builtin-record.c 21 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897
  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/debug.h"
  18. #include "util/session.h"
  19. #include "util/symbol.h"
  20. #include "util/cpumap.h"
  21. #include <unistd.h>
  22. #include <sched.h>
  23. #include <sys/mman.h>
  24. enum write_mode_t {
  25. WRITE_FORCE,
  26. WRITE_APPEND
  27. };
  28. static int *fd[MAX_NR_CPUS][MAX_COUNTERS];
  29. static u64 user_interval = ULLONG_MAX;
  30. static u64 default_interval = 0;
  31. static int nr_cpus = 0;
  32. static unsigned int page_size;
  33. static unsigned int mmap_pages = 128;
  34. static unsigned int user_freq = UINT_MAX;
  35. static int freq = 1000;
  36. static int output;
  37. static int pipe_output = 0;
  38. static const char *output_name = "perf.data";
  39. static int group = 0;
  40. static int realtime_prio = 0;
  41. static bool raw_samples = false;
  42. static bool system_wide = false;
  43. static int profile_cpu = -1;
  44. static pid_t target_pid = -1;
  45. static pid_t target_tid = -1;
  46. static pid_t *all_tids = NULL;
  47. static int thread_num = 0;
  48. static pid_t child_pid = -1;
  49. static bool no_inherit = false;
  50. static enum write_mode_t write_mode = WRITE_FORCE;
  51. static bool call_graph = false;
  52. static bool inherit_stat = false;
  53. static bool no_samples = false;
  54. static bool sample_address = false;
  55. static long samples = 0;
  56. static u64 bytes_written = 0;
  57. static struct pollfd *event_array;
  58. static int nr_poll = 0;
  59. static int nr_cpu = 0;
  60. static int file_new = 1;
  61. static off_t post_processing_offset;
  62. static struct perf_session *session;
  63. struct mmap_data {
  64. int counter;
  65. void *base;
  66. unsigned int mask;
  67. unsigned int prev;
  68. };
  69. static struct mmap_data mmap_array[MAX_NR_CPUS];
  70. static unsigned long mmap_read_head(struct mmap_data *md)
  71. {
  72. struct perf_event_mmap_page *pc = md->base;
  73. long head;
  74. head = pc->data_head;
  75. rmb();
  76. return head;
  77. }
  78. static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
  79. {
  80. struct perf_event_mmap_page *pc = md->base;
  81. /*
  82. * ensure all reads are done before we write the tail out.
  83. */
  84. /* mb(); */
  85. pc->data_tail = tail;
  86. }
  87. static void advance_output(size_t size)
  88. {
  89. bytes_written += size;
  90. }
  91. static void write_output(void *buf, size_t size)
  92. {
  93. while (size) {
  94. int ret = write(output, buf, size);
  95. if (ret < 0)
  96. die("failed to write");
  97. size -= ret;
  98. buf += ret;
  99. bytes_written += ret;
  100. }
  101. }
  102. static int process_synthesized_event(event_t *event,
  103. struct perf_session *self __used)
  104. {
  105. write_output(event, event->header.size);
  106. return 0;
  107. }
  108. static void mmap_read(struct mmap_data *md)
  109. {
  110. unsigned int head = mmap_read_head(md);
  111. unsigned int old = md->prev;
  112. unsigned char *data = md->base + page_size;
  113. unsigned long size;
  114. void *buf;
  115. int diff;
  116. /*
  117. * If we're further behind than half the buffer, there's a chance
  118. * the writer will bite our tail and mess up the samples under us.
  119. *
  120. * If we somehow ended up ahead of the head, we got messed up.
  121. *
  122. * In either case, truncate and restart at head.
  123. */
  124. diff = head - old;
  125. if (diff < 0) {
  126. fprintf(stderr, "WARNING: failed to keep up with mmap data\n");
  127. /*
  128. * head points to a known good entry, start there.
