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 "util.h" 19 #include "event.h" 20 #include "debug.h" 21 #include "evlist.h" 22 #include "symbol.h" 23 #include <elf.h> 24 25 #include "tsc.h" 26 #include "session.h" 27 #include "jit.h" 28 #include "jitdump.h" 29 #include "genelf.h" 30 #include "../builtin.h" 31 32 #include <linux/ctype.h> 33 #include <linux/zalloc.h> 34 35 struct jit_buf_desc { 36 struct perf_data *output; 37 struct perf_session *session; 38 struct machine *machine; 39 union jr_entry *entry; 40 void *buf; 41 uint64_t sample_type; 42 size_t bufsize; 43 FILE *in; 44 bool needs_bswap; /* handles cross-endianness */ 45 bool use_arch_timestamp; 46 void *debug_data; 47 void *unwinding_data; 48 uint64_t unwinding_size; 49 uint64_t unwinding_mapped_size; 50 uint64_t eh_frame_hdr_size; 51 size_t nr_debug_entries; 52 uint32_t code_load_count; 53 u64 bytes_written; 54 struct rb_root code_root; 55 char dir[PATH_MAX]; 56 }; 57 58 struct debug_line_info { 59 unsigned long vma; 60 unsigned int lineno; 61 /* The filename format is unspecified, absolute path, relative etc. */ 62 char const filename[0]; 63 }; 64 65 struct jit_tool { 66 struct perf_tool tool; 67 struct perf_data output; 68 struct perf_data input; 69 u64 bytes_written; 70 }; 71 72 #define hmax(a, b) ((a) > (b) ? (a) : (b)) 73 #define get_jit_tool(t) (container_of(tool, struct jit_tool, tool)) 74 75 static int 76 jit_emit_elf(char *filename, 77 const char *sym, 78 uint64_t code_addr, 79 const void *code, 80 int csize, 81 void *debug, 82 int nr_debug_entries, 83 void *unwinding, 84 uint32_t unwinding_header_size, 85 uint32_t unwinding_size) 86 { 87 int ret, fd; 88 89 if (verbose > 0) 90 fprintf(stderr, "write ELF image %s\n", filename); 91 92 fd = open(filename, O_CREAT|O_TRUNC|O_WRONLY, 0644); 93 if (fd == -1) { 94 pr_warning("cannot create jit ELF %s: %s\n", filename, strerror(errno)); 95 return -1; 96 } 97 98 ret = jit_write_elf(fd, code_addr, sym, (const void *)code, csize, debug, nr_debug_entries, 99 unwinding, unwinding_header_size, unwinding_size); 100 101 close(fd); 102 103 if (ret) 104 unlink(filename); 105 106 return ret; 107 } 108 109 static void 110 jit_close(struct jit_buf_desc *jd) 111 { 112 if (!(jd && jd->in)) 113 return; 114 funlockfile(jd->in); 115 fclose(jd->in); 116 jd->in = NULL; 117 } 118 119 static int 120 jit_validate_events(struct perf_session *session) 121 { 122 struct evsel *evsel; 123 124 /* 125 * check that all events use CLOCK_MONOTONIC 126 */ 127 evlist__for_each_entry(session->evlist, evsel) { 128 if (evsel->core.attr.use_clockid == 0 || evsel->core.attr.clockid != CLOCK_MONOTONIC) 129 return -1; 130 } 131 return 0; 132 } 133 134 static int 135 jit_open(struct jit_buf_desc *jd, const char *name) 136 { 137 struct jitheader header; 138 struct jr_prefix *prefix; 139 ssize_t bs, bsz = 0; 140 void *n, *buf = NULL; 141 int ret, retval = -1; 142 143 jd->in = fopen(name, "r"); 144 if (!jd->in) 145 return -1; 146 147 bsz = hmax(sizeof(header), sizeof(*prefix)); 148 149 buf = malloc(bsz); 150 if (!buf) 151 goto error; 152 153 /* 154 * protect from writer modifying the file while we are reading it 155 */ 156 flockfile(jd->in); 157 158 ret = fread(buf, sizeof(header), 