builtin-record.c 22 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888
  1. /*
  2. * builtin-record.c
  3. *
  4. * Builtin record command: Record the profile of a workload
  5. * (or a CPU, or a PID) into the perf.data output file - for
  6. * later analysis via perf report.
  7. */
  8. #define _FILE_OFFSET_BITS 64
  9. #include "builtin.h"
  10. #include "perf.h"
  11. #include "util/build-id.h"
  12. #include "util/util.h"
  13. #include "util/parse-options.h"
  14. #include "util/parse-events.h"
  15. #include "util/header.h"
  16. #include "util/event.h"
  17. #include "util/evlist.h"
  18. #include "util/evsel.h"
  19. #include "util/debug.h"
  20. #include "util/session.h"
  21. #include "util/symbol.h"
  22. #include "util/cpumap.h"
  23. #include "util/thread_map.h"
  24. #include <unistd.h>
  25. #include <sched.h>
  26. #include <sys/mman.h>
  27. #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  28. enum write_mode_t {
  29. WRITE_FORCE,
  30. WRITE_APPEND
  31. };
  32. static u64 user_interval = ULLONG_MAX;
  33. static u64 default_interval = 0;
  34. static unsigned int page_size;
  35. static unsigned int mmap_pages = UINT_MAX;
  36. static unsigned int user_freq = UINT_MAX;
  37. static int freq = 1000;
  38. static int output;
  39. static int pipe_output = 0;
  40. static const char *output_name = NULL;
  41. static int group = 0;
  42. static int realtime_prio = 0;
  43. static bool nodelay = false;
  44. static bool raw_samples = false;
  45. static bool sample_id_all_avail = true;
  46. static bool system_wide = false;
  47. static pid_t target_pid = -1;
  48. static pid_t target_tid = -1;
  49. static pid_t child_pid = -1;
  50. static bool no_inherit = false;
  51. static enum write_mode_t write_mode = WRITE_FORCE;
  52. static bool call_graph = false;
  53. static bool inherit_stat = false;
  54. static bool no_samples = false;
  55. static bool sample_address = false;
  56. static bool sample_time = false;
  57. static bool no_buildid = false;
  58. static bool no_buildid_cache = false;
  59. static struct perf_evlist *evsel_list;
  60. static long samples = 0;
  61. static u64 bytes_written = 0;
  62. static int file_new = 1;
  63. static off_t post_processing_offset;
  64. static struct perf_session *session;
  65. static const char *cpu_list;
  66. static void advance_output(size_t size)
  67. {
  68. bytes_written += size;
  69. }
  70. static void write_output(void *buf, size_t size)
  71. {
  72. while (size) {
  73. int ret = write(output, buf, size);
  74. if (ret < 0)
  75. die("failed to write");
  76. size -= ret;
  77. buf += ret;
  78. bytes_written += ret;
  79. }
  80. }
  81. static int process_synthesized_event(union perf_event *event,
  82. struct perf_sample *sample __used,
  83. struct perf_session *self __used)
  84. {
  85. write_output(event, event->header.size);
  86. return 0;
  87. }
  88. static void mmap_read(struct perf_mmap *md)
  89. {
  90. unsigned int head = perf_mmap__read_head(md);
  91. unsigned int old = md->prev;
  92. unsigned char *data = md->base + page_size;
  93. unsigned long size;
  94. void *buf;
  95. if (old == head)
  96. return;
  97. samples++;
  98. size = head - old;
  99. if ((old & md->mask) + size != (head & md->mask)) {
  100. buf = &data[old & md->mask];
  101. size = md->mask + 1 - (old & md->mask);
  102. old += size;
  103. write_output(buf, size);
  104. }
  105. buf = &data[old & md->mask];
  106. size = head - old;
  107. old += size;
  108. write_output(buf, size);
  109. md->prev = old;
  110. perf_mmap__write_tail(md, old);
  111. }
  112. static volatile int done = 0;
  113. static volatile int signr = -1;
  114. static void sig_handler(int sig)
  115. {
  116. done = 1;
  117. signr = sig;
  118. }
  119. static void sig_atexit(void)
  120. {
  121. if (child_pid > 0)
  122. kill(child_pid, SIGTERM);
  123. if (signr == -1 || signr == SIGUSR1)
  124. return;
  125. signal(signr, SIG_DFL);
  126. kill(getpid(), signr);
  127. }
  128. static void config_attr(struct perf_evsel *evsel, struct perf_evlist *evlist)
  129. {
  130. struct perf_event_attr *attr = &evsel->attr;
  131. int track = !evsel->idx; /* only the first counter needs these */
  132. attr->inherit = !no_inherit;
  133. attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
  134. PERF_FORMAT_TOTAL_TIME_RUNNING |
  135. PERF_FORMAT_ID;
  136. attr->sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
  137. if (evlist->nr_entries > 1)
  138. attr->sample_type |= PERF_SAMPLE_ID;
  139. /*
  140. * We default some events to a 1 default interval. But keep
  141. * it a weak assumption overridable by the user.
