1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause 3 * 4 * Copyright (c) 2005-2007 Joseph Koshy 5 * Copyright (c) 2007 The FreeBSD Foundation 6 * All rights reserved. 7 * 8 * Portions of this software were developed by A. Joseph Koshy under 9 * sponsorship from the FreeBSD Foundation and Google, Inc. 10 * 11 * Redistribution and use in source and binary forms, with or without 12 * modification, are permitted provided that the following conditions 13 * are met: 14 * 1. Redistributions of source code must retain the above copyright 15 * notice, this list of conditions and the following disclaimer. 16 * 2. Redistributions in binary form must reproduce the above copyright 17 * notice, this list of conditions and the following disclaimer in the 18 * documentation and/or other materials provided with the distribution. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30 * SUCH DAMAGE. 31 */ 32 33 #include <sys/param.h> 34 #include <sys/pmc.h> 35 #include <sys/pmclog.h> 36 37 #include <assert.h> 38 #include <errno.h> 39 #include <pmc.h> 40 #include <pmclog.h> 41 #include <stddef.h> 42 #include <stdlib.h> 43 #include <string.h> 44 #include <strings.h> 45 #include <unistd.h> 46 #include <stdio.h> 47 48 #include <machine/pmc_mdep.h> 49 50 #include "libpmcinternal.h" 51 52 #define PMCLOG_BUFFER_SIZE 512*1024 53 54 /* 55 * API NOTES 56 * 57 * The pmclog(3) API is oriented towards parsing an event stream in 58 * "realtime", i.e., from an data source that may or may not preserve 59 * record boundaries -- for example when the data source is elsewhere 60 * on a network. The API allows data to be fed into the parser zero 61 * or more bytes at a time. 62 * 63 * The state for a log file parser is maintained in a 'struct 64 * pmclog_parse_state'. Parser invocations are done by calling 65 * 'pmclog_read()'; this function will inform the caller when a 66 * complete event is parsed. 67 * 68 * The parser first assembles a complete log file event in an internal 69 * work area (see "ps_saved" below). Once a complete log file event 70 * is read, the parser then parses it and converts it to an event 71 * descriptor usable by the client. We could possibly avoid this two 72 * step process by directly parsing the input log to set fields in the 73 * event record. However the parser's state machine would get 74 * insanely complicated, and this code is unlikely to be used in 75 * performance critical paths. 76 */ 77 78 #define PMCLOG_HEADER_FROM_SAVED_STATE(PS) \ 79 (* ((uint32_t *) &(PS)->ps_saved)) 80 81 #define PMCLOG_INITIALIZE_READER(LE,A) LE = (uint32_t *) &(A) 82 #define PMCLOG_SKIP32(LE) (LE)++ 83 #define PMCLOG_READ32(LE,V) do { \ 84 (V) = *(LE)++; \ 85 } while (0) 86 #define PMCLOG_READ64(LE,V) do { \ 87 uint64_t _v; \ 88 _v = (uint64_t) *(LE)++; \ 89 _v |= ((uint64_t) *(LE)++) << 32; \ 90 (V) = _v; \ 91 } while (0) 92 93 #define PMCLOG_READSTRING(LE,DST,LEN) strlcpy((DST), (char *) (LE), (LEN)) 94 95 /* 96 * Assemble a log record from '*len' octets starting from address '*data'. 