tools/perf/builtin-record.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * builtin-record.c
   4  *
   5  * Builtin record command: Record the profile of a workload
   6  * (or a CPU, or a PID) into the perf.data output file - for
   7  * later analysis via perf report.
   8  */
   9 #include "builtin.h"
  10
  11 #include "util/build-id.h"
  12 #include <subcmd/parse-options.h>
  13 #include "util/parse-events.h"
  14 #include "util/config.h"
  15
  16 #include "util/callchain.h"
  17 #include "util/cgroup.h"
  18 #include "util/header.h"
  19 #include "util/event.h"
  20 #include "util/evlist.h"
  21 #include "util/evsel.h"
  22 #include "util/debug.h"
  23 #include "util/mmap.h"
  24 #include "util/target.h"
  25 #include "util/session.h"
  26 #include "util/tool.h"
  27 #include "util/symbol.h"
  28 #include "util/record.h"
  29 #include "util/cpumap.h"
  30 #include "util/thread_map.h"
  31 #include "util/data.h"
  32 #include "util/perf_regs.h"
  33 #include "util/auxtrace.h"
  34 #include "util/tsc.h"
  35 #include "util/parse-branch-options.h"
  36 #include "util/parse-regs-options.h"
  37 #include "util/perf_api_probe.h"
  38 #include "util/llvm-utils.h"
  39 #include "util/bpf-loader.h"
  40 #include "util/trigger.h"
  41 #include "util/perf-hooks.h"
  42 #include "util/cpu-set-sched.h"
  43 #include "util/synthetic-events.h"
  44 #include "util/time-utils.h"
  45 #include "util/units.h"
  46 #include "util/bpf-event.h"
  47 #include "util/util.h"
  48 #include "util/pfm.h"
  49 #include "util/clockid.h"
  50 #include "asm/bug.h"
  51 #include "perf.h"
  52
  53 #include <errno.h>
  54 #include <inttypes.h>
  55 #include <locale.h>
  56 #include <poll.h>
  57 #include <pthread.h>
  58 #include <unistd.h>
  59 #include <sched.h>
  60 #include <signal.h>
  61 #ifdef HAVE_EVENTFD_SUPPORT
  62 #include <sys/eventfd.h>
  63 #endif
  64 #include <sys/mman.h>
  65 #include <sys/wait.h>
  66 #include <sys/types.h>
  67 #include <sys/stat.h>
  68 #include <fcntl.h>
  69 #include <linux/err.h>
  70 #include <linux/string.h>
  71 #include <linux/time64.h>
  72 #include <linux/zalloc.h>
  73 #include <linux/bitmap.h>
  74 #include <sys/time.h>
  75
  76 struct switch_output {
  77         bool             enabled;
  78         bool             signal;
  79         unsigned long    size;
  80         unsigned long    time;
  81         const char      *str;
  82         bool             set;
  83         char             **filenames;
  84         int              num_files;
  85         int              cur_file;
  86 };
  87
  88 struct record {
  89         struct perf_tool        tool;
  90         struct record_opts      opts;
  91         u64                     bytes_written;
  92         struct perf_data        data;
  93         struct auxtrace_record  *itr;
  94         struct evlist   *evlist;
  95         struct perf_session     *session;
  96         struct evlist           *sb_evlist;
  97         pthread_t               thread_id;
  98         int                     realtime_prio;
  99         bool                    switch_output_event_set;
 100         bool                    no_buildid;
 101         bool                    no_buildid_set;
 102         bool                    no_buildid_cache;
 103         bool                    no_buildid_cache_set;
 104         bool                    buildid_all;
 105         bool                    timestamp_filename;
 106         bool                    timestamp_boundary;
 107         struct switch_output    switch_output;
 108         unsigned long long      samples;
 109         struct mmap_cpu_mask    affinity_mask;
 110         unsigned long           output_max_size;        /* = 0: unlimited */
 111 };
 112
 113 static volatile int done;
 114
 115 static volatile int auxtrace_record__snapshot_started;
 116 static DEFINE_TRIGGER(auxtrace_snapshot_trigger);
 117 static DEFINE_TRIGGER(switch_output_trigger);
 118
 119 static const char *affinity_tags[PERF_AFFINITY_MAX] = {
 120         "SYS", "NODE", "CPU"
 121 };
 122
 123 static bool switch_output_signal(struct record *rec)
 124 {
 125         return rec->switch_output.signal &&
 126                trigger_is_ready(&switch_output_trigger);
 127 }
 128
 129 static bool switch_output_size(struct record *rec)
 130 {
 131         return rec->switch_output.size &&
 132                trigger_is_ready(&switch_output_trigger) &&
 133                (rec->bytes_written >= rec->switch_output.size);
 134 }
 135
 136 static bool switch_output_time(struct record *rec)
 137 {
 138         return rec->switch_output.time &&
 139                trigger_is_ready(&switch_output_trigger);
 140 }
 141
 142 static bool record__output_max_size_exceeded(struct record *rec)
 143 {
 144         return rec->output_max_size &&
 145                (rec->bytes_written >= rec->output_max_size);
 146 }
 147
 148 static int record__write(struct record *rec, struct mmap *map __maybe_unused,
 149                          void *bf, size_t size)
 150 {
 151         struct perf_data_file *file = &rec->session->data->file;
 152
 153         if (perf_data_file__write(file, bf, size) < 0) {
 154                 pr_err("failed to write perf data, error: %m\n");
 155                 return -1;
 156         }
 157
 158         rec->bytes_written += size;
 159
 160         if (record__output_max_size_exceeded(rec) && !done) {
 161                 fprintf(stderr, "[ perf record: perf size limit reached (%" PRIu64 " KB),"
 162                                 " stopping session ]\n",
 163                                 rec->bytes_written >> 10);
 164                 done = 1;
 165         }
 166
 167         if (switch_output_size(rec))
 168                 trigger_hit(&switch_output_trigger);
 169
 170         return 0;
 171 }
 172
 173 static int record__aio_enabled(struct record *rec);
 174 static int record__comp_enabled(struct record *rec);
 175 static size_t zstd_compress(struct perf_session *session, void *dst, size_t dst_size,
 176                             void *src, size_t src_size);
 177
 178 #ifdef HAVE_AIO_SUPPORT
 179 static int record__aio_write(struct aiocb *cblock, int trace_fd,
 180                 void *buf, size_t size, off_t off)
 181 {
 182         int rc;
 183
 184         cblock->aio_fildes = trace_fd;
 185         cblock->aio_buf    = buf;
 186         cblock->aio_nbytes = size;
 187         cblock->aio_offset = off;
 188         cblock->aio_sigevent.sigev_notify = SIGEV_NONE;
 189
 190         do {
 191                 rc = aio_write(cblock);
 192                 if (rc == 0) {
 193                         break;
 194                 } else if (errno != EAGAIN) {
 195                         cblock->aio_fildes = -1;
 196                         pr_err("failed to queue perf data, error: %m\n");
 197                         break;
 198                 }
 199         } while (1);
 200
 201         return rc;
 202 }
 203
 204 static int record__aio_complete(struct mmap *md, struct aiocb *cblock)
 205 {
 206         void *rem_buf;
 207         off_t rem_off;
 208         size_t rem_size;
 209         int rc, aio_errno;
 210         ssize_t aio_ret, written;
 211
 212         aio_errno = aio_error(cblock);
 213         if (aio_errno == EINPROGRESS)
 214                 return 0;
 215
 216         written = aio_ret = aio_return(cblock);
 217         if (aio_ret < 0) {
 218                 if (aio_errno != EINTR)
 219                         pr_err("failed to write perf data, error: %m\n");
 220                 written = 0;
 221         }
 222
 223         rem_size = cblock->aio_nbytes - written;
 224
 225         if (rem_size == 0) {
 226                 cblock->aio_fildes = -1;
 227                 /*
 228                  * md->refcount is incremented in record__aio_pushfn() for
 229                  * every aio write request started in record__aio_push() so
 230                  * decrement it because the request is now complete.
 231                  */
 232                 perf_mmap__put(&md->core);
 233                 rc = 1;
 234         } else {
 235                 /*
 236                  * aio write request may require restart with the
 237                  * reminder if the kernel didn't write whole
 238                  * chunk at once.
 239                  */
 240                 rem_off = cblock->aio_offset + written;
 241                 rem_buf = (void *)(cblock->aio_buf + written);
 242                 record__aio_write(cblock, cblock->aio_fildes,
 243                                 rem_buf, rem_size, rem_off);
 244                 rc = 0;
 245         }
 246
 247         return rc;
 248 }
 249
 250 static int record__aio_sync(struct mmap *md, bool sync_all)
 251 {
 252         struct aiocb **aiocb = md->aio.aiocb;
 253         struct aiocb *cblocks = md->aio.cblocks;
 254         struct timespec timeout = { 0, 1000 * 1000  * 1 }; /* 1ms */
 255         int i, do_suspend;
 256
 257         do {
 258                 do_suspend = 0;
 259                 for (i = 0; i < md->aio.nr_cblocks; ++i) {
 260                         if (cblocks[i].aio_fildes == -1 || record__aio_complete(md, &cblocks[i])) {
 261                                 if (sync_all)
 262                                         aiocb[i] = NULL;
 263                                 else
 264                                         return i;
 265                         } else {
 266                                 /*
 267                                  * Started aio write is not complete yet
 268                                  * so it has to be waited before the
 269                                  * next allocation.
 270                                  */
 271                                 aiocb[i] = &cblocks[i];
 272                                 do_suspend = 1;
 273                         }
 274                 }
 275                 if (!do_suspend)
 276                         return -1;
 277
 278                 while (aio_suspend((const struct aiocb **)aiocb, md->aio.nr_cblocks, &timeout)) {
 279                         if (!(errno == EAGAIN || errno == EINTR))
 280                                 pr_err("failed to sync perf data, error: %m\n");
 281                 }
 282         } while (1);
 283 }
 284
 285 struct record_aio {
 286         struct record   *rec;
 287         void            *data;
 288         size_t          size;
 289 };
 290
 291 static int record__aio_pushfn(struct mmap *map, void *to, void *buf, size_t size)
 292 {
 293         struct record_aio *aio = to;
 294
 295         /*
 296          * map->core.base data pointed by buf is copied into free map->aio.data[] buffer
 297          * to release space in the kernel buffer as fast as possible, calling
 298          * perf_mmap__consume() from perf_mmap__push() function.
 299          *
 300          * That lets the kernel to proceed with storing more profiling data into
 301          * the kernel buffer earlier than other per-cpu kernel buffers are handled.
 302          *
 303          * Coping can be done in two steps in case the chunk of profiling data
 304          * crosses the upper bound of the kernel buffer. In this case we first move
 305          * part of data from map->start till the upper bound and then the reminder
 306          * from the beginning of the kernel buffer till the end of the data chunk.
 307          */
 308
 309         if (record__comp_enabled(aio->rec)) {
 310                 size = zstd_compress(aio->rec->session, aio->data + aio->size,
 311                                      mmap__mmap_len(map) - aio->size,
 312                                      buf, size);
 313         } else {
 314                 memcpy(aio->data + aio->size, buf, size);
 315         }
 316
 317         if (!aio->size) {
 318                 /*
 319                  * Increment map->refcount to guard map->aio.data[] buffer
 320                  * from premature deallocation because map object can be
 321                  * released earlier than aio write request started on
 322                  * map->aio.data[] buffer is complete.
 323                  *
 324                  * perf_mmap__put() is done at record__aio_complete()
 325                  * after started aio request completion or at record__aio_push()
 326                  * if the request failed to start.
 327                  */
 328                 perf_mmap__get(&map->core);
 329         }
 330
 331         aio->size += size;
 332
 333         return size;
 334 }
 335
 336 static int record__aio_push(struct record *rec, struct mmap *map, off_t *off)
 337 {
 338         int ret, idx;
 339         int trace_fd = rec->session->data->file.fd;
 340         struct record_aio aio = { .rec = rec, .size = 0 };
 341
 342         /*
 343          * Call record__aio_sync() to wait till map->aio.data[] buffer
 344          * becomes available after previous aio write operation.
 345          */
 346
 347         idx = record__aio_sync(map, false);
 348         aio.data = map->aio.data[idx];
 349         ret = perf_mmap__push(map, &aio, record__aio_pushfn);
 350         if (ret != 0) /* ret > 0 - no data, ret < 0 - error */
 351                 return ret;
 352
 353         rec->samples++;
 354         ret = record__aio_write(&(map->aio.cblocks[idx]), trace_fd, aio.data, aio.size, *off);
 355         if (!ret) {
 356                 *off += aio.size;
 357                 rec->bytes_written += aio.size;
 358                 if (switch_output_size(rec))
 359                         trigger_hit(&switch_output_trigger);
 360         } else {
 361                 /*
 362                  * Decrement map->refcount incremented in record__aio_pushfn()
 363                  * back if record__aio_write() operation failed to start, otherwise
 364                  * map->refcount is decremented in record__aio_complete() after
 365                  * aio write operation finishes successfully.
