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_TIME
;
279 attr
->sample_type
|= PERF_SAMPLE_RAW
;
280 attr
->sample_type
|= PERF_SAMPLE_CPU
;
285 attr
->inherit
= !no_inherit
;
286 if (target_pid
== -1 && target_tid
== -1 && !system_wide
) {
288 attr
->enable_on_exec
= 1;
291 for (thread_index
= 0; thread_index
< thread_num
; thread_index
++) {
293 fd
[nr_cpu
][counter
][thread_index
] = sys_perf_event_open(attr
,
294 all_tids
[thread_index
], cpu
, group_fd
, 0);
296 if (fd
[nr_cpu
][counter
][thread_index
] < 0) {
299 if (err
== EPERM
|| err
== EACCES
)
300 die("Permission error - are you root?\n"
301 "\t Consider tweaking"
302 " /proc/sys/kernel/perf_event_paranoid.\n");
303 else if (err
== ENODEV
&& cpu_list
) {
304 die("No such device - did you specify"
305 " an out-of-range profile CPU?\n");
309 * If it's cycles then fall back to hrtimer
310 * based cpu-clock-tick sw counter, which
311 * is always available even if no PMU support:
313 if (attr
->type
== PERF_TYPE_HARDWARE
314 && attr
->config
== PERF_COUNT_HW_CPU_CYCLES
) {
317 warning(" ... trying to fall back to cpu-clock-ticks\n");
318 attr
->type
= PERF_TYPE_SOFTWARE
;
319 attr
->config
= PERF_COUNT_SW_CPU_CLOCK
;
323 error("perfcounter syscall returned with %d (%s)\n",
324 fd
[nr_cpu
][counter
][thread_index
], strerror(err
));
326 #if defined(__i386__) || defined(__x86_64__)
327 if (attr
->type
== PERF_TYPE_HARDWARE
&& err
== EOPNOTSUPP
)
328 die("No hardware sampling interrupt available."
329 " No APIC? If so then you can boot the kernel"
330 " with the \"lapic\" boot parameter to"
331 " force-enable it.\n");
334 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
338 h_attr
= get_header_attr(attr
, counter
);
343 if (memcmp(&h_attr
->attr
, attr
, sizeof(*attr
))) {
344 fprintf(stderr
, "incompatible append\n");
349 if (read(fd
[nr_cpu
][counter
][thread_index
], &read_data
, sizeof(read_data
)) == -1) {
350 perror("Unable to read perf file descriptor\n");
354 if (perf_header_attr__add_id(h_attr
, read_data
.id
) < 0) {
355 pr_warning("Not enough memory to add id\n");
359 assert(fd
[nr_cpu
][counter
][thread_index
] >= 0);
360 fcntl(fd
[nr_cpu
][counter
][thread_index
], F_SETFL
, O_NONBLOCK
);
363 * First counter acts as the group leader:
365 if (group
&& group_fd
== -1)
366 group_fd
= fd
[nr_cpu
][counter
][thread_index
];
368 if (counter
|| thread_index
) {
369 ret
= ioctl(fd
[nr_cpu
][counter
][thread_index
],
370 PERF_EVENT_IOC_SET_OUTPUT
,
373 error("failed to set output: %d (%s)\n", errno
,
378 mmap_array
[nr_cpu
].counter
= counter
;
379 mmap_array
[nr_cpu
].prev
= 0;
380 mmap_array
[nr_cpu
].mask
= mmap_pages
*page_size
- 1;
381 mmap_array
[nr_cpu
].base
= mmap(NULL
, (mmap_pages
+1)*page_size
,
382 PROT_READ
|PROT_WRITE
, MAP_SHARED
, fd
[nr_cpu
][counter
][thread_index
], 0);
383 if (mmap_array
[nr_cpu
].base
== MAP_FAILED
) {
384 error("failed to mmap with %d (%s)\n", errno
, strerror(errno
));
388 event_array
[nr_poll
].fd
= fd
[nr_cpu
][counter
][thread_index
];
389 event_array
[nr_poll
].events
= POLLIN
;
393 if (filter
!= NULL
) {
394 ret
= ioctl(fd
[nr_cpu
][counter
][thread_index
],
395 PERF_EVENT_IOC_SET_FILTER
, filter
);
397 error("failed to set filter with %d (%s)\n", errno
,
405 static void open_counters(int cpu
)
410 for (counter
= 0; counter
< nr_counters
; counter
++)
411 create_counter(counter
, cpu
);
416 static int process_buildids(void)
418 u64 size
= lseek(output
, 0, SEEK_CUR
);
423 session
->fd
= output
;
424 return __perf_session__process_events(session
, post_processing_offset
,
425 size
- post_processing_offset
,
426 size
, &build_id__mark_dso_hit_ops
