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;
64 static long samples = 0;
65 static u64 bytes_written = 0;
67 static struct pollfd *event_array;
69 static int nr_poll = 0;
70 static int nr_cpu = 0;
72 static int file_new = 1;
73 static off_t post_processing_offset;
75 static struct perf_session *session;
76 static const char *cpu_list;
85 static struct mmap_data mmap_array[MAX_NR_CPUS];
87 static unsigned long mmap_read_head(struct mmap_data *md)
89 struct perf_event_mmap_page *pc = md->base;
98 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
100 struct perf_event_mmap_page *pc = md->base;
103 * ensure all reads are done before we write the tail out.
106 pc->data_tail = tail;
109 static void advance_output(size_t size)
111 bytes_written += size;
114 static void write_output(void *buf, size_t size)
117 int ret = write(output, buf, size);
120 die("failed to write");
125 bytes_written += ret;
129 static int process_synthesized_event(event_t *event,
130 struct perf_session *self __used)
132 write_output(event, event->header.size);
136 static void mmap_read(struct mmap_data *md)
138 unsigned int head = mmap_read_head(md);
139 unsigned int old = md->prev;
140 unsigned char *data = md->base + page_size;
146 * If we're further behind than half the buffer, there's a chance
147 * the writer will bite our tail and mess up the samples under us.
149 * If we somehow ended up ahead of the head, we got messed up.
151 * In either case, truncate and restart at head.
155 fprintf(stderr, "WARNING: failed to keep up with mmap data\n");
157 * head points to a known good entry, start there.
167 if ((old & md->mask) + size != (head & md->mask)) {
168 buf = &data[old & md->mask];
169 size = md->mask + 1 - (old & md->mask);
172 write_output(buf, size);
175 buf = &data[old & md->mask];
179 write_output(buf, size);
182 mmap_write_tail(md, old);
185 static volatile int done = 0;
186 static volatile int signr = -1;
188 static void sig_handler(int sig)
194 static void sig_atexit(void)
197 kill(child_pid, SIGTERM);
202 signal(signr, SIG_DFL);
203 kill(getpid(), signr);
208 static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
210 struct perf_header_attr *h_attr;
212 if (nr < session->header.attrs) {
213 h_attr = session->header.attr[nr];
215 h_attr = perf_header_attr__new(a);
217 if (perf_header__add_attr(&session->header, h_attr) < 0) {
218 perf_header_attr__delete(h_attr);
226 static void create_counter(int counter, int cpu)
228 char *filter = filters[counter];
229 struct perf_event_attr *attr = attrs + counter;
230 struct perf_header_attr *h_attr;
231 int track = !counter; /* only the first counter needs these */
241 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
242 PERF_FORMAT_TOTAL_TIME_RUNNING |
245 attr->sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
248 attr->sample_type |= PERF_SAMPLE_ID;
251 * We default some events to a 1 default interval. But keep
252 * it a weak assumption overridable by the user.
254 if (!attr->sample_period || (user_freq != UINT_MAX &&
255 user_interval != ULLONG_MAX)) {
257 attr->sample_type |= PERF_SAMPLE_PERIOD;
259 attr->sample_freq = freq;
261 attr->sample_period = default_interval;
266 attr->sample_freq = 0;
269 attr->inherit_stat = 1;
272 attr->sample_type |= PERF_SAMPLE_ADDR;
275 attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
278 attr->sample_type |= PERF_SAMPLE_CPU;
281 attr->sample_type |= PERF_SAMPLE_TIME;
282 attr->sample_type |= PERF_SAMPLE_RAW;
283 attr->sample_type |= PERF_SAMPLE_CPU;
288 attr->inherit = !no_inherit;
289 if (target_pid == -1 && target_tid == -1 && !system_wide) {
291 attr->enable_on_exec = 1;
294 for (thread_index = 0; thread_index < thread_num; thread_index++) {
296 fd[nr_cpu][counter][thread_index] = sys_perf_event_open(attr,
297 all_tids[thread_index], cpu, group_fd, 0);
299 if (fd[nr_cpu][counter][thread_index] < 0) {
302 if (err == EPERM || err == EACCES)
303 die("Permission error - are you root?\n"
304 "\t Consider tweaking"
305 " /proc/sys/kernel/perf_event_paranoid.\n");
306 else if (err == ENODEV && cpu_list) {
307 die("No such device - did you specify"
308 " an out-of-range profile CPU?\n");
312 * If it's cycles then fall back to hrtimer
313 * based cpu-clock-tick sw counter, which
314 * is always available even if no PMU support:
316 if (attr->type == PERF_TYPE_HARDWARE
317 && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
320 warning(" ... trying to fall back to cpu-clock-ticks\n");
321 attr->type = PERF_TYPE_SOFTWARE;
322 attr->config = PERF_COUNT_SW_CPU_CLOCK;
326 error("perfcounter syscall returned with %d (%s)\n",
327 fd[nr_cpu][counter][thread_index], strerror(err));
329 #if defined(__i386__) || defined(__x86_64__)
330 if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
331 die("No hardware sampling interrupt available."
