272 * safe load. 1273 */ 1274 typedef struct dtrace_toxrange { 1275 uintptr_t dtt_base; /* base of toxic range */ 1276 uintptr_t dtt_limit; /* limit of toxic range */ 1277 } dtrace_toxrange_t; 1278 1279 #ifdef illumos 1280 extern uint64_t dtrace_getarg(int, int); 1281 #else 1282 extern uint64_t __noinline dtrace_getarg(int, int); 1283 #endif 1284 extern greg_t dtrace_getfp(void); 1285 extern int dtrace_getipl(void); 1286 extern uintptr_t dtrace_caller(int); 1287 extern uint32_t dtrace_cas32(uint32_t *, uint32_t, uint32_t); 1288 extern void *dtrace_casptr(volatile void *, volatile void *, volatile void *); 1289 extern void dtrace_copyin(uintptr_t, uintptr_t, size_t, volatile uint16_t *); 1290 extern void dtrace_copyinstr(uintptr_t, uintptr_t, size_t, volatile uint16_t *); 1291 extern void dtrace_copyout(uintptr_t, uintptr_t, size_t, volatile uint16_t *); 1292 extern void dtrace_copyoutstr(uintptr_t, uintptr_t, size_t, 1293 volatile uint16_t *); 1294 extern void dtrace_getpcstack(pc_t *, int, int, uint32_t *); 1295 extern ulong_t dtrace_getreg(struct trapframe *, uint_t); 1296 extern int dtrace_getstackdepth(int); 1297 extern void dtrace_getupcstack(uint64_t *, int); 1298 extern void dtrace_getufpstack(uint64_t *, uint64_t *, int); 1299 extern int dtrace_getustackdepth(void); 1300 extern uintptr_t dtrace_fulword(void *); 1301 extern uint8_t dtrace_fuword8(void *); 1302 extern uint16_t dtrace_fuword16(void *); 1303 extern uint32_t dtrace_fuword32(void *); 1304 extern uint64_t dtrace_fuword64(void *); 1305 extern void dtrace_probe_error(dtrace_state_t *, dtrace_epid_t, int, int, 1306 int, uintptr_t); 1307 extern int dtrace_assfail(const char *, const char *, int); 1308 extern int dtrace_attached(void); 1309 #ifdef illumos 1310 extern hrtime_t dtrace_gethrestime(void); 1311 #endif 1312 1313 #ifdef __sparc 1314 extern void dtrace_flush_windows(void); 1315 extern void dtrace_flush_user_windows(void); 1316 extern uint_t dtrace_getotherwin(void); 1317 extern uint_t dtrace_getfprs(void); 1318 #else 1319 extern void dtrace_copy(uintptr_t, uintptr_t, size_t); 1320 extern void dtrace_copystr(uintptr_t, uintptr_t, size_t, volatile uint16_t *); 1321 #endif 1322 1323 /* 1324 * DTrace Assertions 1325 * 1326 * DTrace calls ASSERT and VERIFY from probe context. To assure that a failed 1327 * ASSERT or VERIFY does not induce a markedly more catastrophic failure (e.g., 1328 * one from which a dump cannot be gleaned), DTrace must define its own ASSERT 1329 * and VERIFY macros to be ones that may safely be called from probe context. 1330 * This header file must thus be included by any DTrace component that calls 1331 * ASSERT and/or VERIFY from probe context, and _only_ by those components. 1332 * (The only exception to this is kernel debugging infrastructure at user-level 1333 * that doesn't depend on calling ASSERT.) 1334 */ 1335 #undef ASSERT 1336 #undef VERIFY 1337 #define VERIFY(EX) ((void)((EX) || \ 1338 dtrace_assfail(#EX, __FILE__, __LINE__))) 1339 #ifdef DEBUG 1340 #define ASSERT(EX) ((void)((EX) || \ 1341 dtrace_assfail(#EX, __FILE__, __LINE__))) 1342 #else 1343 #define ASSERT(X) ((void)0) 1344 #endif 1345 1346 #ifdef __cplusplus 1347 } 1348 #endif 1349 1350 #endif /* _SYS_DTRACE_IMPL_H */ 1351