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;
65 static long samples
= 0;
66 static u64 bytes_written
= 0;
68 static struct pollfd
*event_array
;
70 static int nr_poll
= 0;
71 static int nr_cpu
= 0;
73 static int file_new
= 1;
74 static off_t post_processing_offset
;
76 static struct perf_session
*session
;
77 static const char *cpu_list
;
86 static struct mmap_data mmap_array
[MAX_NR_CPUS
];
88 static unsigned long mmap_read_head(struct mmap_data
*md
)
90 struct perf_event_mmap_page
*pc
= md
->base
;
99 static void mmap_write_tail(struct mmap_data
*md
, unsigned long tail
)
101 struct perf_event_mmap_page
*pc
= md
->base
;
104 * ensure all reads are done before we write the tail out.
107 pc
->data_tail
= tail
;
110 static void advance_output(size_t size
)
112 bytes_written
+= size
;
115 static void write_output(void *buf
, size_t size
)
118 int ret
= write(output
, buf
, size
);
121 die("failed to write");
126 bytes_written
+= ret
;
130 static int process_synthesized_event(event_t
*event
,
131 struct perf_session
*self __used
)
133 write_output(event
, event
->header
.size
);
137 static void mmap_read(struct mmap_data
*md
)
139 unsigned int head
= mmap_read_head(md
);
140 unsigned int old
= md
->prev
;
141 unsigned char *data
= md
->base
+ page_size
;
147 * If we're further behind than half the buffer, there's a chance
148 * the writer will bite our tail and mess up the samples under us.
150 * If we somehow ended up ahead of the head, we got messed up.
152 * In either case, truncate and restart at head.
156 fprintf(stderr
, "WARNING: failed to keep up with mmap data\n");
158 * head points to a known good entry, start there.
168 if ((old
& md
->mask
) + size
!= (head
& md
->mask
)) {
169 buf
= &data
[old
& md
->mask
];
170 size
= md
->mask
+ 1 - (old
& md
->mask
);
173 write_output(buf
, size
);
176 buf
= &data
[old
& md
->mask
];
180 write_output(buf
, size
);
183 mmap_write_tail(md
, old
);
186 static volatile int done
= 0;
187 static volatile int signr
= -1;
189 static void sig_handler(int sig
)
195 static void sig_atexit(void)
198 kill(child_pid
, SIGTERM
);
203 signal(signr
, SIG_DFL
);
204 kill(getpid(), signr
);
209 static struct perf_header_attr
*get_header_attr(struct perf_event_attr
*a
, int nr
)
211 struct perf_header_attr
*h_attr
;
213 if (nr
< session
->header
.attrs
) {
214 h_attr
= session
->header
.attr
[nr
];
216 h_attr
= perf_header_attr__new(a
);
218 if (perf_header__add_attr(&session
->header
, h_attr
) < 0) {
219 perf_header_attr__delete(h_attr
);
227 static void create_counter(int counter
, int cpu
)
229 char *filter
= filters
[counter
];
230 struct perf_event_attr
*attr
= attrs
+ counter
;
231 struct perf_header_attr
*h_attr
;
232 int track
= !counter
; /* only the first counter needs these */
242 attr
->read_format
= PERF_FORMAT_TOTAL_TIME_ENABLED
|
243 PERF_FORMAT_TOTAL_TIME_RUNNING
|
246 attr
->sample_type
|= PERF_SAMPLE_IP
| PERF_SAMPLE_TID
;
249 attr
->sample_type
|= PERF_SAMPLE_ID
;
252 * We default some events to a 1 default interval. But keep
253 * it a weak assumption overridable by the user.