 366                  */
 367                 perf_mmap__put(&map->core);
 368         }
 369
 370         return ret;
 371 }
 372
 373 static off_t record__aio_get_pos(int trace_fd)
 374 {
 375         return lseek(trace_fd, 0, SEEK_CUR);
 376 }
 377
 378 static void record__aio_set_pos(int trace_fd, off_t pos)
 379 {
 380         lseek(trace_fd, pos, SEEK_SET);
 381 }
 382
 383 static void record__aio_mmap_read_sync(struct record *rec)
 384 {
 385         int i;
 386         struct evlist *evlist = rec->evlist;
 387         struct mmap *maps = evlist->mmap;
 388
 389         if (!record__aio_enabled(rec))
 390                 return;
 391
 392         for (i = 0; i < evlist->core.nr_mmaps; i++) {
 393                 struct mmap *map = &maps[i];
 394
 395                 if (map->core.base)
 396                         record__aio_sync(map, true);
 397         }
 398 }
 399
 400 static int nr_cblocks_default = 1;
 401 static int nr_cblocks_max = 4;
 402
 403 static int record__aio_parse(const struct option *opt,
 404                              const char *str,
 405                              int unset)
 406 {
 407         struct record_opts *opts = (struct record_opts *)opt->value;
 408
 409         if (unset) {
 410                 opts->nr_cblocks = 0;
 411         } else {
 412                 if (str)
 413                         opts->nr_cblocks = strtol(str, NULL, 0);
 414                 if (!opts->nr_cblocks)
 415                         opts->nr_cblocks = nr_cblocks_default;
 416         }
 417
 418         return 0;
 419 }
 420 #else /* HAVE_AIO_SUPPORT */
 421 static int nr_cblocks_max = 0;
 422
 423 static int record__aio_push(struct record *rec __maybe_unused, struct mmap *map __maybe_unused,
 424                             off_t *off __maybe_unused)
 425 {
 426         return -1;
 427 }
 428
 429 static off_t record__aio_get_pos(int trace_fd __maybe_unused)
 430 {
 431         return -1;
 432 }
 433
 434 static void record__aio_set_pos(int trace_fd __maybe_unused, off_t pos __maybe_unused)
 435 {
 436 }
 437
 438 static void record__aio_mmap_read_sync(struct record *rec __maybe_unused)
 439 {
 440 }
 441 #endif
 442
 443 static int record__aio_enabled(struct record *rec)
 444 {
 445         return rec->opts.nr_cblocks > 0;
 446 }
 447
 448 #define MMAP_FLUSH_DEFAULT 1
 449 static int record__mmap_flush_parse(const struct option *opt,
 450                                     const char *str,
 451                                     int unset)
 452 {
 453         int flush_max;
 454         struct record_opts *opts = (struct record_opts *)opt->value;
 455         static struct parse_tag tags[] = {
 456                         { .tag  = 'B', .mult = 1       },
 457                         { .tag  = 'K', .mult = 1 << 10 },
 458                         { .tag  = 'M', .mult = 1 << 20 },
 459                         { .tag  = 'G', .mult = 1 << 30 },
 460                         { .tag  = 0 },
 461         };
 462
 463         if (unset)
 464                 return 0;
 465
 466         if (str) {
 467                 opts->mmap_flush = parse_tag_value(str, tags);
 468                 if (opts->mmap_flush == (int)-1)
 469                         opts->mmap_flush = strtol(str, NULL, 0);
 470         }
 471
 472         if (!opts->mmap_flush)
 473                 opts->mmap_flush = MMAP_FLUSH_DEFAULT;
 474
 475         flush_max = evlist__mmap_size(opts->mmap_pages);
 476         flush_max /= 4;
 477         if (opts->mmap_flush > flush_max)
 478                 opts->mmap_flush = flush_max;
 479
 480         return 0;
 481 }
 482
 483 #ifdef HAVE_ZSTD_SUPPORT
 484 static unsigned int comp_level_default = 1;
 485
 486 static int record__parse_comp_level(const struct option *opt, const char *str, int unset)
 487 {
 488         struct record_opts *opts = opt->value;
 489
 490         if (unset) {
 491                 opts->comp_level = 0;
 492         } else {
 493                 if (str)
 494                         opts->comp_level = strtol(str, NULL, 0);
 495                 if (!opts->comp_level)
 496                         opts->comp_level = comp_level_default;
 497         }
 498
 499         return 0;
 500 }
 501 #endif
 502 static unsigned int comp_level_max = 22;
 503
 504 static int record__comp_enabled(struct record *rec)
 505 {
 506         return rec->opts.comp_level > 0;
 507 }
 508
 509 static int process_synthesized_event(struct perf_tool *tool,
 510                                      union perf_event *event,
 511                                      struct perf_sample *sample __maybe_unused,
 512                                      struct machine *machine __maybe_unused)
 513 {
 514         struct record *rec = container_of(tool, struct record, tool);
 515         return record__write(rec, NULL, event, event->header.size);
 516 }
 517
 518 static int process_locked_synthesized_event(struct perf_tool *tool,
 519                                      union perf_event *event,
 520                                      struct perf_sample *sample __maybe_unused,
 521                                      struct machine *machine __maybe_unused)
 522 {
 523         static pthread_mutex_t synth_lock = PTHREAD_MUTEX_INITIALIZER;
 524         int ret;
 525
 526         pthread_mutex_lock(&synth_lock);
 527         ret = process_synthesized_event(tool, event, sample, machine);
 528         pthread_mutex_unlock(&synth_lock);
 529         return ret;
 530 }
 531
 532 static int record__pushfn(struct mmap *map, void *to, void *bf, size_t size)
 533 {
 534         struct record *rec = to;
 535
 536         if (record__comp_enabled(rec)) {
 537                 size = zstd_compress(rec->session, map->data, mmap__mmap_len(map), bf, size);
 538                 bf   = map->data;
 539         }
 540
 541         rec->samples++;
 542         return record__write(rec, map, bf, size);
 543 }
 544
 545 static volatile int signr = -1;
 546 static volatile int child_finished;
 547 #ifdef HAVE_EVENTFD_SUPPORT
 548 static int done_fd = -1;
 549 #endif
 550
 551 static void sig_handler(int sig)
 552 {
 553         if (sig == SIGCHLD)
 554                 child_finished = 1;
 555         else
 556                 signr = sig;
 557
 558         done = 1;
 559 #ifdef HAVE_EVENTFD_SUPPORT
 560 {
 561         u64 tmp = 1;
 562         /*
 563          * It is possible for this signal handler to run after done is checked
 564          * in the main loop, but before the perf counter fds are polled. If this
 565          * happens, the poll() will continue to wait even though done is set,
 566          * and will only break out if either another signal is received, or the
 567          * counters are ready for read. To ensure the poll() doesn't sleep when
 568          * done is set, use an eventfd (done_fd) to wake up the poll().
 569          */
 570         if (write(done_fd, &tmp, sizeof(tmp)) < 0)
 571                 pr_err("failed to signal wakeup fd, error: %m\n");
 572 }
 573 #endif // HAVE_EVENTFD_SUPPORT
 574 }
 575
 576 static void sigsegv_handler(int sig)
 577 {
 578         perf_hooks__recover();
 579         sighandler_dump_stack(sig);
 580 }
 581
 582 static void record__sig_exit(void)
 583 {
 584         if (signr == -1)
 585                 return;
 586
 587         signal(signr, SIG_DFL);
 588         raise(signr);
 589 }
 590
 591 #ifdef HAVE_AUXTRACE_SUPPORT
 592
 593 static int record__process_auxtrace(struct perf_tool *tool,
 594                                     struct mmap *map,
 595                                     union perf_event *event, void *data1,
 596                                     size_t len1, void *data2, size_t len2)
 597 {
 598         struct record *rec = container_of(tool, struct record, tool);
 599         struct perf_data *data = &rec->data;
 600         size_t padding;
 601         u8 pad[8] = {0};
 602
 603         if (!perf_data__is_pipe(data) && perf_data__is_single_file(data)) {
 604                 off_t file_offset;
 605                 int fd = perf_data__fd(data);
 606                 int err;
 607
 608                 file_offset = lseek(fd, 0, SEEK_CUR);
 609                 if (file_offset == -1)
 610                         return -1;
 611                 err = auxtrace_index__auxtrace_event(&rec->session->auxtrace_index,
 612                                                      event, file_offset);
 613                 if (err)
 614                         return err;
 615         }
 616
 617         /* event.auxtrace.size includes padding, see __auxtrace_mmap__read() */
 618         padding = (len1 + len2) & 7;
 619         if (padding)
 620                 padding = 8 - padding;
 621
 622         record__write(rec, map, event, event->header.size);
 623         record__write(rec, map, data1, len1);
 624         if (len2)
 625                 record__write(rec, map, data2, len2);
 626         record__write(rec, map, &pad, padding);
 627
 628         return 0;
 629 }
 630
 631 static int record__auxtrace_mmap_read(struct record *rec,
 632                                       struct mmap *map)
 633 {
 634         int ret;
 635
 636         ret = auxtrace_mmap__read(map, rec->itr, &rec->tool,
 637                                   record__process_auxtrace);
 638         if (ret < 0)
 639                 return ret;
 640
 641         if (ret)
 642                 rec->samples++;
 643
 644         return 0;
 645 }
 646
 647 static int record__auxtrace_mmap_read_snapshot(struct record *rec,
 648                                                struct mmap *map)
 649 {
 650         int ret;
 651
 652         ret = auxtrace_mmap__read_snapshot(map, rec->itr, &rec->tool,
 653                                            record__process_auxtrace,
 654                                            rec->opts.auxtrace_snapshot_size);
 655         if (ret < 0)
 656                 return ret;
 657
 658         if (ret)
 659                 rec->samples++;
 660
 661         return 0;
 662 }
 663
 664 static int record__auxtrace_read_snapshot_all(struct record *rec)
 665 {
 666         int i;
 667         int rc = 0;
 668
 669         for (i = 0; i < rec->evlist->core.nr_mmaps; i++) {
 670                 struct mmap *map = &rec->evlist->mmap[i];
 671
 672                 if (!map->auxtrace_mmap.base)
 673                         continue;
 674
 675                 if (record__auxtrace_mmap_read_snapshot(rec, map) != 0) {
 676                         rc = -1;
 677                         goto out;
 678                 }
 679         }
 680 out:
 681         return rc;
 682 }
 683
 684 static void record__read_auxtrace_snapshot(struct record *rec, bool on_exit)
 685 {
 686         pr_debug("Recording AUX area tracing snapshot\n");
 687         if (record__auxtrace_read_snapshot_all(rec) < 0) {
 688                 trigger_error(&auxtrace_snapshot_trigger);
 689         } else {
 690                 if (auxtrace_record__snapshot_finish(rec->itr, on_exit))
 691                         trigger_error(&auxtrace_snapshot_trigger);
 692                 else
 693                         trigger_ready(&auxtrace_snapshot_trigger);
 694         }
 695 }
 696
 697 static int record__auxtrace_snapshot_exit(struct record *rec)
 698 {
 699         if (trigger_is_error(&auxtrace_snapshot_trigger))
 700                 return 0;
 701
 702         if (!auxtrace_record__snapshot_started &&
 703             auxtrace_record__snapshot_start(rec->itr))
 704                 return -1;
 705
 706         record__read_auxtrace_snapshot(rec, true);
 707         if (trigger_is_error(&auxtrace_snapshot_trigger))
 708                 return -1;
 709
 710         return 0;
 711 }
 712
 713 static int record__auxtrace_init(struct record *rec)
 714 {
 715         int err;
 716
 717         if (!rec->itr) {
 718                 rec->itr = auxtrace_record__init(rec->evlist, &err);
 719                 if (err)
 720                         return err;
 721         }
 722
 723         err = auxtrace_parse_snapshot_options(rec->itr, &rec->opts,
 724                                               rec->opts.auxtrace_snapshot_opts);
 725         if (err)
 726                 return err;
 727
 728         err = auxtrace_parse_sample_options(rec->itr, rec->evlist, &rec->opts,
 729                                             rec->opts.auxtrace_sample_opts);
 730         if (err)
 731                 return err;
 732
 733         return auxtrace_parse_filters(rec->evlist);
 734 }
 735
 736 #else
 737
 738 static inline
 739 int record__auxtrace_mmap_read(struct record *rec __maybe_unused,
 740                                struct mmap *map __maybe_unused)
 741 {
 742         return 0;
 743 }
 744
 745 static inline
 746 void record__read_auxtrace_snapshot(struct record *rec __maybe_unused,
 747                                     bool on_exit __maybe_unused)
 748 {
 749 }
 750
 751 static inline
 752 int auxtrace_record__snapshot_start(struct auxtrace_record *itr __maybe_unused)
 753 {
 754         return 0;
 755 }
 756
 757 static inline
 758 int record__auxtrace_snapshot_exit(struct record *rec __maybe_unused)
 759 {
 760         return 0;
 761 }
 762
 763 static int record__auxtrace_init(struct record *rec __maybe_unused)
 764 {
 765         return 0;
 766 }
 767
 768 #endif
 769
 770 static int record__config_text_poke(struct evlist *evlist)
 771 {
 772         struct evsel *evsel;
 773         int err;
 774
 775         /* Nothing to do if text poke is already configured */
 776         evlist__for_each_entry(evlist, evsel) {
 777                 if (evsel->core.attr.text_poke)
 778                         return 0;
 779         }
 780
 781         err = parse_events(evlist, "dummy:u", NULL);
 782         if (err)
 783                 return err;
 784
 785         evsel = evlist__last(evlist);
 786
 787         evsel->core.attr.freq = 0;
 788         evsel->core.attr.sample_period = 1;
 789         evsel->core.attr.text_poke = 1;
 790         evsel->core.attr.ksymbol = 1;
 791
 792         evsel->core.system_wide = true;
 793         evsel->no_aux_samples = true;
 794         evsel->immediate = true;
 795
 796         /* Text poke must be collected on all CPUs */
 797         perf_cpu_map__put(evsel->core.own_cpus);
 798         evsel->core.own_cpus = perf_cpu_map__new(NULL);
 799         perf_cpu_map__put(evsel->core.cpus);
 800         evsel->core.cpus = perf_cpu_map__get(evsel->core.own_cpus);
 801
 802         evsel__set_sample_bit(evsel, TIME);
 803
 804         return 0;
 805 }
 806
 807 static bool record__kcore_readable(struct machine *machine)
 808 {
 809         char kcore[PATH_MAX];
 810         int fd;
 811
 812         scnprintf(kcore, sizeof(kcore), "%s/proc/kcore", machine->root_dir);
 813