  129. */
  130. old = head;
  131. }
  132. if (old != head)
  133. samples++;
  134. size = head - old;
  135. if ((old & md->mask) + size != (head & md->mask)) {
  136. buf = &data[old & md->mask];
  137. size = md->mask + 1 - (old & md->mask);
  138. old += size;
  139. write_output(buf, size);
  140. }
  141. buf = &data[old & md->mask];
  142. size = head - old;
  143. old += size;
  144. write_output(buf, size);
  145. md->prev = old;
  146. mmap_write_tail(md, old);
  147. }
  148. static volatile int done = 0;
  149. static volatile int signr = -1;
  150. static void sig_handler(int sig)
  151. {
  152. done = 1;
  153. signr = sig;
  154. }
  155. static void sig_atexit(void)
  156. {
  157. if (child_pid != -1)
  158. kill(child_pid, SIGTERM);
  159. if (signr == -1)
  160. return;
  161. signal(signr, SIG_DFL);
  162. kill(getpid(), signr);
  163. }
  164. static int group_fd;
  165. static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
  166. {
  167. struct perf_header_attr *h_attr;
  168. if (nr < session->header.attrs) {
  169. h_attr = session->header.attr[nr];
  170. } else {
  171. h_attr = perf_header_attr__new(a);
  172. if (h_attr != NULL)
  173. if (perf_header__add_attr(&session->header, h_attr) < 0) {
  174. perf_header_attr__delete(h_attr);
  175. h_attr = NULL;
  176. }
  177. }
  178. return h_attr;
  179. }
  180. static void create_counter(int counter, int cpu)
  181. {
  182. char *filter = filters[counter];
  183. struct perf_event_attr *attr = attrs + counter;
  184. struct perf_header_attr *h_attr;
  185. int track = !counter; /* only the first counter needs these */
  186. int thread_index;
  187. int ret;
  188. struct {
  189. u64 count;
  190. u64 time_enabled;
  191. u64 time_running;
  192. u64 id;
  193. } read_data;
  194. attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
  195. PERF_FORMAT_TOTAL_TIME_RUNNING |
  196. PERF_FORMAT_ID;
  197. attr->sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
  198. if (nr_counters > 1)
  199. attr->sample_type |= PERF_SAMPLE_ID;
  200. /*
  201. * We default some events to a 1 default interval. But keep
  202. * it a weak assumption overridable by the user.
  203. */
  204. if (!attr->sample_period || (user_freq != UINT_MAX &&
  205. user_interval != ULLONG_MAX)) {
  206. if (freq) {
  207. attr->sample_type |= PERF_SAMPLE_PERIOD;
  208. attr->freq = 1;
  209. attr->sample_freq = freq;
  210. } else {
  211. attr->sample_period = default_interval;
  212. }
  213. }
  214. if (no_samples)
  215. attr->sample_freq = 0;
  216. if (inherit_stat)
  217. attr->inherit_stat = 1;
  218. if (sample_address)
  219. attr->sample_type |= PERF_SAMPLE_ADDR;
  220. if (call_graph)
  221. attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
  222. if (raw_samples) {
  223. attr->sample_type |= PERF_SAMPLE_TIME;
  224. attr->sample_type |= PERF_SAMPLE_RAW;
  225. attr->sample_type |= PERF_SAMPLE_CPU;
  226. }
  227. attr->mmap = track;
  228. attr->comm = track;
  229. attr->inherit = !no_inherit;
  230. if (target_pid == -1 && target_tid == -1 && !system_wide) {
  231. attr->disabled = 1;
  232. attr->enable_on_exec = 1;
  233. }
  234. for (thread_index = 0; thread_index < thread_num; thread_index++) {
  235. try_again:
  236. fd[nr_cpu][counter][thread_index] = sys_perf_event_open(attr,
  237. all_tids[thread_index], cpu, group_fd, 0);
  238. if (fd[nr_cpu][counter][thread_index] < 0) {
  239. int err = errno;
  240. if (err == EPERM || err == EACCES)
  241. die("Permission error - are you root?\n"
  242. "\t Consider tweaking"
  243. " /proc/sys/kernel/perf_event_paranoid.\n");
  244. else if (err == ENODEV && profile_cpu != -1) {
  245. die("No such device - did you specify"
  246. " an out-of-range profile CPU?\n");
  247. }
  248. /*
  249. * If it's cycles then fall back to hrtimer
  250. * based cpu-clock-tick sw counter, which
  251. * is always available even if no PMU support:
  252. */
  253. if (attr->type == PERF_TYPE_HARDWARE
  254. && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
  255. if (verbose)
  256. warning(" ... trying to fall back to cpu-clock-ticks\n");
  257. attr->type = PERF_TYPE_SOFTWARE;
  258. attr->config = PERF_COUNT_SW_CPU_CLOCK;
  259. goto try_again;
  260. }
  261. printf("\n");
  262. error("perfcounter syscall returned with %d (%s)\n",
  263. fd[nr_cpu][counter][thread_index], strerror(err));
  264. #if defined(__i386__) || defined(__x86_64__)
  265. if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
  266. die("No hardware sampling interrupt available."