1, jd->in); 159 if (ret != 1) 160 goto error; 161 162 memcpy(&header, buf, sizeof(header)); 163 164 if (header.magic != JITHEADER_MAGIC) { 165 if (header.magic != JITHEADER_MAGIC_SW) 166 goto error; 167 jd->needs_bswap = true; 168 } 169 170 if (jd->needs_bswap) { 171 header.version = bswap_32(header.version); 172 header.total_size = bswap_32(header.total_size); 173 header.pid = bswap_32(header.pid); 174 header.elf_mach = bswap_32(header.elf_mach); 175 header.timestamp = bswap_64(header.timestamp); 176 header.flags = bswap_64(header.flags); 177 } 178 179 jd->use_arch_timestamp = header.flags & JITDUMP_FLAGS_ARCH_TIMESTAMP; 180 181 if (verbose > 2) 182 pr_debug("version=%u\nhdr.size=%u\nts=0x%llx\npid=%d\nelf_mach=%d\nuse_arch_timestamp=%d\n", 183 header.version, 184 header.total_size, 185 (unsigned long long)header.timestamp, 186 header.pid, 187 header.elf_mach, 188 jd->use_arch_timestamp); 189 190 if (header.version > JITHEADER_VERSION) { 191 pr_err("wrong jitdump version %u, expected " __stringify(JITHEADER_VERSION), 192 header.version); 193 goto error; 194 } 195 196 if (header.flags & JITDUMP_FLAGS_RESERVED) { 197 pr_err("jitdump file contains invalid or unsupported flags 0x%llx\n", 198 (unsigned long long)header.flags & JITDUMP_FLAGS_RESERVED); 199 goto error; 200 } 201 202 if (jd->use_arch_timestamp && !jd->session->time_conv.time_mult) { 203 pr_err("jitdump file uses arch timestamps but there is no timestamp conversion\n"); 204 goto error; 205 } 206 207 /* 208 * validate event is using the correct clockid 209 */ 210 if (!jd->use_arch_timestamp && jit_validate_events(jd->session)) { 211 pr_err("error, jitted code must be sampled with perf record -k 1\n"); 212 goto error; 213 } 214 215 bs = header.total_size - sizeof(header); 216 217 if (bs > bsz) { 218 n = realloc(buf, bs); 219 if (!n) 220 goto error; 221 bsz = bs; 222 buf = n; 223 /* read extra we do not know about */ 224 ret = fread(buf, bs - bsz, 1, jd->in); 225 if (ret != 1) 226 goto error; 227 } 228 /* 229 * keep dirname for generating files and mmap records 230 */ 231 strcpy(jd->dir, name); 232 dirname(jd->dir); 233 234 return 0; 235 error: 236 funlockfile(jd->in); 237 fclose(jd->in); 238 return retval; 239 } 240 241 static union jr_entry * 242 jit_get_next_entry(struct jit_buf_desc *jd) 243 { 244 struct jr_prefix *prefix; 245 union jr_entry *jr; 246 void *addr; 247 size_t bs, size; 248 int id, ret; 249 250 if (!(jd && jd->in)) 251 return NULL; 252 253 if (jd->buf == NULL) { 254 size_t sz = getpagesize(); 255 if (sz < sizeof(*prefix)) 256 sz = sizeof(*prefix); 257 258 jd->buf = malloc(sz); 259 if (jd->buf == NULL) 260 return NULL; 261 262 jd->bufsize = sz; 263 } 264 265 prefix = jd->buf; 266 267 /* 268 * file is still locked at this point 269 */ 270 ret = fread(prefix, sizeof(*prefix), 1, jd->in); 271 if (ret != 1) 272 return NULL; 273 274 if (jd->needs_bswap) { 275 prefix->id = bswap_32(prefix->id); 276 prefix->total_size = bswap_32(prefix->total_size); 277 prefix->timestamp = bswap_64(prefix->timestamp); 278 } 279 id = prefix->id; 280 size = prefix->total_size; 281 282 bs = (size_t)size; 283 if (bs < sizeof(*prefix)) 284 return NULL; 285 286 if (id >= JIT_CODE_MAX) { 287 