 814         fd = open(kcore, O_RDONLY);
 815         if (fd < 0)
 816                 return false;
 817
 818         close(fd);
 819
 820         return true;
 821 }
 822
 823 static int record__kcore_copy(struct machine *machine, struct perf_data *data)
 824 {
 825         char from_dir[PATH_MAX];
 826         char kcore_dir[PATH_MAX];
 827         int ret;
 828
 829         snprintf(from_dir, sizeof(from_dir), "%s/proc", machine->root_dir);
 830
 831         ret = perf_data__make_kcore_dir(data, kcore_dir, sizeof(kcore_dir));
 832         if (ret)
 833                 return ret;
 834
 835         return kcore_copy(from_dir, kcore_dir);
 836 }
 837
 838 static int record__mmap_evlist(struct record *rec,
 839                                struct evlist *evlist)
 840 {
 841         struct record_opts *opts = &rec->opts;
 842         bool auxtrace_overwrite = opts->auxtrace_snapshot_mode ||
 843                                   opts->auxtrace_sample_mode;
 844         char msg[512];
 845
 846         if (opts->affinity != PERF_AFFINITY_SYS)
 847                 cpu__setup_cpunode_map();
 848
 849         if (evlist__mmap_ex(evlist, opts->mmap_pages,
 850                                  opts->auxtrace_mmap_pages,
 851                                  auxtrace_overwrite,
 852                                  opts->nr_cblocks, opts->affinity,
 853                                  opts->mmap_flush, opts->comp_level) < 0) {
 854                 if (errno == EPERM) {
 855                         pr_err("Permission error mapping pages.\n"
 856                                "Consider increasing "
 857                                "/proc/sys/kernel/perf_event_mlock_kb,\n"
 858                                "or try again with a smaller value of -m/--mmap_pages.\n"
 859                                "(current value: %u,%u)\n",
 860                                opts->mmap_pages, opts->auxtrace_mmap_pages);
 861                         return -errno;
 862                 } else {
 863                         pr_err("failed to mmap with %d (%s)\n", errno,
 864                                 str_error_r(errno, msg, sizeof(msg)));
 865                         if (errno)
 866                                 return -errno;
 867                         else
 868                                 return -EINVAL;
 869                 }
 870         }
 871         return 0;
 872 }
 873
 874 static int record__mmap(struct record *rec)
 875 {
 876         return record__mmap_evlist(rec, rec->evlist);
 877 }
 878
 879 static int record__open(struct record *rec)
 880 {
 881         char msg[BUFSIZ];
 882         struct evsel *pos;
 883         struct evlist *evlist = rec->evlist;
 884         struct perf_session *session = rec->session;
 885         struct record_opts *opts = &rec->opts;
 886         int rc = 0;
 887
 888         /*
 889          * For initial_delay or system wide, we need to add a dummy event so
 890          * that we can track PERF_RECORD_MMAP to cover the delay of waiting or
 891          * event synthesis.
 892          */
 893         if (opts->initial_delay || target__has_cpu(&opts->target)) {
 894                 pos = evlist__get_tracking_event(evlist);
 895                 if (!evsel__is_dummy_event(pos)) {
 896                         /* Set up dummy event. */
 897                         if (evlist__add_dummy(evlist))
 898                                 return -ENOMEM;
 899                         pos = evlist__last(evlist);
 900                         evlist__set_tracking_event(evlist, pos);
 901                 }
 902
 903                 /*
 904                  * Enable the dummy event when the process is forked for
 905                  * initial_delay, immediately for system wide.
 906                  */
 907                 if (opts->initial_delay && !pos->immediate)
 908                         pos->core.attr.enable_on_exec = 1;
 909                 else
 910                         pos->immediate = 1;
 911         }
 912
 913         evlist__config(evlist, opts, &callchain_param);
 914
 915         evlist__for_each_entry(evlist, pos) {
 916 try_again:
 917                 if (evsel__open(pos, pos->core.cpus, pos->core.threads) < 0) {
 918                         if (evsel__fallback(pos, errno, msg, sizeof(msg))) {
 919                                 if (verbose > 0)
 920                                         ui__warning("%s\n", msg);
 921                                 goto try_again;
 922                         }
 923                         if ((errno == EINVAL || errno == EBADF) &&
 924                             pos->leader != pos &&
 925                             pos->weak_group) {
 926                                 pos = evlist__reset_weak_group(evlist, pos, true);
 927                                 goto try_again;
 928                         }
 929                         rc = -errno;
 930                         evsel__open_strerror(pos, &opts->target, errno, msg, sizeof(msg));
 931                         ui__error("%s\n", msg);
 932                         goto out;
 933                 }
 934
 935                 pos->supported = true;
 936         }
 937
 938         if (symbol_conf.kptr_restrict && !evlist__exclude_kernel(evlist)) {
 939                 pr_warning(
 940 "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
 941 "check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
 942 "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
 943 "file is not found in the buildid cache or in the vmlinux path.\n\n"
 944 "Samples in kernel modules won't be resolved at all.\n\n"
 945 "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
 946 "even with a suitable vmlinux or kallsyms file.\n\n");
 947         }
 948
 949         if (evlist__apply_filters(evlist, &pos)) {
 950                 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
 951                         pos->filter, evsel__name(pos), errno,
 952                         str_error_r(errno, msg, sizeof(msg)));
 953                 rc = -1;
 954                 goto out;
 955         }
 956
 957         rc = record__mmap(rec);
 958         if (rc)
 959                 goto out;
 960
 961         session->evlist = evlist;
 962         perf_session__set_id_hdr_size(session);
 963 out:
 964         return rc;
 965 }
 966
 967 static int process_sample_event(struct perf_tool *tool,
 968                                 union perf_event *event,
 969                                 struct perf_sample *sample,
 970                                 struct evsel *evsel,
 971                                 struct machine *machine)
 972 {
 973         struct record *rec = container_of(tool, struct record, tool);
 974
 975         if (rec->evlist->first_sample_time == 0)
 976                 rec->evlist->first_sample_time = sample->time;
 977
 978         rec->evlist->last_sample_time = sample->time;
 979
 980         if (rec->buildid_all)
 981                 return 0;
 982
 983         rec->samples++;
 984         return build_id__mark_dso_hit(tool, event, sample, evsel, machine);
 985 }
 986
 987 static int process_buildids(struct record *rec)
 988 {
 989         struct perf_session *session = rec->session;
 990
 991         if (perf_data__size(&rec->data) == 0)
 992                 return 0;
 993
 994         /*
 995          * During this process, it'll load kernel map and replace the
 996          * dso->long_name to a real pathname it found.  In this case
 997          * we prefer the vmlinux path like
 998          *   /lib/modules/3.16.4/build/vmlinux
 999          *
1000          * rather than build-id path (in debug directory).
1001          *   $HOME/.debug/.build-id/f0/6e17aa50adf4d00b88925e03775de107611551
1002          */
1003         symbol_conf.ignore_vmlinux_buildid = true;
1004
1005         /*
1006          * If --buildid-all is given, it marks all DSO regardless of hits,
1007          * so no need to process samples. But if timestamp_boundary is enabled,
1008          * it still needs to walk on all samples to get the timestamps of
1009          * first/last samples.
1010          */
1011         if (rec->buildid_all && !rec->timestamp_boundary)
1012                 rec->tool.sample = NULL;
1013
1014         return perf_session__process_events(session);
1015 }
1016
1017 static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
1018 {
1019         int err;
1020         struct perf_tool *tool = data;
1021         /*
1022          *As for guest kernel when processing subcommand record&report,
1023          *we arrange module mmap prior to guest kernel mmap and trigger
1024          *a preload dso because default guest module symbols are loaded
1025          *from guest kallsyms instead of /lib/modules/XXX/XXX. This
1026          *method is used to avoid symbol missing when the first addr is
1027          *in module instead of in guest kernel.
1028          */
1029         err = perf_event__synthesize_modules(tool, process_synthesized_event,
1030                                              machine);
1031         if (err < 0)
1032                 pr_err("Couldn't record guest kernel [%d]'s reference"
1033                        " relocation symbol.\n", machine->pid);
1034
1035         /*
1036          * We use _stext for guest kernel because guest kernel's /proc/kallsyms
1037          * have no _text sometimes.
1038          */
1039         err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
1040                                                  machine);
1041         if (err < 0)
1042                 pr_err("Couldn't record guest kernel [%d]'s reference"
1043                        " relocation symbol.\n", machine->pid);
1044 }
1045
1046 static struct perf_event_header finished_round_event = {
1047         .size = sizeof(struct perf_event_header),
1048         .type = PERF_RECORD_FINISHED_ROUND,
1049 };
1050
1051 static void record__adjust_affinity(struct record *rec, struct mmap *map)
1052 {
1053         if (rec->opts.affinity != PERF_AFFINITY_SYS &&
1054             !bitmap_equal(rec->affinity_mask.bits, map->affinity_mask.bits,
1055                           rec->affinity_mask.nbits)) {
1056                 bitmap_zero(rec->affinity_mask.bits, rec->affinity_mask.nbits);
1057                 bitmap_or(rec->affinity_mask.bits, rec->affinity_mask.bits,
1058                           map->affinity_mask.bits, rec->affinity_mask.nbits);
1059                 sched_setaffinity(0, MMAP_CPU_MASK_BYTES(&rec->affinity_mask),
1060                                   (cpu_set_t *)rec->affinity_mask.bits);
1061                 if (verbose == 2)
1062                         mmap_cpu_mask__scnprintf(&rec->affinity_mask, "thread");
1063         }
1064 }
1065
1066 static size_t process_comp_header(void *record, size_t increment)
1067 {
1068         struct perf_record_compressed *event = record;
1069         size_t size = sizeof(*event);
1070
1071         if (increment) {
1072                 event->header.size += increment;
1073                 return increment;
1074         }
1075
1076         event->header.type = PERF_RECORD_COMPRESSED;
1077         event->header.size = size;
1078
1079         return size;
1080 }
1081
1082 static size_t zstd_compress(struct perf_session *session, void *dst, size_t dst_size,
1083                             void *src, size_t src_size)
1084 {
1085         size_t compressed;
1086         size_t max_record_size = PERF_SAMPLE_MAX_SIZE - sizeof(struct perf_record_compressed) - 1;
1087
1088         compressed = zstd_compress_stream_to_records(&session->zstd_data, dst, dst_size, src, src_size,
1089                                                      max_record_size, process_comp_header);
1090
1091         session->bytes_transferred += src_size;
1092         session->bytes_compressed  += compressed;
1093
1094         return compressed;
1095 }
1096
1097 static int record__mmap_read_evlist(struct record *rec, struct evlist *evlist,
1098                                     bool overwrite, bool synch)
1099 {
1100         u64 bytes_written = rec->bytes_written;
1101         int i;
1102         int rc = 0;
1103         struct mmap *maps;
1104         int trace_fd = rec->data.file.fd;
1105         off_t off = 0;
1106
1107         if (!evlist)
1108                 return 0;
1109
1110         maps = overwrite ? evlist->overwrite_mmap : evlist->mmap;
1111         if (!maps)
1112                 return 0;
1113
1114         if (overwrite && evlist->bkw_mmap_state != BKW_MMAP_DATA_PENDING)
1115                 return 0;
1116
1117         if (record__aio_enabled(rec))
1118                 off = record__aio_get_pos(trace_fd);
1119
1120         for (i = 0; i < evlist->core.nr_mmaps; i++) {
1121                 u64 flush = 0;
1122                 struct mmap *map = &maps[i];
1123
1124                 if (map->core.base) {
1125                         record__adjust_affinity(rec, map);
1126                         if (synch) {
1127                                 flush = map->core.flush;
1128                                 map->core.flush = 1;
1129                         }
1130                         if (!record__aio_enabled(rec)) {
1131                                 if (perf_mmap__push(map, rec, record__pushfn) < 0) {
1132                                         if (synch)
1133                                                 map->core.flush = flush;
1134                                         rc = -1;
1135                                         goto out;
1136                                 }
1137                         } else {
1138                                 if (record__aio_push(rec, map, &off) < 0) {
1139                                         record__aio_set_pos(trace_fd, off);
1140                                         if (synch)
1141                                                 map->core.flush = flush;
1142                                         rc = -1;
1143                                         goto out;
1144                                 }
1145                         }
1146                         if (synch)
1147                                 map->core.flush = flush;
1148                 }
1149
1150                 if (map->auxtrace_mmap.base && !rec->opts.auxtrace_snapshot_mode &&
1151                     !rec->opts.auxtrace_sample_mode &&
1152                     record__auxtrace_mmap_read(rec, map) != 0) {
1153                         rc = -1;
1154                         goto out;
1155                 }
1156         }
1157
1158         if (record__aio_enabled(rec))
1159                 record__aio_set_pos(trace_fd, off);
1160
1161         /*
1162          * Mark the round finished in case we wrote
1163          * at least one event.