97 * Update 'data' and 'len' to reflect the number of bytes consumed. 98 * 99 * '*data' is potentially an unaligned address and '*len' octets may 100 * not be enough to complete a event record. 101 */ 102 103 static enum pmclog_parser_state 104 pmclog_get_record(struct pmclog_parse_state *ps, char **data, ssize_t *len) 105 { 106 int avail, copylen, recordsize, used; 107 uint32_t h; 108 const int HEADERSIZE = sizeof(uint32_t); 109 char *src, *dst; 110 111 if ((avail = *len) <= 0) 112 return (ps->ps_state = PL_STATE_ERROR); 113 114 src = *data; 115 used = 0; 116 117 if (ps->ps_state == PL_STATE_NEW_RECORD) 118 ps->ps_svcount = 0; 119 120 dst = (char *) &ps->ps_saved + ps->ps_svcount; 121 122 switch (ps->ps_state) { 123 case PL_STATE_NEW_RECORD: 124 125 /* 126 * Transitions: 127 * 128 * Case A: avail < headersize 129 * -> 'expecting header' 130 * 131 * Case B: avail >= headersize 132 * B.1: avail < recordsize 133 * -> 'partial record' 134 * B.2: avail >= recordsize 135 * -> 'new record' 136 */ 137 138 copylen = avail < HEADERSIZE ? avail : HEADERSIZE; 139 bcopy(src, dst, copylen); 140 ps->ps_svcount = used = copylen; 141 142 if (copylen < HEADERSIZE) { 143 ps->ps_state = PL_STATE_EXPECTING_HEADER; 144 goto done; 145 } 146 147 src += copylen; 148 dst += copylen; 149 150 h = PMCLOG_HEADER_FROM_SAVED_STATE(ps); 151 recordsize = PMCLOG_HEADER_TO_LENGTH(h); 152 153 if (recordsize <= 0) 154 goto error; 155 156 if (recordsize <= avail) { /* full record available */ 157 bcopy(src, dst, recordsize - copylen); 158 ps->ps_svcount = used = recordsize; 159 goto done; 160 } 161 162 /* header + a partial record is available */ 163 bcopy(src, dst, avail - copylen); 164 ps->ps_svcount = used = avail; 165 ps->ps_state = PL_STATE_PARTIAL_RECORD; 166 167 break; 168 169 case PL_STATE_EXPECTING_HEADER: 170 171 /* 172 * Transitions: 173 * 174 * Case C: avail+saved < headersize 175 * -> 'expecting header' 176 * 177 * Case D: avail+saved >= headersize 178 * D.1: avail+saved < recordsize 179 * -> 'partial record' 180 * D.2: avail+saved >= recordsize 181 * -> 'new record' 182 * (see PARTIAL_RECORD handling below) 183 */ 184 185 if (avail + ps->ps_svcount < HEADERSIZE) { 186 bcopy(src, dst, avail); 187 ps->ps_svcount += avail; 188 used = avail; 189 break; 190 } 191 192 used = copylen = HEADERSIZE - ps->ps_svcount; 193 bcopy(src, dst, copylen); 194 src += copylen; 195 dst += copylen; 196 avail -= copylen; 197 ps->ps_svcount += copylen; 198 199 /*FALLTHROUGH*/ 200 201 case PL_STATE_PARTIAL_RECORD: 202 203 /* 204 * Transitions: 205 * 206 * Case E: avail+saved < recordsize 207 * -> 'partial record' 208 * 209 * Case F: avail+saved >= recordsize 210 * -> 'new record' 211 */ 212 213 h = PMCLOG_HEADER_FROM_SAVED_STATE(ps); 214 recordsize = PMCLOG_HEADER_TO_LENGTH(h); 215 216 if (recordsize <= 0) 217 goto error; 218 219 if (avail + ps->ps_svcount < recordsize) { 220 copylen = avail; 221 ps->ps_state = PL_STATE_PARTIAL_RECORD; 222 } else { 223 copylen = recordsize - ps->ps_svcount; 224 ps->ps_state = PL_STATE_NEW_RECORD; 225 } 226 227 bcopy(src, dst, copylen); 228 ps->ps_svcount += copylen; 229 used += copylen; 230 break; 231 232 default: 233 goto error; 234 } 235 236 done: 237 *data += used; 238 *len -= used; 239 return ps->ps_state; 240 241 error: 242 ps->ps_state = PL_STATE_ERROR; 243 return ps->ps_state; 244 } 245 246 /* 247 * Get an event from the stream pointed to by '*data'. '*len' 248 * indicates the number of bytes available to parse. Arguments 249 * '*data' and '*len' are updated to indicate the number of bytes 250 * consumed. 251 */ 252 253 static int 254 pmclog_get_event(void *cookie, char **data, ssize_t *len, 255 struct pmclog_ev *ev) 256 { 257 int evlen, pathlen; 258 uint32_t h, *le, npc; 259 enum pmclog_parser_state e; 260 struct pmclog_parse_state *ps; 261 struct pmclog_header *ph; 262 263 ps = (struct pmclog_parse_state *) cookie; 264 265 assert(ps->ps_state != PL_STATE_ERROR); 266 267 if ((e = pmclog_get_record(ps,data,len)) == PL_STATE_ERROR) { 268 ev->pl_state = PMCLOG_ERROR; 269 printf("state error\n"); 270 return -1; 271 } 272 273 if (e != PL_STATE_NEW_RECORD) { 274 ev->pl_state = PMCLOG_REQUIRE_DATA; 275 return -1; 276 } 277 278 PMCLOG_INITIALIZE_READER(le, ps->ps_saved); 279 ev->pl_data = le; 280 ph = (struct pmclog_header *)(uintptr_t)le; 281 282 h = ph->pl_header; 283 if (!PMCLOG_HEADER_CHECK_MAGIC(h)) { 284 printf("bad magic\n"); 285 ps->ps_state = PL_STATE_ERROR; 286 ev->pl_state = PMCLOG_ERROR; 287 return -1; 288 } 289 290 /* copy out the time stamp */ 291 ev->pl_ts.tv_sec = ph->pl_tsc; 292 le += sizeof(*ph)/4; 293 294 evlen = PMCLOG_HEADER_TO_LENGTH(h); 295 296 #define PMCLOG_GET_PATHLEN(P,E,TYPE) do { \ 297 (P) = (E) - offsetof(struct TYPE, pl_pathname); \ 298 if ((P) > PATH_MAX || (P) < 0) \ 299 goto error; \ 300 } while (0) 301 302 #define PMCLOG_GET_CALLCHAIN_SIZE(SZ,E) do { \ 303 (SZ) = ((E) - offsetof(struct pmclog_callchain, pl_pc)) \ 304 / sizeof(uintfptr_t); \ 305 } while (0); 306 307 switch (ev->pl_type = PMCLOG_HEADER_TO_TYPE(h)) { 308 case PMCLOG_TYPE_CALLCHAIN: 309 PMCLOG_READ32(le,ev->pl_u.pl_cc.pl_pid); 310 PMCLOG_READ32(le,ev->pl_u.pl_cc.pl_tid); 311 PMCLOG_READ32(le,ev->pl_u.pl_cc.pl_pmcid); 312 PMCLOG_READ32(le,ev->pl_u.pl_cc.pl_cpuflags); 313 PMCLOG_GET_CALLCHAIN_SIZE(ev->pl_u.pl_cc.pl_npc,evlen); 314 for (npc = 0; npc < ev->pl_u.pl_cc.pl_npc; npc++) 315 PMCLOG_READADDR(le,ev->pl_u.pl_cc.pl_pc[npc]); 316 for (;npc < PMC_CALLCHAIN_DEPTH_MAX; npc++) 317 ev->pl_u.pl_cc.pl_pc[npc] = (uintfptr_t) 0; 318 break; 319 case PMCLOG_TYPE_CLOSELOG: 320 ev->pl_state = PMCLOG_EOF; 321 return (-1); 322 case PMCLOG_TYPE_DROPNOTIFY: 323 /* nothing to do */ 324 break; 325 case PMCLOG_TYPE_INITIALIZE: 326 PMCLOG_READ32(le,ev->pl_u.pl_i.pl_version); 327 PMCLOG_READ32(le,ev->pl_u.pl_i.pl_arch); 328 PMCLOG_READ64(le,ev->pl_u.pl_i.pl_tsc_freq); 329 memcpy(&ev->pl_u.pl_i.pl_ts, le, sizeof(struct timespec)); 330 le += sizeof(struct timespec)/4; 331 PMCLOG_READSTRING(le, ev->pl_u.pl_i.pl_cpuid, PMC_CPUID_LEN); 332 memcpy(ev->pl_u.pl_i.pl_cpuid, le, PMC_CPUID_LEN); 333 ps->ps_cpuid = strdup(ev->pl_u.pl_i.pl_cpuid); 334 ps->ps_version = ev->pl_u.pl_i.pl_version; 335 ps->ps_arch = ev->pl_u.pl_i.pl_arch; 336 ps->ps_initialized = 1; 337 break; 338 case PMCLOG_TYPE_MAP_IN: 339 PMCLOG_GET_PATHLEN(pathlen,evlen,pmclog_map_in); 340 PMCLOG_READ32(le,ev->pl_u.pl_mi.pl_pid); 341 PMCLOG_SKIP32(le); 342 PMCLOG_READADDR(le,ev->pl_u.pl_mi.pl_start); 343 PMCLOG_READSTRING(le, ev->pl_u.pl_mi.pl_pathname, pathlen); 344 break; 345 case PMCLOG_TYPE_MAP_OUT: 346 PMCLOG_READ32(le,ev->pl_u.pl_mo.pl_pid); 347 PMCLOG_SKIP32(le); 348 PMCLOG_READADDR(le,ev->pl_u.pl_mo.pl_start); 349 PMCLOG_READADDR(le,ev->pl_u.pl_mo.pl_end); 350 break; 351 case PMCLOG_TYPE_PMCALLOCATE: 352 PMCLOG_READ32(le,ev->pl_u.pl_a.pl_pmcid); 353 PMCLOG_READ32(le,ev->pl_u.pl_a.pl_event); 354 PMCLOG_READ32(le,ev->pl_u.pl_a.pl_flags); 355 PMCLOG_SKIP32(le); 356 PMCLOG_READ64(le,ev->pl_u.pl_a.pl_rate); 357 358 /* 359 * pl_event could contain either a PMC event code or a PMU 360 * event index. 361 */ 362 if ((ev->pl_u.pl_a.pl_flags & PMC_F_EV_PMU) != 0) 363 ev->pl_u.pl_a.pl_evname = 364 pmc_pmu_event_get_by_idx(ps->ps_cpuid, 365 ev->pl_u.pl_a.pl_event); 366 else if (ev->pl_u.pl_a.pl_event <= PMC_EVENT_LAST) 367 ev->pl_u.pl_a.pl_evname = 368 _pmc_name_of_event(ev->pl_u.pl_a.pl_event, 369 ps->ps_arch); 370 else 371 ev->pl_u.pl_a.pl_evname = NULL; 372 if (ev->pl_u.pl_a.pl_evname == NULL) { 373 printf("unknown event\n"); 374 goto error; 375 } 376 break; 377 case PMCLOG_TYPE_PMCALLOCATEDYN: 378 PMCLOG_READ32(le,ev->pl_u.pl_ad.pl_pmcid); 379 PMCLOG_READ32(le,ev->pl_u.pl_ad.pl_event); 380 PMCLOG_READ32(le,ev->pl_u.pl_ad.pl_flags); 381 PMCLOG_SKIP32(le); 382 PMCLOG_READSTRING(le,ev->pl_u.pl_ad.pl_evname,PMC_NAME_MAX); 383 break; 384 case PMCLOG_TYPE_PMCATTACH: 385 PMCLOG_GET_PATHLEN(pathlen,evlen,pmclog_pmcattach); 386 PMCLOG_READ32(le,ev->pl_u.pl_t.pl_pmcid); 387 PMCLOG_READ32(le,ev->pl_u.pl_t.pl_pid); 388 PMCLOG_READSTRING(le,ev->pl_u.pl_t.pl_pathname,pathlen); 389 break; 390 case PMCLOG_TYPE_PMCDETACH: 391 PMCLOG_READ32(le,ev->pl_u.pl_d.pl_pmcid); 392 