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