1164          */
1165         if (bytes_written != rec->bytes_written)
1166                 rc = record__write(rec, NULL, &finished_round_event, sizeof(finished_round_event));
1167
1168         if (overwrite)
1169                 evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
1170 out:
1171         return rc;
1172 }
1173
1174 static int record__mmap_read_all(struct record *rec, bool synch)
1175 {
1176         int err;
1177
1178         err = record__mmap_read_evlist(rec, rec->evlist, false, synch);
1179         if (err)
1180                 return err;
1181
1182         return record__mmap_read_evlist(rec, rec->evlist, true, synch);
1183 }
1184
1185 static void record__init_features(struct record *rec)
1186 {
1187         struct perf_session *session = rec->session;
1188         int feat;
1189
1190         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1191                 perf_header__set_feat(&session->header, feat);
1192
1193         if (rec->no_buildid)
1194                 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1195
1196         if (!have_tracepoints(&rec->evlist->core.entries))
1197                 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1198
1199         if (!rec->opts.branch_stack)
1200                 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1201
1202         if (!rec->opts.full_auxtrace)
1203                 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1204
1205         if (!(rec->opts.use_clockid && rec->opts.clockid_res_ns))
1206                 perf_header__clear_feat(&session->header, HEADER_CLOCKID);
1207
1208         if (!rec->opts.use_clockid)
1209                 perf_header__clear_feat(&session->header, HEADER_CLOCK_DATA);
1210
1211         perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1212         if (!record__comp_enabled(rec))
1213                 perf_header__clear_feat(&session->header, HEADER_COMPRESSED);
1214
1215         perf_header__clear_feat(&session->header, HEADER_STAT);
1216 }
1217
1218 static void
1219 record__finish_output(struct record *rec)
1220 {
1221         struct perf_data *data = &rec->data;
1222         int fd = perf_data__fd(data);
1223
1224         if (data->is_pipe)
1225                 return;
1226
1227         rec->session->header.data_size += rec->bytes_written;
1228         data->file.size = lseek(perf_data__fd(data), 0, SEEK_CUR);
1229
1230         if (!rec->no_buildid) {
1231                 process_buildids(rec);
1232
1233                 if (rec->buildid_all)
1234                         dsos__hit_all(rec->session);
1235         }
1236         perf_session__write_header(rec->session, rec->evlist, fd, true);
1237
1238         return;
1239 }
1240
1241 static int record__synthesize_workload(struct record *rec, bool tail)
1242 {
1243         int err;
1244         struct perf_thread_map *thread_map;
1245
1246         if (rec->opts.tail_synthesize != tail)
1247                 return 0;
1248
1249         thread_map = thread_map__new_by_tid(rec->evlist->workload.pid);
1250         if (thread_map == NULL)
1251                 return -1;
1252
1253         err = perf_event__synthesize_thread_map(&rec->tool, thread_map,
1254                                                  process_synthesized_event,
1255                                                  &rec->session->machines.host,
1256                                                  rec->opts.sample_address);
1257         perf_thread_map__put(thread_map);
1258         return err;
1259 }
1260
1261 static int record__synthesize(struct record *rec, bool tail);
1262
1263 static int
1264 record__switch_output(struct record *rec, bool at_exit)
1265 {
1266         struct perf_data *data = &rec->data;
1267         int fd, err;
1268         char *new_filename;
1269
1270         /* Same Size:      "2015122520103046"*/
1271         char timestamp[] = "InvalidTimestamp";
1272
1273         record__aio_mmap_read_sync(rec);
1274
1275         record__synthesize(rec, true);
1276         if (target__none(&rec->opts.target))
1277                 record__synthesize_workload(rec, true);
1278
1279         rec->samples = 0;
1280         record__finish_output(rec);
1281         err = fetch_current_timestamp(timestamp, sizeof(timestamp));
1282         if (err) {
1283                 pr_err("Failed to get current timestamp\n");
1284                 return -EINVAL;
1285         }
1286
1287         fd = perf_data__switch(data, timestamp,
1288                                     rec->session->header.data_offset,
1289                                     at_exit, &new_filename);
1290         if (fd >= 0 && !at_exit) {
1291                 rec->bytes_written = 0;
1292                 rec->session->header.data_size = 0;
1293         }
1294
1295         if (!quiet)
1296                 fprintf(stderr, "[ perf record: Dump %s.%s ]\n",
1297                         data->path, timestamp);
1298
1299         if (rec->switch_output.num_files) {
1300                 int n = rec->switch_output.cur_file + 1;
1301
1302                 if (n >= rec->switch_output.num_files)
1303                         n = 0;
1304                 rec->switch_output.cur_file = n;
1305                 if (rec->switch_output.filenames[n]) {
1306                         remove(rec->switch_output.filenames[n]);
1307                         zfree(&rec->switch_output.filenames[n]);
1308                 }
1309                 rec->switch_output.filenames[n] = new_filename;
1310         } else {
1311                 free(new_filename);
1312         }
1313
1314         /* Output tracking events */
1315         if (!at_exit) {
1316                 record__synthesize(rec, false);
1317
1318                 /*
1319                  * In 'perf record --switch-output' without -a,
1320                  * record__synthesize() in record__switch_output() won't
1321                  * generate tracking events because there's no thread_map
1322                  * in evlist. Which causes newly created perf.data doesn't
1323                  * contain map and comm information.
1324                  * Create a fake thread_map and directly call
1325                  * perf_event__synthesize_thread_map() for those events.
1326                  */
1327                 if (target__none(&rec->opts.target))
1328                         record__synthesize_workload(rec, false);
1329         }
1330         return fd;
1331 }
1332
1333 static volatile int workload_exec_errno;
1334
1335 /*
1336  * evlist__prepare_workload will send a SIGUSR1
1337  * if the fork fails, since we asked by setting its
1338  * want_signal to true.
1339  */
1340 static void workload_exec_failed_signal(int signo __maybe_unused,
1341                                         siginfo_t *info,
1342                                         void *ucontext __maybe_unused)
1343 {
1344         workload_exec_errno = info->si_value.sival_int;
1345         done = 1;
1346         child_finished = 1;
1347 }
1348
1349 static void snapshot_sig_handler(int sig);
1350 static void alarm_sig_handler(int sig);
1351
1352 static const struct perf_event_mmap_page *evlist__pick_pc(struct evlist *evlist)
1353 {
1354         if (evlist) {
1355                 if (evlist->mmap && evlist->mmap[0].core.base)
1356                         return evlist->mmap[0].core.base;
1357                 if (evlist->overwrite_mmap && evlist->overwrite_mmap[0].core.base)
1358                         return evlist->overwrite_mmap[0].core.base;
1359         }
1360         return NULL;
1361 }
1362
1363 static const struct perf_event_mmap_page *record__pick_pc(struct record *rec)
1364 {
1365         const struct perf_event_mmap_page *pc = evlist__pick_pc(rec->evlist);
1366         if (pc)
1367                 return pc;
1368         return NULL;
1369 }
1370
1371 static int record__synthesize(struct record *rec, bool tail)
1372 {
1373         struct perf_session *session = rec->session;
1374         struct machine *machine = &session->machines.host;
1375         struct perf_data *data = &rec->data;
1376         struct record_opts *opts = &rec->opts;
1377         struct perf_tool *tool = &rec->tool;
1378         int fd = perf_data__fd(data);
1379         int err = 0;
1380         event_op f = process_synthesized_event;
1381
1382         if (rec->opts.tail_synthesize != tail)
1383                 return 0;
1384
1385         if (data->is_pipe) {
1386                 /*
1387                  * We need to synthesize events first, because some
1388                  * features works on top of them (on report side).
1389                  */
1390                 err = perf_event__synthesize_attrs(tool, rec->evlist,
1391                                                    process_synthesized_event);
1392                 if (err < 0) {
1393                         pr_err("Couldn't synthesize attrs.\n");
1394                         goto out;
1395                 }
1396
1397                 err = perf_event__synthesize_features(tool, session, rec->evlist,
1398                                                       process_synthesized_event);
1399                 if (err < 0) {
1400                         pr_err("Couldn't synthesize features.\n");
1401                         return err;
1402                 }
1403
1404                 if (have_tracepoints(&rec->evlist->core.entries)) {
1405                         /*
1406                          * FIXME err <= 0 here actually means that
1407                          * there were no tracepoints so its not really
1408                          * an error, just that we don't need to
1409                          * synthesize anything.  We really have to
1410                          * return this more properly and also
1411                          * propagate errors that now are calling die()
1412                          */
1413                         err = perf_event__synthesize_tracing_data(tool, fd, rec->evlist,
1414                                                                   process_synthesized_event);
1415                         if (err <= 0) {
1416                                 pr_err("Couldn't record tracing data.\n");
1417                                 goto out;
1418                         }
1419                         rec->bytes_written += err;
1420                 }
1421         }
1422
1423         err = perf_event__synth_time_conv(record__pick_pc(rec), tool,
1424                                           process_synthesized_event, machine);
1425         if (err)
1426                 goto out;
1427
1428         /* Synthesize id_index before auxtrace_info */
1429         if (rec->opts.auxtrace_sample_mode) {
1430                 err = perf_event__synthesize_id_index(tool,
1431                                                       process_synthesized_event,
1432                                                       session->evlist, machine);
1433                 if (err)
1434                         goto out;
1435         }
1436
1437         if (rec->opts.full_auxtrace) {
1438                 err = perf_event__synthesize_auxtrace_info(rec->itr, tool,
1439                                         session, process_synthesized_event);
1440                 if (err)
1441                         goto out;
1442         }
1443
1444         if (!evlist__exclude_kernel(rec->evlist)) {
1445                 err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
1446                                                          machine);
1447                 WARN_ONCE(err < 0, "Couldn't record kernel reference relocation symbol\n"
1448                                    "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
1449                                    "Check /proc/kallsyms permission or run as root.\n");
1450
1451                 err = perf_event__synthesize_modules(tool, process_synthesized_event,
1452                                                      machine);
1453                 WARN_ONCE(err < 0, "Couldn't record kernel module information.\n"
1454                                    "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
1455                                    "Check /proc/modules permission or run as root.\n");
1456         }
1457
1458         if (perf_guest) {
1459                 machines__process_guests(&session->machines,
1460                                          perf_event__synthesize_guest_os, tool);
1461         }
1462
1463         err = perf_event__synthesize_extra_attr(&rec->tool,
1464                                                 rec->evlist,
1465                                                 process_synthesized_event,
1466                                                 data->is_pipe);
1467         if (err)
1468                 goto out;
1469
1470         err = perf_event__synthesize_thread_map2(&rec->tool, rec->evlist->core.threads,
1471                                                  process_synthesized_event,
1472                                                 NULL);
1473         if (err < 0) {
1474                 pr_err("Couldn't synthesize thread map.\n");
1475                 return err;
1476         }
1477
1478         err = perf_event__synthesize_cpu_map(&rec->tool, rec->evlist->core.cpus,
1479                                              process_synthesized_event, NULL);
1480         if (err < 0) {
1481                 pr_err("Couldn't synthesize cpu map.\n");
1482                 return err;
1483         }
1484
1485         err = perf_event__synthesize_bpf_events(session, process_synthesized_event,
1486                                                 machine, opts);
1487         if (err < 0)
1488                 pr_warning("Couldn't synthesize bpf events.\n");
1489
1490         err = perf_event__synthesize_cgroups(tool, process_synthesized_event,
1491                                              machine);
1492         if (err < 0)
1493                 pr_warning("Couldn't synthesize cgroup events.\n");
1494
1495         if (rec->opts.nr_threads_synthesize > 1) {
1496                 perf_set_multithreaded();
1497                 f = process_locked_synthesized_event;
1498         }
1499
1500         err = __machine__synthesize_threads(machine, tool, &opts->target, rec->evlist->core.threads,
1501                                             f, opts->sample_address,
1502                                             rec->opts.nr_threads_synthesize);
1503
1504         if (rec->opts.nr_threads_synthesize > 1)
1505                 perf_set_singlethreaded();
1506
1507 out:
1508         return err;
1509 }
1510
1511 static int record__process_signal_event(union perf_event *event __maybe_unused, void *data)
1512 {
1513         struct record *rec = data;
1514         pthread_kill(rec->thread_id, SIGUSR2);
1515         return 0;
1516 }
1517
1518 static int record__setup_sb_evlist(struct record *rec)
1519 {
1520         struct record_opts *opts = &rec->opts;
1521
1522         if (rec->sb_evlist != NULL) {
1523                 /*
1524                  * We get here if --switch-output-event populated the
1525                  * sb_evlist, so associate a callback that will send a SIGUSR2
1526                  * to the main thread.
