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/cdefs.h> 34 #include <sys/param.h> 35 #include <sys/pmc.h> 36 #include <sys/pmclog.h> 37 38 #include <assert.h> 39 #include <errno.h> 40 #include <pmc.h> 41 #include <pmclog.h> 42 #include <stddef.h> 43 #include <stdlib.h> 44 #include <string.h> 45 #include <strings.h> 46 #include <unistd.h> 47 #include <stdio.h> 48 49 #include <machine/pmc_mdep.h> 50 51 #include "libpmcinternal.h" 52 53 #define PMCLOG_BUFFER_SIZE 512*1024 54 55 /* 56 * API NOTES 57 * 58 * The pmclog(3) API is oriented towards parsing an event stream in 59 * "realtime", i.e., from an data source that may or may not preserve 60 * record boundaries -- for example when the data source is elsewhere 61 * on a network. The API allows data to be fed into the parser zero 62 * or more bytes at a time. 63 * 64 * The state for a log file parser is maintained in a 'struct 65 * pmclog_parse_state'. Parser invocations are done by calling 66 * 'pmclog_read()'; this function will inform the caller when a 67 * complete event is parsed. 68 * 69 * The parser first assembles a complete log file event in an internal 70 * work area (see "ps_saved" below). Once a complete log file event 71 * is read, the parser then parses it and converts it to an event 72 * descriptor usable by the client. We could possibly avoid this two 73 * step process by directly parsing the input log to set fields in the 74 * event record. However the parser's state machine would get 75 * insanely complicated, and this code is unlikely to be used in 76 * performance critical paths. 77 */ 78 79 #define PMCLOG_HEADER_FROM_SAVED_STATE(PS) \ 80 (* ((uint32_t *) &(PS)->ps_saved)) 81 82 #define PMCLOG_INITIALIZE_READER(LE,A) LE = (uint32_t *) &(A) 83 #define PMCLOG_SKIP32(LE) (LE)++ 84 #define PMCLOG_READ32(LE,V) do { \ 85 (V) = *(LE)++; \ 86 } while (0) 87 #define PMCLOG_READ64(LE,V) do { \ 88 uint64_t _v; \ 89 _v = (uint64_t) *(LE)++; \ 90 _v |= ((uint64_t) *(LE)++) << 32; \ 91 (V) = _v; \ 92 } while (0) 93 94 #define PMCLOG_READSTRING(LE,DST,LEN) strlcpy((DST), (char *) (LE), (LEN)) 95 96 /* 97 * Assemble a log record from '*len' octets starting from address '*data'. 98 * Update 'data' and 'len' to reflect the number of bytes consumed. 99 * 100 * '*data' is potentially an unaligned address and '*len' octets may 101 * not be enough to complete a event record. 