);
429 static void atexit_header(void)
432 session
->header
.data_size
+= bytes_written
;
435 perf_header__write(&session
->header
, output
, true);
439 static void event__synthesize_guest_os(struct machine
*machine
, void *data
)
442 char *guest_kallsyms
;
444 struct perf_session
*psession
= data
;
446 if (machine__is_host(machine
))
450 *As for guest kernel when processing subcommand record&report,
451 *we arrange module mmap prior to guest kernel mmap and trigger
452 *a preload dso because default guest module symbols are loaded
453 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
454 *method is used to avoid symbol missing when the first addr is
455 *in module instead of in guest kernel.
457 err
= event__synthesize_modules(process_synthesized_event
,
460 pr_err("Couldn't record guest kernel [%d]'s reference"
461 " relocation symbol.\n", machine
->pid
);
463 if (machine__is_default_guest(machine
))
464 guest_kallsyms
= (char *) symbol_conf
.default_guest_kallsyms
;
466 sprintf(path
, "%s/proc/kallsyms", machine
->root_dir
);
467 guest_kallsyms
= path
;
471 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
472 * have no _text sometimes.
474 err
= event__synthesize_kernel_mmap(process_synthesized_event
,
475 psession
, machine
, "_text");
477 err
= event__synthesize_kernel_mmap(process_synthesized_event
,
478 psession
, machine
, "_stext");
480 pr_err("Couldn't record guest kernel [%d]'s reference"
481 " relocation symbol.\n", machine
->pid
);
484 static struct perf_event_header finished_round_event
= {
485 .size
= sizeof(struct perf_event_header
),
486 .type
= PERF_RECORD_FINISHED_ROUND
,
489 static void mmap_read_all(void)
493 for (i
= 0; i
< nr_cpu
; i
++) {
494 if (mmap_array
[i
].base
)
495 mmap_read(&mmap_array
[i
]);
498 if (perf_header__has_feat(&session
->header
, HEADER_TRACE_INFO
))
499 write_output(&finished_round_event
, sizeof(finished_round_event
));
502 static int __cmd_record(int argc
, const char **argv
)
508 unsigned long waking
= 0;
509 int child_ready_pipe
[2], go_pipe
[2];
510 const bool forks
= argc
> 0;
512 struct machine
*machine
;
514 page_size
= sysconf(_SC_PAGE_SIZE
);
517 signal(SIGCHLD
, sig_handler
);
518 signal(SIGINT
, sig_handler
);
520 if (forks
&& (pipe(child_ready_pipe
) < 0 || pipe(go_pipe
) < 0)) {
521 perror("failed to create pipes");
525 if (!strcmp(output_name
, "-"))
527 else if (!stat(output_name
, &st
) && st
.st_size
) {
528 if (write_mode
== WRITE_FORCE
) {
529 char oldname
[PATH_MAX
];
530 snprintf(oldname
, sizeof(oldname
), "%s.old",
533 rename(output_name
, oldname
);
535 } else if (write_mode
== WRITE_APPEND
) {
536 write_mode
= WRITE_FORCE
;
539 flags
= O_CREAT
|O_RDWR
;
540 if (write_mode
== WRITE_APPEND
)
546 output
= STDOUT_FILENO
;
548 output
= open(output_name
, flags
, S_IRUSR
| S_IWUSR
);
550 perror("failed to create output file");
554 session
= perf_session__new(output_name
, O_WRONLY
,
555 write_mode
== WRITE_FORCE
, false);
556 if (session
== NULL
) {
557 pr_err("Not enough memory for reading perf file header\n");
562 err
= perf_header__read(session
, output
);
567 if (have_tracepoints(attrs
, nr_counters
))
568 perf_header__set_feat(&session
->header
, HEADER_TRACE_INFO
);
570 atexit(atexit_header
);
575 perror("failed to fork");
582 close(child_ready_pipe
[0]);
584 fcntl(go_pipe
[0], F_SETFD
, FD_CLOEXEC
);
587 * Do a dummy execvp to get the PLT entry resolved,
588 * so we avoid the resolver overhead on the real
591 execvp("", (char **)argv
);
594 * Tell the parent we're ready to go
596 close(child_ready_pipe
[1]);
599 * Wait until the parent tells us to go.