255 if (!attr
->sample_period
|| (user_freq
!= UINT_MAX
&&
256 user_interval
!= ULLONG_MAX
)) {
258 attr
->sample_type
|= PERF_SAMPLE_PERIOD
;
260 attr
->sample_freq
= freq
;
262 attr
->sample_period
= default_interval
;
267 attr
->sample_freq
= 0;
270 attr
->inherit_stat
= 1;
272 if (sample_address
) {
273 attr
->sample_type
|= PERF_SAMPLE_ADDR
;
274 attr
->mmap_data
= track
;
278 attr
->sample_type
|= PERF_SAMPLE_CALLCHAIN
;
281 attr
->sample_type
|= PERF_SAMPLE_CPU
;
284 attr
->sample_type
|= PERF_SAMPLE_TIME
;
285 attr
->sample_type
|= PERF_SAMPLE_RAW
;
286 attr
->sample_type
|= PERF_SAMPLE_CPU
;
291 attr
->inherit
= !no_inherit
;
292 if (target_pid
== -1 && target_tid
== -1 && !system_wide
) {
294 attr
->enable_on_exec
= 1;
297 for (thread_index
= 0; thread_index
< thread_num
; thread_index
++) {
299 fd
[nr_cpu
][counter
][thread_index
] = sys_perf_event_open(attr
,
300 all_tids
[thread_index
], cpu
, group_fd
, 0);
302 if (fd
[nr_cpu
][counter
][thread_index
] < 0) {
305 if (err
== EPERM
|| err
== EACCES
)
306 die("Permission error - are you root?\n"
307 "\t Consider tweaking"
308 " /proc/sys/kernel/perf_event_paranoid.\n");
309 else if (err
== ENODEV
&& cpu_list
) {
310 die("No such device - did you specify"
311 " an out-of-range profile CPU?\n");
315 * If it's cycles then fall back to hrtimer
316 * based cpu-clock-tick sw counter, which
317 * is always available even if no PMU support:
319 if (attr
->type
== PERF_TYPE_HARDWARE
320 && attr
->config
== PERF_COUNT_HW_CPU_CYCLES
) {
323 warning(" ... trying to fall back to cpu-clock-ticks\n");
324 attr
->type
= PERF_TYPE_SOFTWARE
;
325 attr
->config
= PERF_COUNT_SW_CPU_CLOCK
;
329 error("perfcounter syscall returned with %d (%s)\n",
330 fd
[nr_cpu
][counter
][thread_index
], strerror(err
));
332 #if defined(__i386__) || defined(__x86_64__)
333 if (attr
->type
== PERF_TYPE_HARDWARE
&& err
== EOPNOTSUPP
)
334 die("No hardware sampling interrupt available."
335 " No APIC? If so then you can boot the kernel"
336 " with the \"lapic\" boot parameter to"
337 " force-enable it.\n");
340 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
344 h_attr
= get_header_attr(attr
, counter
);
349 if (memcmp(&h_attr
->attr
, attr
, sizeof(*attr
))) {
350 fprintf(stderr
, "incompatible append\n");
355 if (read(fd
[nr_cpu
][counter
][thread_index
], &read_data
, sizeof(read_data
)) == -1) {
356 perror("Unable to read perf file descriptor");
360 if (perf_header_attr__add_id(h_attr
, read_data
.id
) < 0) {
361 pr_warning("Not enough memory to add id\n");
365 assert(fd
[nr_cpu
][counter
][thread_index
] >= 0);
366 fcntl(fd
[nr_cpu
][counter
][thread_index
], F_SETFL
, O_NONBLOCK
);
369 * First counter acts as the group leader:
371 if (group
&& group_fd
== -1)
372 group_fd
= fd
[nr_cpu
][counter
][thread_index
];
374 if (counter
|| thread_index
) {
375 ret
= ioctl(fd
[nr_cpu
][counter
][thread_index
],
376 PERF_EVENT_IOC_SET_OUTPUT
,
379 error("failed to set output: %d (%s)\n", errno
,
384 mmap_array
[nr_cpu
].counter
= counter
;
385 mmap_array
[nr_cpu
].prev
= 0;
386 mmap_array
[nr_cpu
].mask
= mmap_pages
*page_size
- 1;
387 mmap_array
[nr_cpu
].base
= mmap(NULL
, (mmap_pages
+1)*page_size
,
388 PROT_READ
|PROT_WRITE
, MAP_SHARED
, fd
[nr_cpu
][counter
][thread_index
], 0);
389 if (mmap_array
[nr_cpu
].base
== MAP_FAILED
) {
390 error("failed to mmap with %d (%s)\n", errno
, strerror(errno
));
394 event_array
[nr_poll
].fd
= fd
[nr_cpu
][counter
][thread_index
];
395 event_array
[nr_poll
].events
= POLLIN
;
399 if (filter
!= NULL
) {
400 ret
= ioctl(fd
[nr_cpu
][counter
][thread_index
],
401 PERF_EVENT_IOC_SET_FILTER
, filter
);
403 error("failed to set filter with %d (%s)\n", errno
,
411 static void open_counters(int cpu
)
416 for (counter
= 0; counter
< nr_counters
; counter
++)
417 create_counter(counter
, cpu
);
422 static int process_buildids(void)
424 u64 size
= lseek(output
, 0, SEEK_CUR
);
429 session
->fd
= output
;
430 return __perf_session__process_events(session
, post_processing_offset
,
431 size
- post_processing_offset
,
432 size
, &build_id__mark_dso_hit_ops
);
435 static void atexit_header(void)
438 session
->header
.data_size
+= bytes_written
;
441 perf_header__write(&session
->header
, output
, true);
442 perf_session__delete(session
);
447 static void event__synthesize_guest_os(struct machine
*machine
, void *data
)
450 struct perf_session
*psession
= data
;
452 if (machine__is_host(machine
))
456 *As for guest kernel when processing subcommand record&report,
457 *we arrange module mmap prior to guest kernel mmap and trigger
458 *a preload dso because default guest module symbols are loaded
459 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
460 *method is used to avoid symbol missing when the first addr is
461 *in module instead of in guest kernel.
