1 /* 2 * builtin-trace.c 3 * 4 * Builtin 'trace' command: 5 * 6 * Display a continuously updated trace of any workload, CPU, specific PID, 7 * system wide, etc. Default format is loosely strace like, but any other 8 * event may be specified using --event. 9 * 10 * Copyright (C) 2012, 2013, 2014, 2015 Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com> 11 * 12 * Initially based on the 'trace' prototype by Thomas Gleixner: 13 * 14 * http://lwn.net/Articles/415728/ ("Announcing a new utility: 'trace'") 15 * 16 * Released under the GPL v2. (and only v2, not any later version) 17 */ 18 19 #include <traceevent/event-parse.h> 20 #include <api/fs/tracing_path.h> 21 #include <bpf/bpf.h> 22 #include "builtin.h" 23 #include "util/cgroup.h" 24 #include "util/color.h" 25 #include "util/debug.h" 26 #include "util/env.h" 27 #include "util/event.h" 28 #include "util/evlist.h" 29 #include <subcmd/exec-cmd.h> 30 #include "util/machine.h" 31 #include "util/path.h" 32 #include "util/session.h" 33 #include "util/thread.h" 34 #include <subcmd/parse-options.h> 35 #include "util/strlist.h" 36 #include "util/intlist.h" 37 #include "util/thread_map.h" 38 #include "util/stat.h" 39 #include "trace/beauty/beauty.h" 40 #include "trace-event.h" 41 #include "util/parse-events.h" 42 #include "util/bpf-loader.h" 43 #include "callchain.h" 44 #include "print_binary.h" 45 #include "string2.h" 46 #include "syscalltbl.h" 47 #include "rb_resort.h" 48 49 #include <errno.h> 50 #include <inttypes.h> 51 #include <poll.h> 52 #include <signal.h> 53 #include <stdlib.h> 54 #include <string.h> 55 #include <linux/err.h> 56 #include <linux/filter.h> 57 #include <linux/kernel.h> 58 #include <linux/random.h> 59 #include <linux/stringify.h> 60 #include <linux/time64.h> 61 #include <fcntl.h> 62 63 #include "sane_ctype.h" 64 65 #ifndef O_CLOEXEC 66 # define O_CLOEXEC 02000000 67 #endif 68 69 #ifndef F_LINUX_SPECIFIC_BASE 70 # define F_LINUX_SPECIFIC_BASE 1024 71 #endif 72 73 struct trace { 74 struct perf_tool tool; 75 struct syscalltbl *sctbl; 76 struct { 77 int max; 78 struct syscall *table; 79 struct { 80 struct perf_evsel *sys_enter, 81 *sys_exit, 82 *augmented; 83 } events; 84 } syscalls; 85 struct record_opts opts; 86 struct perf_evlist *evlist; 87 struct machine *host; 88 struct thread *current; 89 struct cgroup *cgroup; 90 u64 base_time; 91 FILE *output; 92 unsigned long nr_events; 93 unsigned long nr_events_printed; 94 unsigned long max_events; 95 struct strlist *ev_qualifier; 96 struct { 97 size_t nr; 98 int *entries; 99 } ev_qualifier_ids; 100 struct { 101 size_t nr; 102 pid_t *entries; 103 struct bpf_map *map; 104 } filter_pids; 105 double duration_filter; 106 double runtime_ms; 107 struct { 108 u64 vfs_getname, 109 proc_getname; 110 } stats; 111 unsigned int max_stack; 112 unsigned int min_stack; 113 bool raw_augmented_syscalls; 114 bool not_ev_qualifier; 115 bool live; 116 bool full_time; 117 bool sched; 118 bool multiple_threads; 119 bool summary; 120 bool summary_only; 121 bool failure_only; 122 bool show_comm; 123 bool print_sample; 124 bool show_tool_stats; 125 bool trace_syscalls; 126 bool kernel_syscallchains; 127 bool force; 128 bool vfs_getname; 129 int trace_pgfaults; 130 }; 131 132 struct tp_field { 133 int offset; 134 union { 135 u64 (*integer)(struct tp_field *field, struct perf_sample *sample); 136 void *(*pointer)(struct tp_field *field, struct perf_sample *sample); 137 }; 138 }; 139 140 #define TP_UINT_FIELD(bits) \ 141 static u64 tp_field__u##bits(struct tp_field *field, struct perf_sample *sample) \ 142 { \ 143 u##bits value; \ 144 memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \ 145 return value; \ 146 } 147 148 TP_UINT_FIELD(8); 149 TP_UINT_FIELD(16); 150 TP_UINT_FIELD(32); 151 TP_UINT_FIELD(64); 152 153 #define TP_UINT_FIELD__SWAPPED(bits) \ 154 static u64 tp_field__swapped_u##bits(struct tp_field *field, struct perf_sample *sample) \ 155 { \ 156 u##bits value; \ 157 memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \ 158 return bswap_##bits(value);\ 159 } 160 161 TP_UINT_FIELD__SWAPPED(16); 162 TP_UINT_FIELD__SWAPPED(32); 163 TP_UINT_FIELD__SWAPPED(64); 164 165 static int __tp_field__init_uint(struct tp_field *field, int size, int offset, bool needs_swap) 166 { 167 field->offset = offset; 168 169 switch (size) { 170 case 1: 171 field->integer = tp_field__u8; 172 break; 173 case 2: 174 field->integer = needs_swap ? tp_field__swapped_u16 : tp_field__u16; 175 break; 176 case 4: 177 field->integer = needs_swap ? tp_field__swapped_u32 : tp_field__u32; 178 break; 179 case 8: 180 field->integer = needs_swap ? tp_field__swapped_u64 : tp_field__u64; 181 break; 182 default: 183 return -1; 184 } 185 186 return 0; 187 } 188 189 static int tp_field__init_uint(struct tp_field *field, struct tep_format_field *format_field, bool needs_swap) 190 { 191 return __tp_field__init_uint(field, format_field->size, format_field->offset, needs_swap); 192 } 193 194 static void *tp_field__ptr(struct tp_field *field, struct perf_sample *sample) 195 { 196 return sample->raw_data + field->offset; 197 } 198 199 static int __tp_field__init_ptr(struct tp_field *field, int offset) 200 { 201 field->offset = offset; 202 field->pointer = tp_field__ptr; 203 return 0; 204 } 205 206 static int tp_field__init_ptr(struct tp_field *field, struct tep_format_field *format_field) 207 { 208 return __tp_field__init_ptr(field, format_field->offset); 209 } 210 211 struct syscall_tp { 212 struct tp_field id; 213 union { 214 struct tp_field args, ret; 215 }; 216 }; 217 218 static int perf_evsel__init_tp_uint_field(struct perf_evsel *evsel, 219 struct tp_field *field, 220 const char *name) 221 { 222 struct tep_format_field *format_field = perf_evsel__field(evsel, name); 223 224 if (format_field == NULL) 225 return -1; 226 227 return tp_field__init_uint(field, format_field, evsel->needs_swap); 228 } 229 230 #define perf_evsel__init_sc_tp_uint_field(evsel, name) \ 231 ({ struct syscall_tp *sc = evsel->priv;\ 232 perf_evsel__init_tp_uint_field(evsel, &sc->name, #name); }) 233 234 static int perf_evsel__init_tp_ptr_field(struct perf_evsel *evsel, 235 struct tp_field *field, 236 const char *name) 237 { 238 struct tep_format_field *format_field = perf_evsel__field(evsel, name); 239 240 if (format_field == NULL) 241 return -1; 242 243 return tp_field__init_ptr(field, format_field); 244 } 245 246 #define perf_evsel__init_sc_tp_ptr_field(evsel, name) \ 247 ({ struct syscall_tp *sc = evsel->priv;\ 248 perf_evsel__init_tp_ptr_field(evsel, &sc->name, #name); }) 249 250 static void perf_evsel__delete_priv(struct perf_evsel *evsel) 251 { 252 zfree(&evsel->priv); 253 perf_evsel__delete(evsel); 254 } 255 256 static int perf_evsel__init_syscall_tp(struct perf_evsel *evsel) 257 { 258 struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp)); 259 260 if (evsel->priv != NULL) { 261 if (perf_evsel__init_tp_uint_field(evsel, &sc->id, "__syscall_nr") && 262 perf_evsel__init_tp_uint_field(evsel, &sc->id, "nr")) 263 goto out_delete; 264 return 0; 265 } 266 267 return -ENOMEM; 268 out_delete: 269 zfree(&evsel->priv); 270 return -ENOENT; 271 } 272 273 static int perf_evsel__init_augmented_syscall_tp(struct perf_evsel *evsel) 274 { 275 struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp)); 276 277 if (evsel->priv != NULL) { /* field, sizeof_field, offsetof_field */ 278 if (__tp_field__init_uint(&sc->id, sizeof(long), sizeof(long long), evsel->needs_swap)) 279 goto out_delete; 280 281 return 0; 282 } 283 284 return -ENOMEM; 285 out_delete: 286 zfree(&evsel->priv); 287 return -EINVAL; 288 } 289 290 static int perf_evsel__init_augmented_syscall_tp_args(struct perf_evsel *evsel) 291 { 292 struct syscall_tp *sc = evsel->priv; 293 294 return __tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)); 295 } 296 297 static int perf_evsel__init_augmented_syscall_tp_ret(struct perf_evsel *evsel) 298 { 299 struct syscall_tp *sc = evsel->priv; 300 301 return __tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap); 302 } 303 304 static int perf_evsel__init_raw_syscall_tp(struct perf_evsel *evsel, void *handler) 305 { 306 evsel->priv = malloc(sizeof(struct syscall_tp)); 307 if (evsel->priv != NULL) { 308 if (perf_evsel__init_sc_tp_uint_field(evsel, id)) 309 goto out_delete; 310 311 evsel->handler = handler; 312 return 0; 313 } 314 315 return -ENOMEM; 316 317 out_delete: 318 zfree(&evsel->priv); 319 return -ENOENT; 320 } 321 322 static struct perf_evsel *perf_evsel__raw_syscall_newtp(const char *direction, void *handler) 323 { 324 struct perf_evsel *evsel = perf_evsel__newtp("raw_syscalls", direction); 325 326 /* older kernel (e.g., RHEL6) use syscalls:{enter,exit} */ 327 if (IS_ERR(evsel)) 328 evsel = perf_evsel__newtp("syscalls", direction); 329 330 if (IS_ERR(evsel)) 331 return NULL; 332 333 if (perf_evsel__init_raw_syscall_tp(evsel, handler)) 334 goto out_delete; 335 336 return evsel; 337 338 out_delete: 339 perf_evsel__delete_priv(evsel); 340 return NULL; 341 } 342 343 #define perf_evsel__sc_tp_uint(evsel, name, sample) \ 344 ({ struct syscall_tp *fields = evsel->priv; \ 345 fields->name.integer(&fields->name, sample); }) 346 347 #define perf_evsel__sc_tp_ptr(evsel, name, sample) \ 348 ({ struct syscall_tp *fields = evsel->priv; \ 349 fields->name.pointer(&fields->name, sample); }) 350 351 size_t strarray__scnprintf(struct strarray *sa, char *bf, size_t size, const char *intfmt, int val) 352 { 353 int idx = val - sa->offset; 354 355 if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) 356 return scnprintf(bf, size, intfmt, val); 357 358 return scnprintf(bf, size, "%s", sa->entries[idx]); 359 } 360 361 static size_t __syscall_arg__scnprintf_strarray(char *bf, size_t size, 362 const char *intfmt, 363 struct syscall_arg *arg) 364 { 365 return strarray__scnprintf(arg->parm, bf, size, intfmt, arg->val); 366 } 367 368 static size_t syscall_arg__scnprintf_strarray(char *bf, size_t size, 369 struct syscall_arg *arg) 370 { 371 return __syscall_arg__scnprintf_strarray(bf, size, "%d", arg); 372 } 373 374 #define SCA_STRARRAY syscall_arg__scnprintf_strarray 375 376 struct strarrays { 377 int nr_entries; 378 struct strarray **entries; 379 }; 380 381 #define DEFINE_STRARRAYS(array) struct strarrays strarrays__##array = { \ 382 .nr_entries = ARRAY_SIZE(array), \ 383 .entries = array, \ 384 } 385 386 size_t syscall_arg__scnprintf_strarrays(char *bf, size_t size, 387 struct syscall_arg *arg) 388 { 389 struct strarrays *sas = arg->parm; 390 int i; 391 392 for (i = 0; i < sas->nr_entries; ++i) { 393 struct strarray *sa = sas->entries[i]; 394 int idx = arg->val - sa->offset; 395 396 if (idx >= 0 && idx < sa->nr_entries) { 397 if (sa->entries[idx] == NULL) 398 break; 399 return scnprintf(bf, size, "%s", sa->entries[idx]); 400 } 401 } 402 403 return scnprintf(bf, size, "%d", arg->val); 404 } 405 406 #ifndef AT_FDCWD 407 #define AT_FDCWD -100 408 #endif 409 410 static size_t syscall_arg__scnprintf_fd_at(char *bf, size_t size, 411 struct syscall_arg *arg) 412 { 413 int fd = arg->val; 414 415 if (fd == AT_FDCWD) 416 return scnprintf(bf, size, "CWD"); 417 418 return syscall_arg__scnprintf_fd(bf, size, arg); 419 } 420 421 #define SCA_FDAT syscall_arg__scnprintf_fd_at 422 423 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size, 424 struct syscall_arg *arg); 425 426 #define SCA_CLOSE_FD syscall_arg__scnprintf_close_fd 427 428 size_t syscall_arg__scnprintf_hex(char *bf, size_t size, struct syscall_arg *arg) 429 { 430 return scnprintf(bf, size, "%#lx", arg->val); 431 } 432 433 size_t syscall_arg__scnprintf_int(char *bf, size_t size, struct syscall_arg *arg) 434 { 435 return scnprintf(bf, size, "%d", arg->val); 436 } 437 438 size_t syscall_arg__scnprintf_long(char *bf, size_t size, struct syscall_arg *arg) 439 { 440 return scnprintf(bf, size, "%ld", arg->val); 441 } 442 443 static const char *bpf_cmd[] = { 444 "MAP_CREATE", "MAP_LOOKUP_ELEM", "MAP_UPDATE_ELEM", "MAP_DELETE_ELEM", 445 "MAP_GET_NEXT_KEY", "PROG_LOAD", 446 }; 447 static DEFINE_STRARRAY(bpf_cmd); 448 449 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", }; 450 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, 1); 451 452 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", }; 453 static DEFINE_STRARRAY(itimers); 454 455 static const char *keyctl_options[] = { 456 "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN", 457 "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ", 458 "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT", 459 "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT", 460 "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT", 461 }; 462 static DEFINE_STRARRAY(keyctl_options); 463 464 static const char *whences[] = { "SET", "CUR", "END", 465 #ifdef SEEK_DATA 466 "DATA", 467 #endif 468 #ifdef SEEK_HOLE 469 "HOLE", 470 #endif 471 }; 472 static DEFINE_STRARRAY(whences); 473 474 static const char *fcntl_cmds[] = { 475 "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK", 476 "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "GETLK64", 477 "SETLK64", "SETLKW64", "SETOWN_EX", "GETOWN_EX", 478 "GETOWNER_UIDS", 479 }; 480 static DEFINE_STRARRAY(fcntl_cmds); 481 482 static const char *fcntl_linux_specific_cmds[] = { 483 "SETLEASE", "GETLEASE", "NOTIFY", [5] = "CANCELLK", "DUPFD_CLOEXEC", 484 "SETPIPE_SZ", "GETPIPE_SZ", "ADD_SEALS", "GET_SEALS", 485 "GET_RW_HINT", "SET_RW_HINT", "GET_FILE_RW_HINT", "SET_FILE_RW_HINT", 486 }; 487 488 static DEFINE_STRARRAY_OFFSET(fcntl_linux_specific_cmds, F_LINUX_SPECIFIC_BASE); 489 490 static struct strarray *fcntl_cmds_arrays[] = { 491 &strarray__fcntl_cmds, 492 &strarray__fcntl_linux_specific_cmds, 493 }; 494 495 static DEFINE_STRARRAYS(fcntl_cmds_arrays); 496 497 static const char *rlimit_resources[] = { 498 "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE", 499 "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO", 500 "RTTIME", 501 }; 502 static DEFINE_STRARRAY(rlimit_resources); 503 504 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", }; 505 static