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.
8 #define _FILE_OFFSET_BITS 64
14 #include "util/build-id.h"
15 #include "util/util.h"
16 #include "util/parse-options.h"
17 #include "util/parse-events.h"
19 #include "util/header.h"
20 #include "util/event.h"
21 #include "util/debug.h"
22 #include "util/session.h"
23 #include "util/symbol.h"
24 #include "util/cpumap.h"
35 static int *fd[MAX_NR_CPUS][MAX_COUNTERS];
37 static u64 user_interval = ULLONG_MAX;
38 static u64 default_interval = 0;
40 static int nr_cpus = 0;
41 static unsigned int page_size;
42 static unsigned int mmap_pages = 128;
43 static unsigned int user_freq = UINT_MAX;
44 static int freq = 1000;
46 static int pipe_output = 0;
47 static const char *output_name = "perf.data";
49 static int realtime_prio = 0;
50 static bool raw_samples = false;
51 static bool system_wide = false;
52 static pid_t target_pid = -1;
53 static pid_t target_tid = -1;
54 static pid_t *all_tids = NULL;
55 static int thread_num = 0;
56 static pid_t child_pid = -1;
57 static bool no_inherit = false;
58 static enum write_mode_t write_mode = WRITE_FORCE;
59 static bool call_graph = false;
60 static bool inherit_stat = false;
61 static bool no_samples = false;
62 static bool sample_address = false;
63 static bool no_buildid = false;
64 static bool no_buildid_cache = false;
66 static long samples = 0;
67 static u64 bytes_written = 0;
69 static struct pollfd *event_array;
71 static int nr_poll = 0;
72 static int nr_cpu = 0;
74 static int file_new = 1;
75 static off_t post_processing_offset;
77 static struct perf_session *session;
78 static const char *cpu_list;
87 static struct mmap_data mmap_array[MAX_NR_CPUS];
89 static unsigned long mmap_read_head(struct mmap_data *md)
91 struct perf_event_mmap_page *pc = md->base;
100 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
102 struct perf_event_mmap_page *pc = md->base;
105 * ensure all reads are done before we write the tail out.
108 pc->data_tail = tail;
111 static void advance_output(size_t size)
113 bytes_written += size;
116 static void write_output(void *buf, size_t size)
119 int ret = write(output, buf, size);
122 die("failed to write");
127 bytes_written += ret;
131 static int process_synthesized_event(event_t *event,
132 struct perf_session *self __used)
134 write_output(event, event->header.size);
138 static void mmap_read(struct mmap_data *md)
140 unsigned int head = mmap_read_head(md);
141 unsigned int old = md->prev;
142 unsigned char *data = md->base + page_size;
148 * If we're further behind than half the buffer, there's a chance
149 * the writer will bite our tail and mess up the samples under us.
151 * If we somehow ended up ahead of the head, we got messed up.
153 * In either case, truncate and restart at head.
157 fprintf(stderr, "WARNING: failed to keep up with mmap data\n");
159 * head points to a known good entry, start there.
169 if ((old & md->mask) + size != (head & md->mask)) {
170 buf = &data[old & md->mask];
171 size = md->mask + 1 - (old & md->mask);
174 write_output(buf, size);
177 buf = &data[old & md->mask];
181 write_output(buf, size);
184 mmap_write_tail(md, old);
187 static volatile int done = 0;
188 static volatile int signr = -1;
190 static void sig_handler(int sig)
196 static void sig_atexit(void)
199 kill(child_pid, SIGTERM);
204 signal(signr, SIG_DFL);
205 kill(getpid(), signr);
210 static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
212 struct perf_header_attr *h_attr;
214 if (nr < session->header.attrs) {
215 h_attr = session->header.attr[nr];
217 h_attr = perf_header_attr__new(a);
219 if (perf_header__add_attr(&session->header, h_attr) < 0) {
220 perf_header_attr__delete(h_attr);
228 static void create_counter(int counter, int cpu)
230 char *filter = filters[counter];
231 struct perf_event_attr *attr = attrs + counter;
232 struct perf_header_attr *h_attr;
233 int track = !counter; /* only the first counter needs these */
243 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
244 PERF_FORMAT_TOTAL_TIME_RUNNING |
247 attr->sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
250 attr->sample_type |= PERF_SAMPLE_ID;
253 * We default some events to a 1 default interval. But keep
254 * it a weak assumption overridable by the user.