  267. " No APIC? If so then you can boot the kernel"
  268. " with the \"lapic\" boot parameter to"
  269. " force-enable it.\n");
  270. #endif
  271. die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
  272. exit(-1);
  273. }
  274. h_attr = get_header_attr(attr, counter);
  275. if (h_attr == NULL)
  276. die("nomem\n");
  277. if (!file_new) {
  278. if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
  279. fprintf(stderr, "incompatible append\n");
  280. exit(-1);
  281. }
  282. }
  283. if (read(fd[nr_cpu][counter][thread_index], &read_data, sizeof(read_data)) == -1) {
  284. perror("Unable to read perf file descriptor\n");
  285. exit(-1);
  286. }
  287. if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
  288. pr_warning("Not enough memory to add id\n");
  289. exit(-1);
  290. }
  291. assert(fd[nr_cpu][counter][thread_index] >= 0);
  292. fcntl(fd[nr_cpu][counter][thread_index], F_SETFL, O_NONBLOCK);
  293. /*
  294. * First counter acts as the group leader:
  295. */
  296. if (group && group_fd == -1)
  297. group_fd = fd[nr_cpu][counter][thread_index];
  298. if (counter || thread_index) {
  299. ret = ioctl(fd[nr_cpu][counter][thread_index],
  300. PERF_EVENT_IOC_SET_OUTPUT,
  301. fd[nr_cpu][0][0]);
  302. if (ret) {
  303. error("failed to set output: %d (%s)\n", errno,
  304. strerror(errno));
  305. exit(-1);
  306. }
  307. } else {
  308. mmap_array[nr_cpu].counter = counter;
  309. mmap_array[nr_cpu].prev = 0;
  310. mmap_array[nr_cpu].mask = mmap_pages*page_size - 1;
  311. mmap_array[nr_cpu].base = mmap(NULL, (mmap_pages+1)*page_size,
  312. PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter][thread_index], 0);
  313. if (mmap_array[nr_cpu].base == MAP_FAILED) {
  314. error("failed to mmap with %d (%s)\n", errno, strerror(errno));
  315. exit(-1);
  316. }
  317. event_array[nr_poll].fd = fd[nr_cpu][counter][thread_index];
  318. event_array[nr_poll].events = POLLIN;
  319. nr_poll++;
  320. }
  321. if (filter != NULL) {
  322. ret = ioctl(fd[nr_cpu][counter][thread_index],
  323. PERF_EVENT_IOC_SET_FILTER, filter);
  324. if (ret) {
  325. error("failed to set filter with %d (%s)\n", errno,
  326. strerror(errno));
  327. exit(-1);
  328. }
  329. }
  330. }
  331. }
  332. static void open_counters(int cpu)
  333. {
  334. int counter;
  335. group_fd = -1;
  336. for (counter = 0; counter < nr_counters; counter++)
  337. create_counter(counter, cpu);
  338. nr_cpu++;
  339. }
  340. static int process_buildids(void)
  341. {
  342. u64 size = lseek(output, 0, SEEK_CUR);
  343. if (size == 0)
  344. return 0;
  345. session->fd = output;
  346. return __perf_session__process_events(session, post_processing_offset,
  347. size - post_processing_offset,
  348. size, &build_id__mark_dso_hit_ops);
  349. }
  350. static void atexit_header(void)
  351. {
  352. if (!pipe_output) {
  353. session->header.data_size += bytes_written;
  354. process_buildids();
  355. perf_header__write(&session->header, output, true);
  356. }
  357. }
  358. static void event__synthesize_guest_os(struct machine *machine, void *data)
  359. {
  360. int err;
  361. char *guest_kallsyms;
  362. char path[PATH_MAX];
  363. struct perf_session *psession = data;
  364. if (machine__is_host(machine))