DEFINE_STRARRAY(sighow); 506 507 static const char *clockid[] = { 508 "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID", 509 "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME", 510 "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI" 511 }; 512 static DEFINE_STRARRAY(clockid); 513 514 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size, 515 struct syscall_arg *arg) 516 { 517 size_t printed = 0; 518 int mode = arg->val; 519 520 if (mode == F_OK) /* 0 */ 521 return scnprintf(bf, size, "F"); 522 #define P_MODE(n) \ 523 if (mode & n##_OK) { \ 524 printed += scnprintf(bf + printed, size - printed, "%s", #n); \ 525 mode &= ~n##_OK; \ 526 } 527 528 P_MODE(R); 529 P_MODE(W); 530 P_MODE(X); 531 #undef P_MODE 532 533 if (mode) 534 printed += scnprintf(bf + printed, size - printed, "|%#x", mode); 535 536 return printed; 537 } 538 539 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode 540 541 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size, 542 struct syscall_arg *arg); 543 544 #define SCA_FILENAME syscall_arg__scnprintf_filename 545 546 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size, 547 struct syscall_arg *arg) 548 { 549 int printed = 0, flags = arg->val; 550 551 #define P_FLAG(n) \ 552 if (flags & O_##n) { \ 553 printed += scnprintf(bf + printed, size - printed, "%s%s", printed ? "|" : "", #n); \ 554 flags &= ~O_##n; \ 555 } 556 557 P_FLAG(CLOEXEC); 558 P_FLAG(NONBLOCK); 559 #undef P_FLAG 560 561 if (flags) 562 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags); 563 564 return printed; 565 } 566 567 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags 568 569 #ifndef GRND_NONBLOCK 570 #define GRND_NONBLOCK 0x0001 571 #endif 572 #ifndef GRND_RANDOM 573 #define GRND_RANDOM 0x0002 574 #endif 575 576 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size, 577 struct syscall_arg *arg) 578 { 579 int printed = 0, flags = arg->val; 580 581 #define P_FLAG(n) \ 582 if (flags & GRND_##n) { \ 583 printed += scnprintf(bf + printed, size - printed, "%s%s", printed ? "|" : "", #n); \ 584 flags &= ~GRND_##n; \ 585 } 586 587 P_FLAG(RANDOM); 588 P_FLAG(NONBLOCK); 589 #undef P_FLAG 590 591 if (flags) 592 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags); 593 594 return printed; 595 } 596 597 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags 598 599 #define STRARRAY(name, array) \ 600 { .scnprintf = SCA_STRARRAY, \ 601 .parm = &strarray__##array, } 602 603 #include "trace/beauty/arch_errno_names.c" 604 #include "trace/beauty/eventfd.c" 605 #include "trace/beauty/futex_op.c" 606 #include "trace/beauty/futex_val3.c" 607 #include "trace/beauty/mmap.c" 608 #include "trace/beauty/mode_t.c" 609 #include "trace/beauty/msg_flags.c" 610 #include "trace/beauty/open_flags.c" 611 #include "trace/beauty/perf_event_open.c" 612 #include "trace/beauty/pid.c" 613 #include "trace/beauty/sched_policy.c" 614 #include "trace/beauty/seccomp.c" 615 #include "trace/beauty/signum.c" 616 #include "trace/beauty/socket_type.c" 617 #include "trace/beauty/waitid_options.c" 618 619 struct syscall_arg_fmt { 620 size_t (*scnprintf)(char *bf, size_t size, struct syscall_arg *arg); 621 unsigned long (*mask_val)(struct syscall_arg *arg, unsigned long val); 622 void *parm; 623 const char *name; 624 bool show_zero; 625 }; 626 627 static struct syscall_fmt { 628 const char *name; 629 const char *alias; 630 struct syscall_arg_fmt arg[6]; 631 u8 nr_args; 632 bool errpid; 633 bool timeout; 634 bool hexret; 635 } syscall_fmts[] = { 636 { .name = "access", 637 .arg = { [1] = { .scnprintf = SCA_ACCMODE, /* mode */ }, }, }, 638 { .name = "bind", 639 .arg = { [1] = { .scnprintf = SCA_SOCKADDR, /* umyaddr */ }, }, }, 640 { .name = "bpf", 641 .arg = { [0] = STRARRAY(cmd, bpf_cmd), }, }, 642 { .name = "brk", .hexret = true, 643 .arg = { [0] = { .scnprintf = SCA_HEX, /* brk */ }, }, }, 644 { .name = "clock_gettime", 645 .arg = { [0] = STRARRAY(clk_id, clockid), }, }, 646 { .name = "clone", .errpid = true, .nr_args = 5, 647 .arg = { [0] = { .name = "flags", .scnprintf = SCA_CLONE_FLAGS, }, 648 [1] = { .name = "child_stack", .scnprintf = SCA_HEX, }, 649 [2] = { .name = "parent_tidptr", .scnprintf = SCA_HEX, }, 650 [3] = { .name = "child_tidptr", .scnprintf = SCA_HEX, }, 651 [4] = { .name = "tls", .scnprintf = SCA_HEX, }, }, }, 652 { .name = "close", 653 .arg = { [0] = { .scnprintf = SCA_CLOSE_FD, /* fd */ }, }, }, 654 { .name = "connect", 655 .arg = { [1] = { .scnprintf = SCA_SOCKADDR, /* servaddr */ }, }, }, 656 { .name = "epoll_ctl", 657 .arg = { [1] = STRARRAY(op, epoll_ctl_ops), }, }, 658 { .name = "eventfd2", 659 .arg = { [1] = { .scnprintf = SCA_EFD_FLAGS, /* flags */ }, }, }, 660 { .name = "fchmodat", 661 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, }, 662 { .name = "fchownat", 663 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, }, 664 { .name = "fcntl", 665 .arg = { [1] = { .scnprintf = SCA_FCNTL_CMD, /* cmd */ 666 .parm = &strarrays__fcntl_cmds_arrays, 667 .show_zero = true, }, 668 [2] = { .scnprintf = SCA_FCNTL_ARG, /* arg */ }, }, }, 669 { .name = "flock", 670 .arg = { [1] = { .scnprintf = SCA_FLOCK, /* cmd */ }, }, }, 671 { .name = "fstat", .alias = "newfstat", }, 672 { .name = "fstatat", .alias = "newfstatat", }, 673 { .name = "futex", 674 .arg = { [1] = { .scnprintf = SCA_FUTEX_OP, /* op */ }, 675 [5] = { .scnprintf = SCA_FUTEX_VAL3, /* val3 */ }, }, }, 676 { .name = "futimesat", 677 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, }, 678 { .name = "getitimer", 679 .arg = { [0] = STRARRAY(which, itimers), }, }, 680 { .name = "getpid", .errpid = true, }, 681 { .name = "getpgid", .errpid = true, }, 682 { .name = "getppid", .errpid = true, }, 683 { .name = "getrandom", 684 .arg = { [2] = { .scnprintf = SCA_GETRANDOM_FLAGS, /* flags */ }, }, }, 685 { .name = "getrlimit", 686 .arg = { [0] = STRARRAY(resource, rlimit_resources), }, }, 687 { .name = "gettid", .errpid = true, }, 688 { .name = "ioctl", 689 .arg = { 690 #if defined(__i386__) || defined(__x86_64__) 691 /* 692 * FIXME: Make this available to all arches. 693 */ 694 [1] = { .scnprintf = SCA_IOCTL_CMD, /* cmd */ }, 695 [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, }, 696 #else 697 [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, }, 698 #endif 699 { .name = "kcmp", .nr_args = 5, 700 .arg = { [0] = { .name = "pid1", .scnprintf = SCA_PID, }, 701 [1] = { .name = "pid2", .scnprintf = SCA_PID, }, 702 [2] = { .name = "type", .scnprintf = SCA_KCMP_TYPE, }, 703 [3] = { .name = "idx1", .scnprintf = SCA_KCMP_IDX, }, 704 [4] = { .name = "idx2", .scnprintf = SCA_KCMP_IDX, }, }, }, 705 { .name = "keyctl", 706 .arg = { [0] = STRARRAY(option, keyctl_options), }, }, 707 { .name = "kill", 708 .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, }, 709 { .name = "linkat", 710 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, }, 711 { .name = "lseek", 712 .arg = { [2] = STRARRAY(whence, whences), }, }, 713 { .name = "lstat", .alias = "newlstat", }, 714 { .name = "madvise", 715 .arg = { [0] = { .scnprintf = SCA_HEX, /* start */ }, 716 [2] = { .scnprintf = SCA_MADV_BHV, /* behavior */ }, }, }, 717 { .name = "mkdirat", 718 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, }, 719 { .name = "mknodat", 720 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, }, 721 { .name = "mlock", 722 .arg = { [0] = { .scnprintf = SCA_HEX, /* addr */ }, }, }, 723 { .name = "mlockall", 724 .arg = { [0] = { .scnprintf = SCA_HEX, /* addr */ }, }, }, 725 { .name = "mmap", .hexret = true, 726 /* The standard mmap maps to old_mmap on s390x */ 727 #if defined(__s390x__) 728 .alias = "old_mmap", 729 #endif 730 .arg = { [0] = { .scnprintf = SCA_HEX, /* addr */ }, 731 [2] = { .scnprintf = SCA_MMAP_PROT, /* prot */ }, 732 [3] = { .scnprintf = SCA_MMAP_FLAGS, /* flags */ }, }, }, 733 { .name = "mount", 734 .arg = { [0] = { .scnprintf = SCA_FILENAME, /* dev_name */ }, 735 [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */ 736 .mask_val = SCAMV_MOUNT_FLAGS, /* flags */ }, }, }, 737 { .name = "mprotect", 738 .arg = { [0] = { .scnprintf = SCA_HEX, /* start */ }, 739 [2] = { .scnprintf = SCA_MMAP_PROT, /* prot */ }, }, }, 740 { .name = "mq_unlink", 741 .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, }, 742 { .name = "mremap", .hexret = true, 743 .arg = { [0] = { .scnprintf = SCA_HEX, /* addr */ }, 744 [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ }, 745 [4] = { .scnprintf = SCA_HEX, /* new_addr */ }, }, }, 746 { .name = "munlock", 747 .arg = { [0] = { .scnprintf = SCA_HEX, /* addr */ }, }, }, 748 { .name = "munmap", 749 .arg = { [0] = { .scnprintf = SCA_HEX, /* addr */ }, }, }, 750 { .name = "name_to_handle_at", 751 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, }, 752 { .name = "newfstatat", 753 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, }, 754 { .name = "open", 755 .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, }, 756 { .name = "open_by_handle_at", 757 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, 758 [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, }, 759 { .name = "openat", 760 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, 761 [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, }, 762 { .name = "perf_event_open", 763 .arg = { [2] = { .scnprintf = SCA_INT, /* cpu */ }, 764 [3] = { .scnprintf = SCA_FD, /* group_fd */ }, 765 [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, }, 766 { .name = "pipe2", 767 .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, }, 768 { .name = "pkey_alloc", 769 .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS, /* access_rights */ }, }, }, 770 { .name = "pkey_free", 771 .arg = { [0] = { .scnprintf = SCA_INT, /* key */ }, }, }, 772 { .name = "pkey_mprotect", 773 .arg = { [0] = { .scnprintf = SCA_HEX, /* start */ }, 774 [2] = { .scnprintf = SCA_MMAP_PROT, /* prot */ }, 775 [3] = { .scnprintf = SCA_INT, /* pkey */ }, }, }, 776 { .name = "poll", .timeout = true, }, 777 { .name = "ppoll", .timeout = true, }, 778 { .name = "prctl", .alias = "arch_prctl", 779 .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */ }, 780 [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ }, 781 [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, }, 782 { .name = "pread", .alias = "pread64", }, 783 { .name = "preadv", .alias = "pread", }, 784 { .name = "prlimit64", 785 .arg = { [1] = STRARRAY(resource, rlimit_resources), }, }, 786 { .name = "pwrite", .alias = "pwrite64", }, 787 { .name = "readlinkat", 788 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, }, 789 { .name = "recvfrom", 790 .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, }, 791 { .name = "recvmmsg", 792 .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, }, 793 { .name = "recvmsg", 794 .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, }, 795 { .name = "renameat", 796 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, }, 797 { .name = "rt_sigaction", 798 .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, }, 799 { .name = "rt_sigprocmask", 800 .arg = { [0] = STRARRAY(how, sighow), }, }, 801 { .name = "rt_sigqueueinfo", 802 .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, }, 803 { .name = "rt_tgsigqueueinfo", 804 .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, }, 805 { .name = "sched_setscheduler", 806 .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, }, 807 { .name = "seccomp", 808 .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP, /* op */ }, 809 [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, }, 810 { .name = "select", .timeout = true, }, 811 { .name = "sendmmsg", 812 .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, }, 813 { .name = "sendmsg", 814 .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, }, 815 { .name = "sendto", 816 .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, 817 [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, }, 818 { .name = "set_tid_address", .errpid = true, }, 819 { .name = "setitimer", 820 .arg = { [0] = STRARRAY(which, itimers), }, }, 821 { .name = "setrlimit", 822 .arg = { [0] = STRARRAY(resource, rlimit_resources), }, }, 823 { .name = "socket", 824 .arg = { [0] = STRARRAY(family, socket_families), 825 [1] = { .scnprintf = SCA_SK_TYPE, /* type */ }, 826 [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, }, 827 { .name = "socketpair", 828 .arg = { [0] = STRARRAY(family, socket_families), 829 [1] = { .scnprintf = SCA_SK_TYPE, /* type */ }, 830 [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, }, 831 { .name = "stat", .alias = "newstat", }, 832 { .name = "statx", 833 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fdat */ }, 834 [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } , 835 [3] = { .scnprintf = SCA_STATX_MASK, /* mask */ }, }, }, 836 { .name = "swapoff", 837 .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, }, 838 { .name = "swapon", 839 .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, }, 840 { .name = "symlinkat", 841 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, }, 842 { .name = "tgkill", 843 .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, }, 844 { .name = "tkill", 845 .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, }, 846 { .name = "umount2", .alias = "umount", 847 .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, }, 848 { .name = "uname", .alias = "newuname", }, 849 { .name = "unlinkat", 850 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, }, 851 { .name = "utimensat", 852 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, }, 853 { .name = "wait4", .errpid = true, 854 .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, }, 855 { .name = "waitid", .errpid = true, 856 .