463 err
= event__synthesize_modules(process_synthesized_event
,
466 pr_err("Couldn't record guest kernel [%d]'s reference"
467 " relocation symbol.\n", machine
->pid
);
470 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
471 * have no _text sometimes.
473 err
= event__synthesize_kernel_mmap(process_synthesized_event
,
474 psession
, machine
, "_text");
476 err
= event__synthesize_kernel_mmap(process_synthesized_event
,
477 psession
, machine
, "_stext");
479 pr_err("Couldn't record guest kernel [%d]'s reference"
480 " relocation symbol.\n", machine
->pid
);
483 static struct perf_event_header finished_round_event
= {
484 .size
= sizeof(struct perf_event_header
),
485 .type
= PERF_RECORD_FINISHED_ROUND
,
488 static void mmap_read_all(void)
492 for (i
= 0; i
< nr_cpu
; i
++) {
493 if (mmap_array
[i
].base
)
494 mmap_read(&mmap_array
[i
]);
497 if (perf_header__has_feat(&session
->header
, HEADER_TRACE_INFO
))
498 write_output(&finished_round_event
, sizeof(finished_round_event
));
501 static int __cmd_record(int argc
, const char **argv
)
507 unsigned long waking
= 0;
508 int child_ready_pipe
[2], go_pipe
[2];
509 const bool forks
= argc
> 0;
511 struct machine
*machine
;
513 page_size
= sysconf(_SC_PAGE_SIZE
);
516 signal(SIGCHLD
, sig_handler
);
517 signal(SIGINT
, sig_handler
);
519 if (forks
&& (pipe(child_ready_pipe
) < 0 || pipe(go_pipe
) < 0)) {
520 perror("failed to create pipes");
524 if (!strcmp(output_name
, "-"))
526 else if (!stat(output_name
, &st
) && st
.st_size
) {
527 if (write_mode
== WRITE_FORCE
) {
528 char oldname
[PATH_MAX
];
529 snprintf(oldname
, sizeof(oldname
), "%s.old",
532 rename(output_name
, oldname
);
534 } else if (write_mode
== WRITE_APPEND
) {
535 write_mode
= WRITE_FORCE
;
538 flags
= O_CREAT
|O_RDWR
;
539 if (write_mode
== WRITE_APPEND
)
545 output
= STDOUT_FILENO
;
547 output
= open(output_name
, flags
, S_IRUSR
| S_IWUSR
);
549 perror("failed to create output file");
553 session
= perf_session__new(output_name
, O_WRONLY
,
554 write_mode
== WRITE_FORCE
, false);
555 if (session
== NULL
) {
556 pr_err("Not enough memory for reading perf file header\n");
561 err
= perf_header__read(session
, output
);
563 goto out_delete_session
;
566 if (have_tracepoints(attrs
, nr_counters
))
567 perf_header__set_feat(&session
->header
, HEADER_TRACE_INFO
);
570 * perf_session__delete(session) will be called at atexit_header()
572 atexit(atexit_header
);
577 perror("failed to fork");
584 close(child_ready_pipe
[0]);
586 fcntl(go_pipe
[0], F_SETFD
, FD_CLOEXEC
);
589 * Do a dummy execvp to get the PLT entry resolved,
590 * so we avoid the resolver overhead on the real
593 execvp("", (char **)argv
);
596 * Tell the parent we're ready to go
598 close(child_ready_pipe
[1]);
601 * Wait until the parent tells us to go.