  142. */
  143. if (!attr->sample_period || (user_freq != UINT_MAX &&
  144. user_interval != ULLONG_MAX)) {
  145. if (freq) {
  146. attr->sample_type |= PERF_SAMPLE_PERIOD;
  147. attr->freq = 1;
  148. attr->sample_freq = freq;
  149. } else {
  150. attr->sample_period = default_interval;
  151. }
  152. }
  153. if (no_samples)
  154. attr->sample_freq = 0;
  155. if (inherit_stat)
  156. attr->inherit_stat = 1;
  157. if (sample_address) {
  158. attr->sample_type |= PERF_SAMPLE_ADDR;
  159. attr->mmap_data = track;
  160. }
  161. if (call_graph)
  162. attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
  163. if (system_wide)
  164. attr->sample_type |= PERF_SAMPLE_CPU;
  165. if (sample_id_all_avail &&
  166. (sample_time || system_wide || !no_inherit || cpu_list))
  167. attr->sample_type |= PERF_SAMPLE_TIME;
  168. if (raw_samples) {
  169. attr->sample_type |= PERF_SAMPLE_TIME;
  170. attr->sample_type |= PERF_SAMPLE_RAW;
  171. attr->sample_type |= PERF_SAMPLE_CPU;
  172. }
  173. if (nodelay) {
  174. attr->watermark = 0;
  175. attr->wakeup_events = 1;
  176. }
  177. attr->mmap = track;
  178. attr->comm = track;
  179. if (target_pid == -1 && target_tid == -1 && !system_wide) {
  180. attr->disabled = 1;
  181. attr->enable_on_exec = 1;
  182. }
  183. }
  184. static bool perf_evlist__equal(struct perf_evlist *evlist,
  185. struct perf_evlist *other)
  186. {
  187. struct perf_evsel *pos, *pair;
  188. if (evlist->nr_entries != other->nr_entries)
  189. return false;
  190. pair = list_entry(other->entries.next, struct perf_evsel, node);
  191. list_for_each_entry(pos, &evlist->entries, node) {
  192. if (memcmp(&pos->attr, &pair->attr, sizeof(pos->attr) != 0))
  193. return false;
  194. pair = list_entry(pair->node.next, struct perf_evsel, node);
  195. }
  196. return true;
  197. }
  198. static void open_counters(struct perf_evlist *evlist)
  199. {
  200. struct perf_evsel *pos;
  201. if (evlist->cpus->map[0] < 0)
  202. no_inherit = true;
  203. list_for_each_entry(pos, &evlist->entries, node) {
  204. struct perf_event_attr *attr = &pos->attr;
  205. /*
  206. * Check if parse_single_tracepoint_event has already asked for
  207. * PERF_SAMPLE_TIME.
  208. *
  209. * XXX this is kludgy but short term fix for problems introduced by
  210. * eac23d1c that broke 'perf script' by having different sample_types
  211. * when using multiple tracepoint events when we use a perf binary
  212. * that tries to use sample_id_all on an older kernel.
  213. *
  214. * We need to move counter creation to perf_session, support
  215. * different sample_types, etc.