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, }, 857 }; 858 859 static int syscall_fmt__cmp(const void *name, const void *fmtp) 860 { 861 const struct syscall_fmt *fmt = fmtp; 862 return strcmp(name, fmt->name); 863 } 864 865 static struct syscall_fmt *syscall_fmt__find(const char *name) 866 { 867 const int nmemb = ARRAY_SIZE(syscall_fmts); 868 return bsearch(name, syscall_fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp); 869 } 870 871 static struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias) 872 { 873 int i, nmemb = ARRAY_SIZE(syscall_fmts); 874 875 for (i = 0; i < nmemb; ++i) { 876 if (syscall_fmts[i].alias && strcmp(syscall_fmts[i].alias, alias) == 0) 877 return &syscall_fmts[i]; 878 } 879 880 return NULL; 881 } 882 883 /* 884 * is_exit: is this "exit" or "exit_group"? 885 * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter. 886 * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc. 887 */ 888 struct syscall { 889 struct tep_event *tp_format; 890 int nr_args; 891 int args_size; 892 bool is_exit; 893 bool is_open; 894 struct tep_format_field *args; 895 const char *name; 896 struct syscall_fmt *fmt; 897 struct syscall_arg_fmt *arg_fmt; 898 }; 899 900 /* 901 * We need to have this 'calculated' boolean because in some cases we really 902 * don't know what is the duration of a syscall, for instance, when we start 903 * a session and some threads are waiting for a syscall to finish, say 'poll', 904 * in which case all we can do is to print "( ? ) for duration and for the 905 * start timestamp. 906 */ 907 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp) 908 { 909 double duration = (double)t / NSEC_PER_MSEC; 910 size_t printed = fprintf(fp, "("); 911 912 if (!calculated) 913 printed += fprintf(fp, " "); 914 else if (duration >= 1.0) 915 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration); 916 else if (duration >= 0.01) 917 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration); 918 else 919 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration); 920 return printed + fprintf(fp, "): "); 921 } 922 923 /** 924 * filename.ptr: The filename char pointer that will be vfs_getname'd 925 * filename.entry_str_pos: Where to insert the string translated from 926 * filename.ptr by the vfs_getname tracepoint/kprobe. 927 * ret_scnprintf: syscall args may set this to a different syscall return 928 * formatter, for instance, fcntl may return fds, file flags, etc. 929 */ 930 struct thread_trace { 931 u64 entry_time; 932 bool entry_pending; 933 unsigned long nr_events; 934 unsigned long pfmaj, pfmin; 935 char *entry_str; 936 double runtime_ms; 937 size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg); 938 struct { 939 unsigned long ptr; 940 short int entry_str_pos; 941 bool pending_open; 942 unsigned int namelen; 943 char *name; 944 } filename; 945 struct { 946 int max; 947 char **table; 948 } paths; 949 950 struct intlist *syscall_stats; 951 }; 952 953 static struct thread_trace *thread_trace__new(void) 954 { 955 struct thread_trace *ttrace = zalloc(sizeof(struct thread_trace)); 956 957 if (ttrace) 958 ttrace->paths.max = -1; 959 960 ttrace->syscall_stats = intlist__new(NULL); 961 962 return ttrace; 963 } 964 965 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp) 966 { 967 struct thread_trace *ttrace; 968 969 if (thread == NULL) 970 goto fail; 971 972 if (thread__priv(thread) == NULL) 973 thread__set_priv(thread, thread_trace__new()); 974 975 if (thread__priv(thread) == NULL) 976 goto fail; 977 978 ttrace = thread__priv(thread); 979 ++ttrace->nr_events; 980 981 return ttrace; 982 fail: 983 color_fprintf(fp, PERF_COLOR_RED, 984 "WARNING: not enough memory, dropping samples!\n"); 985 return NULL; 986 } 987 988 989 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg, 990 size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg)) 991 { 992 struct thread_trace *ttrace = thread__priv(arg->thread); 993 994 ttrace->ret_scnprintf = ret_scnprintf; 995 } 996 997 #define TRACE_PFMAJ (1 << 0) 998 #define TRACE_PFMIN (1 << 1) 999 1000 static const size_t trace__entry_str_size = 2048; 1001 1002 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname) 1003 { 1004 struct thread_trace *ttrace = thread__priv(thread); 1005 1006 if (fd > ttrace->paths.max) { 1007 char **npath = realloc(ttrace->paths.table, (fd + 1) * sizeof(char *)); 1008 1009 if (npath == NULL) 1010 return -1; 1011 1012 if (ttrace->paths.max != -1) { 1013 memset(npath + ttrace->paths.max + 1, 0, 1014 (fd - ttrace->paths.max) * sizeof(char *)); 1015 } else { 1016 memset(npath, 0, (fd + 1) * sizeof(char *)); 1017 } 1018 1019 ttrace->paths.table = npath; 1020 ttrace->paths.max = fd; 1021 } 1022 1023 ttrace->paths.table[fd] = strdup(pathname); 1024 1025 return ttrace->paths.table[fd] != NULL ? 0 : -1; 1026 } 1027 1028 static int thread__read_fd_path(struct thread *thread, int fd) 1029 { 1030 char linkname[PATH_MAX], pathname[PATH_MAX]; 1031 struct stat st; 1032 int ret; 1033 1034 if (thread->pid_ == thread->tid) { 1035 scnprintf(linkname, sizeof(linkname), 1036 "/proc/%d/fd/%d", thread->pid_, fd); 1037 } else { 1038 scnprintf(linkname, sizeof(linkname), 1039 "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd); 1040 } 1041 1042 if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname)) 1043 return -1; 1044 1045 ret = readlink(linkname, pathname, sizeof(pathname)); 1046 1047 if (ret < 0 || ret > st.st_size) 1048 return -1; 1049 1050 pathname[ret] = '\0'; 1051 return trace__set_fd_pathname(thread, fd, pathname); 1052 } 1053 1054 static const char *thread__fd_path(struct thread *thread, int fd, 1055 struct trace *trace) 1056 { 1057 struct thread_trace *ttrace = thread__priv(thread); 1058 1059 if (ttrace == NULL) 1060 return NULL; 1061 1062 if (fd < 0) 1063 return NULL; 1064 1065 if ((fd > ttrace->paths.max || ttrace->paths.table[fd] == NULL)) { 1066 if (!trace->live) 1067 return NULL; 1068 ++trace->stats.proc_getname; 1069 if (thread__read_fd_path(thread, fd)) 1070 return NULL; 1071 } 1072 1073 return ttrace->paths.table[fd]; 1074 } 1075 1076 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg) 1077 { 1078 int fd = arg->val; 1079 size_t printed = scnprintf(bf, size, "%d", fd); 1080 const char *path = thread__fd_path(arg->thread, fd, arg->trace); 1081 1082 if (path) 1083 printed += scnprintf(bf + printed, size - printed, "<%s>", path); 1084 1085 return printed; 1086 } 1087 1088 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size) 1089 { 1090 size_t printed = scnprintf(bf, size, "%d", fd); 1091 struct thread *thread = machine__find_thread(trace->host, pid, pid); 1092 1093 if (thread) { 1094 const char *path = thread__fd_path(thread, fd, trace); 1095 1096 if (path) 1097 printed += scnprintf(bf + printed, size - printed, "<%s>", path); 1098 1099 thread__put(thread); 1100 } 1101 1102 return printed; 1103 } 1104 1105 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size, 1106 struct syscall_arg *arg) 1107 { 1108 int fd = arg->val; 1109 size_t printed = syscall_arg__scnprintf_fd(bf, size, arg); 1110 struct thread_trace *ttrace = thread__priv(arg->thread); 1111 1112 if (ttrace && fd >= 0 && fd <= ttrace->paths.max) 1113 zfree(&ttrace->paths.table[fd]); 1114 1115 return printed; 1116 } 1117 1118 static void thread__set_filename_pos(struct thread *thread, const char *bf, 1119 unsigned long ptr) 1120 { 1121 struct thread_trace *ttrace = thread__priv(thread); 1122 1123 ttrace->filename.ptr = ptr; 1124 ttrace->filename.entry_str_pos = bf - ttrace->entry_str; 1125 } 1126 1127 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size) 1128 { 1129 struct augmented_arg *augmented_arg = arg->augmented.args; 1130 1131 return scnprintf(bf, size, "%.*s", augmented_arg->size, augmented_arg->value); 1132 } 1133 1134 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size, 1135 struct syscall_arg *arg) 1136 { 1137 unsigned long ptr = arg->val; 1138 1139 if (arg->augmented.args) 1140 return syscall_arg__scnprintf_augmented_string(arg, bf, size); 1141 1142 if (!arg->trace->vfs_getname) 1143 return scnprintf(bf, size, "%#x", ptr); 1144 1145 thread__set_filename_pos(arg->thread, bf, ptr); 1146 return 0; 1147 } 1148 1149 static bool trace__filter_duration(struct trace *trace, double t) 1150 { 1151 return t < (trace->duration_filter * NSEC_PER_MSEC); 1152 } 1153 1154 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp) 1155 { 1156 double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC; 1157 1158 return fprintf(fp, "%10.3f ", ts); 1159 } 1160 1161 /* 1162 * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are 1163 * using ttrace->entry_time for a thread that receives a sys_exit without 1164 * first having received a sys_enter ("poll" issued before tracing session 1165 * starts, lost sys_enter exit due to ring buffer overflow). 1166 */ 1167 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp) 1168 { 1169 if (tstamp > 0) 1170 return __trace__fprintf_tstamp(trace, tstamp, fp); 1171 1172 return fprintf(fp, " ? "); 1173 } 1174 1175 static bool done = false; 1176 static bool interrupted = false; 1177 1178 static void sig_handler(int sig) 1179 { 1180 done = true; 1181 interrupted = sig == SIGINT; 1182 } 1183 1184 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp) 1185 { 1186 size_t printed = 0; 1187 1188 if (trace->multiple_threads) { 1189 if (trace->show_comm) 1190 printed += fprintf(fp, "%.14s/", thread__comm_str(thread)); 1191 printed += fprintf(fp, "%d ", thread->tid); 1192 } 1193 1194 return printed; 1195 } 1196 1197 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread, 1198 u64 duration, bool duration_calculated, u64 tstamp, FILE *fp) 1199 { 1200 size_t printed = trace__fprintf_tstamp(trace, tstamp, fp); 1201 printed += fprintf_duration(duration, duration_calculated, fp); 1202 return printed + trace__fprintf_comm_tid(trace, thread, fp); 1203 } 1204 1205 static int trace__process_event(struct trace *trace, struct machine *machine, 1206 union perf_event *event, struct perf_sample *sample) 1207 { 1208 int ret = 0; 1209 1210 switch (event->header.type) { 1211 case PERF_RECORD_LOST: 1212 color_fprintf(trace->output, PERF_COLOR_RED, 1213 "LOST %" PRIu64 " events!\n", event->lost.lost); 1214 ret = machine__process_lost_event(machine, event, sample); 1215 break; 1216 default: 1217 ret = machine__process_event(machine, event, sample); 1218 break; 1219 } 1220 1221 return ret; 1222 } 1223 1224 static int trace__tool_process(struct perf_tool *tool, 1225 union perf_event *event, 1226 struct perf_sample *sample, 1227 struct machine *machine) 1228 { 1229 struct trace *trace = container_of(tool, struct trace, tool); 1230 return trace__process_event(trace, machine, event, sample); 1231 } 1232 1233 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp) 1234 { 1235 struct machine *machine = vmachine; 1236 1237 if (machine->kptr_restrict_warned) 1238 return NULL; 1239 1240 if (symbol_conf.kptr_restrict) { 1241 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n" 1242 "Check /proc/sys/kernel/kptr_restrict.\n\n" 1243 "Kernel samples will not be resolved.\n"); 1244 machine->kptr_restrict_warned = true; 1245 return NULL; 1246 } 1247 1248 return machine__resolve_kernel_addr(vmachine, addrp, modp); 1249 } 1250 1251 static int trace__symbols_init(struct trace *trace, struct perf_evlist *evlist) 1252 { 1253 int err = symbol__init(NULL); 1254 1255 if (err) 1256 return err; 1257 1258 trace->host = machine__new_host(); 1259 if (trace->host == NULL) 1260 return -ENOMEM; 1261 1262 err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr); 1263 if (err < 0) 1264 goto out; 1265 1266 err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target, 1267 evlist->threads, trace__tool_process, false, 1268 1); 1269 out: 1270 if (err) 1271 symbol__exit(); 1272 1273 return err; 1274 } 1275 1276 static void trace__symbols__exit(struct trace *trace) 1277 { 1278 machine__exit(trace->host); 1279 trace->host = NULL; 1280 1281 symbol__exit(); 1282 } 1283 1284 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args) 1285 { 1286 int idx; 1287 1288 if (nr_args == 6 && sc->fmt && sc->fmt->nr_args != 0) 1289 nr_args = sc->fmt->nr_args; 1290 1291 sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt)); 1292 if (sc->arg_fmt == NULL) 1293 return -1; 1294 1295 for (idx = 0; idx < nr_args; ++idx) { 1296 if (sc->fmt) 1297 sc->arg_fmt[idx] = sc->fmt->arg[idx]; 1298 } 1299 1300 sc->nr_args = nr_args; 1301 return 0; 1302 } 1303 1304 static int syscall__set_arg_fmts(struct syscall *sc) 1305 { 1306 struct tep_format_field *field, *last_field = NULL; 1307 int idx = 0, len; 1308 1309 for (field = sc->args; field; field = field->next, ++idx) { 1310 last_field = field; 1311 1312 if (sc->fmt && sc->fmt->arg[idx].scnprintf) 1313 continue; 1314 1315 if (strcmp(field->type, "const char *") == 0 && 1316 (strcmp(field->name, "filename") == 0 || 1317 strcmp(field->name, "path") == 0 || 1318 strcmp(field->name, "pathname") == 0)) 1319 sc->arg_fmt[idx].scnprintf = SCA_FILENAME; 1320 else if (field->flags & TEP_FIELD_IS_POINTER) 1321 sc->arg_fmt[idx].scnprintf = syscall_arg__scnprintf_hex; 1322 else if (strcmp(field->type, "pid_t") == 0) 1323 sc->arg_fmt[idx].scnprintf = SCA_PID; 1324 else if (strcmp(field->type, "umode_t") == 0) 1325 sc->arg_fmt[idx].scnprintf = SCA_MODE_T; 1326 else if ((strcmp(field->type, "int") == 0 || 1327 strcmp(field->type, "unsigned int") == 0 || 1328 strcmp(field->type, "long") == 0) && 1329 (len = strlen(field->name)) >= 2 && 1330 strcmp(field->name + len - 2, "fd") == 0) { 1331 /* 1332 * /sys/kernel/tracing/events/syscalls/sys_enter* 1333 * egrep 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c 1334 * 65 int 1335 * 23 unsigned int 1336 * 7 unsigned long 1337 */ 1338 sc->arg_fmt[idx].scnprintf = SCA_FD; 1339 } 1340 } 1341 1342 if (last_field) 1343 sc->args_size = last_field->offset + last_field->size; 1344 1345 return 0; 1346 } 1347 1348 static int trace__read_syscall_info(struct trace *trace, int id) 1349 { 1350 char tp_name[128]; 1351 struct syscall *sc; 1352 const char *name = syscalltbl__name(trace->sctbl, id); 1353 1354 if (name == NULL) 1355 return -1; 1356 1357 if (id > trace->syscalls.max) { 1358 struct syscall *nsyscalls = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc)); 1359 1360 if (nsyscalls == NULL) 1361 return -1; 1362 1363 if (trace->syscalls.max != -1) { 1364 memset(nsyscalls + trace->syscalls.max + 1, 0, 1365 (id - trace->syscalls.max) * sizeof(*sc)); 1366 } else { 1367 memset(nsyscalls, 0, (id + 1) * sizeof(*sc)); 1368 } 1369 1370 trace->syscalls.table = nsyscalls; 1371 trace->syscalls.max = id; 1372 } 1373 1374 sc = trace->syscalls.table + id; 1375 sc->name = name; 1376 1377 sc->fmt = syscall_fmt__find(sc->name); 1378 1379 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name); 1380 sc->tp_format = trace_event__tp_format("syscalls", tp_name); 1381 1382 if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) { 1383 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias); 1384 sc->tp_format = trace_event__tp_format("syscalls", tp_name); 1385 } 1386 1387 if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ? 