102 */ 103 104 static enum pmclog_parser_state 105 pmclog_get_record(struct pmclog_parse_state *ps, char **data, ssize_t *len) 106 { 107 int avail, copylen, recordsize, used; 108 uint32_t h; 109 const int HEADERSIZE = sizeof(uint32_t); 110 char *src, *dst; 111 112 if ((avail = *len) <= 0) 113 return (ps->ps_state = PL_STATE_ERROR); 114 115 src = *data; 116 used = 0; 117 118 if (ps->ps_state == PL_STATE_NEW_RECORD) 119 ps->ps_svcount = 0; 120 121 dst = (char *) &ps->ps_saved + ps->ps_svcount; 122 123 switch (ps->ps_state) { 124 case PL_STATE_NEW_RECORD: 125 126 /* 127 * Transitions: 128 * 129 * Case A: avail < headersize 130 * -> 'expecting header' 131 * 132 * Case B: avail >= headersize 133 * B.1: avail < recordsize 134 * -> 'partial record' 135 * B.2: avail >= recordsize 136 * -> 'new record' 137 */ 138 139 copylen = avail < HEADERSIZE ? avail : HEADERSIZE; 140 bcopy(src, dst, copylen); 141 ps->ps_svcount = used = copylen; 142 143 if (copylen < HEADERSIZE) { 144 ps->ps_state = PL_STATE_EXPECTING_HEADER; 145 goto done; 146 } 147 148 src += copylen; 149 dst += copylen; 150 151 h = PMCLOG_HEADER_FROM_SAVED_STATE(ps); 152 recordsize = PMCLOG_HEADER_TO_LENGTH(h); 153 154 if (recordsize <= 0) 155 goto error; 156 157 if (recordsize <= avail) { /* full record available */ 158 bcopy(src, dst, recordsize - copylen); 159 ps->ps_svcount = used = recordsize; 160 goto done; 161 } 162 163 /* header + a partial record is available */ 164 bcopy(src, dst, avail - copylen); 165 ps->ps_svcount = used = avail; 166 ps->ps_state = PL_STATE_PARTIAL_RECORD; 167 168 break; 169 170 case PL_STATE_EXPECTING_HEADER: 171 172 /* 173 * Transitions: 174 * 175 * Case C: avail+saved < headersize 176 * -> 'expecting header' 177 * 178 * Case D: avail+saved >= headersize 179 * D.1: avail+saved < recordsize 180 * -> 'partial record' 181 * D.2: avail+saved >= recordsize 182 * -> 'new record' 183 * (see PARTIAL_RECORD handling below) 184 */ 185 186 if (avail + ps->ps_svcount < HEADERSIZE) { 187 bcopy(src, dst, avail); 188 ps->ps_svcount += avail; 189 used = avail; 190 break; 191 } 192 193 used = copylen = HEADERSIZE - ps->ps_svcount; 194 bcopy(src, dst, copylen); 195 src += copylen; 196 dst += copylen; 197 avail -= copylen; 198 ps->ps_svcount += copylen; 199 200 /*FALLTHROUGH*/ 201 202 case PL_STATE_PARTIAL_RECORD: 203 204 /* 205 * Transitions: 206 * 207 * Case E: avail+saved < recordsize 208 * -> 'partial record' 209 * 210 * Case F: avail+saved >= recordsize 211 * -> 'new record' 212 */ 213 214 h = PMCLOG_HEADER_FROM_SAVED_STATE(ps); 215 recordsize = PMCLOG_HEADER_TO_LENGTH(h); 216 217 if (recordsize <= 0) 218 goto error; 219 220 if (avail + ps->ps_svcount < recordsize) { 221 copylen = avail; 222 ps->ps_state = PL_STATE_PARTIAL_RECORD; 223 } else { 224 copylen = recordsize - ps->ps_svcount; 225 ps->ps_state = PL_STATE_NEW_RECORD; 226 } 227 228 bcopy(src, dst, copylen); 229 ps->ps_svcount += copylen; 230 used += copylen; 231 break; 232 233 default: 234 goto error; 235 } 236 237 done: 238 *data += used; 239 *len -= used; 240 return ps->ps_state; 241 242 error: 243 ps->ps_state = PL_STATE_ERROR; 244 return ps->ps_state; 245 } 246 247 /* 248 * Get an event from the stream pointed to by '*data'. '*len' 249 * indicates the number of bytes available to parse. Arguments 250 * '*data' and '*len' are updated to indicate the number of bytes 251 * consumed. 252 */ 253 254 static int 255 pmclog_get_event(void *cookie, char **data, ssize_t *len, 256 struct pmclog_ev *ev) 257 { 258 int evlen, pathlen; 259 uint32_t h, *le, npc; 260 enum pmclog_parser_state e; 261 struct pmclog_parse_state *ps; 262 struct pmclog_header *ph; 263 264 ps = (struct pmclog_parse_state *) cookie; 265 266 assert(ps->ps_state != PL_STATE_ERROR); 267 268 if ((e = pmclog_get_record(ps,data,len)) == PL_STATE_ERROR) { 269 ev->pl_state = PMCLOG_ERROR; 270 printf("state error\n"); 271 return -1; 272 } 273 274 if (e != PL_STATE_NEW_RECORD) { 275 ev->pl_state = PMCLOG_REQUIRE_DATA; 276 return -1; 277 } 278 279 PMCLOG_INITIALIZE_READER(le, ps->ps_saved); 280 ev->pl_data = le; 281 ph = (struct pmclog_header *)(uintptr_t)le; 282 283 h = ph->pl_header; 284 if (!PMCLOG_HEADER_CHECK_MAGIC(h)) { 285 printf("bad magic\n"); 286 ps->ps_state = PL_STATE_ERROR; 287 ev->pl_state = PMCLOG_ERROR; 288 return -1; 289 } 290 291 /* copy out the time stamp */ 292 ev->pl_ts.tv_sec = ph->pl_tsc; 293 le += sizeof(*ph)/4; 294 295 evlen = PMCLOG_HEADER_TO_LENGTH(h); 296 297 #define PMCLOG_GET_PATHLEN(P,E,TYPE) do { \ 298 (P) = (E) - offsetof(struct TYPE, pl_pathname); \ 299 if ((P) > PATH_MAX || (P) < 0) \ 300 goto error; \ 301 } while (0) 302 303 #define PMCLOG_GET_CALLCHAIN_SIZE(SZ,E) do { \ 304 (SZ) = ((E) - offsetof(struct pmclog_callchain, pl_pc)) \ 305 / sizeof(uintfptr_t); \ 306 } while (0); 307 308 switch (ev->pl_type = PMCLOG_HEADER_TO_TYPE(h)) { 309 case PMCLOG_TYPE_CALLCHAIN: 310 PMCLOG_READ32(le,ev->pl_u.pl_cc.pl_pid); 311 PMCLOG_READ32(le,ev->pl_u.pl_cc.pl_tid); 312 PMCLOG_READ32(le,ev->pl_u.pl_cc.pl_pmcid); 313 PMCLOG_READ32(le,ev->pl_u.pl_cc.pl_cpuflags); 314 PMCLOG_GET_CALLCHAIN_SIZE(ev->pl_u.pl_cc.pl_npc,evlen); 315 for (npc = 0; npc < ev->pl_u.pl_cc.pl_npc; npc++) 316 PMCLOG_READADDR(le,ev->pl_u.pl_cc.pl_pc[npc]); 317 for (;npc < PMC_CALLCHAIN_DEPTH_MAX; npc++) 318 ev->pl_u.pl_cc.pl_pc[npc] = (uintfptr_t) 0; 319 break; 320 case PMCLOG_TYPE_CLOSELOG: 321 ev->pl_state = PMCLOG_EOF; 322 return (-1); 323 case PMCLOG_TYPE_DROPNOTIFY: 324 /* nothing to do */ 325 break; 326 case PMCLOG_TYPE_INITIALIZE: 327 PMCLOG_READ32(le,ev->pl_u.pl_i.pl_version); 328 