  365. return;
  366. /*
  367. *As for guest kernel when processing subcommand record&report,
  368. *we arrange module mmap prior to guest kernel mmap and trigger
  369. *a preload dso because default guest module symbols are loaded
  370. *from guest kallsyms instead of /lib/modules/XXX/XXX. This
  371. *method is used to avoid symbol missing when the first addr is
  372. *in module instead of in guest kernel.
  373. */
  374. err = event__synthesize_modules(process_synthesized_event,
  375. psession, machine);
  376. if (err < 0)
  377. pr_err("Couldn't record guest kernel [%d]'s reference"
  378. " relocation symbol.\n", machine->pid);
  379. if (machine__is_default_guest(machine))
  380. guest_kallsyms = (char *) symbol_conf.default_guest_kallsyms;
  381. else {
  382. sprintf(path, "%s/proc/kallsyms", machine->root_dir);
  383. guest_kallsyms = path;
  384. }
  385. /*
  386. * We use _stext for guest kernel because guest kernel's /proc/kallsyms
  387. * have no _text sometimes.
  388. */
  389. err = event__synthesize_kernel_mmap(process_synthesized_event,
  390. psession, machine, "_text");
  391. if (err < 0)
  392. err = event__synthesize_kernel_mmap(process_synthesized_event,
  393. psession, machine, "_stext");
  394. if (err < 0)
  395. pr_err("Couldn't record guest kernel [%d]'s reference"
  396. " relocation symbol.\n", machine->pid);
  397. }
  398. static struct perf_event_header finished_round_event = {
  399. .size = sizeof(struct perf_event_header),
  400. .type = PERF_RECORD_FINISHED_ROUND,
  401. };
  402. static void mmap_read_all(void)
  403. {
  404. int i;
  405. for (i = 0; i < nr_cpu; i++) {
  406. if (mmap_array[i].base)
  407. mmap_read(&mmap_array[i]);
  408. }
  409. if (perf_header__has_feat(&session->header, HEADER_TRACE_INFO))
  410. write_output(&finished_round_event, sizeof(finished_round_event));
  411. }
  412. static int __cmd_record(int argc, const char **argv)
  413. {
  414. int i, counter;
  415. struct stat st;
  416. pid_t pid = 0;
  417. int flags;
  418. int err;
  419. unsigned long waking = 0;
  420. int child_ready_pipe[2], go_pipe[2];
  421. const bool forks = argc > 0;
  422. char buf;
  423. struct machine *machine;
  424. page_size = sysconf(_SC_PAGE_SIZE);
  425. atexit(sig_atexit);
  426. signal(SIGCHLD, sig_handler);
  427. signal(SIGINT, sig_handler);
  428. if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
  429. perror("failed to create pipes");
  430. exit(-1);
  431. }
  432. if (!strcmp(output_name, "-"))
  433. pipe_output = 1;
  434. else if (!stat(output_name, &st) && st.st_size) {
  435. if (write_mode == WRITE_FORCE) {
  436. char oldname[PATH_MAX];
  437. snprintf(oldname, sizeof(oldname), "%s.old",
  438. output_name);
  439. unlink(oldname);
  440. rename(output_name, oldname);
  441. }
  442. } else if (write_mode == WRITE_APPEND) {
  443. write_mode = WRITE_FORCE;
  444. }
  445. flags = O_CREAT|O_RDWR;
  446. if (write_mode == WRITE_APPEND)
  447. file_new = 0;
  448. else
  449. flags |= O_TRUNC;
  450. if (pipe_output)
  451. output = STDOUT_FILENO;
  452. else
  453. output = open(output_name, flags, S_IRUSR | S_IWUSR);
  454. if (output < 0) {
  455. perror("failed to create output file");
  456. exit(-1);
  457. }