6 : sc->tp_format->format.nr_fields)) 1388 return -1; 1389 1390 if (IS_ERR(sc->tp_format)) 1391 return -1; 1392 1393 sc->args = sc->tp_format->format.fields; 1394 /* 1395 * We need to check and discard the first variable '__syscall_nr' 1396 * or 'nr' that mean the syscall number. It is needless here. 1397 * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels. 1398 */ 1399 if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) { 1400 sc->args = sc->args->next; 1401 --sc->nr_args; 1402 } 1403 1404 sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit"); 1405 sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat"); 1406 1407 return syscall__set_arg_fmts(sc); 1408 } 1409 1410 static int trace__validate_ev_qualifier(struct trace *trace) 1411 { 1412 int err = 0, i; 1413 size_t nr_allocated; 1414 struct str_node *pos; 1415 1416 trace->ev_qualifier_ids.nr = strlist__nr_entries(trace->ev_qualifier); 1417 trace->ev_qualifier_ids.entries = malloc(trace->ev_qualifier_ids.nr * 1418 sizeof(trace->ev_qualifier_ids.entries[0])); 1419 1420 if (trace->ev_qualifier_ids.entries == NULL) { 1421 fputs("Error:\tNot enough memory for allocating events qualifier ids\n", 1422 trace->output); 1423 err = -EINVAL; 1424 goto out; 1425 } 1426 1427 nr_allocated = trace->ev_qualifier_ids.nr; 1428 i = 0; 1429 1430 strlist__for_each_entry(pos, trace->ev_qualifier) { 1431 const char *sc = pos->s; 1432 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1; 1433 1434 if (id < 0) { 1435 id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next); 1436 if (id >= 0) 1437 goto matches; 1438 1439 if (err == 0) { 1440 fputs("Error:\tInvalid syscall ", trace->output); 1441 err = -EINVAL; 1442 } else { 1443 fputs(", ", trace->output); 1444 } 1445 1446 fputs(sc, trace->output); 1447 } 1448 matches: 1449 trace->ev_qualifier_ids.entries[i++] = id; 1450 if (match_next == -1) 1451 continue; 1452 1453 while (1) { 1454 id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next); 1455 if (id < 0) 1456 break; 1457 if (nr_allocated == trace->ev_qualifier_ids.nr) { 1458 void *entries; 1459 1460 nr_allocated += 8; 1461 entries = realloc(trace->ev_qualifier_ids.entries, 1462 nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0])); 1463 if (entries == NULL) { 1464 err = -ENOMEM; 1465 fputs("\nError:\t Not enough memory for parsing\n", trace->output); 1466 goto out_free; 1467 } 1468 trace->ev_qualifier_ids.entries = entries; 1469 } 1470 trace->ev_qualifier_ids.nr++; 1471 trace->ev_qualifier_ids.entries[i++] = id; 1472 } 1473 } 1474 1475 if (err < 0) { 1476 fputs("\nHint:\ttry 'perf list syscalls:sys_enter_*'" 1477 "\nHint:\tand: 'man syscalls'\n", trace->output); 1478 out_free: 1479 zfree(&trace->ev_qualifier_ids.entries); 1480 trace->ev_qualifier_ids.nr = 0; 1481 } 1482 out: 1483 return err; 1484 } 1485 1486 /* 1487 * args is to be interpreted as a series of longs but we need to handle 1488 * 8-byte unaligned accesses. args points to raw_data within the event 1489 * and raw_data is guaranteed to be 8-byte unaligned because it is 1490 * preceded by raw_size which is a u32. So we need to copy args to a temp 1491 * variable to read it. Most notably this avoids extended load instructions 1492 * on unaligned addresses 1493 */ 1494 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx) 1495 { 1496 unsigned long val; 1497 unsigned char *p = arg->args + sizeof(unsigned long) * idx; 1498 1499 memcpy(&val, p, sizeof(val)); 1500 return val; 1501 } 1502 1503 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size, 1504 struct syscall_arg *arg) 1505 { 1506 if (sc->arg_fmt && sc->arg_fmt[arg->idx].name) 1507 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name); 1508 1509 return scnprintf(bf, size, "arg%d: ", arg->idx); 1510 } 1511 1512 /* 1513 * Check if the value is in fact zero, i.e. mask whatever needs masking, such 1514 * as mount 'flags' argument that needs ignoring some magic flag, see comment 1515 * in tools/perf/trace/beauty/mount_flags.c 1516 */ 1517 static unsigned long syscall__mask_val(struct syscall *sc, struct syscall_arg *arg, unsigned long val) 1518 { 1519 if (sc->arg_fmt && sc->arg_fmt[arg->idx].mask_val) 1520 return sc->arg_fmt[arg->idx].mask_val(arg, val); 1521 1522 return val; 1523 } 1524 1525 static size_t syscall__scnprintf_val(struct syscall *sc, char *bf, size_t size, 1526 struct syscall_arg *arg, unsigned long val) 1527 { 1528 if (sc->arg_fmt && sc->arg_fmt[arg->idx].scnprintf) { 1529 arg->val = val; 1530 if (sc->arg_fmt[arg->idx].parm) 1531 arg->parm = sc->arg_fmt[arg->idx].parm; 1532 return sc->arg_fmt[arg->idx].scnprintf(bf, size, arg); 1533 } 1534 return scnprintf(bf, size, "%ld", val); 1535 } 1536 1537 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size, 1538 unsigned char *args, void *augmented_args, int augmented_args_size, 1539 struct trace *trace, struct thread *thread) 1540 { 1541 size_t printed = 0; 1542 unsigned long val; 1543 u8 bit = 1; 1544 struct syscall_arg arg = { 1545 .args = args, 1546 .augmented = { 1547 .size = augmented_args_size, 1548 .args = augmented_args, 1549 }, 1550 .idx = 0, 1551 .mask = 0, 1552 .trace = trace, 1553 .thread = thread, 1554 }; 1555 struct thread_trace *ttrace = thread__priv(thread); 1556 1557 /* 1558 * Things like fcntl will set this in its 'cmd' formatter to pick the 1559 * right formatter for the return value (an fd? file flags?), which is 1560 * not needed for syscalls that always return a given type, say an fd. 1561 */ 1562 ttrace->ret_scnprintf = NULL; 1563 1564 if (sc->args != NULL) { 1565 struct tep_format_field *field; 1566 1567 for (field = sc->args; field; 1568 field = field->next, ++arg.idx, bit <<= 1) { 1569 if (arg.mask & bit) 1570 continue; 1571 1572 val = syscall_arg__val(&arg, arg.idx); 1573 /* 1574 * Some syscall args need some mask, most don't and 1575 * return val untouched. 1576 */ 1577 val = syscall__mask_val(sc, &arg, val); 1578 1579 /* 1580 * Suppress this argument if its value is zero and 1581 * and we don't have a string associated in an 1582 * strarray for it. 1583 */ 1584 if (val == 0 && 1585 !(sc->arg_fmt && 1586 (sc->arg_fmt[arg.idx].show_zero || 1587 sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY || 1588 sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) && 1589 sc->arg_fmt[arg.idx].parm)) 1590 continue; 1591 1592 printed += scnprintf(bf + printed, size - printed, 1593 "%s%s: ", printed ? ", " : "", field->name); 1594 printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val); 1595 } 1596 } else if (IS_ERR(sc->tp_format)) { 1597 /* 1598 * If we managed to read the tracepoint /format file, then we 1599 * may end up not having any args, like with gettid(), so only 1600 * print the raw args when we didn't manage to read it. 1601 */ 1602 while (arg.idx < sc->nr_args) { 1603 if (arg.mask & bit) 1604 goto next_arg; 1605 val = syscall_arg__val(&arg, arg.idx); 1606 if (printed) 1607 printed += scnprintf(bf + printed, size - printed, ", "); 1608 printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg); 1609 printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val); 1610 next_arg: 1611 ++arg.idx; 1612 bit <<= 1; 1613 } 1614 } 1615 1616 return printed; 1617 } 1618 1619 typedef int (*tracepoint_handler)(struct trace *trace, struct perf_evsel *evsel, 1620 union perf_event *event, 1621 struct perf_sample *sample); 1622 1623 static struct syscall *trace__syscall_info(struct trace *trace, 1624 struct perf_evsel *evsel, int id) 1625 { 1626 1627 if (id < 0) { 1628 1629 /* 1630 * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried 1631 * before that, leaving at a higher verbosity level till that is 1632 * explained. Reproduced with plain ftrace with: 1633 * 1634 * echo 1 > /t/events/raw_syscalls/sys_exit/enable 1635 * grep "NR -1 " /t/trace_pipe 1636 * 1637 * After generating some load on the machine. 1638 */ 1639 if (verbose > 1) { 1640 static u64 n; 1641 fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n", 1642 id, perf_evsel__name(evsel), ++n); 1643 } 1644 return NULL; 1645 } 1646 1647 if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL) && 1648 trace__read_syscall_info(trace, id)) 1649 goto out_cant_read; 1650 1651 if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL)) 1652 goto out_cant_read; 1653 1654 return &trace->syscalls.table[id]; 1655 1656 out_cant_read: 1657 if (verbose > 0) { 1658 fprintf(trace->output, "Problems reading syscall %d", id); 1659 if (id <= trace->syscalls.max && trace->syscalls.table[id].name != NULL) 1660 fprintf(trace->output, "(%s)", trace->syscalls.table[id].name); 1661 fputs(" information\n", trace->output); 1662 } 1663 return NULL; 1664 } 1665 1666 static void thread__update_stats(struct thread_trace *ttrace, 1667 int id, struct perf_sample *sample) 1668 { 1669 struct int_node *inode; 1670 struct stats *stats; 1671 u64 duration = 0; 1672 1673 inode = intlist__findnew(ttrace->syscall_stats, id); 1674 if (inode == NULL) 1675 return; 1676 1677 stats = inode->priv; 1678 if (stats == NULL) { 1679 stats = malloc(sizeof(struct stats)); 1680 if (stats == NULL) 1681 return; 1682 init_stats(stats); 1683 inode->priv = stats; 1684 } 1685 1686 if (ttrace->entry_time && sample->time > ttrace->entry_time) 1687 duration = sample->time - ttrace->entry_time; 1688 1689 update_stats(stats, duration); 1690 } 1691 1692 static int trace__printf_interrupted_entry(struct trace *trace) 1693 { 1694 struct thread_trace *ttrace; 1695 size_t printed; 1696 1697 if (trace->failure_only || trace->current == NULL) 1698 return 0; 1699 1700 ttrace = thread__priv(trace->current); 1701 1702 if (!ttrace->entry_pending) 1703 return 0; 1704 1705 printed = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output); 1706 printed += fprintf(trace->output, "%-70s) ...\n", ttrace->entry_str); 1707 ttrace->entry_pending = false; 1708 1709 ++trace->nr_events_printed; 1710 1711 return printed; 1712 } 1713 1714 static int trace__fprintf_sample(struct trace *trace, struct perf_evsel *evsel, 1715 struct perf_sample *sample, struct thread *thread) 1716 { 1717 int printed = 0; 1718 1719 if (trace->print_sample) { 1720 double ts = (double)sample->time / NSEC_PER_MSEC; 1721 1722 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n", 1723 perf_evsel__name(evsel), ts, 1724 thread__comm_str(thread), 1725 sample->pid, sample->tid, sample->cpu); 1726 } 1727 1728 return printed; 1729 } 1730 1731 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, bool raw_augmented) 1732 { 1733 void *augmented_args = NULL; 1734 /* 1735 * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter 1736 * and there we get all 6 syscall args plus the tracepoint common 1737 * fields (sizeof(long)) and the syscall_nr (another long). So we check 1738 * if that is the case and if so don't look after the sc->args_size, 1739 * but always after the full raw_syscalls:sys_enter payload, which is 1740 * fixed. 1741 * 1742 * We'll revisit this later to pass s->args_size to the BPF augmenter 1743 * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it 1744 * copies only what we need for each syscall, like what happens when we 1745 * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace 1746 * traffic to just what is needed for each syscall. 1747 */ 1748 int args_size = raw_augmented ? (8 * (int)sizeof(long)) : sc->args_size; 1749 1750 *augmented_args_size = sample->raw_size - args_size; 1751 if (*augmented_args_size > 0) 1752 augmented_args = sample->raw_data + args_size; 1753 1754 return augmented_args; 1755 } 1756 1757 static int trace__sys_enter(struct trace *trace, struct perf_evsel *evsel, 1758 union perf_event *event __maybe_unused, 1759 struct perf_sample *sample) 1760 { 1761 char *msg; 1762 void *args; 1763 size_t printed = 0; 1764 struct thread *thread; 1765 int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1; 1766 int augmented_args_size = 0; 1767 void *augmented_args = NULL; 1768 struct syscall *sc = trace__syscall_info(trace, evsel, id); 1769 struct thread_trace *ttrace; 1770 1771 if (sc == NULL) 1772 return -1; 1773 1774 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid); 1775 ttrace = thread__trace(thread, trace->output); 1776 if (ttrace == NULL) 1777 goto out_put; 1778 1779 trace__fprintf_sample(trace, evsel, sample, thread); 1780 1781 args = perf_evsel__sc_tp_ptr(evsel, args, sample); 1782 1783 if (ttrace->entry_str == NULL) { 1784 ttrace->entry_str = malloc(trace__entry_str_size); 1785 if (!ttrace->entry_str) 1786 goto out_put; 1787 } 1788 1789 if (!