  216. */
  217. bool time_needed = attr->sample_type & PERF_SAMPLE_TIME;
  218. config_attr(pos, evlist);
  219. retry_sample_id:
  220. attr->sample_id_all = sample_id_all_avail ? 1 : 0;
  221. try_again:
  222. if (perf_evsel__open(pos, evlist->cpus, evlist->threads, group) < 0) {
  223. int err = errno;
  224. if (err == EPERM || err == EACCES) {
  225. ui__warning_paranoid();
  226. exit(EXIT_FAILURE);
  227. } else if (err == ENODEV && cpu_list) {
  228. die("No such device - did you specify"
  229. " an out-of-range profile CPU?\n");
  230. } else if (err == EINVAL && sample_id_all_avail) {
  231. /*
  232. * Old kernel, no attr->sample_id_type_all field
  233. */
  234. sample_id_all_avail = false;
  235. if (!sample_time && !raw_samples && !time_needed)
  236. attr->sample_type &= ~PERF_SAMPLE_TIME;
  237. goto retry_sample_id;
  238. }
  239. /*
  240. * If it's cycles then fall back to hrtimer
  241. * based cpu-clock-tick sw counter, which
  242. * is always available even if no PMU support:
  243. */
  244. if (attr->type == PERF_TYPE_HARDWARE
  245. && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
  246. if (verbose)
  247. ui__warning("The cycles event is not supported, "
  248. "trying to fall back to cpu-clock-ticks\n");
  249. attr->type = PERF_TYPE_SOFTWARE;
  250. attr->config = PERF_COUNT_SW_CPU_CLOCK;
  251. goto try_again;
  252. }
  253. if (err == ENOENT) {
  254. ui__warning("The %s event is not supported.\n",
  255. event_name(pos));
  256. exit(EXIT_FAILURE);
  257. }
  258. printf("\n");
  259. error("sys_perf_event_open() syscall returned with %d (%s). /bin/dmesg may provide additional information.\n",
  260. err, strerror(err));
  261. #if defined(__i386__) || defined(__x86_64__)
  262. if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
  263. die("No hardware sampling interrupt available."
  264. " No APIC? If so then you can boot the kernel"
  265. " with the \"lapic\" boot parameter to"
  266. " force-enable it.\n");
  267. #endif
  268. die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
  269. }
  270. }
  271. if (perf_evlist__set_filters(evlist)) {
  272. error("failed to set filter with %d (%s)\n", errno,
  273. strerror(errno));
  274. exit(-1);
  275. }
  276. if (perf_evlist__mmap(evlist, mmap_pages, false) < 0)
  277. die("failed to mmap with %d (%s)\n", errno, strerror(errno));
  278. if (file_new)
  279. session->evlist = evlist;
  280. else {
  281. if (!perf_evlist__equal(session->evlist, evlist)) {
  282. fprintf(stderr, "incompatible append\n");
  283. exit(-1);
  284. }
  285. }
  286. perf_session__update_sample_type(session);
  287. }
  288. static int process_buildids(void)
  289. {
  290. u64 size = lseek(output, 0, SEEK_CUR);
  291. if (size == 0)
  292. return 0;
  293. session->fd = output;
  294. return __perf_session__process_events(session, post_processing_offset,
  295. size - post_processing_offset,
  296. size, &build_id__mark_dso_hit_ops);
  297. }
  298. static void atexit_header(void)
  299. {
  300. if (!pipe_output) {
  301. session->header.data_size += bytes_written;
  302. if (!no_buildid)
  303. process_buildids();
  304. perf_session__write_header(session, evsel_list, output, true);
  305. perf_session__delete(session);
  306. perf_evlist__delete(evsel_list);
  307. symbol__exit();
  308. }
  309. }
  310. static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
  311. {
  312. int err;
  313. struct perf_session *psession = data;
  314. if (machine__is_host(machine))
  315. return;
  316. /*
  317. *As for guest kernel when processing subcommand record&report,
  318. *we arrange module mmap prior to guest kernel mmap and trigger
  319. *a preload dso because default guest module symbols are loaded
  320. *from guest kallsyms instead of /lib/modules/XXX/XXX. This
  321. *method is used to avoid symbol missing when the first addr is
  322. *in module instead of in guest kernel.