256 if (!attr->sample_period || (user_freq != UINT_MAX &&
257 user_interval != ULLONG_MAX)) {
259 attr->sample_type |= PERF_SAMPLE_PERIOD;
261 attr->sample_freq = freq;
263 attr->sample_period = default_interval;
268 attr->sample_freq = 0;
271 attr->inherit_stat = 1;
273 if (sample_address) {
274 attr->sample_type |= PERF_SAMPLE_ADDR;
275 attr->mmap_data = track;
279 attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
282 attr->sample_type |= PERF_SAMPLE_CPU;
285 attr->sample_type |= PERF_SAMPLE_TIME;
286 attr->sample_type |= PERF_SAMPLE_RAW;
287 attr->sample_type |= PERF_SAMPLE_CPU;
292 attr->inherit = !no_inherit;
293 if (target_pid == -1 && target_tid == -1 && !system_wide) {
295 attr->enable_on_exec = 1;
298 for (thread_index = 0; thread_index < thread_num; thread_index++) {
300 fd[nr_cpu][counter][thread_index] = sys_perf_event_open(attr,
301 all_tids[thread_index], cpu, group_fd, 0);
303 if (fd[nr_cpu][counter][thread_index] < 0) {
306 if (err == EPERM || err == EACCES)
307 die("Permission error - are you root?\n"
308 "\t Consider tweaking"
309 " /proc/sys/kernel/perf_event_paranoid.\n");
310 else if (err == ENODEV && cpu_list) {
311 die("No such device - did you specify"
312 " an out-of-range profile CPU?\n");
316 * If it's cycles then fall back to hrtimer
317 * based cpu-clock-tick sw counter, which
318 * is always available even if no PMU support:
320 if (attr->type == PERF_TYPE_HARDWARE
321 && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
324 warning(" ... trying to fall back to cpu-clock-ticks\n");
325 attr->type = PERF_TYPE_SOFTWARE;
326 attr->config = PERF_COUNT_SW_CPU_CLOCK;
330 error("sys_perf_event_open() syscall returned with %d (%s). /bin/dmesg may provide additional information.\n",
331 fd[nr_cpu][counter][thread_index], strerror(err));
333 #if defined(__i386__) || defined(__x86_64__)
334 if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
335 die("No hardware sampling interrupt available."
336 " No APIC? If so then you can boot the kernel"
337 " with the \"lapic\" boot parameter to"
338 " force-enable it.\n");
341 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
345 h_attr = get_header_attr(attr, counter);
350 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
351 fprintf(stderr, "incompatible append\n");
356 if (read(fd[nr_cpu][counter][thread_index], &read_data, sizeof(read_data)) == -1) {
357 perror("Unable to read perf file descriptor");
361 if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
362 pr_warning("Not enough memory to add id\n");
366 assert(fd[nr_cpu][counter][thread_index] >= 0);
367 fcntl(fd[nr_cpu][counter][thread_index], F_SETFL, O_NONBLOCK);
370 * First counter acts as the group leader:
372 if (group && group_fd == -1)
373 group_fd = fd[nr_cpu][counter][thread_index];
375 if (counter || thread_index) {
376 ret = ioctl(fd[nr_cpu][counter][thread_index],
377 PERF_EVENT_IOC_SET_OUTPUT,
380 error("failed to set output: %d (%s)\n", errno,
385 mmap_array[nr_cpu].counter = counter;
386 mmap_array[nr_cpu].prev = 0;
387 mmap_array[nr_cpu].mask = mmap_pages*page_size - 1;
388 mmap_array[nr_cpu].base = mmap(NULL, (mmap_pages+1)*page_size,
389 PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter][thread_index], 0);
390 if (mmap_array[nr_cpu].base == MAP_FAILED) {
391 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
395 event_array[nr_poll].fd = fd[nr_cpu][counter][thread_index];
396 event_array[nr_poll].events = POLLIN;
400 if (filter != NULL) {
401 ret = ioctl(fd[nr_cpu][counter][thread_index],
402 PERF_EVENT_IOC_SET_FILTER, filter);
404 error("failed to set filter with %d (%s)\n", errno,
412 static void open_counters(int cpu)
417 for (counter = 0; counter < nr_counters; counter++)
418 create_counter(counter, cpu);