(trace->duration_filter || trace->summary_only || trace->min_stack)) 1790 trace__printf_interrupted_entry(trace); 1791 /* 1792 * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible 1793 * arguments, even if the syscall being handled, say "openat", uses only 4 arguments 1794 * this breaks syscall__augmented_args() check for augmented args, as we calculate 1795 * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file, 1796 * so when handling, say the openat syscall, we end up getting 6 args for the 1797 * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly 1798 * thinking that the extra 2 u64 args are the augmented filename, so just check 1799 * here and avoid using augmented syscalls when the evsel is the raw_syscalls one. 1800 */ 1801 if (evsel != trace->syscalls.events.sys_enter) 1802 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls); 1803 ttrace->entry_time = sample->time; 1804 msg = ttrace->entry_str; 1805 printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name); 1806 1807 printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed, 1808 args, augmented_args, augmented_args_size, trace, thread); 1809 1810 if (sc->is_exit) { 1811 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) { 1812 trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output); 1813 fprintf(trace->output, "%-70s)\n", ttrace->entry_str); 1814 } 1815 } else { 1816 ttrace->entry_pending = true; 1817 /* See trace__vfs_getname & trace__sys_exit */ 1818 ttrace->filename.pending_open = false; 1819 } 1820 1821 if (trace->current != thread) { 1822 thread__put(trace->current); 1823 trace->current = thread__get(thread); 1824 } 1825 err = 0; 1826 out_put: 1827 thread__put(thread); 1828 return err; 1829 } 1830 1831 static int trace__fprintf_sys_enter(struct trace *trace, struct perf_evsel *evsel, 1832 struct perf_sample *sample) 1833 { 1834 struct thread_trace *ttrace; 1835 struct thread *thread; 1836 int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1; 1837 struct syscall *sc = trace__syscall_info(trace, evsel, id); 1838 char msg[1024]; 1839 void *args, *augmented_args = NULL; 1840 int augmented_args_size; 1841 1842 if (sc == NULL) 1843 return -1; 1844 1845 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid); 1846 ttrace = thread__trace(thread, trace->output); 1847 /* 1848 * We need to get ttrace just to make sure it is there when syscall__scnprintf_args() 1849 * and the rest of the beautifiers accessing it via struct syscall_arg touches it. 1850 */ 1851 if (ttrace == NULL) 1852 goto out_put; 1853 1854 args = perf_evsel__sc_tp_ptr(evsel, args, sample); 1855 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls); 1856 syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread); 1857 fprintf(trace->output, "%s", msg); 1858 err = 0; 1859 out_put: 1860 thread__put(thread); 1861 return err; 1862 } 1863 1864 static int trace__resolve_callchain(struct trace *trace, struct perf_evsel *evsel, 1865 struct perf_sample *sample, 1866 struct callchain_cursor *cursor) 1867 { 1868 struct addr_location al; 1869 int max_stack = evsel->attr.sample_max_stack ? 1870 evsel->attr.sample_max_stack : 1871 trace->max_stack; 1872 int err; 1873 1874 if (machine__resolve(trace->host, &al, sample) < 0) 1875 return -1; 1876 1877 err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack); 1878 addr_location__put(&al); 1879 return err; 1880 } 1881 1882 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample) 1883 { 1884 /* TODO: user-configurable print_opts */ 1885 const unsigned int print_opts = EVSEL__PRINT_SYM | 1886 EVSEL__PRINT_DSO | 1887 EVSEL__PRINT_UNKNOWN_AS_ADDR; 1888 1889 return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, trace->output); 1890 } 1891 1892 static const char *errno_to_name(struct perf_evsel *evsel, int err) 1893 { 1894 struct perf_env *env = perf_evsel__env(evsel); 1895 const char *arch_name = perf_env__arch(env); 1896 1897 return arch_syscalls__strerrno(arch_name, err); 1898 } 1899 1900 static int trace__sys_exit(struct trace *trace, struct perf_evsel *evsel, 1901 union perf_event *event __maybe_unused, 1902 struct perf_sample *sample) 1903 { 1904 long ret; 1905 u64 duration = 0; 1906 bool duration_calculated = false; 1907 struct thread *thread; 1908 int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0; 1909 struct syscall *sc = trace__syscall_info(trace, evsel, id); 1910 struct thread_trace *ttrace; 1911 1912 if (sc == NULL) 1913 return -1; 1914 1915 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid); 1916 ttrace = thread__trace(thread, trace->output); 1917 if (ttrace == NULL) 1918 goto out_put; 1919 1920 trace__fprintf_sample(trace, evsel, sample, thread); 1921 1922 if (trace->summary) 1923 thread__update_stats(ttrace, id, sample); 1924 1925 ret = perf_evsel__sc_tp_uint(evsel, ret, sample); 1926 1927 if (sc->is_open && ret >= 0 && ttrace->filename.pending_open) { 1928 trace__set_fd_pathname(thread, ret, ttrace->filename.name); 1929 ttrace->filename.pending_open = false; 1930 ++trace->stats.vfs_getname; 1931 } 1932 1933 if (ttrace->entry_time) { 1934 duration = sample->time - ttrace->entry_time; 1935 if (trace__filter_duration(trace, duration)) 1936 goto out; 1937 duration_calculated = true; 1938 } else if (trace->duration_filter) 1939 goto out; 1940 1941 if (sample->callchain) { 1942 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor); 1943 if (callchain_ret == 0) { 1944 if (callchain_cursor.nr < trace->min_stack) 1945 goto out; 1946 callchain_ret = 1; 1947 } 1948 } 1949 1950 if (trace->summary_only || (ret >= 0 && trace->failure_only)) 1951 goto out; 1952 1953 trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output); 1954 1955 if (ttrace->entry_pending) { 1956 fprintf(trace->output, "%-70s", ttrace->entry_str); 1957 } else { 1958 fprintf(trace->output, " ... ["); 1959 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued"); 1960 fprintf(trace->output, "]: %s()", sc->name); 1961 } 1962 1963 if (sc->fmt == NULL) { 1964 if (ret < 0) 1965 goto errno_print; 1966 signed_print: 1967 fprintf(trace->output, ") = %ld", ret); 1968 } else if (ret < 0) { 1969 errno_print: { 1970 char bf[STRERR_BUFSIZE]; 1971 const char *emsg = str_error_r(-ret, bf, sizeof(bf)), 1972 *e = errno_to_name(evsel, -ret); 1973 1974 fprintf(trace->output, ") = -1 %s %s", e, emsg); 1975 } 1976 } else if (ret == 0 && sc->fmt->timeout) 1977 fprintf(trace->output, ") = 0 Timeout"); 1978 else if (ttrace->ret_scnprintf) { 1979 char bf[1024]; 1980 struct syscall_arg arg = { 1981 .val = ret, 1982 .thread = thread, 1983 .trace = trace, 1984 }; 1985 ttrace->ret_scnprintf(bf, sizeof(bf), &arg); 1986 ttrace->ret_scnprintf = NULL; 1987 fprintf(trace->output, ") = %s", bf); 1988 } else if (sc->fmt->hexret) 1989 fprintf(trace->output, ") = %#lx", ret); 1990 else if (sc->fmt->errpid) { 1991 struct thread *child = machine__find_thread(trace->host, ret, ret); 1992 1993 if (child != NULL) { 1994 fprintf(trace->output, ") = %ld", ret); 1995 if (child->comm_set) 1996 fprintf(trace->output, " (%s)", thread__comm_str(child)); 1997 thread__put(child); 1998 } 1999 } else 2000 goto signed_print; 2001 2002 fputc('\n', trace->output); 2003 2004 /* 2005 * We only consider an 'event' for the sake of --max-events a non-filtered 2006 * sys_enter + sys_exit and other tracepoint events. 2007 */ 2008 if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX) 2009 interrupted = true; 2010 2011 if (callchain_ret > 0) 2012 trace__fprintf_callchain(trace, sample); 2013 else if (callchain_ret < 0) 2014 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel)); 2015 out: 2016 ttrace->entry_pending = false; 2017 err = 0; 2018 out_put: 2019 thread__put(thread); 2020 return err; 2021 } 2022 2023 static int trace__vfs_getname(struct trace *trace, struct perf_evsel *evsel, 2024 union perf_event *event __maybe_unused, 2025 struct perf_sample *sample) 2026 { 2027 struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid); 2028 struct thread_trace *ttrace; 2029 size_t filename_len, entry_str_len, to_move; 2030 ssize_t remaining_space; 2031 char *pos; 2032 const char *filename = perf_evsel__rawptr(evsel, sample, "pathname"); 2033 2034 if (!thread) 2035 goto out; 2036 2037 ttrace = thread__priv(thread); 2038 if (!ttrace) 2039 goto out_put; 2040 2041 filename_len = strlen(filename); 2042 if (filename_len == 0) 2043 goto out_put; 2044 2045 if (ttrace->filename.namelen < filename_len) { 2046 char *f = realloc(ttrace->filename.name, filename_len + 1); 2047 2048 if (f == NULL) 2049 goto out_put; 2050 2051 ttrace->filename.namelen = filename_len; 2052 ttrace->filename.name = f; 2053 } 2054 2055 strcpy(ttrace->filename.name, filename); 2056 ttrace->filename.pending_open = true; 2057 2058 if (!ttrace->filename.ptr) 2059 goto out_put; 2060 2061 entry_str_len = strlen(ttrace->entry_str); 2062 remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */ 2063 if (remaining_space <= 0) 2064 goto out_put; 2065 2066 if (filename_len > (size_t)remaining_space) { 2067 filename += filename_len - remaining_space; 2068 filename_len = remaining_space; 2069 } 2070 2071 to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */ 2072 pos = ttrace->entry_str + ttrace->filename.entry_str_pos; 2073 memmove(pos + filename_len, pos, to_move); 2074 memcpy(pos, filename, filename_len); 2075 2076 ttrace->filename.ptr = 0; 2077 ttrace->filename.entry_str_pos = 0; 2078 out_put: 2079 thread__put(thread); 2080 out: 2081 return 0; 2082 } 2083 2084 static int trace__sched_stat_runtime(struct trace *trace, struct perf_evsel *evsel, 2085 union perf_event *event __maybe_unused, 2086 struct perf_sample *sample) 2087 { 2088 u64 runtime = perf_evsel__intval(evsel, sample, "runtime"); 2089 double runtime_ms = (double)runtime / NSEC_PER_MSEC; 2090 struct thread *thread = machine__findnew_thread(trace->host, 2091 sample->pid, 2092 sample->tid); 2093 struct thread_trace *ttrace = thread__trace(thread, trace->output); 2094 2095 if (ttrace == NULL) 2096 goto out_dump; 2097 2098 ttrace->runtime_ms += runtime_ms; 2099 trace->runtime_ms += runtime_ms; 2100 out_put: 2101 thread__put(thread); 2102 return 0; 2103 2104 out_dump: 2105 fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n", 2106 evsel->name, 2107 perf_evsel__strval(evsel, sample, "comm"), 2108 (pid_t)perf_evsel__intval(evsel, sample, "pid"), 2109 runtime, 2110 perf_evsel__intval(evsel, sample, "vruntime")); 2111 goto out_put; 2112 } 2113 2114 static int bpf_output__printer(enum binary_printer_ops op, 2115 unsigned int val, void *extra __maybe_unused, FILE *fp) 2116 { 2117 unsigned char ch = (unsigned char)val; 2118 2119 switch (op) { 2120 case BINARY_PRINT_CHAR_DATA: 2121 return fprintf(fp, "%c", isprint(ch) ? ch : '.'); 2122 case BINARY_PRINT_DATA_BEGIN: 2123 case BINARY_PRINT_LINE_BEGIN: 2124 case BINARY_PRINT_ADDR: 2125 case BINARY_PRINT_NUM_DATA: 2126 case BINARY_PRINT_NUM_PAD: 2127 case BINARY_PRINT_SEP: 2128 case BINARY_PRINT_CHAR_PAD: 2129 case BINARY_PRINT_LINE_END: 2130 case BINARY_PRINT_DATA_END: 2131 default: 2132 break; 2133 } 2134 2135 return 0; 2136 } 2137 2138 static void bpf_output__fprintf(struct trace *trace, 2139 struct perf_sample *sample) 2140 { 2141 binary__fprintf(sample->raw_data, sample->raw_size, 8, 2142 bpf_output__printer, NULL, trace->output); 2143 ++trace->nr_events_printed; 2144 } 2145 2146 static int trace__event_handler(struct trace *trace, struct perf_evsel *evsel, 2147 union perf_event *event __maybe_unused, 2148 struct perf_sample *sample) 2149 { 2150 struct thread *thread; 2151 int callchain_ret = 0; 2152 /* 2153 * Check if we called perf_evsel__disable(evsel) due to, for instance, 2154 * this event's max_events having been hit and this is an entry coming 2155 * from the ring buffer that we should discard, since the max events 2156 * have already been considered/printed. 2157 */ 2158 if (evsel->disabled) 2159 return 0; 2160 2161 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid); 2162 2163 if (sample->callchain) { 2164 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor); 2165 if (callchain_ret == 0) { 2166 if (callchain_cursor.nr < trace->min_stack) 2167 goto out; 2168 callchain_ret = 1; 2169 } 2170 } 2171 2172 trace__printf_interrupted_entry(trace); 2173 trace__fprintf_tstamp(trace, sample->time, trace->output); 2174 2175 if (trace->trace_syscalls) 2176 fprintf(trace->output, "( ): "); 2177 2178 if (thread) 2179 trace__fprintf_comm_tid(trace, thread, trace->output); 2180 2181 if (evsel == trace->syscalls.events.augmented) { 2182 int id = perf_evsel__sc_tp_uint(evsel, id, sample); 2183 struct syscall *sc = trace__syscall_info(trace, evsel, id); 2184 2185 if (sc) { 2186 fprintf(trace->output, "%s(", sc->name); 2187 trace__fprintf_sys_enter(trace, evsel, sample); 2188 fputc(')', trace->output); 2189 goto newline; 2190 } 2191 2192 /* 2193 * XXX: Not having the associated syscall info or not finding/adding 2194 * the thread should never happen, but if it does... 2195 * fall thru and print it as a bpf_output event. 2196 */ 2197 } 2198 2199 fprintf(trace->output, "%s:", evsel->name); 2200 2201 if (perf_evsel__is_bpf_output(evsel)) { 2202 bpf_output__fprintf(trace, sample); 2203 } else if (evsel->tp_format) { 2204 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) || 2205 trace__fprintf_sys_enter(trace, evsel, sample)) { 2206 event_format__fprintf(evsel->tp_format, sample->cpu, 2207 sample->raw_data, sample->raw_size, 2208 trace->output); 2209 ++trace->nr_events_printed; 2210 2211 if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) { 2212 perf_evsel__disable(evsel); 2213 perf_evsel__close(evsel); 2214 } 2215 } 2216 } 2217 2218 newline: 2219 fprintf(trace->output, "\n"); 2220 2221 if (callchain_ret > 0) 2222 trace__fprintf_callchain(trace, sample); 2223 else if (callchain_ret < 0) 2224 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel)); 2225 out: 2226 thread__put(thread); 2227 return 0; 2228 } 2229 2230 static void print_location(FILE *f, struct perf_sample *sample, 2231 struct addr_location *al, 2232 bool print_dso, bool print_sym) 2233 { 2234 2235 if ((verbose > 0 || print_dso) && al->map) 2236 fprintf(f, "%s@", al->map->dso->long_name); 2237 2238 if ((verbose > 0 || print_sym) && al->sym) 2239 fprintf(f, "%s+0x%" PRIx64, al->sym->name, 2240 al->addr - al->sym->start); 2241 else if (al->map) 2242 fprintf(f, "0x%" PRIx64, al->addr); 2243 else 2244 fprintf(f, "0x%" PRIx64, sample->addr); 2245 } 2246 2247 static int trace__pgfault(struct trace *trace, 2248 struct perf_evsel *evsel, 2249 union perf_event *event __maybe_unused, 2250 struct perf_sample *sample) 2251 { 2252 struct thread *thread; 2253 struct addr_location al; 2254 char map_type = 'd'; 2255 struct thread_trace *ttrace; 2256 int err = -1; 2257 int callchain_ret = 0; 2258 2259 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid); 2260 2261 if (sample->callchain) { 2262 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor); 2263 if (callchain_ret == 0) { 2264 if (callchain_cursor.nr < trace->min_stack) 2265 goto out_put; 2266 callchain_ret = 1; 2267 } 2268 } 2269 2270 ttrace = thread__trace(thread, trace->output); 2271 if (ttrace == NULL) 2272 goto out_put; 2273 2274 if (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ) 2275 ttrace->pfmaj++; 2276 else 2277 ttrace->pfmin++; 2278 2279 if (trace->summary_only) 2280 goto out; 2281 2282 thread__find_symbol(thread, sample->cpumode, sample->ip, &al); 2283 2284 trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output); 2285 2286 fprintf(trace->output, "%sfault [", 2287 evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ? 