332 " No APIC? If so then you can boot the kernel"
333 " with the \"lapic\" boot parameter to"
334 " force-enable it.\n");
337 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
341 h_attr = get_header_attr(attr, counter);
346 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
347 fprintf(stderr, "incompatible append\n");
352 if (read(fd[nr_cpu][counter][thread_index], &read_data, sizeof(read_data)) == -1) {
353 perror("Unable to read perf file descriptor\n");
357 if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
358 pr_warning("Not enough memory to add id\n");
362 assert(fd[nr_cpu][counter][thread_index] >= 0);
363 fcntl(fd[nr_cpu][counter][thread_index], F_SETFL, O_NONBLOCK);
366 * First counter acts as the group leader:
368 if (group && group_fd == -1)
369 group_fd = fd[nr_cpu][counter][thread_index];
371 if (counter || thread_index) {
372 ret = ioctl(fd[nr_cpu][counter][thread_index],
373 PERF_EVENT_IOC_SET_OUTPUT,
376 error("failed to set output: %d (%s)\n", errno,
381 mmap_array[nr_cpu].counter = counter;
382 mmap_array[nr_cpu].prev = 0;
383 mmap_array[nr_cpu].mask = mmap_pages*page_size - 1;
384 mmap_array[nr_cpu].base = mmap(NULL, (mmap_pages+1)*page_size,
385 PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter][thread_index], 0);
386 if (mmap_array[nr_cpu].base == MAP_FAILED) {
387 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
391 event_array[nr_poll].fd = fd[nr_cpu][counter][thread_index];
392 event_array[nr_poll].events = POLLIN;
396 if (filter != NULL) {
397 ret = ioctl(fd[nr_cpu][counter][thread_index],
398 PERF_EVENT_IOC_SET_FILTER, filter);
400 error("failed to set filter with %d (%s)\n", errno,
408 static void open_counters(int cpu)
413 for (counter = 0; counter < nr_counters; counter++)
414 create_counter(counter, cpu);
419 static int process_buildids(void)
421 u64 size = lseek(output, 0, SEEK_CUR);
426 session->fd = output;
427 return __perf_session__process_events(session, post_processing_offset,
428 size - post_processing_offset,
429 size, &build_id__mark_dso_hit_ops);
432 static void atexit_header(void)
435 session->header.data_size += bytes_written;
438 perf_header__write(&session->header, output, true);
442 static void event__synthesize_guest_os(struct machine *machine, void *data)
445 char *guest_kallsyms;
447 struct perf_session *psession = data;
449 if (machine__is_host(machine))
453 *As for guest kernel when processing subcommand record&report,
454 *we arrange module mmap prior to guest kernel mmap and trigger
455 *a preload dso because default guest module symbols are loaded
456 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
457 *method is used to avoid symbol missing when the first addr is
458 *in module instead of in guest kernel.
460 err = event__synthesize_modules(process_synthesized_event,
463 pr_err("Couldn't record guest kernel [%d]'s reference"
464 " relocation symbol.\n", machine->pid);
466 if (machine__is_default_guest(machine))
467 guest_kallsyms = (char *) symbol_conf.default_guest_kallsyms;
469 sprintf(path, "%s/proc/kallsyms", machine->root_dir);
470 guest_kallsyms = path;
474 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
475 * have no _text sometimes.