603 if (read(go_pipe
[0], &buf
, 1) == -1)
604 perror("unable to read pipe");
606 execvp(argv
[0], (char **)argv
);
612 if (!system_wide
&& target_tid
== -1 && target_pid
== -1)
613 all_tids
[0] = child_pid
;
615 close(child_ready_pipe
[1]);
618 * wait for child to settle
620 if (read(child_ready_pipe
[0], &buf
, 1) == -1) {
621 perror("unable to read pipe");
624 close(child_ready_pipe
[0]);
627 nr_cpus
= read_cpu_map(cpu_list
);
629 perror("failed to collect number of CPUs");
633 if (!system_wide
&& no_inherit
&& !cpu_list
) {
636 for (i
= 0; i
< nr_cpus
; i
++)
637 open_counters(cpumap
[i
]);
641 err
= perf_header__write_pipe(output
);
644 } else if (file_new
) {
645 err
= perf_header__write(&session
->header
, output
, false);
650 post_processing_offset
= lseek(output
, 0, SEEK_CUR
);
653 err
= event__synthesize_attrs(&session
->header
,
654 process_synthesized_event
,
657 pr_err("Couldn't synthesize attrs.\n");
661 err
= event__synthesize_event_types(process_synthesized_event
,
664 pr_err("Couldn't synthesize event_types.\n");
668 if (have_tracepoints(attrs
, nr_counters
)) {
670 * FIXME err <= 0 here actually means that
671 * there were no tracepoints so its not really
672 * an error, just that we don't need to
673 * synthesize anything. We really have to
674 * return this more properly and also
675 * propagate errors that now are calling die()
677 err
= event__synthesize_tracing_data(output
, attrs
,
679 process_synthesized_event
,
682 pr_err("Couldn't record tracing data.\n");
689 machine
= perf_session__find_host_machine(session
);
691 pr_err("Couldn't find native kernel information.\n");
695 err
= event__synthesize_kernel_mmap(process_synthesized_event
,
696 session
, machine
, "_text");
698 err
= event__synthesize_kernel_mmap(process_synthesized_event
,
699 session
, machine
, "_stext");
701 pr_err("Couldn't record kernel reference relocation symbol.\n");
705 err
= event__synthesize_modules(process_synthesized_event
,
708 pr_err("Couldn't record kernel reference relocation symbol.\n");
712 perf_session__process_machines(session
, event__synthesize_guest_os
);
715 event__synthesize_thread(target_tid
, process_synthesized_event
,
718 event__synthesize_threads(process_synthesized_event
, session
);
721 struct sched_param param
;
723 param
.sched_priority
= realtime_prio
;
724 if (sched_setscheduler(0, SCHED_FIFO
, ¶m
)) {
725 pr_err("Could not set realtime priority.\n");
742 if (hits
== samples
) {
745 err
= poll(event_array
, nr_poll
, -1);
750 for (i
= 0; i
< nr_cpu
; i
++) {
752 counter
< nr_counters
;
757 ioctl(fd
[i
][counter
][thread
],
758 PERF_EVENT_IOC_DISABLE
);
767 fprintf(stderr
, "[ perf record: Woken up %ld times to write data ]\n", waking
);
770 * Approximate RIP event size: 24 bytes.