2288 "maj" : "min"); 2289 2290 print_location(trace->output, sample, &al, false, true); 2291 2292 fprintf(trace->output, "] => "); 2293 2294 thread__find_symbol(thread, sample->cpumode, sample->addr, &al); 2295 2296 if (!al.map) { 2297 thread__find_symbol(thread, sample->cpumode, sample->addr, &al); 2298 2299 if (al.map) 2300 map_type = 'x'; 2301 else 2302 map_type = '?'; 2303 } 2304 2305 print_location(trace->output, sample, &al, true, false); 2306 2307 fprintf(trace->output, " (%c%c)\n", map_type, al.level); 2308 2309 if (callchain_ret > 0) 2310 trace__fprintf_callchain(trace, sample); 2311 else if (callchain_ret < 0) 2312 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel)); 2313 2314 ++trace->nr_events_printed; 2315 out: 2316 err = 0; 2317 out_put: 2318 thread__put(thread); 2319 return err; 2320 } 2321 2322 static void trace__set_base_time(struct trace *trace, 2323 struct perf_evsel *evsel, 2324 struct perf_sample *sample) 2325 { 2326 /* 2327 * BPF events were not setting PERF_SAMPLE_TIME, so be more robust 2328 * and don't use sample->time unconditionally, we may end up having 2329 * some other event in the future without PERF_SAMPLE_TIME for good 2330 * reason, i.e. we may not be interested in its timestamps, just in 2331 * it taking place, picking some piece of information when it 2332 * appears in our event stream (vfs_getname comes to mind). 2333 */ 2334 if (trace->base_time == 0 && !trace->full_time && 2335 (evsel->attr.sample_type & PERF_SAMPLE_TIME)) 2336 trace->base_time = sample->time; 2337 } 2338 2339 static int trace__process_sample(struct perf_tool *tool, 2340 union perf_event *event, 2341 struct perf_sample *sample, 2342 struct perf_evsel *evsel, 2343 struct machine *machine __maybe_unused) 2344 { 2345 struct trace *trace = container_of(tool, struct trace, tool); 2346 struct thread *thread; 2347 int err = 0; 2348 2349 tracepoint_handler handler = evsel->handler; 2350 2351 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid); 2352 if (thread && thread__is_filtered(thread)) 2353 goto out; 2354 2355 trace__set_base_time(trace, evsel, sample); 2356 2357 if (handler) { 2358 ++trace->nr_events; 2359 handler(trace, evsel, event, sample); 2360 } 2361 out: 2362 thread__put(thread); 2363 return err; 2364 } 2365 2366 static int trace__record(struct trace *trace, int argc, const char **argv) 2367 { 2368 unsigned int rec_argc, i, j; 2369 const char **rec_argv; 2370 const char * const record_args[] = { 2371 "record", 2372 "-R", 2373 "-m", "1024", 2374 "-c", "1", 2375 }; 2376 2377 const char * const sc_args[] = { "-e", }; 2378 unsigned int sc_args_nr = ARRAY_SIZE(sc_args); 2379 const char * const majpf_args[] = { "-e", "major-faults" }; 2380 unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args); 2381 const char * const minpf_args[] = { "-e", "minor-faults" }; 2382 unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args); 2383 2384 /* +1 is for the event string below */ 2385 rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 1 + 2386 majpf_args_nr + minpf_args_nr + argc; 2387 rec_argv = calloc(rec_argc + 1, sizeof(char *)); 2388 2389 if (rec_argv == NULL) 2390 return -ENOMEM; 2391 2392 j = 0; 2393 for (i = 0; i < ARRAY_SIZE(record_args); i++) 2394 rec_argv[j++] = record_args[i]; 2395 2396 if (trace->trace_syscalls) { 2397 for (i = 0; i < sc_args_nr; i++) 2398 rec_argv[j++] = sc_args[i]; 2399 2400 /* event string may be different for older kernels - e.g., RHEL6 */ 2401 if (is_valid_tracepoint("raw_syscalls:sys_enter")) 2402 rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit"; 2403 else if (is_valid_tracepoint("syscalls:sys_enter")) 2404 rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit"; 2405 else { 2406 pr_err("Neither raw_syscalls nor syscalls events exist.\n"); 2407 free(rec_argv); 2408 return -1; 2409 } 2410 } 2411 2412 if (trace->trace_pgfaults & TRACE_PFMAJ) 2413 for (i = 0; i < majpf_args_nr; i++) 2414 rec_argv[j++] = majpf_args[i]; 2415 2416 if (trace->trace_pgfaults & TRACE_PFMIN) 2417 for (i = 0; i < minpf_args_nr; i++) 2418 rec_argv[j++] = minpf_args[i]; 2419 2420 for (i = 0; i < (unsigned int)argc; i++) 2421 rec_argv[j++] = argv[i]; 2422 2423 return cmd_record(j, rec_argv); 2424 } 2425 2426 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp); 2427 2428 static bool perf_evlist__add_vfs_getname(struct perf_evlist *evlist) 2429 { 2430 struct perf_evsel *evsel = perf_evsel__newtp("probe", "vfs_getname"); 2431 2432 if (IS_ERR(evsel)) 2433 return false; 2434 2435 if (perf_evsel__field(evsel, "pathname") == NULL) { 2436 perf_evsel__delete(evsel); 2437 return false; 2438 } 2439 2440 evsel->handler = trace__vfs_getname; 2441 perf_evlist__add(evlist, evsel); 2442 return true; 2443 } 2444 2445 static struct perf_evsel *perf_evsel__new_pgfault(u64 config) 2446 { 2447 struct perf_evsel *evsel; 2448 struct perf_event_attr attr = { 2449 .type = PERF_TYPE_SOFTWARE, 2450 .mmap_data = 1, 2451 }; 2452 2453 attr.config = config; 2454 attr.sample_period = 1; 2455 2456 event_attr_init(&attr); 2457 2458 evsel = perf_evsel__new(&attr); 2459 if (evsel) 2460 evsel->handler = trace__pgfault; 2461 2462 return evsel; 2463 } 2464 2465 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample) 2466 { 2467 const u32 type = event->header.type; 2468 struct perf_evsel *evsel; 2469 2470 if (type != PERF_RECORD_SAMPLE) { 2471 trace__process_event(trace, trace->host, event, sample); 2472 return; 2473 } 2474 2475 evsel = perf_evlist__id2evsel(trace->evlist, sample->id); 2476 if (evsel == NULL) { 2477 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id); 2478 return; 2479 } 2480 2481 trace__set_base_time(trace, evsel, sample); 2482 2483 if (evsel->attr.type == PERF_TYPE_TRACEPOINT && 2484 sample->raw_data == NULL) { 2485 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n", 2486 perf_evsel__name(evsel), sample->tid, 2487 sample->cpu, sample->raw_size); 2488 } else { 2489 tracepoint_handler handler = evsel->handler; 2490 handler(trace, evsel, event, sample); 2491 } 2492 2493 if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX) 2494 interrupted = true; 2495 } 2496 2497 static int trace__add_syscall_newtp(struct trace *trace) 2498 { 2499 int ret = -1; 2500 struct perf_evlist *evlist = trace->evlist; 2501 struct perf_evsel *sys_enter, *sys_exit; 2502 2503 sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter); 2504 if (sys_enter == NULL) 2505 goto out; 2506 2507 if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args)) 2508 goto out_delete_sys_enter; 2509 2510 sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit); 2511 if (sys_exit == NULL) 2512 goto out_delete_sys_enter; 2513 2514 if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret)) 2515 goto out_delete_sys_exit; 2516 2517 perf_evsel__config_callchain(sys_enter, &trace->opts, &callchain_param); 2518 perf_evsel__config_callchain(sys_exit, &trace->opts, &callchain_param); 2519 2520 perf_evlist__add(evlist, sys_enter); 2521 perf_evlist__add(evlist, sys_exit); 2522 2523 if (callchain_param.enabled && !trace->kernel_syscallchains) { 2524 /* 2525 * We're interested only in the user space callchain 2526 * leading to the syscall, allow overriding that for 2527 * debugging reasons using --kernel_syscall_callchains 2528 */ 2529 sys_exit->attr.exclude_callchain_kernel = 1; 2530 } 2531 2532 trace->syscalls.events.sys_enter = sys_enter; 2533 trace->syscalls.events.sys_exit = sys_exit; 2534 2535 ret = 0; 2536 out: 2537 return ret; 2538 2539 out_delete_sys_exit: 2540 perf_evsel__delete_priv(sys_exit); 2541 out_delete_sys_enter: 2542 perf_evsel__delete_priv(sys_enter); 2543 goto out; 2544 } 2545 2546 static int trace__set_ev_qualifier_filter(struct trace *trace) 2547 { 2548 int err = -1; 2549 struct perf_evsel *sys_exit; 2550 char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier, 2551 trace->ev_qualifier_ids.nr, 2552 trace->ev_qualifier_ids.entries); 2553 2554 if (filter == NULL) 2555 goto out_enomem; 2556 2557 if (!perf_evsel__append_tp_filter(trace->syscalls.events.sys_enter, 2558 filter)) { 2559 sys_exit = trace->syscalls.events.sys_exit; 2560 err = perf_evsel__append_tp_filter(sys_exit, filter); 2561 } 2562 2563 free(filter); 2564 out: 2565 return err; 2566 out_enomem: 2567 errno = ENOMEM; 2568 goto out; 2569 } 2570 2571 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused, 2572 size_t npids __maybe_unused, pid_t *pids __maybe_unused) 2573 { 2574 int err = 0; 2575 #ifdef HAVE_LIBBPF_SUPPORT 2576 bool value = true; 2577 int map_fd = bpf_map__fd(map); 2578 size_t i; 2579 2580 for (i = 0; i < npids; ++i) { 2581 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY); 2582 if (err) 2583 break; 2584 } 2585 #endif 2586 return err; 2587 } 2588 2589 static int trace__set_filter_loop_pids(struct trace *trace) 2590 { 2591 unsigned int nr = 1, err; 2592 pid_t pids[32] = { 2593 getpid(), 2594 }; 2595 struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]); 2596 2597 while (thread && nr < ARRAY_SIZE(pids)) { 2598 struct thread *parent = machine__find_thread(trace->host, thread->ppid, thread->ppid); 2599 2600 if (parent == NULL) 2601 break; 2602 2603 if (!strcmp(thread__comm_str(parent), "sshd")) { 2604 pids[nr++] = parent->tid; 2605 break; 2606 } 2607 thread = parent; 2608 } 2609 2610 err = perf_evlist__set_tp_filter_pids(trace->evlist, nr, pids); 2611 if (!err && trace->filter_pids.map) 2612 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids); 2613 2614 return err; 2615 } 2616 2617 static int trace__set_filter_pids(struct trace *trace) 2618 { 2619 int err = 0; 2620 /* 2621 * Better not use !target__has_task() here because we need to cover the 2622 * case where no threads were specified in the command line, but a 2623 * workload was, and in that case we will fill in the thread_map when 2624 * we fork the workload in perf_evlist__prepare_workload. 2625 */ 2626 if (trace->filter_pids.nr > 0) { 2627 err = perf_evlist__set_tp_filter_pids(trace->evlist, trace->filter_pids.nr, 2628 trace->filter_pids.entries); 2629 if (!err && trace->filter_pids.map) { 2630 err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr, 2631 trace->filter_pids.entries); 2632 } 2633 } else if (thread_map__pid(trace->evlist->threads, 0) == -1) { 2634 err = trace__set_filter_loop_pids(trace); 2635 } 2636 2637 return err; 2638 } 2639 2640 static int trace__run(struct trace *trace, int argc, const char **argv) 2641 { 2642 struct perf_evlist *evlist = trace->evlist; 2643 struct perf_evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL; 2644 int err = -1, i; 2645 unsigned long before; 2646 const bool forks = argc > 0; 2647 bool draining = false; 2648 2649 trace->live = true; 2650 2651 if (trace->trace_syscalls && trace__add_syscall_newtp(trace)) 2652 goto out_error_raw_syscalls; 2653 2654 if (trace->trace_syscalls) 2655 trace->vfs_getname = perf_evlist__add_vfs_getname(evlist); 2656 2657 if ((trace->trace_pgfaults & TRACE_PFMAJ)) { 2658 pgfault_maj = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ); 2659 if (pgfault_maj == NULL) 2660 goto out_error_mem; 2661 perf_evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param); 2662 perf_evlist__add(evlist, pgfault_maj); 2663 } 2664 2665 if ((trace->trace_pgfaults & TRACE_PFMIN)) { 2666 pgfault_min = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN); 2667 if (pgfault_min == NULL) 2668 goto out_error_mem; 2669 perf_evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param); 2670 perf_evlist__add(evlist, pgfault_min); 2671 } 2672 2673 if (trace->sched && 2674 perf_evlist__add_newtp(evlist, "sched", "sched_stat_runtime", 2675 trace__sched_stat_runtime)) 2676 goto out_error_sched_stat_runtime; 2677 2678 /* 2679 * If a global cgroup was set, apply it to all the events without an 2680 * explicit cgroup. I.e.: 2681 * 2682 * trace -G A -e sched:*switch 2683 * 2684 * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc 2685 * _and_ sched:sched_switch to the 'A' cgroup, while: 2686 * 2687 * trace -e sched:*switch -G A 2688 * 2689 * will only set the sched:sched_switch event to the 'A' cgroup, all the 2690 * other events (raw_syscalls:sys_{enter,exit}, etc are left "without" 2691 * a cgroup (on the root cgroup, sys wide, etc). 2692 * 2693 * Multiple cgroups: 2694 * 2695 * trace -G A -e sched:*switch -G B 2696 * 2697 * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes 2698 * to the 'B' cgroup. 2699 * 2700 * evlist__set_default_cgroup() grabs a reference of the passed cgroup 2701 * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL. 2702 */ 2703 if (trace->cgroup) 2704 evlist__set_default_cgroup(trace->evlist, trace->cgroup); 2705 2706 err = perf_evlist__create_maps(evlist, &trace->opts.target); 2707 if (err < 0) { 2708 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n"); 2709 goto out_delete_evlist; 2710 } 2711 2712 err = trace__symbols_init(trace, evlist); 2713 if (err < 0) { 2714 fprintf(trace->output, "Problems initializing symbol libraries!\n"); 2715 goto out_delete_evlist; 2716 } 2717 2718 perf_evlist__config(evlist, &trace->opts, &callchain_param); 2719 2720 signal(SIGCHLD, sig_handler); 2721 signal(SIGINT, sig_handler); 2722 2723 if (forks) { 2724 err = perf_evlist__prepare_workload(evlist, &trace->opts.target, 2725 argv, false, NULL); 2726 if (err < 0) { 2727 fprintf(trace->output, "Couldn't run the workload!