  458. session = perf_session__new(output_name, O_WRONLY,
  459. write_mode == WRITE_FORCE, false);
  460. if (session == NULL) {
  461. pr_err("Not enough memory for reading perf file header\n");
  462. return -1;
  463. }
  464. if (!file_new) {
  465. err = perf_header__read(session, output);
  466. if (err < 0)
  467. return err;
  468. }
  469. if (have_tracepoints(attrs, nr_counters))
  470. perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
  471. atexit(atexit_header);
  472. if (forks) {
  473. child_pid = fork();
  474. if (pid < 0) {
  475. perror("failed to fork");
  476. exit(-1);
  477. }
  478. if (!child_pid) {
  479. if (pipe_output)
  480. dup2(2, 1);
  481. close(child_ready_pipe[0]);
  482. close(go_pipe[1]);
  483. fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
  484. /*
  485. * Do a dummy execvp to get the PLT entry resolved,
  486. * so we avoid the resolver overhead on the real
  487. * execvp call.
  488. */
  489. execvp("", (char **)argv);
  490. /*
  491. * Tell the parent we're ready to go
  492. */
  493. close(child_ready_pipe[1]);
  494. /*
  495. * Wait until the parent tells us to go.
  496. */
  497. if (read(go_pipe[0], &buf, 1) == -1)
  498. perror("unable to read pipe");
  499. execvp(argv[0], (char **)argv);
  500. perror(argv[0]);
  501. exit(-1);
  502. }
  503. if (!system_wide && target_tid == -1 && target_pid == -1)
  504. all_tids[0] = child_pid;
  505. close(child_ready_pipe[1]);
  506. close(go_pipe[0]);
  507. /*
  508. * wait for child to settle
  509. */
  510. if (read(child_ready_pipe[0], &buf, 1) == -1) {
  511. perror("unable to read pipe");
  512. exit(-1);
  513. }
  514. close(child_ready_pipe[0]);
  515. }
  516. if ((!system_wide && no_inherit) || profile_cpu != -1) {
  517. open_counters(profile_cpu);
  518. } else {
  519. nr_cpus = read_cpu_map();
  520. for (i = 0; i < nr_cpus; i++)
  521. open_counters(cpumap[i]);
  522. }
  523. if (pipe_output) {
  524. err = perf_header__write_pipe(output);
  525. if (err < 0)
  526. return err;
  527. } else if (file_new) {
  528. err = perf_header__write(&session->header, output, false);
  529. if (err < 0)
  530. return err;
  531. }
  532. post_processing_offset = lseek(output, 0, SEEK_CUR);
  533. if (pipe_output) {
  534. err = event__synthesize_attrs(&session->header,
  535. process_synthesized_event,
  536. session);
  537. if (err < 0) {
  538. pr_err("Couldn't synthesize attrs.\n");
  539. return err;
  540. }
  541. err = event__synthesize_event_types(process_synthesized_event,
  542. session);
  543. if (err < 0) {
  544. pr_err("Couldn't synthesize event_types.\n");
  545. return err;
  546. }
  547. if (have_tracepoints(attrs, nr_counters)) {
  548. /*
  549. * FIXME err <= 0 here actually means that
  550. * there were no tracepoints so its not really
  551. * an error, just that we don't need to
  552. * synthesize anything. We really have to
  553. * return this more properly and also
  554. * propagate errors that now are calling die()
  555. */
  556. err = event__synthesize_tracing_data(output, attrs,
  557. nr_counters,
  558. process_synthesized_event,
  559. session);
  560. if (err <= 0) {
  561. pr_err("Couldn't record tracing data.\n");
  562. return err;
  563. }
  564. advance_output(err);
  565. }