pr_warning("next_entry: unknown record type %d, skipping\n", id); 288 } 289 if (bs > jd->bufsize) { 290 void *n; 291 n = realloc(jd->buf, bs); 292 if (!n) 293 return NULL; 294 jd->buf = n; 295 jd->bufsize = bs; 296 } 297 298 addr = ((void *)jd->buf) + sizeof(*prefix); 299 300 ret = fread(addr, bs - sizeof(*prefix), 1, jd->in); 301 if (ret != 1) 302 return NULL; 303 304 jr = (union jr_entry *)jd->buf; 305 306 switch(id) { 307 case JIT_CODE_DEBUG_INFO: 308 if (jd->needs_bswap) { 309 uint64_t n; 310 jr->info.code_addr = bswap_64(jr->info.code_addr); 311 jr->info.nr_entry = bswap_64(jr->info.nr_entry); 312 for (n = 0 ; n < jr->info.nr_entry; n++) { 313 jr->info.entries[n].addr = bswap_64(jr->info.entries[n].addr); 314 jr->info.entries[n].lineno = bswap_32(jr->info.entries[n].lineno); 315 jr->info.entries[n].discrim = bswap_32(jr->info.entries[n].discrim); 316 } 317 } 318 break; 319 case JIT_CODE_UNWINDING_INFO: 320 if (jd->needs_bswap) { 321 jr->unwinding.unwinding_size = bswap_64(jr->unwinding.unwinding_size); 322 jr->unwinding.eh_frame_hdr_size = bswap_64(jr->unwinding.eh_frame_hdr_size); 323 jr->unwinding.mapped_size = bswap_64(jr->unwinding.mapped_size); 324 } 325 break; 326 case JIT_CODE_CLOSE: 327 break; 328 case JIT_CODE_LOAD: 329 if (jd->needs_bswap) { 330 jr->load.pid = bswap_32(jr->load.pid); 331 jr->load.tid = bswap_32(jr->load.tid); 332 jr->load.vma = bswap_64(jr->load.vma); 333 jr->load.code_addr = bswap_64(jr->load.code_addr); 334 jr->load.code_size = bswap_64(jr->load.code_size); 335 jr->load.code_index= bswap_64(jr->load.code_index); 336 } 337 jd->code_load_count++; 338 break; 339 case JIT_CODE_MOVE: 340 if (jd->needs_bswap) { 341 jr->move.pid = bswap_32(jr->move.pid); 342 jr->move.tid = bswap_32(jr->move.tid); 343 jr->move.vma = bswap_64(jr->move.vma); 344 jr->move.old_code_addr = bswap_64(jr->move.old_code_addr); 345 jr->move.new_code_addr = bswap_64(jr->move.new_code_addr); 346 jr->move.code_size = bswap_64(jr->move.code_size); 347 jr->move.code_index = bswap_64(jr->move.code_index); 348 } 349 break; 350 case JIT_CODE_MAX: 351 default: 352 /* skip unknown record (we have read them) */ 353 break; 354 } 355 return jr; 356 } 357 358 static int 359 jit_inject_event(struct jit_buf_desc *jd, union perf_event *event) 360 { 361 ssize_t size; 362 363 size = perf_data__write(jd->output, event, event->header.size); 364 if (size < 0) 365 return -1; 366 367 jd->bytes_written += size; 368 return 0; 369 } 370 371 static uint64_t convert_timestamp(struct jit_buf_desc *jd, uint64_t timestamp) 372 { 373 struct perf_tsc_conversion tc; 374 375 if (!jd->use_arch_timestamp) 376 return timestamp; 377 378 tc.time_shift = jd->session->time_conv.time_shift; 379 tc.time_mult = jd->session->time_conv.time_mult; 380 tc.time_zero = jd->session->time_conv.time_zero; 381 382 if (!tc.time_mult) 383 return 0; 384 385 return tsc_to_perf_time(timestamp, &tc); 386 } 387 388 static int jit_repipe_code_load(struct jit_buf_desc *jd, union jr_entry *jr) 389 { 390 struct perf_sample sample; 391 union perf_event *event; 392 struct perf_tool *tool = jd->session->tool; 393 uint64_t code, addr; 394 uintptr_t uaddr; 395 char *filename; 396 struct stat st; 397 size_t size; 398 u16 idr_size; 