  323. */
  324. err = perf_event__synthesize_modules(process_synthesized_event,
  325. psession, machine);
  326. if (err < 0)
  327. pr_err("Couldn't record guest kernel [%d]'s reference"
  328. " relocation symbol.\n", machine->pid);
  329. /*
  330. * We use _stext for guest kernel because guest kernel's /proc/kallsyms
  331. * have no _text sometimes.
  332. */
  333. err = perf_event__synthesize_kernel_mmap(process_synthesized_event,
  334. psession, machine, "_text");
  335. if (err < 0)
  336. err = perf_event__synthesize_kernel_mmap(process_synthesized_event,
  337. psession, machine,
  338. "_stext");
  339. if (err < 0)
  340. pr_err("Couldn't record guest kernel [%d]'s reference"
  341. " relocation symbol.\n", machine->pid);
  342. }
  343. static struct perf_event_header finished_round_event = {
  344. .size = sizeof(struct perf_event_header),
  345. .type = PERF_RECORD_FINISHED_ROUND,
  346. };
  347. static void mmap_read_all(void)
  348. {
  349. int i;
  350. for (i = 0; i < evsel_list->nr_mmaps; i++) {
  351. if (evsel_list->mmap[i].base)
  352. mmap_read(&evsel_list->mmap[i]);
  353. }
  354. if (perf_header__has_feat(&session->header, HEADER_TRACE_INFO))
  355. write_output(&finished_round_event, sizeof(finished_round_event));
  356. }
  357. static int __cmd_record(int argc, const char **argv)
  358. {
  359. int i;
  360. struct stat st;
  361. int flags;
  362. int err;
  363. unsigned long waking = 0;
  364. int child_ready_pipe[2], go_pipe[2];
  365. const bool forks = argc > 0;
  366. char buf;
  367. struct machine *machine;
  368. page_size = sysconf(_SC_PAGE_SIZE);
  369. atexit(sig_atexit);
  370. signal(SIGCHLD, sig_handler);
  371. signal(SIGINT, sig_handler);
  372. signal(SIGUSR1, sig_handler);
  373. if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
  374. perror("failed to create pipes");
  375. exit(-1);
  376. }
  377. if (!output_name) {
  378. if (!fstat(STDOUT_FILENO, &st) && S_ISFIFO(st.st_mode))
  379. pipe_output = 1;
  380. else
  381. output_name = "perf.data";
  382. }
  383. if (output_name) {
  384. if (!strcmp(output_name, "-"))
  385. pipe_output = 1;
  386. else if (!stat(output_name, &st) && st.st_size) {
  387. if (write_mode == WRITE_FORCE) {
  388. char oldname[PATH_MAX];
  389. snprintf(oldname, sizeof(oldname), "%s.old",
  390. output_name);
  391. unlink(oldname);
  392. rename(output_name, oldname);
  393. }
  394. } else if (write_mode == WRITE_APPEND) {
  395. write_mode = WRITE_FORCE;
  396. }
  397. }
  398. flags = O_CREAT|O_RDWR;
  399. if (write_mode == WRITE_APPEND)
  400. file_new = 0;
  401. else
  402. flags |= O_TRUNC;
  403. if (pipe_output)
  404. output = STDOUT_FILENO;
  405. else
  406. output = open(output_name, flags, S_IRUSR | S_IWUSR);
  407. if (output < 0) {
  408. perror("failed to create output file");
  409. exit(-1);
  410. }
  411. session = perf_session__new(output_name, O_WRONLY,
  412. write_mode == WRITE_FORCE, false, NULL);
  413. if (session == NULL) {
  414. pr_err("Not enough memory for reading perf file header\n");
  415. return -1;
  416. }
  417. if (!no_buildid)
  418. perf_header__set_feat(&session->header, HEADER_BUILD_ID);
  419. if (!file_new) {
  420. err = perf_session__read_header(session, output);
  421. if (err < 0)
  422. goto out_delete_session;
  423. }
  424. if (have_tracepoints(&evsel_list->entries))
  425. perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
  426. /* 512 kiB: default amount of unprivileged mlocked memory */
  427. if (mmap_pages == UINT_MAX)
  428. mmap_pages = (512 * 1024) / page_size;
  429. if (forks) {
  430. child_pid = fork();
  431. if (child_pid < 0) {
  432. perror("failed to fork");
  433. exit(-1);
  434. }
  435. if (!child_pid) {
  436. if (pipe_output)
  437. dup2(2, 1);
  438. close(child_ready_pipe[0]);
  439. close(go_pipe[1]);
  440. fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
  441. /*
  442. * Do a dummy execvp to get the PLT entry resolved,
  443. * so we avoid the resolver overhead on the real
  444. * execvp call.