PMCLOG_READ32(le,ev->pl_u.pl_i.pl_arch); 329 PMCLOG_READ64(le,ev->pl_u.pl_i.pl_tsc_freq); 330 memcpy(&ev->pl_u.pl_i.pl_ts, le, sizeof(struct timespec)); 331 le += sizeof(struct timespec)/4; 332 PMCLOG_READSTRING(le, ev->pl_u.pl_i.pl_cpuid, PMC_CPUID_LEN); 333 memcpy(ev->pl_u.pl_i.pl_cpuid, le, PMC_CPUID_LEN); 334 ps->ps_cpuid = strdup(ev->pl_u.pl_i.pl_cpuid); 335 ps->ps_version = ev->pl_u.pl_i.pl_version; 336 ps->ps_arch = ev->pl_u.pl_i.pl_arch; 337 ps->ps_initialized = 1; 338 break; 339 case PMCLOG_TYPE_MAP_IN: 340 PMCLOG_GET_PATHLEN(pathlen,evlen,pmclog_map_in); 341 PMCLOG_READ32(le,ev->pl_u.pl_mi.pl_pid); 342 PMCLOG_SKIP32(le); 343 PMCLOG_READADDR(le,ev->pl_u.pl_mi.pl_start); 344 PMCLOG_READSTRING(le, ev->pl_u.pl_mi.pl_pathname, pathlen); 345 break; 346 case PMCLOG_TYPE_MAP_OUT: 347 PMCLOG_READ32(le,ev->pl_u.pl_mo.pl_pid); 348 PMCLOG_SKIP32(le); 349 PMCLOG_READADDR(le,ev->pl_u.pl_mo.pl_start); 350 PMCLOG_READADDR(le,ev->pl_u.pl_mo.pl_end); 351 break; 352 case PMCLOG_TYPE_PMCALLOCATE: 353 PMCLOG_READ32(le,ev->pl_u.pl_a.pl_pmcid); 354 PMCLOG_READ32(le,ev->pl_u.pl_a.pl_event); 355 PMCLOG_READ32(le,ev->pl_u.pl_a.pl_flags); 356 PMCLOG_SKIP32(le); 357 PMCLOG_READ64(le,ev->pl_u.pl_a.pl_rate); 358 359 /* 360 * pl_event could contain either a PMC event code or a PMU 361 * event index. 362 */ 363 if ((ev->pl_u.pl_a.pl_flags & PMC_F_EV_PMU) != 0) 364 ev->pl_u.pl_a.pl_evname = 365 pmc_pmu_event_get_by_idx(ps->ps_cpuid, 366 ev->pl_u.pl_a.pl_event); 367 else if (ev->pl_u.pl_a.pl_event <= PMC_EVENT_LAST) 368 ev->pl_u.pl_a.pl_evname = 369 _pmc_name_of_event(ev->pl_u.pl_a.pl_event, 370 ps->ps_arch); 371 else 372 ev->pl_u.pl_a.pl_evname = NULL; 373 if (ev->pl_u.pl_a.pl_evname == NULL) { 374 printf("unknown event\n"); 375 goto error; 376 } 377 break; 378 case PMCLOG_TYPE_PMCALLOCATEDYN: 379 PMCLOG_READ32(le,ev->pl_u.pl_ad.pl_pmcid); 380 PMCLOG_READ32(le,ev->pl_u.pl_ad.pl_event); 381 PMCLOG_READ32(le,ev->pl_u.pl_ad.pl_flags); 382 PMCLOG_SKIP32(le); 383 PMCLOG_READSTRING(le,ev->pl_u.pl_ad.pl_evname,PMC_NAME_MAX); 384 break; 385 case PMCLOG_TYPE_PMCATTACH: 386 PMCLOG_GET_PATHLEN(pathlen,evlen,pmclog_pmcattach); 387 PMCLOG_READ32(le,ev->pl_u.pl_t.pl_pmcid); 388 PMCLOG_READ32(le,ev->pl_u.pl_t.pl_pid); 389 PMCLOG_READSTRING(le,ev->pl_u.pl_t.pl_pathname,pathlen); 390 break; 391 case PMCLOG_TYPE_PMCDETACH: 392 PMCLOG_READ32(le,ev->pl_u.pl_d.pl_pmcid); 393 PMCLOG_READ32(le,ev->pl_u.pl_d.pl_pid); 394 break; 395 case PMCLOG_TYPE_PROCCSW: 396 PMCLOG_READ64(le,ev->pl_u.pl_c.pl_value); 397 PMCLOG_READ32(le,ev->pl_u.pl_c.pl_pmcid); 398 PMCLOG_READ32(le,ev->pl_u.pl_c.pl_pid); 399 PMCLOG_READ32(le,ev->pl_u.pl_c.pl_tid); 400 break; 401 case