399 const char *sym; 400 uint32_t count; 401 int ret, csize, usize; 402 pid_t pid, tid; 403 struct { 404 u32 pid, tid; 405 u64 time; 406 } *id; 407 408 pid = jr->load.pid; 409 tid = jr->load.tid; 410 csize = jr->load.code_size; 411 usize = jd->unwinding_mapped_size; 412 addr = jr->load.code_addr; 413 sym = (void *)((unsigned long)jr + sizeof(jr->load)); 414 code = (unsigned long)jr + jr->load.p.total_size - csize; 415 count = jr->load.code_index; 416 idr_size = jd->machine->id_hdr_size; 417 418 event = calloc(1, sizeof(*event) + idr_size); 419 if (!event) 420 return -1; 421 422 filename = event->mmap2.filename; 423 size = snprintf(filename, PATH_MAX, "%s/jitted-%d-%u.so", 424 jd->dir, 425 pid, 426 count); 427 428 size++; /* for \0 */ 429 430 size = PERF_ALIGN(size, sizeof(u64)); 431 uaddr = (uintptr_t)code; 432 ret = jit_emit_elf(filename, sym, addr, (const void *)uaddr, csize, jd->debug_data, jd->nr_debug_entries, 433 jd->unwinding_data, jd->eh_frame_hdr_size, jd->unwinding_size); 434 435 if (jd->debug_data && jd->nr_debug_entries) { 436 zfree(&jd->debug_data); 437 jd->nr_debug_entries = 0; 438 } 439 440 if (jd->unwinding_data && jd->eh_frame_hdr_size) { 441 zfree(&jd->unwinding_data); 442 jd->eh_frame_hdr_size = 0; 443 jd->unwinding_mapped_size = 0; 444 jd->unwinding_size = 0; 445 } 446 447 if (ret) { 448 free(event); 449 return -1; 450 } 451 if (stat(filename, &st)) 452 memset(&st, 0, sizeof(st)); 453 454 event->mmap2.header.type = PERF_RECORD_MMAP2; 455 event->mmap2.header.misc = PERF_RECORD_MISC_USER; 456 event->mmap2.header.size = (sizeof(event->mmap2) - 457 (sizeof(event->mmap2.filename) - size) + idr_size); 458 459 event->mmap2.pgoff = GEN_ELF_TEXT_OFFSET; 460 event->mmap2.start = addr; 461 event->mmap2.len = usize ? ALIGN_8(csize) + usize : csize; 462 event->mmap2.pid = pid; 463 event->mmap2.tid = tid; 464 event->mmap2.ino = st.st_ino; 465 event->mmap2.maj = major(st.st_dev); 466 event->mmap2.min = minor(st.st_dev); 467 event->mmap2.prot = st.st_mode; 468 event->mmap2.flags = MAP_SHARED; 469 event->mmap2.ino_generation = 1; 470 471 id = (void *)((unsigned long)event + event->mmap.header.size - idr_size); 472 if (jd->sample_type & PERF_SAMPLE_TID) { 473 id->pid = pid; 474 id->tid = tid; 475 } 476 if (jd->sample_type & PERF_SAMPLE_TIME) 477 id->time = convert_timestamp(jd, jr->load.p.timestamp); 478 479 /* 480 * create pseudo sample to induce dso hit increment 481 * use first address as sample address 482 */ 483 memset(&sample, 0, sizeof(sample)); 484 sample.cpumode = PERF_RECORD_MISC_USER; 485 sample.pid = pid; 486 sample.tid = tid; 487 sample.time = id->time; 488 sample.ip = addr; 489 490 ret = perf_event__process_mmap2(tool, event, &sample, jd->machine); 491 if (ret) 492 return ret; 493 494 ret = jit_inject_event(jd, event); 495 /* 496 * mark dso as use to generate buildid in the header 497 */ 498 if (!ret) 499 build_id__mark_dso_hit(tool, event, &sample, NULL, jd->machine); 500 501 return ret; 502 } 503 504 static int jit_repipe_code_move(struct jit_buf_desc *jd, union jr_entry *jr) 505 { 506 struct perf_sample sample; 507 union perf_event *event; 508 struct perf_tool *tool = jd->session->tool; 509 char *filename; 510 size_t size; 511 