  445. */
  446. execvp("", (char **)argv);
  447. /*
  448. * Tell the parent we're ready to go
  449. */
  450. close(child_ready_pipe[1]);
  451. /*
  452. * Wait until the parent tells us to go.
  453. */
  454. if (read(go_pipe[0], &buf, 1) == -1)
  455. perror("unable to read pipe");
  456. execvp(argv[0], (char **)argv);
  457. perror(argv[0]);
  458. kill(getppid(), SIGUSR1);
  459. exit(-1);
  460. }
  461. if (!system_wide && target_tid == -1 && target_pid == -1)
  462. evsel_list->threads->map[0] = child_pid;
  463. close(child_ready_pipe[1]);
  464. close(go_pipe[0]);
  465. /*
  466. * wait for child to settle
  467. */
  468. if (read(child_ready_pipe[0], &buf, 1) == -1) {
  469. perror("unable to read pipe");
  470. exit(-1);
  471. }
  472. close(child_ready_pipe[0]);
  473. }
  474. open_counters(evsel_list);
  475. /*
  476. * perf_session__delete(session) will be called at atexit_header()
  477. */
  478. atexit(atexit_header);
  479. if (pipe_output) {
  480. err = perf_header__write_pipe(output);
  481. if (err < 0)
  482. return err;
  483. } else if (file_new) {
  484. err = perf_session__write_header(session, evsel_list,
  485. output, false);
  486. if (err < 0)
  487. return err;
  488. }
  489. post_processing_offset = lseek(output, 0, SEEK_CUR);
  490. if (pipe_output) {
  491. err = perf_session__synthesize_attrs(session,
  492. process_synthesized_event);
  493. if (err < 0) {
  494. pr_err("Couldn't synthesize attrs.\n");
  495. return err;
  496. }
  497. err = perf_event__synthesize_event_types(process_synthesized_event,
  498. session);
  499. if (err < 0) {
  500. pr_err("Couldn't synthesize event_types.\n");
  501. return err;
  502. }
  503. if (have_tracepoints(&evsel_list->entries)) {
  504. /*
  505. * FIXME err <= 0 here actually means that
  506. * there were no tracepoints so its not really
  507. * an error, just that we don't need to
  508. * synthesize anything. We really have to
  509. * return this more properly and also
  510. * propagate errors that now are calling die()
  511. */
  512. err = perf_event__synthesize_tracing_data(output, evsel_list,
  513. process_synthesized_event,
  514. session);
  515. if (err <= 0) {
  516. pr_err("Couldn't record tracing data.\n");
  517. return err;
  518. }
  519. advance_output(err);
  520. }
  521. }
  522. machine = perf_session__find_host_machine(session);
  523. if (!machine) {
  524. pr_err("Couldn't find native kernel information.\n");
  525. return -1;
  526. }
  527. err = perf_event__synthesize_kernel_mmap(process_synthesized_event,
  528. session, machine, "_text");
  529. if (err < 0)
  530. err = perf_event__synthesize_kernel_mmap(process_synthesized_event,
  531. session, machine, "_stext");
  532. if (err < 0)
  533. pr_err("Couldn't record kernel reference relocation symbol\n"
  534. "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
  535. "Check /proc/kallsyms permission or run as root.\n");
  536. err = perf_event__synthesize_modules(process_synthesized_event,
  537. session, machine);
  538. if (err < 0)
  539. pr_err("Couldn't record kernel module information.\n"
  540. "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
  541. "Check /proc/modules permission or run as root.\n");
  542. if (perf_guest)
  543. perf_session__process_machines(session,
  544. perf_event__synthesize_guest_os);
  545. if (!system_wide)
  546. perf_event__synthesize_thread_map(evsel_list->threads,
  547. process_synthesized_event,
  548. session);
  549. else
  550. perf_event__synthesize_threads(process_synthesized_event,
  551. session);