1527                  */
1528                 evlist__set_cb(rec->sb_evlist, record__process_signal_event, rec);
1529                 rec->thread_id = pthread_self();
1530         }
1531 #ifdef HAVE_LIBBPF_SUPPORT
1532         if (!opts->no_bpf_event) {
1533                 if (rec->sb_evlist == NULL) {
1534                         rec->sb_evlist = evlist__new();
1535
1536                         if (rec->sb_evlist == NULL) {
1537                                 pr_err("Couldn't create side band evlist.\n.");
1538                                 return -1;
1539                         }
1540                 }
1541
1542                 if (evlist__add_bpf_sb_event(rec->sb_evlist, &rec->session->header.env)) {
1543                         pr_err("Couldn't ask for PERF_RECORD_BPF_EVENT side band events.\n.");
1544                         return -1;
1545                 }
1546         }
1547 #endif
1548         if (evlist__start_sb_thread(rec->sb_evlist, &rec->opts.target)) {
1549                 pr_debug("Couldn't start the BPF side band thread:\nBPF programs starting from now on won't be annotatable\n");
1550                 opts->no_bpf_event = true;
1551         }
1552
1553         return 0;
1554 }
1555
1556 static int record__init_clock(struct record *rec)
1557 {
1558         struct perf_session *session = rec->session;
1559         struct timespec ref_clockid;
1560         struct timeval ref_tod;
1561         u64 ref;
1562
1563         if (!rec->opts.use_clockid)
1564                 return 0;
1565
1566         if (rec->opts.use_clockid && rec->opts.clockid_res_ns)
1567                 session->header.env.clock.clockid_res_ns = rec->opts.clockid_res_ns;
1568
1569         session->header.env.clock.clockid = rec->opts.clockid;
1570
1571         if (gettimeofday(&ref_tod, NULL) != 0) {
1572                 pr_err("gettimeofday failed, cannot set reference time.\n");
1573                 return -1;
1574         }
1575
1576         if (clock_gettime(rec->opts.clockid, &ref_clockid)) {
1577                 pr_err("clock_gettime failed, cannot set reference time.\n");
1578                 return -1;
1579         }
1580
1581         ref = (u64) ref_tod.tv_sec * NSEC_PER_SEC +
1582               (u64) ref_tod.tv_usec * NSEC_PER_USEC;
1583
1584         session->header.env.clock.tod_ns = ref;
1585
1586         ref = (u64) ref_clockid.tv_sec * NSEC_PER_SEC +
1587               (u64) ref_clockid.tv_nsec;
1588
1589         session->header.env.clock.clockid_ns = ref;
1590         return 0;
1591 }
1592
1593 static void hit_auxtrace_snapshot_trigger(struct record *rec)
1594 {
1595         if (trigger_is_ready(&auxtrace_snapshot_trigger)) {
1596                 trigger_hit(&auxtrace_snapshot_trigger);
1597                 auxtrace_record__snapshot_started = 1;
1598                 if (auxtrace_record__snapshot_start(rec->itr))
1599                         trigger_error(&auxtrace_snapshot_trigger);
1600         }
1601 }
1602
1603 static int __cmd_record(struct record *rec, int argc, const char **argv)
1604 {
1605         int err;
1606         int status = 0;
1607         unsigned long waking = 0;
1608         const bool forks = argc > 0;
1609         struct perf_tool *tool = &rec->tool;
1610         struct record_opts *opts = &rec->opts;
1611         struct perf_data *data = &rec->data;
1612         struct perf_session *session;
1613         bool disabled = false, draining = false;
1614         int fd;
1615         float ratio = 0;
1616         enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
1617
1618         atexit(record__sig_exit);
1619         signal(SIGCHLD, sig_handler);
1620         signal(SIGINT, sig_handler);
1621         signal(SIGTERM, sig_handler);
1622         signal(SIGSEGV, sigsegv_handler);
1623
1624         if (rec->opts.record_namespaces)
1625                 tool->namespace_events = true;
1626
1627         if (rec->opts.record_cgroup) {
1628 #ifdef HAVE_FILE_HANDLE
1629                 tool->cgroup_events = true;
1630 #else
1631                 pr_err("cgroup tracking is not supported\n");
1632                 return -1;
1633 #endif
1634         }
1635
1636         if (rec->opts.auxtrace_snapshot_mode || rec->switch_output.enabled) {
1637                 signal(SIGUSR2, snapshot_sig_handler);
1638                 if (rec->opts.auxtrace_snapshot_mode)
1639                         trigger_on(&auxtrace_snapshot_trigger);
1640                 if (rec->switch_output.enabled)
1641                         trigger_on(&switch_output_trigger);
1642         } else {
1643                 signal(SIGUSR2, SIG_IGN);
1644         }
1645
1646         session = perf_session__new(data, false, tool);
1647         if (IS_ERR(session)) {
1648                 pr_err("Perf session creation failed.\n");
1649                 return PTR_ERR(session);
1650         }
1651
1652         fd = perf_data__fd(data);
1653         rec->session = session;
1654
1655         if (zstd_init(&session->zstd_data, rec->opts.comp_level) < 0) {
1656                 pr_err("Compression initialization failed.\n");
1657                 return -1;
1658         }
1659 #ifdef HAVE_EVENTFD_SUPPORT
1660         done_fd = eventfd(0, EFD_NONBLOCK);
1661         if (done_fd < 0) {
1662                 pr_err("Failed to create wakeup eventfd, error: %m\n");
1663                 status = -1;
1664                 goto out_delete_session;
1665         }
1666         err = evlist__add_pollfd(rec->evlist, done_fd);
1667         if (err < 0) {
1668                 pr_err("Failed to add wakeup eventfd to poll list\n");
1669                 status = err;
1670                 goto out_delete_session;
1671         }
1672 #endif // HAVE_EVENTFD_SUPPORT
1673
1674         session->header.env.comp_type  = PERF_COMP_ZSTD;
1675         session->header.env.comp_level = rec->opts.comp_level;
1676
1677         if (rec->opts.kcore &&
1678             !record__kcore_readable(&session->machines.host)) {
1679                 pr_err("ERROR: kcore is not readable.\n");
1680                 return -1;
1681         }
1682
1683         if (record__init_clock(rec))
1684                 return -1;
1685
1686         record__init_features(rec);
1687
1688         if (forks) {
1689                 err = evlist__prepare_workload(rec->evlist, &opts->target, argv, data->is_pipe,
1690                                                workload_exec_failed_signal);
1691                 if (err < 0) {
1692                         pr_err("Couldn't run the workload!\n");
1693                         status = err;
1694                         goto out_delete_session;
1695                 }
1696         }
1697
1698         /*
1699          * If we have just single event and are sending data
1700          * through pipe, we need to force the ids allocation,
1701          * because we synthesize event name through the pipe
1702          * and need the id for that.
1703          */
1704         if (data->is_pipe && rec->evlist->core.nr_entries == 1)
1705                 rec->opts.sample_id = true;
1706
1707         if (record__open(rec) != 0) {
1708                 err = -1;
1709                 goto out_child;
1710         }
1711         session->header.env.comp_mmap_len = session->evlist->core.mmap_len;
1712
1713         if (rec->opts.kcore) {
1714                 err = record__kcore_copy(&session->machines.host, data);
1715                 if (err) {
1716                         pr_err("ERROR: Failed to copy kcore\n");
1717                         goto out_child;
1718                 }
1719         }
1720
1721         err = bpf__apply_obj_config();
1722         if (err) {
1723                 char errbuf[BUFSIZ];
1724
1725                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
1726                 pr_err("ERROR: Apply config to BPF failed: %s\n",
1727                          errbuf);
1728                 goto out_child;
1729         }
1730
1731         /*
1732          * Normally perf_session__new would do this, but it doesn't have the
1733          * evlist.
1734          */
1735         if (rec->tool.ordered_events && !evlist__sample_id_all(rec->evlist)) {
1736                 pr_warning("WARNING: No sample_id_all support, falling back to unordered processing\n");
1737                 rec->tool.ordered_events = false;
1738         }
1739
1740         if (!rec->evlist->nr_groups)
1741                 perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
1742
1743         if (data->is_pipe) {
1744                 err = perf_header__write_pipe(fd);
1745                 if (err < 0)
1746                         goto out_child;
1747         } else {
1748                 err = perf_session__write_header(session, rec->evlist, fd, false);
1749                 if (err < 0)
1750                         goto out_child;
1751         }
1752
1753         err = -1;
1754         if (!rec->no_buildid
1755             && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
1756                 pr_err("Couldn't generate buildids. "
1757                        "Use --no-buildid to profile anyway.\n");
1758                 goto out_child;
1759         }
1760
1761         err = record__setup_sb_evlist(rec);
1762         if (err)
1763                 goto out_child;
1764
1765         err = record__synthesize(rec, false);
1766         if (err < 0)
1767                 goto out_child;
1768
1769         if (rec->realtime_prio) {
1770                 struct sched_param param;
1771
1772                 param.sched_priority = rec->realtime_prio;
1773                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1774                         pr_err("Could not set realtime priority.\n");
1775                         err = -1;
1776                         goto out_child;
1777                 }
1778         }
1779
1780         /*
1781          * When perf is starting the traced process, all the events
1782          * (apart from group members) have enable_on_exec=1 set,
1783          * so don't spoil it by prematurely enabling them.
1784          */
1785         if (!target__none(&opts->target) && !opts->initial_delay)
1786                 evlist__enable(rec->evlist);
1787
1788         /*
1789          * Let the child rip
1790          */
1791         if (forks) {
1792                 struct machine *machine = &session->machines.host;
1793                 union perf_event *event;
1794                 pid_t tgid;
1795
1796                 event = malloc(sizeof(event->comm) + machine->id_hdr_size);
1797                 if (event == NULL) {
1798                         err = -ENOMEM;
1799                         goto out_child;
1800                 }
1801
1802                 /*
1803                  * Some H/W events are generated before COMM event
1804                  * which is emitted during exec(), so perf script
1805                  * cannot see a correct process name for those events.
1806                  * Synthesize COMM event to prevent it.
1807                  */
1808                 tgid = perf_event__synthesize_comm(tool, event,
1809                                                    rec->evlist->workload.pid,
1810                                                    process_synthesized_event,
1811                                                    machine);
1812                 free(event);
1813
1814                 if (tgid == -1)
1815                         goto out_child;
1816
1817                 event = malloc(sizeof(event->namespaces) +
1818                                (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
1819                                machine->id_hdr_size);
1820                 if (event == NULL) {
1821                         err = -ENOMEM;
1822                         goto out_child;
1823                 }
1824
1825                 /*
1826                  * Synthesize NAMESPACES event for the command specified.
1827                  */
1828                 perf_event__synthesize_namespaces(tool, event,
1829                                                   rec->evlist->workload.pid,
1830                                                   tgid, process_synthesized_event,
1831                                                   machine);
1832                 free(event);
1833
1834                 evlist__start_workload(rec->evlist);
1835         }
1836
1837         if (evlist__initialize_ctlfd(rec->evlist, opts->ctl_fd, opts->ctl_fd_ack))
1838                 goto out_child;
1839
1840         if (opts->initial_delay) {
1841                 pr_info(EVLIST_DISABLED_MSG);
1842                 if (opts->initial_delay > 0) {
1843                         usleep(opts->initial_delay * USEC_PER_MSEC);
1844                         evlist__enable(rec->evlist);
1845                         pr_info(EVLIST_ENABLED_MSG);
1846                 }
1847         }
1848
1849         trigger_ready(&auxtrace_snapshot_trigger);
1850         trigger_ready(&switch_output_trigger);
1851         perf_hooks__invoke_record_start();
1852         for (;;) {
1853                 unsigned long long hits = rec->samples;
1854
1855                 /*
1856                  * rec->evlist->bkw_mmap_state is possible to be
1857                  * BKW_MMAP_EMPTY here: when done == true and
1858                  * hits != rec->samples in previous round.
1859                  *
1860                  * evlist__toggle_bkw_mmap ensure we never
1861                  * convert BKW_MMAP_EMPTY to BKW_MMAP_DATA_PENDING.