\n"); 2728 goto out_delete_evlist; 2729 } 2730 } 2731 2732 err = perf_evlist__open(evlist); 2733 if (err < 0) 2734 goto out_error_open; 2735 2736 err = bpf__apply_obj_config(); 2737 if (err) { 2738 char errbuf[BUFSIZ]; 2739 2740 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf)); 2741 pr_err("ERROR: Apply config to BPF failed: %s\n", 2742 errbuf); 2743 goto out_error_open; 2744 } 2745 2746 err = trace__set_filter_pids(trace); 2747 if (err < 0) 2748 goto out_error_mem; 2749 2750 if (trace->ev_qualifier_ids.nr > 0) { 2751 err = trace__set_ev_qualifier_filter(trace); 2752 if (err < 0) 2753 goto out_errno; 2754 2755 pr_debug("event qualifier tracepoint filter: %s\n", 2756 trace->syscalls.events.sys_exit->filter); 2757 } 2758 2759 err = perf_evlist__apply_filters(evlist, &evsel); 2760 if (err < 0) 2761 goto out_error_apply_filters; 2762 2763 err = perf_evlist__mmap(evlist, trace->opts.mmap_pages); 2764 if (err < 0) 2765 goto out_error_mmap; 2766 2767 if (!target__none(&trace->opts.target) && !trace->opts.initial_delay) 2768 perf_evlist__enable(evlist); 2769 2770 if (forks) 2771 perf_evlist__start_workload(evlist); 2772 2773 if (trace->opts.initial_delay) { 2774 usleep(trace->opts.initial_delay * 1000); 2775 perf_evlist__enable(evlist); 2776 } 2777 2778 trace->multiple_threads = thread_map__pid(evlist->threads, 0) == -1 || 2779 evlist->threads->nr > 1 || 2780 perf_evlist__first(evlist)->attr.inherit; 2781 2782 /* 2783 * Now that we already used evsel->attr to ask the kernel to setup the 2784 * events, lets reuse evsel->attr.sample_max_stack as the limit in 2785 * trace__resolve_callchain(), allowing per-event max-stack settings 2786 * to override an explicitly set --max-stack global setting. 2787 */ 2788 evlist__for_each_entry(evlist, evsel) { 2789 if (evsel__has_callchain(evsel) && 2790 evsel->attr.sample_max_stack == 0) 2791 evsel->attr.sample_max_stack = trace->max_stack; 2792 } 2793 again: 2794 before = trace->nr_events; 2795 2796 for (i = 0; i < evlist->nr_mmaps; i++) { 2797 union perf_event *event; 2798 struct perf_mmap *md; 2799 2800 md = &evlist->mmap[i]; 2801 if (perf_mmap__read_init(md) < 0) 2802 continue; 2803 2804 while ((event = perf_mmap__read_event(md)) != NULL) { 2805 struct perf_sample sample; 2806 2807 ++trace->nr_events; 2808 2809 err = perf_evlist__parse_sample(evlist, event, &sample); 2810 if (err) { 2811 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err); 2812 goto next_event; 2813 } 2814 2815 trace__handle_event(trace, event, &sample); 2816 next_event: 2817 perf_mmap__consume(md); 2818 2819 if (interrupted) 2820 goto out_disable; 2821 2822 if (done && !draining) { 2823 perf_evlist__disable(evlist); 2824 draining = true; 2825 } 2826 } 2827 perf_mmap__read_done(md); 2828 } 2829 2830 if (trace->nr_events == before) { 2831 int timeout = done ? 100 : -1; 2832 2833 if (!draining && perf_evlist__poll(evlist, timeout) > 0) { 2834 if (perf_evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0) 2835 draining = true; 2836 2837 goto again; 2838 } 2839 } else { 2840 goto again; 2841 } 2842 2843 out_disable: 2844 thread__zput(trace->current); 2845 2846 perf_evlist__disable(evlist); 2847 2848 if (!err) { 2849 if (trace->summary) 2850 trace__fprintf_thread_summary(trace, trace->output); 2851 2852 if (trace->show_tool_stats) { 2853 fprintf(trace->output, "Stats:\n " 2854 " vfs_getname : %" PRIu64 "\n" 2855 " proc_getname: %" PRIu64 "\n", 2856 trace->stats.vfs_getname, 2857 trace->stats.proc_getname); 2858 } 2859 } 2860 2861 out_delete_evlist: 2862 trace__symbols__exit(trace); 2863 2864 perf_evlist__delete(evlist); 2865 cgroup__put(trace->cgroup); 2866 trace->evlist = NULL; 2867 trace->live = false; 2868 return err; 2869 { 2870 char errbuf[BUFSIZ]; 2871 2872 out_error_sched_stat_runtime: 2873 tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime"); 2874 goto out_error; 2875 2876 out_error_raw_syscalls: 2877 tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)"); 2878 goto out_error; 2879 2880 out_error_mmap: 2881 perf_evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf)); 2882 goto out_error; 2883 2884 out_error_open: 2885 perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf)); 2886 2887 out_error: 2888 fprintf(trace->output, "%s\n", errbuf); 2889 goto out_delete_evlist; 2890 2891 out_error_apply_filters: 2892 fprintf(trace->output, 2893 "Failed to set filter \"%s\" on event %s with %d (%s)\n", 2894 evsel->filter, perf_evsel__name(evsel), errno, 2895 str_error_r(errno, errbuf, sizeof(errbuf))); 2896 goto out_delete_evlist; 2897 } 2898 out_error_mem: 2899 fprintf(trace->output, "Not enough memory to run!\n"); 2900 goto out_delete_evlist; 2901 2902 out_errno: 2903 fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno)); 2904 goto out_delete_evlist; 2905 } 2906 2907 static int trace__replay(struct trace *trace) 2908 { 2909 const struct perf_evsel_str_handler handlers[] = { 2910 { "probe:vfs_getname", trace__vfs_getname, }, 2911 }; 2912 struct perf_data data = { 2913 .file = { 2914 .path = input_name, 2915 }, 2916 .mode = PERF_DATA_MODE_READ, 2917 .force = trace->force, 2918 }; 2919 struct perf_session *session; 2920 struct perf_evsel *evsel; 2921 int err = -1; 2922 2923 trace->tool.sample = trace__process_sample; 2924 trace->tool.mmap = perf_event__process_mmap; 2925 trace->tool.mmap2 = perf_event__process_mmap2; 2926 trace->tool.comm = perf_event__process_comm; 2927 trace->tool.exit = perf_event__process_exit; 2928 trace->tool.fork = perf_event__process_fork; 2929 trace->tool.attr = perf_event__process_attr; 2930 trace->tool.tracing_data = perf_event__process_tracing_data; 2931 trace->tool.build_id = perf_event__process_build_id; 2932 trace->tool.namespaces = perf_event__process_namespaces; 2933 2934 trace->tool.ordered_events = true; 2935 trace->tool.ordering_requires_timestamps = true; 2936 2937 /* add tid to output */ 2938 trace->multiple_threads = true; 2939 2940 session = perf_session__new(&data, false, &trace->tool); 2941 if (session == NULL) 2942 return -1; 2943 2944 if (trace->opts.target.pid) 2945 symbol_conf.pid_list_str = strdup(trace->opts.target.pid); 2946 2947 if (trace->opts.target.tid) 2948 symbol_conf.tid_list_str = strdup(trace->opts.target.tid); 2949 2950 if (symbol__init(&session->header.env) < 0) 2951 goto out; 2952 2953 trace->host = &session->machines.host; 2954 2955 err = perf_session__set_tracepoints_handlers(session, handlers); 2956 if (err) 2957 goto out; 2958 2959 evsel = perf_evlist__find_tracepoint_by_name(session->evlist, 2960 "raw_syscalls:sys_enter"); 2961 /* older kernels have syscalls tp versus raw_syscalls */ 2962 if (evsel == NULL) 2963 evsel = perf_evlist__find_tracepoint_by_name(session->evlist, 2964 "syscalls:sys_enter"); 2965 2966 if (evsel && 2967 (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 || 2968 perf_evsel__init_sc_tp_ptr_field(evsel, args))) { 2969 pr_err("Error during initialize raw_syscalls:sys_enter event\n"); 2970 goto out; 2971 } 2972 2973 evsel = perf_evlist__find_tracepoint_by_name(session->evlist, 2974 "raw_syscalls:sys_exit"); 2975 if (evsel == NULL) 2976 evsel = perf_evlist__find_tracepoint_by_name(session->evlist, 2977 "syscalls:sys_exit"); 2978 if (evsel && 2979 (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 || 2980 perf_evsel__init_sc_tp_uint_field(evsel, ret))) { 2981 pr_err("Error during initialize raw_syscalls:sys_exit event\n"); 2982 goto out; 2983 } 2984 2985 evlist__for_each_entry(session->evlist, evsel) { 2986 if (evsel->attr.type == PERF_TYPE_SOFTWARE && 2987 (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ || 2988 evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN || 2989 evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS)) 2990 evsel->handler = trace__pgfault; 2991 } 2992 2993 setup_pager(); 2994 2995 err = perf_session__process_events(session); 2996 if (err) 2997 pr_err("Failed to process events, error %d", err); 2998 2999 else if (trace->summary) 3000 trace__fprintf_thread_summary(trace, trace->output); 3001 3002 out: 3003 perf_session__delete(session); 3004 3005 return err; 3006 } 3007 3008 static size_t trace__fprintf_threads_header(FILE *fp) 3009 { 3010 size_t printed; 3011 3012 printed = fprintf(fp, "\n Summary of events:\n\n"); 3013 3014 return printed; 3015 } 3016 3017 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs, 3018 struct stats *stats; 3019 double msecs; 3020 int syscall; 3021 ) 3022 { 3023 struct int_node *source = rb_entry(nd, struct int_node, rb_node); 3024 struct stats *stats = source->priv; 3025 3026 entry->syscall = source->i; 3027 entry->stats = stats; 3028 entry->msecs = stats ? (u64)stats->n * (avg_stats(stats) / NSEC_PER_MSEC) : 0; 3029 } 3030 3031 static size_t thread__dump_stats(struct thread_trace *ttrace, 3032 struct trace *trace, FILE *fp) 3033 { 3034 size_t printed = 0; 3035 struct syscall *sc; 3036 struct rb_node *nd; 3037 DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats); 3038 3039 if (syscall_stats == NULL) 3040 return 0; 3041 3042 printed += fprintf(fp, "\n"); 3043 3044 printed += fprintf(fp, " syscall calls total min avg max stddev\n"); 3045 printed += fprintf(fp, " (msec) (msec) (msec) (msec) (%%)\n"); 3046 printed += fprintf(fp, " --------------- -------- --------- --------- --------- --------- ------\n"); 3047 3048 resort_rb__for_each_entry(nd, syscall_stats) { 3049 struct stats *stats = syscall_stats_entry->stats; 3050 if (stats) { 3051 double min = (double)(stats->min) / NSEC_PER_MSEC; 3052 double max = (double)(stats->max) / NSEC_PER_MSEC; 3053 double avg = avg_stats(stats); 3054 double pct; 3055 u64 n = (u64) stats->n; 3056 3057 pct = avg ? 100.0 * stddev_stats(stats)/avg : 0.0; 3058 avg /= NSEC_PER_MSEC; 3059 3060 sc = &trace->syscalls.table[syscall_stats_entry->syscall]; 3061 printed += fprintf(fp, " %-15s", sc->name); 3062 printed += fprintf(fp, " %8" PRIu64 " %9.3f %9.3f %9.3f", 3063 n, syscall_stats_entry->msecs, min, avg); 3064 printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct); 3065 } 3066 } 3067 3068 resort_rb__delete(syscall_stats); 3069 printed += fprintf(fp, "\n\n"); 3070 3071 return printed; 3072 } 3073 3074 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace) 3075 { 3076 size_t printed = 0; 3077 struct thread_trace *ttrace = thread__priv(thread); 3078 double ratio; 3079 3080 if (ttrace == NULL) 3081 return 0; 3082 3083 ratio = (double)ttrace->nr_events / trace->nr_events * 100.0; 3084 3085 printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid); 3086 printed += fprintf(fp, "%lu events, ", ttrace->nr_events); 3087 printed += fprintf(fp, "%.1f%%", ratio); 3088 if (ttrace->pfmaj) 3089 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj); 3090 if (ttrace->pfmin) 3091 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin); 3092 if (trace->sched) 3093 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms); 3094 else if (fputc('\n', fp) != EOF) 3095 ++printed; 3096 3097 printed += thread__dump_stats(ttrace, trace, fp); 3098 3099 return printed; 3100 } 3101 3102 static unsigned long thread__nr_events(struct thread_trace *ttrace) 3103 { 3104 return ttrace ? ttrace->nr_events : 0; 3105 } 3106 3107 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)), 3108 struct thread *thread; 3109 ) 3110 { 3111 entry->thread = rb_entry(nd, struct thread, rb_node); 3112 } 3113 3114 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp) 3115 { 3116 size_t printed = trace__fprintf_threads_header(fp); 3117 struct rb_node *nd; 3118 int i; 3119 3120 for (i = 0; i < THREADS__TABLE_SIZE; i++) { 3121 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i); 3122 3123 if (threads == NULL) { 3124 fprintf(fp, "%s", "Error sorting output by nr_events!\n"); 3125 return 0; 3126 } 3127 3128 resort_rb__for_each_entry(nd, threads) 3129 printed += trace__fprintf_thread(fp, threads_entry->thread, trace); 3130 3131 resort_rb__delete(threads); 3132 } 3133 return printed; 3134 } 3135 3136 static int trace__set_duration(const struct option *opt, const char *str, 3137 int unset __maybe_unused) 3138 { 3139 struct trace *trace = opt->value; 3140 3141 trace->duration_filter = atof(str); 3142 return 0; 3143 } 3144 3145 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str, 3146 int unset __maybe_unused) 3147 { 3148 int ret = -1; 3149 size_t i; 3150 struct trace *trace = opt->value; 3151 /* 3152 * FIXME: introduce a intarray class, plain parse csv and create a 3153 * { int nr, int entries[] } struct... 3154 */ 3155 struct intlist *list = intlist__new(str); 3156 3157 if (list == NULL) 3158 return -1; 3159 3160 i = trace->filter_pids.nr = intlist__nr_entries(list) + 1; 3161 trace->filter_pids.entries = calloc(i, sizeof(pid_t)); 3162 3163 if (trace->filter_pids.entries == NULL) 3164 goto out; 3165 3166 trace->filter_pids.entries[0] = getpid(); 3167 3168 for (i = 1; i < trace->filter_pids.nr; ++i) 3169 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i; 3170 3171 intlist__delete(list); 3172 ret = 0; 3173 out: 3174 return ret; 3175 } 3176 3177 static int trace__open_output(struct trace *trace, const char *filename) 3178 { 3179 struct stat st; 3180 3181 if (!stat(filename, &st) && st.st_size) { 3182 char oldname[PATH_MAX]; 3183 3184 scnprintf(oldname, sizeof(oldname), "%s.old", filename); 3185 unlink(oldname); 3186 rename(filename, oldname); 3187 } 3188 3189 trace->output = fopen(filename, "w"); 3190 3191 return trace->output == NULL ? -errno : 0; 3192 } 3193 3194 static int parse_pagefaults(const struct option *opt, const char *str, 3195 int unset __maybe_unused) 3196 { 3197 int *trace_pgfaults = opt->value; 3198 3199 if (strcmp(str, "all") == 0) 3200 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN; 3201 else if (strcmp(str, "maj") == 0) 3202 *trace_pgfaults |= TRACE_PFMAJ; 3203 else if (strcmp(str, "min") == 0) 3204 *trace_pgfaults |= TRACE_PFMIN; 3205 else 3206 return -1; 3207 3208 return 0; 3209 } 3210 3211 static void evlist__set_evsel_handler(struct perf_evlist *evlist, void *handler) 3212 { 3213 struct perf_evsel *evsel; 3214 3215 evlist__for_each_entry(evlist, evsel) 3216 evsel->handler = handler; 3217 } 3218 3219 static int evlist__set_syscall_tp_fields(struct perf_evlist *evlist) 3220 { 3221 struct perf_evsel *evsel; 3222 3223 evlist__for_each_entry(evlist, evsel) { 3224 if (evsel->priv || !evsel->tp_format) 3225 continue; 3226 3227 if (strcmp(evsel->tp_format->system, "syscalls")) 3228 continue; 3229 3230 if (perf_evsel__init_syscall_tp(evsel)) 3231 return -1; 3232 3233 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) { 3234 struct syscall_tp *sc = evsel->priv; 3235 3236 if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64))) 3237 return -1; 3238 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) { 3239 struct syscall_tp *sc = evsel->priv; 3240 3241 if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap)) 3242 return -1; 3243 } 3244 } 3245 3246 return 0; 3247 } 3248 3249 /* 3250 * XXX: Hackish, just splitting the combined -e+--event (syscalls 3251 * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use 3252 * existing facilities unchanged (trace->ev_qualifier + parse_options()). 