struct stat st; 512 int usize; 513 u16 idr_size; 514 int ret; 515 pid_t pid, tid; 516 struct { 517 u32 pid, tid; 518 u64 time; 519 } *id; 520 521 pid = jr->move.pid; 522 tid = jr->move.tid; 523 usize = jd->unwinding_mapped_size; 524 idr_size = jd->machine->id_hdr_size; 525 526 /* 527 * +16 to account for sample_id_all (hack) 528 */ 529 event = calloc(1, sizeof(*event) + 16); 530 if (!event) 531 return -1; 532 533 filename = event->mmap2.filename; 534 size = snprintf(filename, PATH_MAX, "%s/jitted-%d-%"PRIu64, 535 jd->dir, 536 pid, 537 jr->move.code_index); 538 539 size++; /* for \0 */ 540 541 if (stat(filename, &st)) 542 memset(&st, 0, sizeof(st)); 543 544 size = PERF_ALIGN(size, sizeof(u64)); 545 546 event->mmap2.header.type = PERF_RECORD_MMAP2; 547 event->mmap2.header.misc = PERF_RECORD_MISC_USER; 548 event->mmap2.header.size = (sizeof(event->mmap2) - 549 (sizeof(event->mmap2.filename) - size) + idr_size); 550 event->mmap2.pgoff = GEN_ELF_TEXT_OFFSET; 551 event->mmap2.start = jr->move.new_code_addr; 552 event->mmap2.len = usize ? ALIGN_8(jr->move.code_size) + usize 553 : jr->move.code_size; 554 event->mmap2.pid = pid; 555 event->mmap2.tid = tid; 556 event->mmap2.ino = st.st_ino; 557 event->mmap2.maj = major(st.st_dev); 558 event->mmap2.min = minor(st.st_dev); 559 event->mmap2.prot = st.st_mode; 560 event->mmap2.flags = MAP_SHARED; 561 event->mmap2.ino_generation = 1; 562 563 id = (void *)((unsigned long)event + event->mmap.header.size - idr_size); 564 if (jd->sample_type & PERF_SAMPLE_TID) { 565 id->pid = pid; 566 id->tid = tid; 567 } 568 if (jd->sample_type & PERF_SAMPLE_TIME) 569 id->time = convert_timestamp(jd, jr->load.p.timestamp); 570 571 /* 572 * create pseudo sample to induce dso hit increment 573 * use first address as sample address 574 */ 575 memset(&sample, 0, sizeof(sample)); 576 sample.cpumode = PERF_RECORD_MISC_USER; 577 sample.pid = pid; 578 sample.tid = tid; 579 sample.time = id->time; 580 sample.ip = jr->move.new_code_addr; 581 582 ret = perf_event__process_mmap2(tool, event, &sample, jd->machine); 583 if (ret) 584 return ret; 585 586 ret = jit_inject_event(jd, event); 587 if (!ret) 588 build_id__mark_dso_hit(tool, event, &sample, NULL, jd->machine); 589 590 return ret; 591 } 592 593 static int jit_repipe_debug_info(struct jit_buf_desc *jd, union jr_entry *jr) 594 { 595 void *data; 596 size_t sz; 597 598 if (!(jd && jr)) 599 return -1; 600 601 sz = jr->prefix.total_size - sizeof(jr->info); 602 data = malloc(sz); 603 if (!data) 604 return -1; 605 606 memcpy(data, &jr->info.entries, sz); 607 608 jd->debug_data = data; 609 610 /* 611 * we must use nr_entry instead of size here because 612 * we cannot distinguish actual entry from padding otherwise 613 */ 614 jd->nr_debug_entries = jr->info.nr_entry; 615 616 return 0; 617 } 618 619 static int 620 jit_repipe_unwinding_info(struct jit_buf_desc *jd, union jr_entry *jr) 621 { 622 void *unwinding_data; 623 uint32_t unwinding_data_size; 624 625 if (!