1862                  */
1863                 if (trigger_is_hit(&switch_output_trigger) || done || draining)
1864                         evlist__toggle_bkw_mmap(rec->evlist, BKW_MMAP_DATA_PENDING);
1865
1866                 if (record__mmap_read_all(rec, false) < 0) {
1867                         trigger_error(&auxtrace_snapshot_trigger);
1868                         trigger_error(&switch_output_trigger);
1869                         err = -1;
1870                         goto out_child;
1871                 }
1872
1873                 if (auxtrace_record__snapshot_started) {
1874                         auxtrace_record__snapshot_started = 0;
1875                         if (!trigger_is_error(&auxtrace_snapshot_trigger))
1876                                 record__read_auxtrace_snapshot(rec, false);
1877                         if (trigger_is_error(&auxtrace_snapshot_trigger)) {
1878                                 pr_err("AUX area tracing snapshot failed\n");
1879                                 err = -1;
1880                                 goto out_child;
1881                         }
1882                 }
1883
1884                 if (trigger_is_hit(&switch_output_trigger)) {
1885                         /*
1886                          * If switch_output_trigger is hit, the data in
1887                          * overwritable ring buffer should have been collected,
1888                          * so bkw_mmap_state should be set to BKW_MMAP_EMPTY.
1889                          *
1890                          * If SIGUSR2 raise after or during record__mmap_read_all(),
1891                          * record__mmap_read_all() didn't collect data from
1892                          * overwritable ring buffer. Read again.
1893                          */
1894                         if (rec->evlist->bkw_mmap_state == BKW_MMAP_RUNNING)
1895                                 continue;
1896                         trigger_ready(&switch_output_trigger);
1897
1898                         /*
1899                          * Reenable events in overwrite ring buffer after
1900                          * record__mmap_read_all(): we should have collected
1901                          * data from it.
1902                          */
1903                         evlist__toggle_bkw_mmap(rec->evlist, BKW_MMAP_RUNNING);
1904
1905                         if (!quiet)
1906                                 fprintf(stderr, "[ perf record: dump data: Woken up %ld times ]\n",
1907                                         waking);
1908                         waking = 0;
1909                         fd = record__switch_output(rec, false);
1910                         if (fd < 0) {
1911                                 pr_err("Failed to switch to new file\n");
1912                                 trigger_error(&switch_output_trigger);
1913                                 err = fd;
1914                                 goto out_child;
1915                         }
1916
1917                         /* re-arm the alarm */
1918                         if (rec->switch_output.time)
1919                                 alarm(rec->switch_output.time);
1920                 }
1921
1922                 if (hits == rec->samples) {
1923                         if (done || draining)
1924                                 break;
1925                         err = evlist__poll(rec->evlist, -1);
1926                         /*
1927                          * Propagate error, only if there's any. Ignore positive
1928                          * number of returned events and interrupt error.
1929                          */
1930                         if (err > 0 || (err < 0 && errno == EINTR))
1931                                 err = 0;
1932                         waking++;
1933
1934                         if (evlist__filter_pollfd(rec->evlist, POLLERR | POLLHUP) == 0)
1935                                 draining = true;
1936                 }
1937
1938                 if (evlist__ctlfd_process(rec->evlist, &cmd) > 0) {
1939                         switch (cmd) {
1940                         case EVLIST_CTL_CMD_ENABLE:
1941                                 pr_info(EVLIST_ENABLED_MSG);
1942                                 break;
1943                         case EVLIST_CTL_CMD_DISABLE:
1944                                 pr_info(EVLIST_DISABLED_MSG);
1945                                 break;
1946                         case EVLIST_CTL_CMD_SNAPSHOT:
1947                                 hit_auxtrace_snapshot_trigger(rec);
1948                                 evlist__ctlfd_ack(rec->evlist);
1949                                 break;
1950                         case EVLIST_CTL_CMD_ACK:
1951                         case EVLIST_CTL_CMD_UNSUPPORTED:
1952                         default:
1953                                 break;
1954                         }
1955                 }
1956
1957                 /*
1958                  * When perf is starting the traced process, at the end events
1959                  * die with the process and we wait for that. Thus no need to
1960                  * disable events in this case.
1961                  */
1962                 if (done && !disabled && !target__none(&opts->target)) {
1963                         trigger_off(&auxtrace_snapshot_trigger);
1964                         evlist__disable(rec->evlist);
1965                         disabled = true;
1966                 }
1967         }
1968
1969         trigger_off(&auxtrace_snapshot_trigger);
1970         trigger_off(&switch_output_trigger);
1971
1972         if (opts->auxtrace_snapshot_on_exit)
1973                 record__auxtrace_snapshot_exit(rec);
1974
1975         if (forks && workload_exec_errno) {
1976                 char msg[STRERR_BUFSIZE];
1977                 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
1978                 pr_err("Workload failed: %s\n", emsg);
1979                 err = -1;
1980                 goto out_child;
1981         }
1982
1983         if (!quiet)
1984                 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
1985
1986         if (target__none(&rec->opts.target))
1987                 record__synthesize_workload(rec, true);
1988
1989 out_child:
1990         evlist__finalize_ctlfd(rec->evlist);
1991         record__mmap_read_all(rec, true);
1992         record__aio_mmap_read_sync(rec);
1993
1994         if (rec->session->bytes_transferred && rec->session->bytes_compressed) {
1995                 ratio = (float)rec->session->bytes_transferred/(float)rec->session->bytes_compressed;
1996                 session->header.env.comp_ratio = ratio + 0.5;
1997         }
1998
1999         if (forks) {
2000                 int exit_status;
2001
2002                 if (!child_finished)
2003                         kill(rec->evlist->workload.pid, SIGTERM);
2004
2005                 wait(&exit_status);
2006
2007                 if (err < 0)
2008                         status = err;
2009                 else if (WIFEXITED(exit_status))
2010                         status = WEXITSTATUS(exit_status);
2011                 else if (WIFSIGNALED(exit_status))
2012                         signr = WTERMSIG(exit_status);
2013         } else
2014                 status = err;
2015
2016         record__synthesize(rec, true);
2017         /* this will be recalculated during process_buildids() */
2018         rec->samples = 0;
2019
2020         if (!err) {
2021                 if (!rec->timestamp_filename) {
2022                         record__finish_output(rec);
2023                 } else {
2024                         fd = record__switch_output(rec, true);
2025                         if (fd < 0) {
2026                                 status = fd;
2027                                 goto out_delete_session;
2028                         }
2029                 }
2030         }
2031
2032         perf_hooks__invoke_record_end();
2033
2034         if (!err && !quiet) {
2035                 char samples[128];
2036                 const char *postfix = rec->timestamp_filename ?
2037                                         ".<timestamp>" : "";
2038
2039                 if (rec->samples && !rec->opts.full_auxtrace)
2040                         scnprintf(samples, sizeof(samples),
2041                                   " (%" PRIu64 " samples)", rec->samples);
2042                 else
2043                         samples[0] = '\0';
2044
2045                 fprintf(stderr, "[ perf record: Captured and wrote %.3f MB %s%s%s",
2046                         perf_data__size(data) / 1024.0 / 1024.0,
2047                         data->path, postfix, samples);
2048                 if (ratio) {
2049                         fprintf(stderr, ", compressed (original %.3f MB, ratio is %.3f)",
2050                                         rec->session->bytes_transferred / 1024.0 / 1024.0,
2051                                         ratio);
2052                 }
2053                 fprintf(stderr, " ]\n");
2054         }
2055
2056 out_delete_session:
2057 #ifdef HAVE_EVENTFD_SUPPORT
2058         if (done_fd >= 0)
2059                 close(done_fd);
2060 #endif
2061         zstd_fini(&session->zstd_data);
2062         perf_session__delete(session);
2063
2064         if (!opts->no_bpf_event)
2065                 evlist__stop_sb_thread(rec->sb_evlist);
2066         return status;
2067 }
2068
2069 static void callchain_debug(struct callchain_param *callchain)
2070 {
2071         static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF", "LBR" };
2072
2073         pr_debug("callchain: type %s\n", str[callchain->record_mode]);
2074
2075         if (callchain->record_mode == CALLCHAIN_DWARF)
2076                 pr_debug("callchain: stack dump size %d\n",
2077                          callchain->dump_size);
2078 }
2079
2080 int record_opts__parse_callchain(struct record_opts *record,
2081                                  struct callchain_param *callchain,
2082                                  const char *arg, bool unset)
2083 {
2084         int ret;
2085         callchain->enabled = !unset;
2086
2087         /* --no-call-graph */
2088         if (unset) {
2089                 callchain->record_mode = CALLCHAIN_NONE;
2090                 pr_debug("callchain: disabled\n");
2091                 return 0;
2092         }
2093
2094         ret = parse_callchain_record_opt(arg, callchain);
2095         if (!ret) {
2096                 /* Enable data address sampling for DWARF unwind. */
2097                 if (callchain->record_mode == CALLCHAIN_DWARF)
2098                         record->sample_address = true;
2099                 callchain_debug(callchain);
2100         }
2101
2102         return ret;
2103 }
2104
2105 int record_parse_callchain_opt(const struct option *opt,
2106                                const char *arg,
2107                                int unset)
2108 {
2109         return record_opts__parse_callchain(opt->value, &callchain_param, arg, unset);
2110 }
2111
2112 int record_callchain_opt(const struct option *opt,
2113                          const char *arg __maybe_unused,
2114                          int unset __maybe_unused)
2115 {
2116         struct callchain_param *callchain = opt->value;
2117
2118         callchain->enabled = true;
2119
2120         if (callchain->record_mode == CALLCHAIN_NONE)
2121                 callchain->record_mode = CALLCHAIN_FP;
2122
2123         callchain_debug(callchain);
2124         return 0;
2125 }
2126
2127 static int perf_record_config(const char *var, const char *value, void *cb)
2128 {
2129         struct record *rec = cb;
2130
2131         if (!strcmp(var, "record.build-id")) {
2132                 if (!strcmp(value, "cache"))
2133                         rec->no_buildid_cache = false;
2134                 else if (!strcmp(value, "no-cache"))
2135                         rec->no_buildid_cache = true;
2136                 else if (!strcmp(value, "skip"))
2137                         rec->no_buildid = true;
2138                 else
2139                         return -1;
2140                 return 0;
2141         }
2142         if (!strcmp(var, "record.call-graph")) {
2143                 var = "call-graph.record-mode";
2144                 return perf_default_config(var, value, cb);
2145         }
2146 #ifdef HAVE_AIO_SUPPORT
2147         if (!strcmp(var, "record.aio")) {
2148                 rec->opts.nr_cblocks = strtol(value, NULL, 0);
2149                 if (!rec->opts.nr_cblocks)
2150                         rec->opts.nr_cblocks = nr_cblocks_default;
2151         }
2152 #endif
2153
2154         return 0;
2155 }
2156
2157
2158 static int record__parse_affinity(const struct option *opt, const char *str, int unset)
2159 {
2160         struct record_opts *opts = (struct record_opts *)opt->value;
2161
2162         if (unset || !str)
2163                 return 0;
2164
2165         if (!strcasecmp(str, "node"))
2166                 opts->affinity = PERF_AFFINITY_NODE;
2167         else if (!strcasecmp(str, "cpu"))
2168                 opts->affinity = PERF_AFFINITY_CPU;
2169
2170         return 0;
2171 }
2172
2173 static int parse_output_max_size(const struct option *opt,
2174                                  const char *str, int unset)
2175 {
2176         unsigned long *s = (unsigned long *)opt->value;
2177         static struct parse_tag tags_size[] = {
2178                 { .tag  = 'B', .mult = 1       },
2179                 { .tag  = 'K', .mult = 1 << 10 },
2180                 { .tag  = 'M', .mult = 1 << 20 },
2181                 { .tag  = 'G', .mult = 1 << 30 },
2182                 { .tag  = 0 },
2183         };
2184         unsigned long val;
2185
2186         if (unset) {
2187                 *s = 0;
2188                 return 0;
2189         }
2190
2191         val = parse_tag_value(str, tags_size);
2192         if (val != (unsigned long) -1) {
2193                 *s = val;
2194                 return 0;
2195         }
2196
2197         return -1;
2198 }
2199
2200 static int record__parse_mmap_pages(const struct option *opt,
2201                                     const char *str,
2202                                     int unset __maybe_unused)
2203 {
2204         struct record_opts *opts = opt->value;
2205         char *s, *p;
2206         unsigned int mmap_pages;
2207         int ret;
2208
2209         if (!str)
2210                 return -EINVAL;
2211
2212         s = strdup(str);
2213         if (!s)
2214                 return -ENOMEM;
2215
2216         p = strchr(s, ',');
2217         if (p)
2218                 *p = '\0';
2219
2220         if (*s) {
2221                 ret = __evlist__parse_mmap_pages(&mmap_pages, s);
2222                 if (ret)
2223                         goto out_free;
2224                 opts->mmap_pages = mmap_pages;
2225         }
2226
2227         if (!p) {
2228                 ret = 0;
2229                 goto out_free;
2230         }
2231
2232         ret = __evlist__parse_mmap_pages(&mmap_pages, p + 1);
2233         if (ret)
2234                 goto out_free;
2235
2236         opts->auxtrace_mmap_pages = mmap_pages;
2237
2238 out_free:
2239         free(s);
2240         return ret;
2241 }
2242
2243 static int parse_control_option(const struct option *opt,
2244                                 const char *str,
2245                                 int unset __maybe_unused)
2246 {
2247         struct record_opts *opts = opt->value;
2248
2249         return evlist__parse_control(str, &opts->ctl_fd, &opts->ctl_fd_ack, &opts->ctl_fd_close);
2250 }
2251
2252 static void switch_output_size_warn(struct record *rec)
2253 {
2254         u64 wakeup_size = evlist__mmap_size(rec->opts.mmap_pages);
2255         struct switch_output *s = &rec->switch_output;
2256
2257         wakeup_size /= 2;
2258
2259         if (s->size < wakeup_size) {
2260                 char buf[100];
2261
2262                 unit_number__scnprintf(buf, sizeof(buf), wakeup_size);
2263                 pr_warning("WARNING: switch-output data size lower than "
2264                            "wakeup kernel buffer size (%s) "
2265                            "expect bigger perf.data sizes\n", buf);
2266         }
2267 }
2268
2269 static int switch_output_setup(struct record *rec)
2270 {
2271         struct switch_output *s = &rec->switch_output;
2272         static struct parse_tag tags_size[] = {
2273                 { .tag  = 'B', .mult = 1       },
2274                 { .tag  = 'K', .mult = 1 << 10 },
2275                 { .tag  = 'M', .mult = 1 << 20 },
2276                 { .tag  = 'G', .mult = 1 << 30 },
2277                 { .tag  = 0 },
2278         };
2279         static struct parse_tag tags_time[] = {
2280                 { .tag  = 's', .mult = 1        },
2281                 { .tag  = 'm', .mult = 60       },
2282                 { .tag  = 'h', .mult = 60*60    },
2283                 { .tag  = 'd', .mult = 60*60*24 },
2284                 { .tag  = 0 },
2285         };
2286         unsigned long val;
2287
2288         /*
2289          * If we're using --switch-output-events, then we imply its
2290          * --switch-output=signal, as we'll send a SIGUSR2 from the side band
2291          *  thread to its parent.