423 static int process_buildids(void)
425 u64 size = lseek(output, 0, SEEK_CUR);
430 session->fd = output;
431 return __perf_session__process_events(session, post_processing_offset,
432 size - post_processing_offset,
433 size, &build_id__mark_dso_hit_ops);
436 static void atexit_header(void)
439 session->header.data_size += bytes_written;
443 perf_header__write(&session->header, output, true);
444 perf_session__delete(session);
449 static void event__synthesize_guest_os(struct machine *machine, void *data)
452 struct perf_session *psession = data;
454 if (machine__is_host(machine))
458 *As for guest kernel when processing subcommand record&report,
459 *we arrange module mmap prior to guest kernel mmap and trigger
460 *a preload dso because default guest module symbols are loaded
461 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
462 *method is used to avoid symbol missing when the first addr is
463 *in module instead of in guest kernel.
465 err = event__synthesize_modules(process_synthesized_event,
468 pr_err("Couldn't record guest kernel [%d]'s reference"
469 " relocation symbol.\n", machine->pid);
472 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
473 * have no _text sometimes.
475 err = event__synthesize_kernel_mmap(process_synthesized_event,
476 psession, machine, "_text");
478 err = event__synthesize_kernel_mmap(process_synthesized_event,
479 psession, machine, "_stext");
481 pr_err("Couldn't record guest kernel [%d]'s reference"
482 " relocation symbol.\n", machine->pid);
485 static struct perf_event_header finished_round_event = {
486 .size = sizeof(struct perf_event_header),
487 .type = PERF_RECORD_FINISHED_ROUND,
490 static void mmap_read_all(void)
494 for (i = 0; i < nr_cpu; i++) {
495 if (mmap_array[i].base)
496 mmap_read(&mmap_array[i]);
499 if (perf_header__has_feat(&session->header, HEADER_TRACE_INFO))
500 write_output(&finished_round_event, sizeof(finished_round_event));
503 static int __cmd_record(int argc, const char **argv)
509 unsigned long waking = 0;
510 int child_ready_pipe[2], go_pipe[2];
511 const bool forks = argc > 0;
513 struct machine *machine;
515 page_size = sysconf(_SC_PAGE_SIZE);
518 signal(SIGCHLD, sig_handler);
519 signal(SIGINT, sig_handler);
521 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
522 perror("failed to create pipes");
526 if (!strcmp(output_name, "-"))
528 else if (!stat(output_name, &st) && st.st_size) {
529 if (write_mode == WRITE_FORCE) {
530 char oldname[PATH_MAX];
531 snprintf(oldname, sizeof(oldname), "%s.old",
534 rename(output_name, oldname);
536 } else if (write_mode == WRITE_APPEND) {
537 write_mode = WRITE_FORCE;
540 flags = O_CREAT|O_RDWR;
541 if (write_mode == WRITE_APPEND)
547 output = STDOUT_FILENO;
549 output = open(output_name, flags, S_IRUSR | S_IWUSR);
551 perror("failed to create output file");
555 session = perf_session__new(output_name, O_WRONLY,
556 write_mode == WRITE_FORCE, false);
557 if (session == NULL) {
558 pr_err("Not enough memory for reading perf file header\n");
563 perf_header__set_feat(&session->header, HEADER_BUILD_ID);
566 err = perf_header__read(session, output);
568 goto out_delete_session;
571 if (have_tracepoints(attrs, nr_counters))
572 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
575 * perf_session__delete(session) will be called at atexit_header()
577 atexit(atexit_header);
582 perror("failed to fork");
589 close(child_ready_pipe[0]);
591 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
594 * Do a dummy execvp to get the PLT entry resolved,
595 * so we avoid the resolver overhead on the real
598 execvp("", (char **)argv);
601 * Tell the parent we're ready to go
603 close(child_ready_pipe[1]);
606 * Wait until the parent tells us to go.