(jd && jr)) 626 return -1; 627 628 unwinding_data_size = jr->prefix.total_size - sizeof(jr->unwinding); 629 unwinding_data = malloc(unwinding_data_size); 630 if (!unwinding_data) 631 return -1; 632 633 memcpy(unwinding_data, &jr->unwinding.unwinding_data, 634 unwinding_data_size); 635 636 jd->eh_frame_hdr_size = jr->unwinding.eh_frame_hdr_size; 637 jd->unwinding_size = jr->unwinding.unwinding_size; 638 jd->unwinding_mapped_size = jr->unwinding.mapped_size; 639 jd->unwinding_data = unwinding_data; 640 641 return 0; 642 } 643 644 static int 645 jit_process_dump(struct jit_buf_desc *jd) 646 { 647 union jr_entry *jr; 648 int ret = 0; 649 650 while ((jr = jit_get_next_entry(jd))) { 651 switch(jr->prefix.id) { 652 case JIT_CODE_LOAD: 653 ret = jit_repipe_code_load(jd, jr); 654 break; 655 case JIT_CODE_MOVE: 656 ret = jit_repipe_code_move(jd, jr); 657 break; 658 case JIT_CODE_DEBUG_INFO: 659 ret = jit_repipe_debug_info(jd, jr); 660 break; 661 case JIT_CODE_UNWINDING_INFO: 662 ret = jit_repipe_unwinding_info(jd, jr); 663 break; 664 default: 665 ret = 0; 666 continue; 667 } 668 } 669 return ret; 670 } 671 672 static int 673 jit_inject(struct jit_buf_desc *jd, char *path) 674 { 675 int ret; 676 677 if (verbose > 0) 678 fprintf(stderr, "injecting: %s\n", path); 679 680 ret = jit_open(jd, path); 681 if (ret) 682 return -1; 683 684 ret = jit_process_dump(jd); 685 686 jit_close(jd); 687 688 if (verbose > 0) 689 fprintf(stderr, "injected: %s (%d)\n", path, ret); 690 691 return 0; 692 } 693 694 /* 695 * File must be with pattern .../jit-XXXX.dump 696 * where XXXX is the PID of the process which did the mmap() 697 * as captured in the RECORD_MMAP record 698 */ 699 static int 700 jit_detect(char *mmap_name, pid_t pid) 701 { 702 char *p; 703 char *end = NULL; 704 pid_t pid2; 705 706 if (verbose > 2) 707 fprintf(stderr, "jit marker trying : %s\n", mmap_name); 708 /* 709 * get file name 710 */ 711 p = strrchr(mmap_name, '/'); 712 if (!p) 713 return -1; 714 715 /* 716 * match prefix 717 */ 718 if (strncmp(p, "/jit-", 5)) 719 return -1; 720 721 /* 722 * skip prefix 723 */ 724 p += 5; 725 726 /* 727 * must be followed by a pid 728 */ 729 if (!isdigit(*p)) 730 return -1; 731 732 pid2 = (int)strtol(p, &end, 10); 733 if (!end) 734 return -1; 735 736 /* 737 * pid does not match mmap pid 738 * pid==0 in system-wide mode (synthesized) 739 */ 740 if (pid && pid2 != pid) 741 return -1; 742 /* 743 * validate suffix 744 */ 745 if (strcmp(end, ".dump")) 746 return -1; 747 748 if (verbose > 0) 749 fprintf(stderr, "jit marker found: %s\n", mmap_name); 750 751 return 0; 752 } 753 754 int 755 jit_process(struct perf_session *session, 756 struct perf_data *output, 757 struct machine *machine, 758 char *filename, 759 pid_t pid, 760 u64 *nbytes) 761 { 762 struct evsel *first; 763 struct jit_buf_desc jd; 764 int ret; 765 766 /* 767 * first, detect marker mmap (i.e., the jitdump mmap) 768 */ 769 if (jit_detect(filename, pid)) 770 return 0; 771 772 memset(&jd, 0, sizeof(jd)); 773 774 jd.session = session; 775 jd.output = output; 776 jd.machine = machine; 777 778 /* 779 * track sample_type to compute id_all layout 780 * perf sets the same sample type to all events as of now 781 */ 782 first = perf_evlist__first(session->evlist); 783 jd.sample_type = first->core.attr.sample_type; 784 785 *nbytes = 0; 786 787 ret = jit_inject(&jd, filename); 788 if (!ret) { 789 *nbytes = jd.bytes_written; 790 ret = 1; 791 } 792 793 return ret; 794 } 795