1 // SPDX-License-Identifier: GPL-2.0 2 #include <sys/sysmacros.h> 3 #include <sys/types.h> 4 #include <errno.h> 5 #include <libgen.h> 6 #include <stdio.h> 7 #include <stdlib.h> 8 #include <string.h> 9 #include <fcntl.h> 10 #include <unistd.h> 11 #include <inttypes.h> 12 #include <byteswap.h> 13 #include <sys/stat.h> 14 #include <sys/mman.h> 15 #include <linux/stringify.h> 16 17 #include "build-id.h" 18 #include "event.h" 19 #include "debug.h" 20 #include "evlist.h" 21 #include "symbol.h" 22 #include <elf.h> 23 24 #include "tsc.h" 25 #include "session.h" 26 #include "jit.h" 27 #include "jitdump.h" 28 #include "genelf.h" 29 30 #include <linux/ctype.h> 31 #include <linux/zalloc.h> 32 33 struct jit_buf_desc { 34 struct perf_data *output; 35 struct perf_session *session; 36 struct machine *machine; 37 union jr_entry *entry; 38 void *buf; 39 uint64_t sample_type; 40 size_t bufsize; 41 FILE *in; 42 bool needs_bswap; /* handles cross-endianness */ 43 bool use_arch_timestamp; 44 void *debug_data; 45 void *unwinding_data; 46 uint64_t unwinding_size; 47 uint64_t unwinding_mapped_size; 48 uint64_t eh_frame_hdr_size; 49 size_t nr_debug_entries; 50 uint32_t code_load_count; 51 u64 bytes_written; 52 struct rb_root code_root; 53 char dir[PATH_MAX]; 54 }; 55 56 struct debug_line_info { 57 unsigned long vma; 58 unsigned int lineno; 59 /* The filename format is unspecified, absolute path, relative etc. */ 60 char const filename[0]; 61 }; 62 63 struct jit_tool { 64 struct perf_tool tool; 65 struct perf_data output; 66 struct perf_data input; 67 u64 bytes_written; 68 }; 69 70 #define hmax(a, b) ((a) > (b) ? (a) : (b)) 71 #define get_jit_tool(t) (container_of(tool, struct jit_tool, tool)) 72 73 static int 74 jit_emit_elf(char *filename, 75 const char *sym, 76 uint64_t code_addr, 77 const void *code, 78 int csize, 79 void *debug, 80 int nr_debug_entries, 81 void *unwinding, 82 uint32_t unwinding_header_size, 83 uint32_t unwinding_size) 84 { 85 int ret, fd; 86 87 if (verbose > 0) 88 fprintf(stderr, "write ELF image %s\n", filename); 89 90 fd = open(filename, O_CREAT|O_TRUNC|O_WRONLY, 0644); 91 if (fd == -1) { 92 pr_warning("cannot create jit ELF %s: %s\n", filename, strerror(errno)); 93 return -1; 94 } 95 96 ret = jit_write_elf(fd, code_addr, sym, (const void *)code, csize, debug, nr_debug_entries, 97 unwinding, unwinding_header_size, unwinding_size); 98 99 close(fd); 100 101 if (ret) 102 unlink(filename); 103 104 return ret; 105 } 106 107 static void 108 jit_close(struct jit_buf_desc *jd) 109 { 110 if (!(jd && jd->in)) 111 return; 112 funlockfile(jd->in); 113 fclose(jd->in); 114 jd->in = NULL; 115 } 116 117 static int 118 jit_validate_events(struct perf_session *session) 119 { 120 struct evsel *evsel; 121 122 /* 123 * check that all events use CLOCK_MONOTONIC 124 */ 125 evlist__for_each_entry(session->evlist, evsel) { 126 if (evsel->core.attr.use_clockid == 0 || evsel->core.attr.clockid != CLOCK_MONOTONIC) 127 return -1; 128 } 129 return 0; 130 } 131 132 static int 133 jit_open(struct jit_buf_desc *jd, const char *name) 134 { 135 struct jitheader header; 136 struct jr_prefix *prefix; 137 ssize_t bs, bsz = 0; 138 void *n, *buf = NULL; 139 int ret, retval = -1; 140 141 jd->in = fopen(name, "r"); 142 if (!jd->in) 143 return -1; 144 145 bsz = hmax(sizeof(header), sizeof(*prefix)); 146 147 buf = malloc(bsz); 148 if (!buf) 149 goto error; 150 151 /* 152 * protect from writer modifying the file while we are reading it 153 */ 154 flockfile(jd->in); 155 156 ret = fread(buf, sizeof(header), 1, jd->in); 157 if (ret != 1) 158 goto error; 159 160 memcpy(&header, buf, sizeof(header)); 161 162 if (header.magic != JITHEADER_MAGIC) { 163 if (header.magic != JITHEADER_MAGIC_SW) 164 goto error; 165 jd->needs_bswap = true; 166 } 167 168 if (jd->needs_bswap) { 169 header.version = bswap_32(header.version); 170 header.total_size = bswap_32(header.total_size); 171 header.pid = bswap_32(header.pid); 172 header.elf_mach = bswap_32(header.elf_mach); 173 header.timestamp = bswap_64(header.timestamp); 174 header.flags = bswap_64(header.flags); 175 } 176 177 jd->use_arch_timestamp = header.flags & JITDUMP_FLAGS_ARCH_TIMESTAMP; 178 179 if (verbose > 2) 180 pr_debug("version=%u\nhdr.size=%u\nts=0x%llx\npid=%d\nelf_mach=%d\nuse_arch_timestamp=%d\n", 181 header.version, 182 header.total_size, 183 (unsigned long long)header.timestamp, 184 header.pid, 185 header.elf_mach, 186 jd->use_arch_timestamp); 187 188 if (header.version > JITHEADER_VERSION) { 189 pr_err("wrong jitdump version %u, expected " __stringify(JITHEADER_VERSION), 190 header.version); 191 goto error; 192 } 193 194 if (header.flags & JITDUMP_FLAGS_RESERVED) { 195 pr_err("jitdump file contains invalid or unsupported flags 0x%llx\n", 196 (unsigned long long)header.flags & JITDUMP_FLAGS_RESERVED); 197 goto error; 198 } 199 200 if (jd->use_arch_timestamp && !jd->session->time_conv.time_mult) { 201 pr_err("jitdump file uses arch timestamps but there is no timestamp conversion\n"); 202 goto error; 203 } 204 205 /* 206 * validate event is using the correct clockid 207 */ 208 if (!jd->use_arch_timestamp && jit_validate_events(jd->session)) { 209 pr_err("error, jitted code must be sampled with perf record -k 1\n"); 210 goto error; 211 } 212 213 bs = header.total_size - sizeof(header); 214 215 if (bs > bsz) { 216 n = realloc(buf, bs); 217 if (!n) 218 goto error; 219 bsz = bs; 220 buf = n; 221 /* read extra we do not know about */ 222 ret = fread(buf, bs - bsz, 1, jd->in); 223 if (ret != 1) 224 goto error; 225 } 226 /* 227 * keep dirname for generating files and mmap records 228 */ 229 strcpy(jd->dir, name); 230 dirname(jd->dir); 231 232 return 0; 233 error: 234 funlockfile(jd->in); 235 fclose(jd->in); 236 return retval; 237 } 238 239 static union jr_entry * 240 jit_get_next_entry(struct jit_buf_desc *jd) 241 { 242 struct jr_prefix *prefix; 243 union jr_entry *jr; 244 void *addr; 245 size_t bs, size; 246 int id, ret; 247 248 if (!(jd && jd->in)) 249 return NULL; 250 251 if (jd->buf == NULL) { 252 size_t sz = getpagesize(); 253 if (sz < sizeof(*prefix)) 254 sz = sizeof(*prefix); 255 256 jd->buf = malloc(sz); 257 if (jd->buf == NULL) 258 return NULL; 259 260 jd->bufsize = sz; 261 } 262 263 prefix = jd->buf; 264 265 /* 266 * file is still locked at this point 267 */ 268 ret = fread(prefix, sizeof(*prefix), 1, jd->in); 269 if (ret != 1) 270 return NULL; 271 272 if (jd->needs_bswap) { 273 prefix->id = bswap_32(prefix->id); 274 prefix->total_size = bswap_32(prefix->total_size); 275 prefix->timestamp = bswap_64(prefix->timestamp); 276 } 277 id = prefix->id; 278 size = prefix->total_size; 279 280 bs = (size_t)size; 281 if (bs < sizeof(*prefix)) 282 return NULL; 283 284 if (id >= JIT_CODE_MAX) { 285 pr_warning("next_entry: unknown record type %d, skipping\n", id); 286 } 287 if (bs > jd->bufsize) { 288 void *n; 289 n = realloc(jd->buf, bs); 290 if (!n) 291 return NULL; 292 jd->buf = n; 293 jd->bufsize = bs; 294 } 295 296 addr = ((void *)jd->buf) + sizeof(*prefix); 297 298 ret = fread(addr, bs - sizeof(*prefix), 1, jd->in); 299 if (ret != 1) 300 return NULL; 301 302 jr = (union jr_entry *)jd->buf; 303 304 switch(id) { 305 case JIT_CODE_DEBUG_INFO: 306 if (jd->needs_bswap) { 307 uint64_t n; 308 jr->info.code_addr = bswap_64(jr->info.code_addr); 309 jr->info.nr_entry = bswap_64(jr->info.nr_entry); 310 for (n = 0 ; n < jr->info.nr_entry; n++) { 311 jr->info.entries[n].addr = bswap_64(jr->info.entries[n].addr); 312 jr->info.entries[n].lineno = bswap_32(jr->info.entries[n].lineno); 313 jr->info.entries[n].discrim = bswap_32(jr->info.entries[n].discrim); 314 } 315 } 316 break; 317 case JIT_CODE_UNWINDING_INFO: 318 if (jd->needs_bswap) { 319 jr->unwinding.unwinding_size = bswap_64(jr->unwinding.unwinding_size); 320 jr->unwinding.eh_frame_hdr_size = bswap_64(jr->unwinding.eh_frame_hdr_size); 321 jr->unwinding.mapped_size = bswap_64(jr->unwinding.mapped_size); 322 } 323 break; 324 case JIT_CODE_CLOSE: 325 break; 326 case JIT_CODE_LOAD: 327 if (jd->needs_bswap) { 328 jr->load.pid = bswap_32(jr->load.pid); 329 jr->load.tid = bswap_32(jr->load.tid); 330 jr->load.vma = bswap_64(jr->load.vma); 331 jr->load.code_addr = bswap_64(jr->load.code_addr); 332 jr->load.code_size = bswap_64(jr->load.code_size); 333 jr->load.code_index= bswap_64(jr->load.code_index); 334 } 335 jd->code_load_count++; 336 break; 337 case JIT_CODE_MOVE: 338 if (jd->needs_bswap) { 339 jr->move.pid = bswap_32(jr->move.pid); 340 jr->move.tid = bswap_32(jr->move.tid); 341 jr->move.vma = bswap_64(jr->move.vma); 342 jr->move.old_code_addr = bswap_64(jr->move.old_code_addr); 343 jr->move.new_code_addr = bswap_64(jr->move.new_code_addr); 344 jr->move.code_size = bswap_64(jr->move.code_size); 345 jr->move.code_index = bswap_64(jr->move.code_index); 346 } 347 break; 348 case JIT_CODE_MAX: 349 default: 350 /* skip unknown record (we have read them) */ 351 break; 352 } 353 return jr; 354 } 355 356 static int 357 jit_inject_event(struct jit_buf_desc *jd, union perf_event *event) 358 { 359 ssize_t size; 360 361 size = perf_data__write(jd->output, event, event->header.size); 362 if (size < 0) 363 return -1; 364 365 jd->bytes_written += size; 366 return 0; 367 } 368 369 static uint64_t convert_timestamp(struct jit_buf_desc *jd, uint64_t timestamp) 370 { 371 struct perf_tsc_conversion tc; 372 373 if (!jd->use_arch_timestamp) 374 return timestamp; 375 376 tc.time_shift = jd->session->time_conv.time_shift; 377 tc.time_mult = jd->session->time_conv.time_mult; 378 tc.time_zero = jd->session->time_conv.time_zero; 379 380 if (!tc.time_mult) 381 return 0; 382 383 return tsc_to_perf_time(timestamp, &tc); 384 } 385 386 static int jit_repipe_code_load(struct jit_buf_desc *jd, union jr_entry *jr) 387 { 388 struct perf_sample sample; 389 union perf_event *event; 390 struct perf_tool *tool = jd->session->tool; 391 uint64_t code, addr; 392 uintptr_t uaddr; 393 char *filename; 394 struct stat st; 395 size_t size; 396 u16 idr_size; 397 const char *sym; 398 uint64_t count; 399 int ret, csize, usize; 400 pid_t pid, tid; 401 struct { 402 u32 pid, tid; 403 u64 time; 404 } *id; 405 406 pid = jr->load.pid; 407 tid = jr->load.tid; 408 csize = jr->load.code_size; 409 usize = jd->unwinding_mapped_size; 410 addr = jr->load.code_addr; 411 sym = (void *)((unsigned long)jr + sizeof(jr->load)); 412 code = (unsigned long)jr + jr->load.p.total_size - csize; 413 count = jr->load.code_index; 414 idr_size = jd->machine->id_hdr_size; 415 416 event = calloc(1, sizeof(*event) + idr_size); 417 if (!event) 418 return -1; 419 420 filename = event->mmap2.filename; 421 size = snprintf(filename, PATH_MAX, "%s/jitted-%d-%" PRIu64 ".so", 422 jd->dir, 423 pid, 424 count); 425 426 size++; /* for \0 */ 427 428 size = PERF_ALIGN(size, sizeof(u64)); 429 uaddr = (uintptr_t)code; 430 ret = jit_emit_elf(filename, sym, addr, (const void *)uaddr, csize, jd->debug_data, jd->nr_debug_entries, 431 jd->unwinding_data, jd->eh_frame_hdr_size, jd->unwinding_size); 432 433 if (jd->debug_data && jd->nr_debug_entries) { 434 zfree(&jd->debug_data); 435 jd->nr_debug_entries = 0; 436 } 437 438 if (jd->unwinding_data && jd->eh_frame_hdr_size) { 439 zfree(&jd->unwinding_data); 440 jd->eh_frame_hdr_size = 0; 441 jd->unwinding_mapped_size = 0; 442 jd->unwinding_size = 0; 443 } 444 445 if (ret) { 446 free(event); 447 return -1; 448 } 449 if (stat(filename, &st)) 450 memset(&st, 0, sizeof(st)); 451 452 event->mmap2.header.type = PERF_RECORD_MMAP2; 453 event->mmap2.header.misc = PERF_RECORD_MISC_USER; 454 event->mmap2.header.size = (sizeof(event->mmap2) - 455 (sizeof(event->mmap2.filename) - size) + idr_size); 456 457 event->mmap2.pgoff = GEN_ELF_TEXT_OFFSET; 458 event->mmap2.start = addr; 459 event->mmap2.len = usize ? ALIGN_8(csize) + usize : csize; 460 event->mmap2.pid = pid; 461 event->mmap2.tid = tid; 462 event->mmap2.ino = st.st_ino; 463 event->mmap2.maj = major(st.st_dev); 464 event->mmap2.min = minor(st.st_dev); 465 event->mmap2.prot = st.st_mode; 466 event->mmap2.flags = MAP_SHARED; 467 event->mmap2.ino_generation = 1; 468 469 id = (void *)((unsigned long)event + event->mmap.header.size - idr_size); 470 if (jd->sample_type & PERF_SAMPLE_TID) { 471 id->pid = pid; 472 id->tid = tid; 473 } 474 if (jd->sample_type & PERF_SAMPLE_TIME) 475 id->time = convert_timestamp(jd, jr->load.p.timestamp); 476 477 /* 478 * create pseudo sample to induce dso hit increment 479 * use first address as sample address 480 */ 481 memset(&sample, 0, sizeof(sample)); 482 sample.cpumode = PERF_RECORD_MISC_USER; 483 sample.pid = pid; 484 sample.tid = tid; 485 sample.time = id->time; 486 sample.ip = addr; 487 488 ret = perf_event__process_mmap2(tool, event, &sample, jd->machine); 489 if (ret) 490 return ret; 491 492 ret = jit_inject_event(jd, event); 493 /* 494 * mark dso as use to generate buildid in the header 495 */ 496 if (!ret) 497 build_id__mark_dso_hit(tool, event, &sample, NULL, jd->machine); 498 499 return ret; 500 } 501 502 static int jit_repipe_code_move(struct jit_buf_desc *jd, union jr_entry *jr) 503 { 504 struct perf_sample sample; 505 union