608 if (read(go_pipe[0], &buf, 1) == -1)
609 perror("unable to read pipe");
611 execvp(argv[0], (char **)argv);
617 if (!system_wide && target_tid == -1 && target_pid == -1)
618 all_tids[0] = child_pid;
620 close(child_ready_pipe[1]);
623 * wait for child to settle
625 if (read(child_ready_pipe[0], &buf, 1) == -1) {
626 perror("unable to read pipe");
629 close(child_ready_pipe[0]);
632 nr_cpus = read_cpu_map(cpu_list);
634 perror("failed to collect number of CPUs");
638 if (!system_wide && no_inherit && !cpu_list) {
641 for (i = 0; i < nr_cpus; i++)
642 open_counters(cpumap[i]);
646 err = perf_header__write_pipe(output);
649 } else if (file_new) {
650 err = perf_header__write(&session->header, output, false);
655 post_processing_offset = lseek(output, 0, SEEK_CUR);
658 err = event__synthesize_attrs(&session->header,
659 process_synthesized_event,
662 pr_err("Couldn't synthesize attrs.\n");
666 err = event__synthesize_event_types(process_synthesized_event,
669 pr_err("Couldn't synthesize event_types.\n");
673 if (have_tracepoints(attrs, nr_counters)) {
675 * FIXME err <= 0 here actually means that
676 * there were no tracepoints so its not really
677 * an error, just that we don't need to
678 * synthesize anything. We really have to
679 * return this more properly and also
680 * propagate errors that now are calling die()
682 err = event__synthesize_tracing_data(output, attrs,
684 process_synthesized_event,
687 pr_err("Couldn't record tracing data.\n");
694 machine = perf_session__find_host_machine(session);
696 pr_err("Couldn't find native kernel information.\n");
700 err = event__synthesize_kernel_mmap(process_synthesized_event,
701 session, machine, "_text");
703 err = event__synthesize_kernel_mmap(process_synthesized_event,
704 session, machine, "_stext");
706 pr_err("Couldn't record kernel reference relocation symbol\n"
707 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
708 "Check /proc/kallsyms permission or run as root.\n");
710 err = event__synthesize_modules(process_synthesized_event,
713 pr_err("Couldn't record kernel module information.\n"
714 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
715 "Check /proc/modules permission or run as root.\n");
718 perf_session__process_machines(session, event__synthesize_guest_os);
721 event__synthesize_thread(target_tid, process_synthesized_event,
724 event__synthesize_threads(process_synthesized_event, session);
727 struct sched_param param;
729 param.sched_priority = realtime_prio;
730 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
731 pr_err("Could not set realtime priority.\n");
748 if (hits == samples) {
751 err = poll(event_array, nr_poll, -1);
756 for (i = 0; i < nr_cpu; i++) {
758 counter < nr_counters;
763 ioctl(fd[i][counter][thread],
764 PERF_EVENT_IOC_DISABLE);
773 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
776 * Approximate RIP event size: 24 bytes.