2292          */
2293         if (rec->switch_output_event_set)
2294                 goto do_signal;
2295
2296         if (!s->set)
2297                 return 0;
2298
2299         if (!strcmp(s->str, "signal")) {
2300 do_signal:
2301                 s->signal = true;
2302                 pr_debug("switch-output with SIGUSR2 signal\n");
2303                 goto enabled;
2304         }
2305
2306         val = parse_tag_value(s->str, tags_size);
2307         if (val != (unsigned long) -1) {
2308                 s->size = val;
2309                 pr_debug("switch-output with %s size threshold\n", s->str);
2310                 goto enabled;
2311         }
2312
2313         val = parse_tag_value(s->str, tags_time);
2314         if (val != (unsigned long) -1) {
2315                 s->time = val;
2316                 pr_debug("switch-output with %s time threshold (%lu seconds)\n",
2317                          s->str, s->time);
2318                 goto enabled;
2319         }
2320
2321         return -1;
2322
2323 enabled:
2324         rec->timestamp_filename = true;
2325         s->enabled              = true;
2326
2327         if (s->size && !rec->opts.no_buffering)
2328                 switch_output_size_warn(rec);
2329
2330         return 0;
2331 }
2332
2333 static const char * const __record_usage[] = {
2334         "perf record [<options>] [<command>]",
2335         "perf record [<options>] -- <command> [<options>]",
2336         NULL
2337 };
2338 const char * const *record_usage = __record_usage;
2339
2340 static int build_id__process_mmap(struct perf_tool *tool, union perf_event *event,
2341                                   struct perf_sample *sample, struct machine *machine)
2342 {
2343         /*
2344          * We already have the kernel maps, put in place via perf_session__create_kernel_maps()
2345          * no need to add them twice.
2346          */
2347         if (!(event->header.misc & PERF_RECORD_MISC_USER))
2348                 return 0;
2349         return perf_event__process_mmap(tool, event, sample, machine);
2350 }
2351
2352 static int build_id__process_mmap2(struct perf_tool *tool, union perf_event *event,
2353                                    struct perf_sample *sample, struct machine *machine)
2354 {
2355         /*
2356          * We already have the kernel maps, put in place via perf_session__create_kernel_maps()
2357          * no need to add them twice.
2358          */
2359         if (!(event->header.misc & PERF_RECORD_MISC_USER))
2360                 return 0;
2361
2362         return perf_event__process_mmap2(tool, event, sample, machine);
2363 }
2364
2365 /*
2366  * XXX Ideally would be local to cmd_record() and passed to a record__new
2367  * because we need to have access to it in record__exit, that is called
2368  * after cmd_record() exits, but since record_options need to be accessible to
2369  * builtin-script, leave it here.
2370  *
2371  * At least we don't ouch it in all the other functions here directly.
2372  *
2373  * Just say no to tons of global variables, sigh.
2374  */
2375 static struct record record = {
2376         .opts = {
2377                 .sample_time         = true,
2378                 .mmap_pages          = UINT_MAX,
2379                 .user_freq           = UINT_MAX,
2380                 .user_interval       = ULLONG_MAX,
2381                 .freq                = 4000,
2382                 .target              = {
2383                         .uses_mmap   = true,
2384                         .default_per_cpu = true,
2385                 },
2386                 .mmap_flush          = MMAP_FLUSH_DEFAULT,
2387                 .nr_threads_synthesize = 1,
2388                 .ctl_fd              = -1,
2389                 .ctl_fd_ack          = -1,
2390         },
2391         .tool = {
2392                 .sample         = process_sample_event,
2393                 .fork           = perf_event__process_fork,
2394                 .exit           = perf_event__process_exit,
2395                 .comm           = perf_event__process_comm,
2396                 .namespaces     = perf_event__process_namespaces,
2397                 .mmap           = build_id__process_mmap,
2398                 .mmap2          = build_id__process_mmap2,
2399                 .ordered_events = true,
2400         },
2401 };
2402
2403 const char record_callchain_help[] = CALLCHAIN_RECORD_HELP
2404         "\n\t\t\t\tDefault: fp";
2405
2406 static bool dry_run;
2407
2408 /*
2409  * XXX Will stay a global variable till we fix builtin-script.c to stop messing
2410  * with it and switch to use the library functions in perf_evlist that came
2411  * from builtin-record.c, i.e. use record_opts,
2412  * evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
2413  * using pipes, etc.
2414  */
2415 static struct option __record_options[] = {
2416         OPT_CALLBACK('e', "event", &record.evlist, "event",
2417                      "event selector. use 'perf list' to list available events",
2418                      parse_events_option),
2419         OPT_CALLBACK(0, "filter", &record.evlist, "filter",
2420                      "event filter", parse_filter),
2421         OPT_CALLBACK_NOOPT(0, "exclude-perf", &record.evlist,
2422                            NULL, "don't record events from perf itself",
2423                            exclude_perf),
2424         OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
2425                     "record events on existing process id"),
2426         OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
2427                     "record events on existing thread id"),
2428         OPT_INTEGER('r', "realtime", &record.realtime_prio,
2429                     "collect data with this RT SCHED_FIFO priority"),
2430         OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
2431                     "collect data without buffering"),
2432         OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
2433                     "collect raw sample records from all opened counters"),
2434         OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
2435                             "system-wide collection from all CPUs"),
2436         OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
2437                     "list of cpus to monitor"),
2438         OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
2439         OPT_STRING('o', "output", &record.data.path, "file",
2440                     "output file name"),
2441         OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
2442                         &record.opts.no_inherit_set,
2443                         "child tasks do not inherit counters"),
2444         OPT_BOOLEAN(0, "tail-synthesize", &record.opts.tail_synthesize,
2445                     "synthesize non-sample events at the end of output"),
2446         OPT_BOOLEAN(0, "overwrite", &record.opts.overwrite, "use overwrite mode"),
2447         OPT_BOOLEAN(0, "no-bpf-event", &record.opts.no_bpf_event, "do not record bpf events"),
2448         OPT_BOOLEAN(0, "strict-freq", &record.opts.strict_freq,
2449                     "Fail if the specified frequency can't be used"),
2450         OPT_CALLBACK('F', "freq", &record.opts, "freq or 'max'",
2451                      "profile at this frequency",
2452                       record__parse_freq),
2453         OPT_CALLBACK('m', "mmap-pages", &record.opts, "pages[,pages]",
2454                      "number of mmap data pages and AUX area tracing mmap pages",
2455                      record__parse_mmap_pages),
2456         OPT_CALLBACK(0, "mmap-flush", &record.opts, "number",
2457                      "Minimal number of bytes that is extracted from mmap data pages (default: 1)",
2458                      record__mmap_flush_parse),
2459         OPT_BOOLEAN(0, "group", &record.opts.group,
2460                     "put the counters into a counter group"),
2461         OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
2462                            NULL, "enables call-graph recording" ,
2463                            &record_callchain_opt),
2464         OPT_CALLBACK(0, "call-graph", &record.opts,
2465                      "record_mode[,record_size]", record_callchain_help,
2466                      &record_parse_callchain_opt),
2467         OPT_INCR('v', "verbose", &verbose,
2468                     "be more verbose (show counter open errors, etc)"),
2469         OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
2470         OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
2471                     "per thread counts"),
2472         OPT_BOOLEAN('d', "data", &record.opts.sample_address, "Record the sample addresses"),
2473         OPT_BOOLEAN(0, "phys-data", &record.opts.sample_phys_addr,
2474                     "Record the sample physical addresses"),
2475         OPT_BOOLEAN(0, "data-page-size", &record.opts.sample_data_page_size,
2476                     "Record the sampled data address data page size"),
2477         OPT_BOOLEAN(0, "sample-cpu", &record.opts.sample_cpu, "Record the sample cpu"),
2478         OPT_BOOLEAN_SET('T', "timestamp", &record.opts.sample_time,
2479                         &record.opts.sample_time_set,
2480                         "Record the sample timestamps"),
2481         OPT_BOOLEAN_SET('P', "period", &record.opts.period, &record.opts.period_set,
2482                         "Record the sample period"),
2483         OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
2484                     "don't sample"),
2485         OPT_BOOLEAN_SET('N', "no-buildid-cache", &record.no_buildid_cache,
2486                         &record.no_buildid_cache_set,
2487                         "do not update the buildid cache"),
2488         OPT_BOOLEAN_SET('B', "no-buildid", &record.no_buildid,
2489                         &record.no_buildid_set,
2490                         "do not collect buildids in perf.data"),
2491         OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
2492                      "monitor event in cgroup name only",
2493                      parse_cgroups),
2494         OPT_INTEGER('D', "delay", &record.opts.initial_delay,
2495                   "ms to wait before starting measurement after program start (-1: start with events disabled)"),
2496         OPT_BOOLEAN(0, "kcore", &record.opts.kcore, "copy /proc/kcore"),
2497         OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
2498                    "user to profile"),
2499
2500         OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
2501                      "branch any", "sample any taken branches",
2502                      parse_branch_stack),
2503
2504         OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
2505                      "branch filter mask", "branch stack filter modes",
2506                      parse_branch_stack),
2507         OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
2508                     "sample by weight (on special events only)"),
2509         OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
2510                     "sample transaction flags (special events only)"),
2511         OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
2512                     "use per-thread mmaps"),
2513         OPT_CALLBACK_OPTARG('I', "intr-regs", &record.opts.sample_intr_regs, NULL, "any register",
2514                     "sample selected machine registers on interrupt,"
2515                     " use '-I?' to list register names", parse_intr_regs),
2516         OPT_CALLBACK_OPTARG(0, "user-regs", &record.opts.sample_user_regs, NULL, "any register",
2517                     "sample selected machine registers on interrupt,"
2518                     " use '--user-regs=?' to list register names", parse_user_regs),
2519         OPT_BOOLEAN(0, "running-time", &record.opts.running_time,
2520                     "Record running/enabled time of read (:S) events"),
2521         OPT_CALLBACK('k', "clockid", &record.opts,
2522         "clockid", "clockid to use for events, see clock_gettime()",
2523         parse_clockid),
2524         OPT_STRING_OPTARG('S', "snapshot", &record.opts.auxtrace_snapshot_opts,
2525                           "opts", "AUX area tracing Snapshot Mode", ""),
2526         OPT_STRING_OPTARG(0, "aux-sample", &record.opts.auxtrace_sample_opts,
2527                           "opts", "sample AUX area", ""),
2528         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
2529                         "per thread proc mmap processing timeout in ms"),
2530         OPT_BOOLEAN(0, "namespaces", &record.opts.record_namespaces,
2531                     "Record namespaces events"),
2532         OPT_BOOLEAN(0, "all-cgroups", &record.opts.record_cgroup,
2533                     "Record cgroup events"),
2534         OPT_BOOLEAN_SET(0, "switch-events", &record.opts.record_switch_events,
2535                         &record.opts.record_switch_events_set,
2536                         "Record context switch events"),
2537         OPT_BOOLEAN_FLAG(0, "all-kernel", &record.opts.all_kernel,
2538                          "Configure all used events to run in kernel space.",
2539                          PARSE_OPT_EXCLUSIVE),
2540         OPT_BOOLEAN_FLAG(0, "all-user", &record.opts.all_user,
2541                          "Configure all used events to run in user space.",
2542                          PARSE_OPT_EXCLUSIVE),
2543         OPT_BOOLEAN(0, "kernel-callchains", &record.opts.kernel_callchains,
2544                     "collect kernel callchains"),
2545         OPT_BOOLEAN(0, "user-callchains", &record.opts.user_callchains,
2546                     "collect user callchains"),