3253 * 3254 * It'd be better to introduce a parse_options() variant that would return a 3255 * list with the terms it didn't match to an event... 3256 */ 3257 static int trace__parse_events_option(const struct option *opt, const char *str, 3258 int unset __maybe_unused) 3259 { 3260 struct trace *trace = (struct trace *)opt->value; 3261 const char *s = str; 3262 char *sep = NULL, *lists[2] = { NULL, NULL, }; 3263 int len = strlen(str) + 1, err = -1, list, idx; 3264 char *strace_groups_dir = system_path(STRACE_GROUPS_DIR); 3265 char group_name[PATH_MAX]; 3266 struct syscall_fmt *fmt; 3267 3268 if (strace_groups_dir == NULL) 3269 return -1; 3270 3271 if (*s == '!') { 3272 ++s; 3273 trace->not_ev_qualifier = true; 3274 } 3275 3276 while (1) { 3277 if ((sep = strchr(s, ',')) != NULL) 3278 *sep = '\0'; 3279 3280 list = 0; 3281 if (syscalltbl__id(trace->sctbl, s) >= 0 || 3282 syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) { 3283 list = 1; 3284 goto do_concat; 3285 } 3286 3287 fmt = syscall_fmt__find_by_alias(s); 3288 if (fmt != NULL) { 3289 list = 1; 3290 s = fmt->name; 3291 } else { 3292 path__join(group_name, sizeof(group_name), strace_groups_dir, s); 3293 if (access(group_name, R_OK) == 0) 3294 list = 1; 3295 } 3296 do_concat: 3297 if (lists[list]) { 3298 sprintf(lists[list] + strlen(lists[list]), ",%s", s); 3299 } else { 3300 lists[list] = malloc(len); 3301 if (lists[list] == NULL) 3302 goto out; 3303 strcpy(lists[list], s); 3304 } 3305 3306 if (!sep) 3307 break; 3308 3309 *sep = ','; 3310 s = sep + 1; 3311 } 3312 3313 if (lists[1] != NULL) { 3314 struct strlist_config slist_config = { 3315 .dirname = strace_groups_dir, 3316 }; 3317 3318 trace->ev_qualifier = strlist__new(lists[1], &slist_config); 3319 if (trace->ev_qualifier == NULL) { 3320 fputs("Not enough memory to parse event qualifier", trace->output); 3321 goto out; 3322 } 3323 3324 if (trace__validate_ev_qualifier(trace)) 3325 goto out; 3326 trace->trace_syscalls = true; 3327 } 3328 3329 err = 0; 3330 3331 if (lists[0]) { 3332 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event", 3333 "event selector. use 'perf list' to list available events", 3334 parse_events_option); 3335 err = parse_events_option(&o, lists[0], 0); 3336 } 3337 out: 3338 if (sep) 3339 *sep = ','; 3340 3341 return err; 3342 } 3343 3344 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset) 3345 { 3346 struct trace *trace = opt->value; 3347 3348 if (!list_empty(&trace->evlist->entries)) 3349 return parse_cgroups(opt, str, unset); 3350 3351 trace->cgroup = evlist__findnew_cgroup(trace->evlist, str); 3352 3353 return 0; 3354 } 3355 3356 static struct bpf_map *bpf__find_map_by_name(const char *name) 3357 { 3358 struct bpf_object *obj, *tmp; 3359 3360 bpf_object__for_each_safe(obj, tmp) { 3361 struct bpf_map *map = bpf_object__find_map_by_name(obj, name); 3362 if (map) 3363 return map; 3364 3365 } 3366 3367 return NULL; 3368 } 3369 3370 static void trace__set_bpf_map_filtered_pids(struct trace *trace) 3371 { 3372 trace->filter_pids.map = bpf__find_map_by_name("pids_filtered"); 3373 } 3374 3375 int cmd_trace(int argc, const char **argv) 3376 { 3377 const char *trace_usage[] = { 3378 "perf trace [<options>] [<command>]", 3379 "perf trace [<options>] -- <command> [<options>]", 3380 "perf trace record [<options>] [<command>]", 3381 "perf trace record [<options>] -- <command> [<options>]", 3382 NULL 3383 }; 3384 struct trace trace = { 3385 .syscalls = { 3386 . max = -1, 3387 }, 3388 .opts = { 3389 .target = { 3390 .uid = UINT_MAX, 3391 .uses_mmap = true, 3392 }, 3393 .user_freq = UINT_MAX, 3394 .user_interval = ULLONG_MAX, 3395 .no_buffering = true, 3396 .mmap_pages = UINT_MAX, 3397 }, 3398 .output = stderr, 3399 .show_comm = true, 3400 .trace_syscalls = false, 3401 .kernel_syscallchains = false, 3402 .max_stack = UINT_MAX, 3403 .max_events = ULONG_MAX, 3404 }; 3405 const char *output_name = NULL; 3406 const struct option trace_options[] = { 3407 OPT_CALLBACK('e', "event", &trace, "event", 3408 "event/syscall selector. use 'perf list' to list available events", 3409 trace__parse_events_option), 3410 OPT_BOOLEAN(0, "comm", &trace.show_comm, 3411 "show the thread COMM next to its id"), 3412 OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"), 3413 OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace", 3414 trace__parse_events_option), 3415 OPT_STRING('o', "output", &output_name, "file", "output file name"), 3416 OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"), 3417 OPT_STRING('p', "pid", &trace.opts.target.pid, "pid", 3418 "trace events on existing process id"), 3419 OPT_STRING('t', "tid", &trace.opts.target.tid, "tid", 3420 "trace events on existing thread id"), 3421 OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids", 3422 "pids to filter (by the kernel)", trace__set_filter_pids_from_option), 3423 OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide, 3424 "system-wide collection from all CPUs"), 3425 OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu", 3426 "list of cpus to monitor"), 3427 OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit, 3428 "child tasks do not inherit counters"), 3429 OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages", 3430 "number of mmap data pages", 3431 perf_evlist__parse_mmap_pages), 3432 OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user", 3433 "user to profile"), 3434 OPT_CALLBACK(0, "duration", &trace, "float", 3435 "show only events with duration > N.M ms", 3436 trace__set_duration), 3437 OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"), 3438 OPT_INCR('v', "verbose", &verbose, "be more verbose"), 3439 OPT_BOOLEAN('T', "time", &trace.full_time, 3440 "Show full timestamp, not time relative to first start"), 3441 OPT_BOOLEAN(0, "failure", &trace.failure_only, 3442 "Show only syscalls that failed"), 3443 OPT_BOOLEAN('s', "summary", &trace.summary_only, 3444 "Show only syscall summary with statistics"), 3445 OPT_BOOLEAN('S', "with-summary", &trace.summary, 3446 "Show all syscalls and summary with statistics"), 3447 OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min", 3448 "Trace pagefaults", parse_pagefaults, "maj"), 3449 OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"), 3450 OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"), 3451 OPT_CALLBACK(0, "call-graph", &trace.opts, 3452 "record_mode[,record_size]", record_callchain_help, 3453 &record_parse_callchain_opt), 3454 OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains, 3455 "Show the kernel callchains on the syscall exit path"), 3456 OPT_ULONG(0, "max-events", &trace.max_events, 3457 "Set the maximum number of events to print, exit after that is reached. "), 3458 OPT_UINTEGER(0, "min-stack", &trace.min_stack, 3459 "Set the minimum stack depth when parsing the callchain, " 3460 "anything below the specified depth will be ignored."), 3461 OPT_UINTEGER(0, "max-stack", &trace.max_stack, 3462 "Set the maximum stack depth when parsing the callchain, " 3463 "anything beyond the specified depth will be ignored. " 3464 "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)), 3465 OPT_BOOLEAN(0, "print-sample", &trace.print_sample, 3466 "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"), 3467 OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout, 3468 "per thread proc mmap processing timeout in ms"), 3469 OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only", 3470 trace__parse_cgroups), 3471 OPT_UINTEGER('D', "delay", &trace.opts.initial_delay, 3472 "ms to wait before starting measurement after program " 3473 "start"), 3474 OPT_END() 3475 }; 3476 bool __maybe_unused max_stack_user_set = true; 3477 bool mmap_pages_user_set = true; 3478 struct perf_evsel *evsel; 3479 const char * const trace_subcommands[] = { "record", NULL }; 3480 int err = -1; 3481 char bf[BUFSIZ]; 3482 3483 signal(SIGSEGV, sighandler_dump_stack); 3484 signal(SIGFPE, sighandler_dump_stack); 3485 3486 trace.evlist = perf_evlist__new(); 3487 trace.sctbl = syscalltbl__new(); 3488 3489 if (trace.evlist == NULL || trace.sctbl == NULL) { 3490 pr_err("Not enough memory to run!\n"); 3491 err = -ENOMEM; 3492 goto out; 3493 } 3494 3495 argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands, 3496 trace_usage, PARSE_OPT_STOP_AT_NON_OPTION); 3497 3498 if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) { 3499 usage_with_options_msg(trace_usage, trace_options, 3500 "cgroup monitoring only available in system-wide mode"); 3501 } 3502 3503 evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__"); 3504 if (IS_ERR(evsel)) { 3505 bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf)); 3506 pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf); 3507 goto out; 3508 } 3509 3510 if (evsel) { 3511 trace.syscalls.events.augmented = evsel; 3512 trace__set_bpf_map_filtered_pids(&trace); 3513 } 3514 3515 err = bpf__setup_stdout(trace.evlist); 3516 if (err) { 3517 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf)); 3518 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf); 3519 goto out; 3520 } 3521 3522 err = -1; 3523 3524 if (trace.trace_pgfaults) { 3525 trace.opts.sample_address = true; 3526 trace.opts.sample_time = true; 3527 } 3528 3529 if (trace.opts.mmap_pages == UINT_MAX) 3530 mmap_pages_user_set = false; 3531 3532 if (trace.max_stack == UINT_MAX) { 3533 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack(); 3534 max_stack_user_set = false; 3535 } 3536 3537 #ifdef HAVE_DWARF_UNWIND_SUPPORT 3538 if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) { 3539 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false); 3540 } 3541 #endif 3542 3543 if (callchain_param.enabled) { 3544 if (!mmap_pages_user_set && geteuid() == 0) 3545 trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4; 3546 3547 symbol_conf.use_callchain = true; 3548 } 3549 3550 if (trace.evlist->nr_entries > 0) { 3551 evlist__set_evsel_handler(trace.evlist, trace__event_handler); 3552 if (evlist__set_syscall_tp_fields(trace.evlist)) { 3553 perror("failed to set syscalls:* tracepoint fields"); 3554 goto out; 3555 } 3556 } 3557 3558 /* 3559 * If we are augmenting syscalls, then combine what we put in the 3560 * __augmented_syscalls__ BPF map with what is in the 3561 * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF, 3562 * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit. 3563 * 3564 * We'll switch to look at two BPF maps, one for sys_enter and the 3565 * other for sys_exit when we start augmenting the sys_exit paths with 3566 * buffers that are being copied from kernel to userspace, think 'read' 3567 * syscall. 3568 */ 3569 if (trace.syscalls.events.augmented) { 3570 evsel = trace.syscalls.events.augmented; 3571 3572 if (perf_evsel__init_augmented_syscall_tp(evsel) || 3573 perf_evsel__init_augmented_syscall_tp_args(evsel)) 3574 goto out; 3575 evsel->handler = trace__sys_enter; 3576 3577 evlist__for_each_entry(trace.evlist, evsel) { 3578 bool raw_syscalls_sys_exit = strcmp(perf_evsel__name(evsel), "raw_syscalls:sys_exit") == 0; 3579 3580 if (raw_syscalls_sys_exit) { 3581 trace.raw_augmented_syscalls = true; 3582 goto init_augmented_syscall_tp; 3583 } 3584 3585 if (strstarts(perf_evsel__name(evsel), "syscalls:sys_exit_")) { 3586 init_augmented_syscall_tp: 3587 perf_evsel__init_augmented_syscall_tp(evsel); 3588 perf_evsel__init_augmented_syscall_tp_ret(evsel); 3589 evsel->handler = trace__sys_exit; 3590 } 3591 } 3592 } 3593 3594 if ((argc >= 1) && (strcmp(argv[0], "record") == 0)) 3595 return trace__record(&trace, argc-1, &argv[1]); 3596 3597 /* summary_only implies summary option, but don't overwrite summary if set */ 3598 if (trace.summary_only) 3599 trace.summary = trace.summary_only; 3600 3601 if (!trace.trace_syscalls && !trace.trace_pgfaults && 3602 trace.evlist->nr_entries == 0 /* Was --events used? */) { 3603 trace.trace_syscalls = true; 3604 } 3605 3606 if (output_name != NULL) { 3607 err = trace__open_output(&trace, output_name); 3608 if (err < 0) { 3609 perror("failed to create output file"); 3610 goto out; 3611 } 3612 } 3613 3614 err = target__validate(&trace.opts.target); 3615 if (err) { 3616 target__strerror(&trace.opts.target, err, bf, sizeof(bf)); 3617 fprintf(trace.output, "%s", bf); 3618 goto out_close; 3619 } 3620 3621 err = target__parse_uid(&trace.opts.target); 3622 if (err) { 3623 target__strerror(&trace.opts.target, err, bf, sizeof(bf)); 3624 fprintf(trace.output, "%s", bf); 3625 goto out_close; 3626 } 3627 3628 if (!argc && target__none(&trace.opts.target)) 3629 trace.opts.target.system_wide = true; 3630 3631 if (input_name) 3632 err = trace__replay(&trace); 3633 else 3634 err = trace__run(&trace, argc, argv); 3635 3636 out_close: 3637 if (output_name != NULL) 3638 fclose(trace.output); 3639 out: 3640 return err; 3641 } 3642