  552. if (realtime_prio) {
  553. struct sched_param param;
  554. param.sched_priority = realtime_prio;
  555. if (sched_setscheduler(0, SCHED_FIFO, &param)) {
  556. pr_err("Could not set realtime priority.\n");
  557. exit(-1);
  558. }
  559. }
  560. /*
  561. * Let the child rip
  562. */
  563. if (forks)
  564. close(go_pipe[1]);
  565. for (;;) {
  566. int hits = samples;
  567. int thread;
  568. mmap_read_all();
  569. if (hits == samples) {
  570. if (done)
  571. break;
  572. err = poll(evsel_list->pollfd, evsel_list->nr_fds, -1);
  573. waking++;
  574. }
  575. if (done) {
  576. for (i = 0; i < evsel_list->cpus->nr; i++) {
  577. struct perf_evsel *pos;
  578. list_for_each_entry(pos, &evsel_list->entries, node) {
  579. for (thread = 0;
  580. thread < evsel_list->threads->nr;
  581. thread++)
  582. ioctl(FD(pos, i, thread),
  583. PERF_EVENT_IOC_DISABLE);
  584. }
  585. }
  586. }
  587. }
  588. if (quiet || signr == SIGUSR1)
  589. return 0;
  590. fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
  591. /*
  592. * Approximate RIP event size: 24 bytes.
  593. */
  594. fprintf(stderr,
  595. "[ perf record: Captured and wrote %.3f MB %s (~%" PRIu64 " samples) ]\n",
  596. (double)bytes_written / 1024.0 / 1024.0,
  597. output_name,
  598. bytes_written / 24);
  599. return 0;
  600. out_delete_session:
  601. perf_session__delete(session);
  602. return err;
  603. }
  604. static const char * const record_usage[] = {
  605. "perf record [<options>] [<command>]",
  606. "perf record [<options>] -- <command> [<options>]",
  607. NULL
  608. };
  609. static bool force, append_file;
  610. const struct option record_options[] = {
  611. OPT_CALLBACK('e', "event", &evsel_list, "event",
  612. "event selector. use 'perf list' to list available events",
  613. parse_events),
  614. OPT_CALLBACK(0, "filter", &evsel_list, "filter",
  615. "event filter", parse_filter),
  616. OPT_INTEGER('p', "pid", &target_pid,
  617. "record events on existing process id"),
  618. OPT_INTEGER('t', "tid", &target_tid,
  619. "record events on existing thread id"),
  620. OPT_INTEGER('r', "realtime", &realtime_prio,
  621. "collect data with this RT SCHED_FIFO priority"),
  622. OPT_BOOLEAN('D', "no-delay", &nodelay,
  623. "collect data without buffering"),
  624. OPT_BOOLEAN('R', "raw-samples", &raw_samples,
  625. "collect raw sample records from all opened counters"),
  626. OPT_BOOLEAN('a', "all-cpus", &system_wide,
  627. "system-wide collection from all CPUs"),
  628. OPT_BOOLEAN('A', "append", &append_file,
  629. "append to the output file to do incremental profiling"),
  630. OPT_STRING('C', "cpu", &cpu_list, "cpu",
  631. "list of cpus to monitor"),
  632. OPT_BOOLEAN('f', "force", &force,
  633. "overwrite existing data file (deprecated)"),
  634. OPT_U64('c', "count", &user_interval, "event period to sample"),
  635. OPT_STRING('o', "output", &output_name, "file",
  636. "output file name"),
  637. OPT_BOOLEAN('i', "no-inherit", &no_inherit,
  638. "child tasks do not inherit counters"),
  639. OPT_UINTEGER('F', "freq", &user_freq, "profile at this frequency"),
  640. OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
  641. OPT_BOOLEAN('g', "call-graph", &call_graph,
  642. "do call-graph (stack chain/backtrace) recording"),
  643. OPT_INCR('v', "verbose", &verbose,
  644. "be more verbose (show counter open errors, etc)"),
  645. OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
  646. OPT_BOOLEAN('s', "stat", &inherit_stat,
  647. "per thread counts"),
  648. OPT_BOOLEAN('d', "data", &sample_address,