2547         OPT_STRING(0, "clang-path", &llvm_param.clang_path, "clang path",
2548                    "clang binary to use for compiling BPF scriptlets"),
2549         OPT_STRING(0, "clang-opt", &llvm_param.clang_opt, "clang options",
2550                    "options passed to clang when compiling BPF scriptlets"),
2551         OPT_STRING(0, "vmlinux", &symbol_conf.vmlinux_name,
2552                    "file", "vmlinux pathname"),
2553         OPT_BOOLEAN(0, "buildid-all", &record.buildid_all,
2554                     "Record build-id of all DSOs regardless of hits"),
2555         OPT_BOOLEAN(0, "timestamp-filename", &record.timestamp_filename,
2556                     "append timestamp to output filename"),
2557         OPT_BOOLEAN(0, "timestamp-boundary", &record.timestamp_boundary,
2558                     "Record timestamp boundary (time of first/last samples)"),
2559         OPT_STRING_OPTARG_SET(0, "switch-output", &record.switch_output.str,
2560                           &record.switch_output.set, "signal or size[BKMG] or time[smhd]",
2561                           "Switch output when receiving SIGUSR2 (signal) or cross a size or time threshold",
2562                           "signal"),
2563         OPT_CALLBACK_SET(0, "switch-output-event", &record.sb_evlist, &record.switch_output_event_set, "switch output event",
2564                          "switch output event selector. use 'perf list' to list available events",
2565                          parse_events_option_new_evlist),
2566         OPT_INTEGER(0, "switch-max-files", &record.switch_output.num_files,
2567                    "Limit number of switch output generated files"),
2568         OPT_BOOLEAN(0, "dry-run", &dry_run,
2569                     "Parse options then exit"),
2570 #ifdef HAVE_AIO_SUPPORT
2571         OPT_CALLBACK_OPTARG(0, "aio", &record.opts,
2572                      &nr_cblocks_default, "n", "Use <n> control blocks in asynchronous trace writing mode (default: 1, max: 4)",
2573                      record__aio_parse),
2574 #endif
2575         OPT_CALLBACK(0, "affinity", &record.opts, "node|cpu",
2576                      "Set affinity mask of trace reading thread to NUMA node cpu mask or cpu of processed mmap buffer",
2577                      record__parse_affinity),
2578 #ifdef HAVE_ZSTD_SUPPORT
2579         OPT_CALLBACK_OPTARG('z', "compression-level", &record.opts, &comp_level_default,
2580                             "n", "Compressed records using specified level (default: 1 - fastest compression, 22 - greatest compression)",
2581                             record__parse_comp_level),
2582 #endif
2583         OPT_CALLBACK(0, "max-size", &record.output_max_size,
2584                      "size", "Limit the maximum size of the output file", parse_output_max_size),
2585         OPT_UINTEGER(0, "num-thread-synthesize",
2586                      &record.opts.nr_threads_synthesize,
2587                      "number of threads to run for event synthesis"),
2588 #ifdef HAVE_LIBPFM
2589         OPT_CALLBACK(0, "pfm-events", &record.evlist, "event",
2590                 "libpfm4 event selector. use 'perf list' to list available events",
2591                 parse_libpfm_events_option),
2592 #endif
2593         OPT_CALLBACK(0, "control", &record.opts, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
2594                      "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events,\n"
2595                      "\t\t\t  'snapshot': AUX area tracing snapshot).\n"
2596                      "\t\t\t  Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
2597                      "\t\t\t  Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
2598                       parse_control_option),
2599         OPT_END()
2600 };
2601
2602 struct option *record_options = __record_options;
2603
2604 int cmd_record(int argc, const char **argv)
2605 {
2606         int err;
2607         struct record *rec = &record;
2608         char errbuf[BUFSIZ];
2609
2610         setlocale(LC_ALL, "");
2611
2612 #ifndef HAVE_LIBBPF_SUPPORT
2613 # define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, "NO_LIBBPF=1", c)
2614         set_nobuild('\0', "clang-path", true);
2615         set_nobuild('\0', "clang-opt", true);
2616 # undef set_nobuild
2617 #endif
2618
2619 #ifndef HAVE_BPF_PROLOGUE
2620 # if !defined (HAVE_DWARF_SUPPORT)
2621 #  define REASON  "NO_DWARF=1"
2622 # elif !defined (HAVE_LIBBPF_SUPPORT)
2623 #  define REASON  "NO_LIBBPF=1"
2624 # else
2625 #  define REASON  "this architecture doesn't support BPF prologue"
2626 # endif
2627 # define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, REASON, c)
2628         set_nobuild('\0', "vmlinux", true);
2629 # undef set_nobuild
2630 # undef REASON
2631 #endif
2632
2633         rec->opts.affinity = PERF_AFFINITY_SYS;
2634
2635         rec->evlist = evlist__new();
2636         if (rec->evlist == NULL)
2637                 return -ENOMEM;
2638
2639         err = perf_config(perf_record_config, rec);
2640         if (err)
2641                 return err;
2642
2643         argc = parse_options(argc, argv, record_options, record_usage,
2644                             PARSE_OPT_STOP_AT_NON_OPTION);
2645         if (quiet)
2646                 perf_quiet_option();
2647
2648         /* Make system wide (-a) the default target. */
2649         if (!argc && target__none(&rec->opts.target))
2650                 rec->opts.target.system_wide = true;
2651
2652         if (nr_cgroups && !rec->opts.target.system_wide) {
2653                 usage_with_options_msg(record_usage, record_options,
2654                         "cgroup monitoring only available in system-wide mode");
2655
2656         }
2657
2658         if (rec->opts.kcore)
2659                 rec->data.is_dir = true;
2660
2661         if (rec->opts.comp_level != 0) {
2662                 pr_debug("Compression enabled, disabling build id collection at the end of the session.\n");
2663                 rec->no_buildid = true;
2664         }
2665
2666         if (rec->opts.record_switch_events &&
2667             !perf_can_record_switch_events()) {
2668                 ui__error("kernel does not support recording context switch events\n");
2669                 parse_options_usage(record_usage, record_options, "switch-events", 0);
2670                 err = -EINVAL;
2671                 goto out_opts;
2672         }
2673
2674         if (switch_output_setup(rec)) {
2675                 parse_options_usage(record_usage, record_options, "switch-output", 0);
2676                 err = -EINVAL;
2677                 goto out_opts;
2678         }
2679
2680         if (rec->switch_output.time) {
2681                 signal(SIGALRM, alarm_sig_handler);
2682                 alarm(rec->switch_output.time);
2683         }
2684
2685         if (rec->switch_output.num_files) {
2686                 rec->switch_output.filenames = calloc(sizeof(char *),
2687                                                       rec->switch_output.num_files);
2688                 if (!rec->switch_output.filenames) {
2689                         err = -EINVAL;
2690                         goto out_opts;
2691                 }
2692         }
2693
2694         /*
2695          * Allow aliases to facilitate the lookup of symbols for address
2696          * filters. Refer to auxtrace_parse_filters().
2697          */
2698         symbol_conf.allow_aliases = true;
2699
2700         symbol__init(NULL);
2701
2702         if (rec->opts.affinity != PERF_AFFINITY_SYS) {
2703                 rec->affinity_mask.nbits = cpu__max_cpu();
2704                 rec->affinity_mask.bits = bitmap_alloc(rec->affinity_mask.nbits);
2705                 if (!rec->affinity_mask.bits) {
2706                         pr_err("Failed to allocate thread mask for %zd cpus\n", rec->affinity_mask.nbits);
2707                         err = -ENOMEM;
2708                         goto out_opts;
2709                 }
2710                 pr_debug2("thread mask[%zd]: empty\n", rec->affinity_mask.nbits);
2711         }
2712
2713         err = record__auxtrace_init(rec);
2714         if (err)
2715                 goto out;
2716
2717         if (dry_run)
2718                 goto out;
2719
2720         err = bpf__setup_stdout(rec->evlist);
2721         if (err) {
2722                 bpf__strerror_setup_stdout(rec->evlist, err, errbuf, sizeof(errbuf));
2723                 pr_err("ERROR: Setup BPF stdout failed: %s\n",
2724                          errbuf);
2725                 goto out;
2726         }
2727
2728         err = -ENOMEM;
2729
2730         if (rec->no_buildid_cache || rec->no_buildid) {
2731                 disable_buildid_cache();
2732         } else if (rec->switch_output.enabled) {
2733                 /*
2734                  * In 'perf record --switch-output', disable buildid
2735                  * generation by default to reduce data file switching
2736                  * overhead. Still generate buildid if they are required
2737                  * explicitly using
2738                  *
2739                  *  perf record --switch-output --no-no-buildid \
2740                  *              --no-no-buildid-cache
2741                  *
2742                  * Following code equals to:
2743                  *
2744                  * if ((rec->no_buildid || !rec->no_buildid_set) &&
2745                  *     (rec->no_buildid_cache || !rec->no_buildid_cache_set))
2746                  *         disable_buildid_cache();
2747                  */
2748                 bool disable = true;
2749
2750                 if (rec->no_buildid_set && !rec->no_buildid)
2751                         disable = false;
2752                 if (rec->no_buildid_cache_set && !rec->no_buildid_cache)
2753                         disable = false;
2754                 if (disable) {
2755                         rec->no_buildid = true;
2756                         rec->no_buildid_cache = true;
2757                         disable_buildid_cache();
2758                 }
2759         }
2760
2761         if (record.opts.overwrite)
2762                 record.opts.tail_synthesize = true;
2763
2764         if (rec->evlist->core.nr_entries == 0 &&
2765             __evlist__add_default(rec->evlist, !record.opts.no_samples) < 0) {
2766                 pr_err("Not enough memory for event selector list\n");
2767                 goto out;
2768         }
2769
2770         if (rec->opts.target.tid && !rec->opts.no_inherit_set)
2771                 rec->opts.no_inherit = true;
2772
2773         err = target__validate(&rec->opts.target);
2774         if (err) {
2775                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
2776                 ui__warning("%s\n", errbuf);
2777         }
2778
2779         err = target__parse_uid(&rec->opts.target);
2780         if (err) {
2781                 int saved_errno = errno;
2782
2783                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
2784                 ui__error("%s", errbuf);
2785
2786                 err = -saved_errno;
2787                 goto out;
2788         }
2789
2790         /* Enable ignoring missing threads when -u/-p option is defined. */
2791         rec->opts.ignore_missing_thread = rec->opts.target.uid != UINT_MAX || rec->opts.target.pid;
2792
2793         err = -ENOMEM;
2794         if (evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
2795                 usage_with_options(record_usage, record_options);
2796
2797         err = auxtrace_record__options(rec->itr, rec->evlist, &rec->opts);
2798         if (err)
2799                 goto out;
2800
2801         /*
2802          * We take all buildids when the file contains
2803          * AUX area tracing data because we do not decode the
2804          * trace because it would take too long.
2805          */
2806         if (rec->opts.full_auxtrace)
2807                 rec->buildid_all = true;
2808
2809         if (rec->opts.text_poke) {
2810                 err = record__config_text_poke(rec->evlist);
2811                 if (err) {
2812                         pr_err("record__config_text_poke failed, error %d\n", err);
2813                         goto out;
2814                 }
2815         }
2816
2817         if (record_opts__config(&rec->opts)) {
2818                 err = -EINVAL;
2819                 goto out;
2820         }
2821
2822         if (rec->opts.nr_cblocks > nr_cblocks_max)
2823                 rec->opts.nr_cblocks = nr_cblocks_max;
2824         pr_debug("nr_cblocks: %d\n", rec->opts.nr_cblocks);
2825
2826         pr_debug("affinity: %s\n", affinity_tags[rec->opts.affinity]);
2827         pr_debug("mmap flush: %d\n", rec->opts.mmap_flush);
2828
2829         if (rec->opts.comp_level > comp_level_max)
2830                 rec->opts.comp_level = comp_level_max;
2831         pr_debug("comp level: %d\n", rec->opts.comp_level);
2832
2833         err = __cmd_record(&record, argc, argv);
2834 out:
2835         bitmap_free(rec->affinity_mask.bits);
2836         evlist__delete(rec->evlist);
2837         symbol__exit();
2838         auxtrace_record__free(rec->itr);
2839 out_opts:
2840         evlist__close_control(rec->opts.ctl_fd, rec->opts.ctl_fd_ack, &rec->opts.ctl_fd_close);
2841         return err;
2842 }
2843
2844 static void snapshot_sig_handler(int sig __maybe_unused)
2845 {
2846         struct record *rec = &record;
2847
2848         hit_auxtrace_snapshot_trigger(rec);
2849
2850         if (switch_output_signal(rec))
2851                 trigger_hit(&switch_output_trigger);
2852 }
2853
2854 static void alarm_sig_handler(int sig __maybe_unused)
2855 {
2856         struct record *rec = &record;
2857
2858         if (switch_output_time(rec))
2859                 trigger_hit(&switch_output_trigger);
2860 }