PMCLOG_TYPE_PROCEXEC: 402 PMCLOG_GET_PATHLEN(pathlen,evlen,pmclog_procexec); 403 PMCLOG_READ32(le,ev->pl_u.pl_x.pl_pid); 404 PMCLOG_READ32(le,ev->pl_u.pl_x.pl_pmcid); 405 PMCLOG_READADDR(le,ev->pl_u.pl_x.pl_baseaddr); 406 PMCLOG_READADDR(le,ev->pl_u.pl_x.pl_dynaddr); 407 PMCLOG_READSTRING(le,ev->pl_u.pl_x.pl_pathname,pathlen); 408 break; 409 case PMCLOG_TYPE_PROCEXIT: 410 PMCLOG_READ32(le,ev->pl_u.pl_e.pl_pmcid); 411 PMCLOG_READ32(le,ev->pl_u.pl_e.pl_pid); 412 PMCLOG_READ64(le,ev->pl_u.pl_e.pl_value); 413 break; 414 case PMCLOG_TYPE_PROCFORK: 415 PMCLOG_READ32(le,ev->pl_u.pl_f.pl_oldpid); 416 PMCLOG_READ32(le,ev->pl_u.pl_f.pl_newpid); 417 break; 418 case PMCLOG_TYPE_SYSEXIT: 419 PMCLOG_READ32(le,ev->pl_u.pl_se.pl_pid); 420 break; 421 case PMCLOG_TYPE_USERDATA: 422 PMCLOG_READ32(le,ev->pl_u.pl_u.pl_userdata); 423 break; 424 case PMCLOG_TYPE_THR_CREATE: 425 PMCLOG_READ32(le,ev->pl_u.pl_tc.pl_tid); 426 PMCLOG_READ32(le,ev->pl_u.pl_tc.pl_pid); 427 PMCLOG_READ32(le,ev->pl_u.pl_tc.pl_flags); 428 PMCLOG_SKIP32(le); 429 memcpy(ev->pl_u.pl_tc.pl_tdname, le, MAXCOMLEN+1); 430 break; 431 case PMCLOG_TYPE_THR_EXIT: 432 PMCLOG_READ32(le,ev->pl_u.pl_te.pl_tid); 433 break; 434 case PMCLOG_TYPE_PROC_CREATE: 435 PMCLOG_READ32(le,ev->pl_u.pl_pc.pl_pid); 436 PMCLOG_READ32(le,ev->pl_u.pl_pc.pl_flags); 437 memcpy(ev->pl_u.pl_pc.pl_pcomm, le, MAXCOMLEN+1); 438 break; 439 default: /* unknown record type */ 440 ps->ps_state = PL_STATE_ERROR; 441 ev->pl_state = PMCLOG_ERROR; 442 return (-1); 443 } 444 445 ev->pl_offset = (ps->ps_offset += evlen); 446 ev->pl_count = (ps->ps_count += 1); 447 ev->pl_len = evlen; 448 ev->pl_state = PMCLOG_OK; 449 return 0; 450 451 error: 452 ev->pl_state = PMCLOG_ERROR; 453 ps->ps_state = PL_STATE_ERROR; 454 return -1; 455 } 456 457 /* 458 * Extract and return the next event from the byte stream. 459 * 460 * Returns 0 and sets the event's state to PMCLOG_OK in case an event 461 * was successfully parsed. Otherwise this function returns -1 and 462 * sets the event's state to one of PMCLOG_REQUIRE_DATA (if more data 463 * is needed) or PMCLOG_EOF (if an EOF was seen) or PMCLOG_ERROR if 464 * a parse error was encountered. 465 */ 466 467 int 468 pmclog_read(void *cookie, struct pmclog_ev *ev) 469 { 470 int retval; 471 ssize_t nread; 472 struct pmclog_parse_state *ps; 473 474 ps = (struct pmclog_parse_state *) cookie; 475 476 if (ps->ps_state == PL_STATE_ERROR) { 477 ev->pl_state = PMCLOG_ERROR; 478 return -1; 479 } 480 481 /* 482 * If there isn't enough data left for a new event try and get 483 * more data. 484 */ 485 if (ps->ps_len == 0) { 486 ev->pl_state = PMCLOG_REQUIRE_DATA; 487 488 /* 489 * If