477 err = event__synthesize_kernel_mmap(process_synthesized_event,
478 psession, machine, "_text");
480 err = event__synthesize_kernel_mmap(process_synthesized_event,
481 psession, machine, "_stext");
483 pr_err("Couldn't record guest kernel [%d]'s reference"
484 " relocation symbol.\n", machine->pid);
487 static struct perf_event_header finished_round_event = {
488 .size = sizeof(struct perf_event_header),
489 .type = PERF_RECORD_FINISHED_ROUND,
492 static void mmap_read_all(void)
496 for (i = 0; i < nr_cpu; i++) {
497 if (mmap_array[i].base)
498 mmap_read(&mmap_array[i]);
501 if (perf_header__has_feat(&session->header, HEADER_TRACE_INFO))
502 write_output(&finished_round_event, sizeof(finished_round_event));
505 static int __cmd_record(int argc, const char **argv)
511 unsigned long waking = 0;
512 int child_ready_pipe[2], go_pipe[2];
513 const bool forks = argc > 0;
515 struct machine *machine;
517 page_size = sysconf(_SC_PAGE_SIZE);
520 signal(SIGCHLD, sig_handler);
521 signal(SIGINT, sig_handler);
523 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
524 perror("failed to create pipes");
528 if (!strcmp(output_name, "-"))
530 else if (!stat(output_name, &st) && st.st_size) {
531 if (write_mode == WRITE_FORCE) {
532 char oldname[PATH_MAX];
533 snprintf(oldname, sizeof(oldname), "%s.old",
536 rename(output_name, oldname);
538 } else if (write_mode == WRITE_APPEND) {
539 write_mode = WRITE_FORCE;
542 flags = O_CREAT|O_RDWR;
543 if (write_mode == WRITE_APPEND)
549 output = STDOUT_FILENO;
551 output = open(output_name, flags, S_IRUSR | S_IWUSR);
553 perror("failed to create output file");
557 session = perf_session__new(output_name, O_WRONLY,
558 write_mode == WRITE_FORCE, false);
559 if (session == NULL) {
560 pr_err("Not enough memory for reading perf file header\n");
565 err = perf_header__read(session, output);
570 if (have_tracepoints(attrs, nr_counters))
571 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
573 atexit(atexit_header);
578 perror("failed to fork");
585 close(child_ready_pipe[0]);
587 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
590 * Do a dummy execvp to get the PLT entry resolved,
591 * so we avoid the resolver overhead on the real
594 execvp("", (char **)argv);
597 * Tell the parent we're ready to go
599 close(child_ready_pipe[1]);
602 * Wait until the parent tells us to go.
604 if (read(go_pipe[0], &buf, 1) == -1)
605 perror("unable to read pipe");
607 execvp(argv[0], (char **)argv);
613 if (!system_wide && target_tid == -1 && target_pid == -1)
614 all_tids[0] = child_pid;
616 close(child_ready_pipe[1]);
619 * wait for child to settle
621 if (read(child_ready_pipe[0], &buf, 1) == -1) {
622 perror("unable to read pipe");
625 close(child_ready_pipe[0]);
628 nr_cpus = read_cpu_map(cpu_list);
630 perror("failed to collect number of CPUs\n");
634 if (!system_wide && no_inherit && !cpu_list) {
637 for (i = 0; i < nr_cpus; i++)
638 open_counters(cpumap[i]);
642 err = perf_header__write_pipe(output);
645 } else if (file_new) {
646 err = perf_header__write(&session->header, output, false);
651 post_processing_offset = lseek(output, 0, SEEK_CUR);
654 err = event__synthesize_attrs(&session->header,
655 process_synthesized_event,
658 pr_err("Couldn't synthesize attrs.\n");
662 err = event__synthesize_event_types(process_synthesized_event,
665 pr_err("Couldn't synthesize event_types.\n");
669 if (have_tracepoints(attrs, nr_counters)) {
671 * FIXME err <= 0 here actually means that
672 * there were no tracepoints so its not really
673 * an error, just that we don't need to
674 * synthesize anything. We really have to
675 * return this more properly and also
676 * propagate errors that now are calling die()
678 err = event__synthesize_tracing_data(output, attrs,
680 process_synthesized_event,
683 pr_err("Couldn't record tracing data.\n");
690 machine = perf_session__find_host_machine(session);
692 pr_err("Couldn't find native kernel information.\n");
696 err = event__synthesize_kernel_mmap(process_synthesized_event,
697 session, machine, "_text");
699 err = event__synthesize_kernel_mmap(process_synthesized_event,
700 session, machine, "_stext");
702 pr_err("Couldn't record kernel reference relocation symbol.\n");
706 err = event__synthesize_modules(process_synthesized_event,
709 pr_err("Couldn't record kernel reference relocation symbol.\n");
713 perf_session__process_machines(session, event__synthesize_guest_os);
715 if (!system_wide && cpu_list)
716 event__synthesize_thread(target_tid, process_synthesized_event,
719 event__synthesize_threads(process_synthesized_event, session);
722 struct sched_param param;
724 param.sched_priority = realtime_prio;
725 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
726 pr_err("Could not set realtime priority.\n");
743 if (hits == samples) {
746 err = poll(event_array, nr_poll, -1);
751 for (i = 0; i < nr_cpu; i++) {
753 counter < nr_counters;
758 ioctl(fd[i][counter][thread],
759 PERF_EVENT_IOC_DISABLE);
765 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
768 * Approximate RIP event size: 24 bytes.