  649. "Sample addresses"),
  650. OPT_BOOLEAN('T', "timestamp", &sample_time, "Sample timestamps"),
  651. OPT_BOOLEAN('n', "no-samples", &no_samples,
  652. "don't sample"),
  653. OPT_BOOLEAN('N', "no-buildid-cache", &no_buildid_cache,
  654. "do not update the buildid cache"),
  655. OPT_BOOLEAN('B', "no-buildid", &no_buildid,
  656. "do not collect buildids in perf.data"),
  657. OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
  658. "monitor event in cgroup name only",
  659. parse_cgroups),
  660. OPT_END()
  661. };
  662. int cmd_record(int argc, const char **argv, const char *prefix __used)
  663. {
  664. int err = -ENOMEM;
  665. struct perf_evsel *pos;
  666. evsel_list = perf_evlist__new(NULL, NULL);
  667. if (evsel_list == NULL)
  668. return -ENOMEM;
  669. argc = parse_options(argc, argv, record_options, record_usage,
  670. PARSE_OPT_STOP_AT_NON_OPTION);
  671. if (!argc && target_pid == -1 && target_tid == -1 &&
  672. !system_wide && !cpu_list)
  673. usage_with_options(record_usage, record_options);
  674. if (force && append_file) {
  675. fprintf(stderr, "Can't overwrite and append at the same time."
  676. " You need to choose between -f and -A");
  677. usage_with_options(record_usage, record_options);
  678. } else if (append_file) {
  679. write_mode = WRITE_APPEND;
  680. } else {
  681. write_mode = WRITE_FORCE;
  682. }
  683. if (nr_cgroups && !system_wide) {
  684. fprintf(stderr, "cgroup monitoring only available in"
  685. " system-wide mode\n");
  686. usage_with_options(record_usage, record_options);
  687. }
  688. symbol__init();
  689. if (symbol_conf.kptr_restrict)
  690. pr_warning(
  691. "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
  692. "check /proc/sys/kernel/kptr_restrict.\n\n"
  693. "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
  694. "file is not found in the buildid cache or in the vmlinux path.\n\n"
  695. "Samples in kernel modules won't be resolved at all.\n\n"
  696. "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
  697. "even with a suitable vmlinux or kallsyms file.\n\n");
  698. if (no_buildid_cache || no_buildid)
  699. disable_buildid_cache();
  700. if (evsel_list->nr_entries == 0 &&
  701. perf_evlist__add_default(evsel_list) < 0) {
  702. pr_err("Not enough memory for event selector list\n");
  703. goto out_symbol_exit;
  704. }
  705. if (target_pid != -1)
  706. target_tid = target_pid;
  707. if (perf_evlist__create_maps(evsel_list, target_pid,
  708. target_tid, cpu_list) < 0)
  709. usage_with_options(record_usage, record_options);
  710. list_for_each_entry(pos, &evsel_list->entries, node) {
  711. if (perf_evsel__alloc_fd(pos, evsel_list->cpus->nr,
  712. evsel_list->threads->nr) < 0)
  713. goto out_free_fd;
  714. if (perf_header__push_event(pos->attr.config, event_name(pos)))
  715. goto out_free_fd;
  716. }
  717. if (perf_evlist__alloc_pollfd(evsel_list) < 0)
  718. goto out_free_fd;
  719. if (user_interval != ULLONG_MAX)
  720. default_interval = user_interval;
  721. if (user_freq != UINT_MAX)
  722. freq = user_freq;
  723. /*
  724. * User specified count overrides default frequency.
  725. */
  726. if (default_interval)
  727. freq = 0;
  728. else if (freq) {
  729. default_interval = freq;
  730. } else {
  731. fprintf(stderr, "frequency and count are zero, aborting\n");
  732. err = -EINVAL;
  733. goto out_free_fd;
  734. }
  735. err = __cmd_record(argc, argv);
  736. out_free_fd:
  737. perf_evlist__delete_maps(evsel_list);
  738. out_symbol_exit:
  739. symbol__exit();
  740. return err;
  741. }