PMCLOG_READ32(le,ev->pl_u.pl_d.pl_pid); 393 break; 394 case PMCLOG_TYPE_PROCCSW: 395 PMCLOG_READ64(le,ev->pl_u.pl_c.pl_value); 396 PMCLOG_READ32(le,ev->pl_u.pl_c.pl_pmcid); 397 PMCLOG_READ32(le,ev->pl_u.pl_c.pl_pid); 398 PMCLOG_READ32(le,ev->pl_u.pl_c.pl_tid); 399 break; 400 case PMCLOG_TYPE_PROCEXEC: 401 PMCLOG_GET_PATHLEN(pathlen,evlen,pmclog_procexec); 402 PMCLOG_READ32(le,ev->pl_u.pl_x.pl_pid); 403 PMCLOG_READ32(le,ev->pl_u.pl_x.pl_pmcid); 404 PMCLOG_READADDR(le,ev->pl_u.pl_x.pl_baseaddr); 405 PMCLOG_READADDR(le,ev->pl_u.pl_x.pl_dynaddr); 406 PMCLOG_READSTRING(le,ev->pl_u.pl_x.pl_pathname,pathlen); 407 break; 408 case PMCLOG_TYPE_PROCEXIT: 409 PMCLOG_READ32(le,ev->pl_u.pl_e.pl_pmcid); 410 PMCLOG_READ32(le,ev->pl_u.pl_e.pl_pid); 411 PMCLOG_READ64(le,ev->pl_u.pl_e.pl_value); 412 break; 413 case PMCLOG_TYPE_PROCFORK: 414 PMCLOG_READ32(le,ev->pl_u.pl_f.pl_oldpid); 415 PMCLOG_READ32(le,ev->pl_u.pl_f.pl_newpid); 416 break; 417 case PMCLOG_TYPE_SYSEXIT: 418 PMCLOG_READ32(le,ev->pl_u.pl_se.pl_pid); 419 break; 420 case PMCLOG_TYPE_USERDATA: 421 PMCLOG_READ32(le,ev->pl_u.pl_u.pl_userdata); 422 break; 423 case PMCLOG_TYPE_THR_CREATE: 424 PMCLOG_READ32(le,ev->pl_u.pl_tc.pl_tid); 425 PMCLOG_READ32(le,ev->pl_u.pl_tc.pl_pid); 426 PMCLOG_READ32(le,ev->pl_u.pl_tc.pl_flags); 427 PMCLOG_SKIP32(le); 428 memcpy(ev->pl_u.pl_tc.pl_tdname, le, MAXCOMLEN+1); 429 break; 430 case PMCLOG_TYPE_THR_EXIT: 431 PMCLOG_READ32(le,ev->pl_u.pl_te.pl_tid); 432 break; 433 case PMCLOG_TYPE_PROC_CREATE: 434 PMCLOG_READ32(le,ev->pl_u.pl_pc.pl_pid); 435 PMCLOG_READ32(le,ev->pl_u.pl_pc.pl_flags); 436 memcpy(ev->pl_u.pl_pc.pl_pcomm, le, MAXCOMLEN+1); 437 break; 438 default: /* unknown record type */ 439 ps->ps_state = PL_STATE_ERROR; 440 ev->pl_state = PMCLOG_ERROR; 441 return (-1); 442 } 443 444 ev->pl_offset = (ps->ps_offset += evlen); 445 ev->pl_count = (ps->ps_count += 1); 446 ev->pl_len = evlen; 447 ev->pl_state = PMCLOG_OK; 448 return 0; 449 450 error: 451 ev->pl_state = PMCLOG_ERROR; 452 ps->ps_state = PL_STATE_ERROR; 453 return -1; 454 } 455 456 /* 457 * Extract and return the next event from the byte stream. 458 * 459 * Returns 0 and sets the event's state to PMCLOG_OK in case an event 460 * was successfully parsed. Otherwise this function returns -1 and 461 * sets the event's state to one of PMCLOG_REQUIRE_DATA (if more data 462 * is needed) or PMCLOG_EOF (if an EOF was seen) or PMCLOG_ERROR if 463 * a parse error was encountered. 