  566. }
  567. machine = perf_session__find_host_machine(session);
  568. if (!machine) {
  569. pr_err("Couldn't find native kernel information.\n");
  570. return -1;
  571. }
  572. err = event__synthesize_kernel_mmap(process_synthesized_event,
  573. session, machine, "_text");
  574. if (err < 0)
  575. err = event__synthesize_kernel_mmap(process_synthesized_event,
  576. session, machine, "_stext");
  577. if (err < 0) {
  578. pr_err("Couldn't record kernel reference relocation symbol.\n");
  579. return err;
  580. }
  581. err = event__synthesize_modules(process_synthesized_event,
  582. session, machine);
  583. if (err < 0) {
  584. pr_err("Couldn't record kernel reference relocation symbol.\n");
  585. return err;
  586. }
  587. if (perf_guest)
  588. perf_session__process_machines(session, event__synthesize_guest_os);
  589. if (!system_wide && profile_cpu == -1)
  590. event__synthesize_thread(target_tid, process_synthesized_event,
  591. session);
  592. else
  593. event__synthesize_threads(process_synthesized_event, session);
  594. if (realtime_prio) {
  595. struct sched_param param;
  596. param.sched_priority = realtime_prio;
  597. if (sched_setscheduler(0, SCHED_FIFO, &param)) {
  598. pr_err("Could not set realtime priority.\n");
  599. exit(-1);
  600. }
  601. }
  602. /*
  603. * Let the child rip
  604. */
  605. if (forks)
  606. close(go_pipe[1]);
  607. for (;;) {
  608. int hits = samples;
  609. int thread;
  610. mmap_read_all();
  611. if (hits == samples) {
  612. if (done)
  613. break;
  614. err = poll(event_array, nr_poll, -1);
  615. waking++;
  616. }
  617. if (done) {
  618. for (i = 0; i < nr_cpu; i++) {
  619. for (counter = 0;
  620. counter < nr_counters;
  621. counter++) {
  622. for (thread = 0;
  623. thread < thread_num;
  624. thread++)
  625. ioctl(fd[i][counter][thread],
  626. PERF_EVENT_IOC_DISABLE);
  627. }
  628. }
  629. }
  630. }
  631. fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
  632. /*
  633. * Approximate RIP event size: 24 bytes.
  634. */
  635. fprintf(stderr,
  636. "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
  637. (double)bytes_written / 1024.0 / 1024.0,
  638. output_name,
  639. bytes_written / 24);
  640. return 0;
  641. }
  642. static const char * const record_usage[] = {
  643. "perf record [<options>] [<command>]",
  644. "perf record [<options>] -- <command> [<options>]",
  645. NULL
  646. };
  647. static bool force, append_file;
  648. static const struct option options[] = {
  649. OPT_CALLBACK('e', "event", NULL, "event",
  650. "event selector. use 'perf list' to list available events",
  651. parse_events),
  652. OPT_CALLBACK(0, "filter", NULL, "filter",
  653. "event filter", parse_filter),
  654. OPT_INTEGER('p', "pid", &target_pid,
  655. "record events on existing process id"),
  656. OPT_INTEGER('t', "tid", &target_tid,
  657. "record events on existing thread id"),
  658. OPT_INTEGER('r', "realtime", &realtime_prio,
  659. "collect data with this RT SCHED_FIFO priority"),
  660. OPT_BOOLEAN('R', "raw-samples", &raw_samples,
  661. "collect raw sample records from all opened counters"),
  662. OPT_BOOLEAN('a', "all-cpus", &system_wide,
  663. "system-wide collection from all CPUs"),
  664. OPT_BOOLEAN('A', "append", &append_file,