779 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
780 (double)bytes_written / 1024.0 / 1024.0,
787 perf_session__delete(session);
791 static const char * const record_usage[] = {
792 "perf record [<options>] [<command>]",
793 "perf record [<options>] -- <command> [<options>]",
797 static bool force, append_file;
799 const struct option record_options[] = {
800 OPT_CALLBACK('e', "event", NULL, "event",
801 "event selector. use 'perf list' to list available events",
803 OPT_CALLBACK(0, "filter", NULL, "filter",
804 "event filter", parse_filter),
805 OPT_INTEGER('p', "pid", &target_pid,
806 "record events on existing process id"),
807 OPT_INTEGER('t', "tid", &target_tid,
808 "record events on existing thread id"),
809 OPT_INTEGER('r', "realtime", &realtime_prio,
810 "collect data with this RT SCHED_FIFO priority"),
811 OPT_BOOLEAN('R', "raw-samples", &raw_samples,
812 "collect raw sample records from all opened counters"),
813 OPT_BOOLEAN('a', "all-cpus", &system_wide,
814 "system-wide collection from all CPUs"),
815 OPT_BOOLEAN('A', "append", &append_file,
816 "append to the output file to do incremental profiling"),
817 OPT_STRING('C', "cpu", &cpu_list, "cpu",
818 "list of cpus to monitor"),
819 OPT_BOOLEAN('f', "force", &force,
820 "overwrite existing data file (deprecated)"),
821 OPT_U64('c', "count", &user_interval, "event period to sample"),
822 OPT_STRING('o', "output", &output_name, "file",
824 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
825 "child tasks do not inherit counters"),
826 OPT_UINTEGER('F', "freq", &user_freq, "profile at this frequency"),
827 OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
828 OPT_BOOLEAN('g', "call-graph", &call_graph,
829 "do call-graph (stack chain/backtrace) recording"),
830 OPT_INCR('v', "verbose", &verbose,
831 "be more verbose (show counter open errors, etc)"),
832 OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
833 OPT_BOOLEAN('s', "stat", &inherit_stat,
834 "per thread counts"),
835 OPT_BOOLEAN('d', "data", &sample_address,
837 OPT_BOOLEAN('n', "no-samples", &no_samples,
839 OPT_BOOLEAN('N', "no-buildid-cache", &no_buildid_cache,
840 "do not update the buildid cache"),
841 OPT_BOOLEAN('B', "no-buildid", &no_buildid,
842 "do not collect buildids in perf.data"),
846 int cmd_record(int argc, const char **argv, const char *prefix __used)
848 int i, j, err = -ENOMEM;
850 argc = parse_options(argc, argv, record_options, record_usage,
851 PARSE_OPT_STOP_AT_NON_OPTION);
852 if (!argc && target_pid == -1 && target_tid == -1 &&
853 !system_wide && !cpu_list)
854 usage_with_options(record_usage, record_options);
856 if (force && append_file) {
857 fprintf(stderr, "Can't overwrite and append at the same time."
858 " You need to choose between -f and -A");
859 usage_with_options(record_usage, record_options);
860 } else if (append_file) {
861 write_mode = WRITE_APPEND;
863 write_mode = WRITE_FORCE;
868 if (no_buildid_cache || no_buildid)
869 disable_buildid_cache();
873 attrs[0].type = PERF_TYPE_HARDWARE;
874 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
877 if (target_pid != -1) {
878 target_tid = target_pid;
879 thread_num = find_all_tid(target_pid, &all_tids);
880 if (thread_num <= 0) {
881 fprintf(stderr, "Can't find all threads of pid %d\n",
883 usage_with_options(record_usage, record_options);
886 all_tids=malloc(sizeof(pid_t));
888 goto out_symbol_exit;
890 all_tids[0] = target_tid;
894 for (i = 0; i < MAX_NR_CPUS; i++) {
895 for (j = 0; j < MAX_COUNTERS; j++) {
896 fd[i][j] = malloc(sizeof(int)*thread_num);
901 event_array = malloc(
902 sizeof(struct pollfd)*MAX_NR_CPUS*MAX_COUNTERS*thread_num);
906 if (user_interval != ULLONG_MAX)
907 default_interval = user_interval;
908 if (user_freq != UINT_MAX)
912 * User specified count overrides default frequency.
914 if (default_interval)
917 default_interval = freq;
919 fprintf(stderr, "frequency and count are zero, aborting\n");
921 goto out_free_event_array;
924 err = __cmd_record(argc, argv);
926 out_free_event_array:
929 for (i = 0; i < MAX_NR_CPUS; i++) {
930 for (j = 0; j < MAX_COUNTERS; j++)