perf_event *event; 506 struct perf_tool *tool = jd->session->tool; 507 char *filename; 508 size_t size; 509 struct stat st; 510 int usize; 511 u16 idr_size; 512 int ret; 513 pid_t pid, tid; 514 struct { 515 u32 pid, tid; 516 u64 time; 517 } *id; 518 519 pid = jr->move.pid; 520 tid = jr->move.tid; 521 usize = jd->unwinding_mapped_size; 522 idr_size = jd->machine->id_hdr_size; 523 524 /* 525 * +16 to account for sample_id_all (hack) 526 */ 527 event = calloc(1, sizeof(*event) + 16); 528 if (!event) 529 return -1; 530 531 filename = event->mmap2.filename; 532 size = snprintf(filename, PATH_MAX, "%s/jitted-%d-%" PRIu64 ".so", 533 jd->dir, 534 pid, 535 jr->move.code_index); 536 537 size++; /* for \0 */ 538 539 if (stat(filename, &st)) 540 memset(&st, 0, sizeof(st)); 541 542 size = PERF_ALIGN(size, sizeof(u64)); 543 544 event->mmap2.header.type = PERF_RECORD_MMAP2; 545 event->mmap2.header.misc = PERF_RECORD_MISC_USER; 546 event->mmap2.header.size = (sizeof(event->mmap2) - 547 (sizeof(event->mmap2.filename) - size) + idr_size); 548 event->mmap2.pgoff = GEN_ELF_TEXT_OFFSET; 549 event->mmap2.start = jr->move.new_code_addr; 550 event->mmap2.len = usize ? ALIGN_8(jr->move.code_size) + usize 551 : jr->move.code_size; 552 event->mmap2.pid = pid; 553 event->mmap2.tid = tid; 554 event->mmap2.ino = st.st_ino; 555 event->mmap2.maj = major(st.st_dev); 556 event->mmap2.min = minor(st.st_dev); 557 event->mmap2.prot = st.st_mode; 558 event->mmap2.flags = MAP_SHARED; 559 event->mmap2.ino_generation = 1; 560 561 id = (void *)((unsigned long)event + event->mmap.header.size - idr_size); 562 if (jd->sample_type & PERF_SAMPLE_TID) { 563 id->pid = pid; 564 id->tid = tid; 565 } 566 if (jd->sample_type & PERF_SAMPLE_TIME) 567 id->time = convert_timestamp(jd, jr->load.p.timestamp); 568 569 /* 570 * create pseudo sample to induce dso hit increment 571 * use first address as sample address 572 */ 573 memset(&sample, 0, sizeof(sample)); 574 sample.cpumode = PERF_RECORD_MISC_USER; 575 sample.pid = pid; 576 sample.tid = tid; 577 sample.time = id->time; 578 sample.ip = jr->move.new_code_addr; 579 580 ret = perf_event__process_mmap2(tool, event, &sample, jd->machine); 581 if (ret) 582 return ret; 583 584 ret = jit_inject_event(jd, event); 585 if (!ret) 586 build_id__mark_dso_hit(tool, event, &sample, NULL, jd->machine); 587 588 return ret; 589 } 590 591 static int jit_repipe_debug_info(struct jit_buf_desc *jd, union jr_entry *jr) 592 { 593 void *data; 594 size_t sz; 595 596 if (!(jd && jr)) 597 return -1; 598 599 sz = jr->prefix.total_size - sizeof(jr->info); 600 data = malloc(sz); 601 if (!data) 602 return -1; 603 604 memcpy(data, &jr->info.entries, sz); 605 606 jd->debug_data = data; 607 608 /* 609 * we must use nr_entry instead of size here because 610 * we cannot distinguish actual entry from padding otherwise 611 */ 612 jd->nr_debug_entries = jr->info.nr_entry; 613 614 return 0; 615 } 616 617 static int 618 jit_repipe_unwinding_info(struct jit_buf_desc *jd, union jr_entry *jr) 619 { 620 void *unwinding_data; 621 uint32_t unwinding_data_size; 622 623 if (!