  665. "append to the output file to do incremental profiling"),
  666. OPT_INTEGER('C', "profile_cpu", &profile_cpu,
  667. "CPU to profile on"),
  668. OPT_BOOLEAN('f', "force", &force,
  669. "overwrite existing data file (deprecated)"),
  670. OPT_U64('c', "count", &user_interval, "event period to sample"),
  671. OPT_STRING('o', "output", &output_name, "file",
  672. "output file name"),
  673. OPT_BOOLEAN('i', "no-inherit", &no_inherit,
  674. "child tasks do not inherit counters"),
  675. OPT_UINTEGER('F', "freq", &user_freq, "profile at this frequency"),
  676. OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
  677. OPT_BOOLEAN('g', "call-graph", &call_graph,
  678. "do call-graph (stack chain/backtrace) recording"),
  679. OPT_INCR('v', "verbose", &verbose,
  680. "be more verbose (show counter open errors, etc)"),
  681. OPT_BOOLEAN('s', "stat", &inherit_stat,
  682. "per thread counts"),
  683. OPT_BOOLEAN('d', "data", &sample_address,
  684. "Sample addresses"),
  685. OPT_BOOLEAN('n', "no-samples", &no_samples,
  686. "don't sample"),
  687. OPT_END()
  688. };
  689. int cmd_record(int argc, const char **argv, const char *prefix __used)
  690. {
  691. int i,j;
  692. argc = parse_options(argc, argv, options, record_usage,
  693. PARSE_OPT_STOP_AT_NON_OPTION);
  694. if (!argc && target_pid == -1 && target_tid == -1 &&
  695. !system_wide && profile_cpu == -1)
  696. usage_with_options(record_usage, options);
  697. if (force && append_file) {
  698. fprintf(stderr, "Can't overwrite and append at the same time."
  699. " You need to choose between -f and -A");
  700. usage_with_options(record_usage, options);
  701. } else if (append_file) {
  702. write_mode = WRITE_APPEND;
  703. } else {
  704. write_mode = WRITE_FORCE;
  705. }
  706. symbol__init();
  707. if (!nr_counters) {
  708. nr_counters = 1;
  709. attrs[0].type = PERF_TYPE_HARDWARE;
  710. attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
  711. }
  712. if (target_pid != -1) {
  713. target_tid = target_pid;
  714. thread_num = find_all_tid(target_pid, &all_tids);
  715. if (thread_num <= 0) {
  716. fprintf(stderr, "Can't find all threads of pid %d\n",
  717. target_pid);
  718. usage_with_options(record_usage, options);
  719. }
  720. } else {
  721. all_tids=malloc(sizeof(pid_t));
  722. if (!all_tids)
  723. return -ENOMEM;
  724. all_tids[0] = target_tid;
  725. thread_num = 1;
  726. }
  727. for (i = 0; i < MAX_NR_CPUS; i++) {
  728. for (j = 0; j < MAX_COUNTERS; j++) {
  729. fd[i][j] = malloc(sizeof(int)*thread_num);
  730. if (!fd[i][j])
  731. return -ENOMEM;
  732. }
  733. }
  734. event_array = malloc(
  735. sizeof(struct pollfd)*MAX_NR_CPUS*MAX_COUNTERS*thread_num);
  736. if (!event_array)
  737. return -ENOMEM;
  738. if (user_interval != ULLONG_MAX)
  739. default_interval = user_interval;
  740. if (user_freq != UINT_MAX)
  741. freq = user_freq;
  742. /*
  743. * User specified count overrides default frequency.
  744. */
  745. if (default_interval)
  746. freq = 0;
  747. else if (freq) {
  748. default_interval = freq;
  749. } else {
  750. fprintf(stderr, "frequency and count are zero, aborting\n");
  751. exit(EXIT_FAILURE);
  752. }
  753. return __cmd_record(argc, argv);
  754. }