464 */ 465 466 int 467 pmclog_read(void *cookie, struct pmclog_ev *ev) 468 { 469 int retval; 470 ssize_t nread; 471 struct pmclog_parse_state *ps; 472 473 ps = (struct pmclog_parse_state *) cookie; 474 475 if (ps->ps_state == PL_STATE_ERROR) { 476 ev->pl_state = PMCLOG_ERROR; 477 return -1; 478 } 479 480 /* 481 * If there isn't enough data left for a new event try and get 482 * more data. 483 */ 484 if (ps->ps_len == 0) { 485 ev->pl_state = PMCLOG_REQUIRE_DATA; 486 487 /* 488 * If we have a valid file descriptor to read from, attempt 489 * to read from that. This read may return with an error, 490 * (which may be EAGAIN or other recoverable error), or 491 * can return EOF. 492 */ 493 if (ps->ps_fd != PMCLOG_FD_NONE) { 494 refill: 495 nread = read(ps->ps_fd, ps->ps_buffer, 496 PMCLOG_BUFFER_SIZE); 497 498 if (nread <= 0) { 499 if (nread == 0) 500 ev->pl_state = PMCLOG_EOF; 501 else if (errno != EAGAIN) /* not restartable */ 502 ev->pl_state = PMCLOG_ERROR; 503 return -1; 504 } 505 506 ps->ps_len = nread; 507 ps->ps_data = ps->ps_buffer; 508 } else { 509 return -1; 510 } 511 } 512 513 assert(ps->ps_len > 0); 514 515 516 /* Retrieve one event from the byte stream. */ 517 retval = pmclog_get_event(ps, &ps->ps_data, &ps->ps_len, ev); 518 /* 519 * If we need more data and we have a configured fd, try read 520 * from it. 521 */ 522 if (retval < 0 && ev->pl_state == PMCLOG_REQUIRE_DATA && 523 ps->ps_fd != -1) { 524 assert(ps->ps_len == 0); 525 goto refill; 526 } 527 528 return retval; 529 } 530 531 /* 532 * Feed data to a memory based parser. 533 * 534 * The memory area pointed to by 'data' needs to be valid till the 535 * next error return from pmclog_next_event(). 536 */ 537 538 int 539 pmclog_feed(void *cookie, char *data, int len) 540 { 541 struct pmclog_parse_state *ps; 542 543 ps = (struct pmclog_parse_state *) cookie; 544 545 if (len < 0 || /* invalid length */ 546 ps->ps_buffer || /* called for a file parser */ 547 ps->ps_len != 0) /* unnecessary call */ 548 return -1; 549 550 ps->ps_data = data; 551 ps->ps_len = len; 552 553 return 0; 554 } 555 556 /* 557 * Allocate and initialize parser state. 558 */ 559 560 void * 561 pmclog_open(int fd) 562 { 563 struct pmclog_parse_state *ps; 564 565 if ((ps = (struct pmclog_parse_state *) malloc(sizeof(*ps))) == NULL) 566 return NULL; 567 568 ps->ps_state = PL_STATE_NEW_RECORD; 569 ps->ps_arch = -1; 570 ps->ps_initialized = 0; 571 ps->ps_count = 0; 572 ps->ps_offset = (off_t) 0; 573 bzero(&ps->ps_saved, sizeof(ps->ps_saved)); 574 ps->ps_cpuid = NULL; 575 ps->ps_svcount = 0; 576 ps->ps_fd = fd; 577 ps->ps_data = NULL; 578 ps->ps_buffer = NULL; 579 ps->ps_len = 0; 580 581 /* allocate space for a work area */ 582 if (ps->ps_fd != PMCLOG_FD_NONE) { 583 if ((ps->ps_buffer = malloc(PMCLOG_BUFFER_SIZE)) == NULL) { 584 free(ps); 585 return NULL; 586 } 587 } 588 589 return ps; 590 } 591 592 593 /* 594 * Free up parser state. 595 */ 596 597 void 598 pmclog_close(void *cookie) 599 { 600 struct pmclog_parse_state *ps; 601 602 ps = (struct pmclog_parse_state *) cookie; 603 604 if (ps->ps_buffer) 605 free(ps->ps_buffer); 606 607 free(ps); 608 } 609