(jd && jr)) 624 return -1; 625 626 unwinding_data_size = jr->prefix.total_size - sizeof(jr->unwinding); 627 unwinding_data = malloc(unwinding_data_size); 628 if (!unwinding_data) 629 return -1; 630 631 memcpy(unwinding_data, &jr->unwinding.unwinding_data, 632 unwinding_data_size); 633 634 jd->eh_frame_hdr_size = jr->unwinding.eh_frame_hdr_size; 635 jd->unwinding_size = jr->unwinding.unwinding_size; 636 jd->unwinding_mapped_size = jr->unwinding.mapped_size; 637 jd->unwinding_data = unwinding_data; 638 639 return 0; 640 } 641 642 static int 643 jit_process_dump(struct jit_buf_desc *jd) 644 { 645 union jr_entry *jr; 646 int ret = 0; 647 648 while ((jr = jit_get_next_entry(jd))) { 649 switch(jr->prefix.id) { 650 case JIT_CODE_LOAD: 651 ret = jit_repipe_code_load(jd, jr); 652 break; 653 case JIT_CODE_MOVE: 654 ret = jit_repipe_code_move(jd, jr); 655 break; 656 case JIT_CODE_DEBUG_INFO: 657 ret = jit_repipe_debug_info(jd, jr); 658 break; 659 case JIT_CODE_UNWINDING_INFO: 660 ret = jit_repipe_unwinding_info(jd, jr); 661 break; 662 default: 663 ret = 0; 664 continue; 665 } 666 } 667 return ret; 668 } 669 670 static int 671 jit_inject(struct jit_buf_desc *jd, char *path) 672 { 673 int ret; 674 675 if (verbose > 0) 676 fprintf(stderr, "injecting: %s\n", path); 677 678 ret = jit_open(jd, path); 679 if (ret) 680 return -1; 681 682 ret = jit_process_dump(jd); 683 684 jit_close(jd); 685 686 if (verbose > 0) 687 fprintf(stderr, "injected: %s (%d)\n", path, ret); 688 689 return 0; 690 } 691 692 /* 693 * File must be with pattern .../jit-XXXX.dump 694 * where XXXX is the PID of the process which did the mmap() 695 * as captured in the RECORD_MMAP record 696 */ 697 static int 698 jit_detect(char *mmap_name, pid_t pid) 699 { 700 char *p; 701 char *end = NULL; 702 pid_t pid2; 703 704 if (verbose > 2) 705 fprintf(stderr, "jit marker trying : %s\n", mmap_name); 706 /* 707 * get file name 708 */ 709 p = strrchr(mmap_name, '/'); 710 if (!p) 711 return -1; 712 713 /* 714 * match prefix 715 */ 716 if (strncmp(p, "/jit-", 5)) 717 return -1; 718 719 /* 720 * skip prefix 721 */ 722 p += 5; 723 724 /* 725 * must be followed by a pid 726 */ 727 if (!isdigit(*p)) 728 return -1; 729 730 pid2 = (int)strtol(p, &end, 10); 731 if (!end) 732 return -1; 733 734 /* 735 * pid does not match mmap pid 736 * pid==0 in system-wide mode (synthesized) 737 */ 738 if (pid && pid2 != pid) 739 return -1; 740 /* 741 * validate suffix 742 */ 743 if (strcmp(end, ".dump")) 744 return -1; 745 746 if (verbose > 0) 747 fprintf(stderr, "jit marker found: %s\n", mmap_name); 748 749 return 0; 750 } 751 752 int 753 jit_process(struct perf_session *session, 754 struct perf_data *output, 755 struct machine *machine, 756 char *filename, 757 pid_t pid, 758 u64 *nbytes) 759 { 760 struct evsel *first; 761 struct jit_buf_desc jd; 762 int ret; 763 764 /* 765 * first, detect marker mmap (i.e., the jitdump mmap) 766 */ 767 if (jit_detect(filename, pid)) 768 return 0; 769 770 memset(&jd, 0, sizeof(jd)); 771 772 jd.session = session; 773 jd.output = output; 774 jd.machine = machine; 775 776 /* 777 * track sample_type to compute id_all layout 778 * perf sets the same sample type to all events as of now 779 */ 780 first = evlist__first(session->evlist); 781 jd.sample_type = first->core.attr.sample_type; 782 783 *nbytes = 0; 784 785 ret = jit_inject(&jd, filename); 786 if (!ret) { 787 *nbytes = jd.bytes_written; 788 ret = 1; 789 } 790 791 return ret; 792 } 793