771 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
772 (double)bytes_written / 1024.0 / 1024.0,
779 static const char * const record_usage[] = {
780 "perf record [<options>] [<command>]",
781 "perf record [<options>] -- <command> [<options>]",
785 static bool force, append_file;
787 static const struct option options[] = {
788 OPT_CALLBACK('e', "event", NULL, "event",
789 "event selector. use 'perf list' to list available events",
791 OPT_CALLBACK(0, "filter", NULL, "filter",
792 "event filter", parse_filter),
793 OPT_INTEGER('p', "pid", &target_pid,
794 "record events on existing process id"),
795 OPT_INTEGER('t', "tid", &target_tid,
796 "record events on existing thread id"),
797 OPT_INTEGER('r', "realtime", &realtime_prio,
798 "collect data with this RT SCHED_FIFO priority"),
799 OPT_BOOLEAN('R', "raw-samples", &raw_samples,
800 "collect raw sample records from all opened counters"),
801 OPT_BOOLEAN('a', "all-cpus", &system_wide,
802 "system-wide collection from all CPUs"),
803 OPT_BOOLEAN('A', "append", &append_file,
804 "append to the output file to do incremental profiling"),
805 OPT_STRING('C', "cpu", &cpu_list, "cpu",
806 "list of cpus to monitor"),
807 OPT_BOOLEAN('f', "force", &force,
808 "overwrite existing data file (deprecated)"),
809 OPT_U64('c', "count", &user_interval, "event period to sample"),
810 OPT_STRING('o', "output", &output_name, "file",
812 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
813 "child tasks do not inherit counters"),
814 OPT_UINTEGER('F', "freq", &user_freq, "profile at this frequency"),
815 OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
816 OPT_BOOLEAN('g', "call-graph", &call_graph,
817 "do call-graph (stack chain/backtrace) recording"),
818 OPT_INCR('v', "verbose", &verbose,
819 "be more verbose (show counter open errors, etc)"),
820 OPT_BOOLEAN('s', "stat", &inherit_stat,
821 "per thread counts"),
822 OPT_BOOLEAN('d', "data", &sample_address,
824 OPT_BOOLEAN('n', "no-samples", &no_samples,
829 int cmd_record(int argc, const char **argv, const char *prefix __used)
833 argc = parse_options(argc, argv, options, record_usage,
834 PARSE_OPT_STOP_AT_NON_OPTION);
835 if (!argc && target_pid == -1 && target_tid == -1 &&
836 !system_wide && !cpu_list)
837 usage_with_options(record_usage, options);
839 if (force && append_file) {
840 fprintf(stderr, "Can't overwrite and append at the same time."
841 " You need to choose between -f and -A");
842 usage_with_options(record_usage, options);
843 } else if (append_file) {
844 write_mode = WRITE_APPEND;
846 write_mode = WRITE_FORCE;
853 attrs[0].type = PERF_TYPE_HARDWARE;
854 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
857 if (target_pid != -1) {
858 target_tid = target_pid;
859 thread_num = find_all_tid(target_pid, &all_tids);
860 if (thread_num <= 0) {
861 fprintf(stderr, "Can't find all threads of pid %d\n",
863 usage_with_options(record_usage, options);
866 all_tids=malloc(sizeof(pid_t));
870 all_tids[0] = target_tid;
874 for (i = 0; i < MAX_NR_CPUS; i++) {
875 for (j = 0; j < MAX_COUNTERS; j++) {
876 fd[i][j] = malloc(sizeof(int)*thread_num);
881 event_array = malloc(
882 sizeof(struct pollfd)*MAX_NR_CPUS*MAX_COUNTERS*thread_num);
886 if (user_interval != ULLONG_MAX)
887 default_interval = user_interval;
888 if (user_freq != UINT_MAX)
892 * User specified count overrides default frequency.
894 if (default_interval)
897 default_interval = freq;
899 fprintf(stderr, "frequency and count are zero, aborting\n");
903 return __cmd_record(argc, argv);