we have a valid file descriptor to read from, attempt 490 * to read from that. This read may return with an error, 491 * (which may be EAGAIN or other recoverable error), or 492 * can return EOF. 493 */ 494 if (ps->ps_fd != PMCLOG_FD_NONE) { 495 refill: 496 nread = read(ps->ps_fd, ps->ps_buffer, 497 PMCLOG_BUFFER_SIZE); 498 499 if (nread <= 0) { 500 if (nread == 0) 501 ev->pl_state = PMCLOG_EOF; 502 else if (errno != EAGAIN) /* not restartable */ 503 ev->pl_state = PMCLOG_ERROR; 504 return -1; 505 } 506 507 ps->ps_len = nread; 508 ps->ps_data = ps->ps_buffer; 509 } else { 510 return -1; 511 } 512 } 513 514 assert(ps->ps_len > 0); 515 516 517 /* Retrieve one event from the byte stream. */ 518 retval = pmclog_get_event(ps, &ps->ps_data, &ps->ps_len, ev); 519 /* 520 * If we need more data and we have a configured fd, try read 521 * from it. 522 */ 523 if (retval < 0 && ev->pl_state == PMCLOG_REQUIRE_DATA && 524 ps->ps_fd != -1) { 525 assert(ps->ps_len == 0); 526 goto refill; 527 } 528 529 return retval; 530 } 531 532 /* 533 * Feed data to a memory based parser. 534 * 535 * The memory area pointed to by 'data' needs to be valid till the 536 * next error return from pmclog_next_event(). 537 */ 538 539 int 540 pmclog_feed(void *cookie, char *data, int len) 541 { 542 struct pmclog_parse_state *ps; 543 544 ps = (struct pmclog_parse_state *) cookie; 545 546 if (len < 0 || /* invalid length */ 547 ps->ps_buffer || /* called for a file parser */ 548 ps->ps_len != 0) /* unnecessary call */ 549 return -1; 550 551 ps->ps_data = data; 552 ps->ps_len = len; 553 554 return 0; 555 } 556 557 /* 558 * Allocate and initialize parser state. 559 */ 560 561 void * 562 pmclog_open(int fd) 563 { 564 struct pmclog_parse_state *ps; 565 566 if ((ps = (struct pmclog_parse_state *) malloc(sizeof(*ps))) == NULL) 567 return NULL; 568 569 ps->ps_state = PL_STATE_NEW_RECORD; 570 ps->ps_arch = -1; 571 ps->ps_initialized = 0; 572 ps->ps_count = 0; 573 ps->ps_offset = (off_t) 0; 574 bzero(&ps->ps_saved, sizeof(ps->ps_saved)); 575 ps->ps_cpuid = NULL; 576 ps->ps_svcount = 0; 577 ps->ps_fd = fd; 578 ps->ps_data = NULL; 579 ps->ps_buffer = NULL; 580 ps->ps_len = 0; 581 582 /* allocate space for a work area */ 583 if (ps->ps_fd != PMCLOG_FD_NONE) { 584 if ((ps->ps_buffer = malloc(PMCLOG_BUFFER_SIZE)) == NULL) { 585 free(ps); 586 return NULL; 587 } 588 } 589 590 return ps; 591 } 592 593 594 /* 595 * Free up parser state. 596 */ 597 598 void 599 pmclog_close(void *cookie) 600 { 601 struct pmclog_parse_state *ps; 602 603 ps = (struct pmclog_parse_state *) cookie; 604 605 if (ps->ps_buffer) 606 free(ps->ps_buffer); 607 608 free(ps); 609 } 610