773 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
774 (double)bytes_written
/ 1024.0 / 1024.0,
781 perf_session__delete(session
);
785 static const char * const record_usage
[] = {
786 "perf record [<options>] [<command>]",
787 "perf record [<options>] -- <command> [<options>]",
791 static bool force
, append_file
;
793 const struct option record_options
[] = {
794 OPT_CALLBACK('e', "event", NULL
, "event",
795 "event selector. use 'perf list' to list available events",
797 OPT_CALLBACK(0, "filter", NULL
, "filter",
798 "event filter", parse_filter
),
799 OPT_INTEGER('p', "pid", &target_pid
,
800 "record events on existing process id"),
801 OPT_INTEGER('t', "tid", &target_tid
,
802 "record events on existing thread id"),
803 OPT_INTEGER('r', "realtime", &realtime_prio
,
804 "collect data with this RT SCHED_FIFO priority"),
805 OPT_BOOLEAN('R', "raw-samples", &raw_samples
,
806 "collect raw sample records from all opened counters"),
807 OPT_BOOLEAN('a', "all-cpus", &system_wide
,
808 "system-wide collection from all CPUs"),
809 OPT_BOOLEAN('A', "append", &append_file
,
810 "append to the output file to do incremental profiling"),
811 OPT_STRING('C', "cpu", &cpu_list
, "cpu",
812 "list of cpus to monitor"),
813 OPT_BOOLEAN('f', "force", &force
,
814 "overwrite existing data file (deprecated)"),
815 OPT_U64('c', "count", &user_interval
, "event period to sample"),
816 OPT_STRING('o', "output", &output_name
, "file",
818 OPT_BOOLEAN('i', "no-inherit", &no_inherit
,
819 "child tasks do not inherit counters"),
820 OPT_UINTEGER('F', "freq", &user_freq
, "profile at this frequency"),
821 OPT_UINTEGER('m', "mmap-pages", &mmap_pages
, "number of mmap data pages"),
822 OPT_BOOLEAN('g', "call-graph", &call_graph
,
823 "do call-graph (stack chain/backtrace) recording"),
824 OPT_INCR('v', "verbose", &verbose
,
825 "be more verbose (show counter open errors, etc)"),
826 OPT_BOOLEAN('q', "quiet", &quiet
, "don't print any message"),
827 OPT_BOOLEAN('s', "stat", &inherit_stat
,
828 "per thread counts"),
829 OPT_BOOLEAN('d', "data", &sample_address
,
831 OPT_BOOLEAN('n', "no-samples", &no_samples
,
833 OPT_BOOLEAN('N', "no-buildid-cache", &no_buildid
,
834 "do not update the buildid cache"),
838 int cmd_record(int argc
, const char **argv
, const char *prefix __used
)
840 int i
, j
, err
= -ENOMEM
;
842 argc
= parse_options(argc
, argv
, record_options
, record_usage
,
843 PARSE_OPT_STOP_AT_NON_OPTION
);
844 if (!argc
&& target_pid
== -1 && target_tid
== -1 &&
845 !system_wide
&& !cpu_list
)
846 usage_with_options(record_usage
, record_options
);
848 if (force
&& append_file
) {
849 fprintf(stderr
, "Can't overwrite and append at the same time."
850 " You need to choose between -f and -A");
851 usage_with_options(record_usage
, record_options
);
852 } else if (append_file
) {
853 write_mode
= WRITE_APPEND
;
855 write_mode
= WRITE_FORCE
;
860 disable_buildid_cache();
864 attrs
[0].type
= PERF_TYPE_HARDWARE
;
865 attrs
[0].config
= PERF_COUNT_HW_CPU_CYCLES
;
868 if (target_pid
!= -1) {
869 target_tid
= target_pid
;
870 thread_num
= find_all_tid(target_pid
, &all_tids
);
871 if (thread_num
<= 0) {
872 fprintf(stderr
, "Can't find all threads of pid %d\n",
874 usage_with_options(record_usage
, record_options
);
877 all_tids
=malloc(sizeof(pid_t
));
879 goto out_symbol_exit
;
881 all_tids
[0] = target_tid
;
885 for (i
= 0; i
< MAX_NR_CPUS
; i
++) {
886 for (j
= 0; j
< MAX_COUNTERS
; j
++) {
887 fd
[i
][j
] = malloc(sizeof(int)*thread_num
);
892 event_array
= malloc(
893 sizeof(struct pollfd
)*MAX_NR_CPUS
*MAX_COUNTERS
*thread_num
);
897 if (user_interval
!= ULLONG_MAX
)
898 default_interval
= user_interval
;
899 if (user_freq
!= UINT_MAX
)
903 * User specified count overrides default frequency.
905 if (default_interval
)
908 default_interval
= freq
;
910 fprintf(stderr
, "frequency and count are zero, aborting\n");
912 goto out_free_event_array
;
915 err
= __cmd_record(argc
, argv
);
917 out_free_event_array
:
920 for (i
= 0; i
< MAX_NR_CPUS
; i
++) {
921 for (j
= 0; j
< MAX_COUNTERS
; j
++)