601 if (read(go_pipe
[0], &buf
, 1) == -1)
602 perror("unable to read pipe");
604 execvp(argv
[0], (char **)argv
);
610 if (!system_wide
&& target_tid
== -1 && target_pid
== -1)
611 all_tids
[0] = child_pid
;
613 close(child_ready_pipe
[1]);
616 * wait for child to settle
618 if (read(child_ready_pipe
[0], &buf
, 1) == -1) {
619 perror("unable to read pipe");
622 close(child_ready_pipe
[0]);
625 nr_cpus
= read_cpu_map(cpu_list
);
627 perror("failed to collect number of CPUs\n");
631 if (!system_wide
&& no_inherit
&& !cpu_list
) {
634 for (i
= 0; i
< nr_cpus
; i
++)
635 open_counters(cpumap
[i
]);
639 err
= perf_header__write_pipe(output
);
642 } else if (file_new
) {
643 err
= perf_header__write(&session
->header
, output
, false);
648 post_processing_offset
= lseek(output
, 0, SEEK_CUR
);
651 err
= event__synthesize_attrs(&session
->header
,
652 process_synthesized_event
,
655 pr_err("Couldn't synthesize attrs.\n");
659 err
= event__synthesize_event_types(process_synthesized_event
,
662 pr_err("Couldn't synthesize event_types.\n");
666 if (have_tracepoints(attrs
, nr_counters
)) {
668 * FIXME err <= 0 here actually means that
669 * there were no tracepoints so its not really
670 * an error, just that we don't need to
671 * synthesize anything. We really have to
672 * return this more properly and also
673 * propagate errors that now are calling die()
675 err
= event__synthesize_tracing_data(output
, attrs
,
677 process_synthesized_event
,
680 pr_err("Couldn't record tracing data.\n");
687 machine
= perf_session__find_host_machine(session
);
689 pr_err("Couldn't find native kernel information.\n");
693 err
= event__synthesize_kernel_mmap(process_synthesized_event
,
694 session
, machine
, "_text");
696 err
= event__synthesize_kernel_mmap(process_synthesized_event
,
697 session
, machine
, "_stext");
699 pr_err("Couldn't record kernel reference relocation symbol.\n");
703 err
= event__synthesize_modules(process_synthesized_event
,
706 pr_err("Couldn't record kernel reference relocation symbol.\n");
710 perf_session__process_machines(session
, event__synthesize_guest_os
);
712 if (!system_wide
&& cpu_list
)
713 event__synthesize_thread(target_tid
, process_synthesized_event
,
716 event__synthesize_threads(process_synthesized_event
, session
);
719 struct sched_param param
;
721 param
.sched_priority
= realtime_prio
;
722 if (sched_setscheduler(0, SCHED_FIFO
, ¶m
)) {
723 pr_err("Could not set realtime priority.\n");
740 if (hits
== samples
) {
743 err
= poll(event_array
, nr_poll
, -1);
748 for (i
= 0; i
< nr_cpu
; i
++) {
750 counter
< nr_counters
;
755 ioctl(fd
[i
][counter
][thread
],
756 PERF_EVENT_IOC_DISABLE
);
762 fprintf(stderr
, "[ perf record: Woken up %ld times to write data ]\n", waking
);
765 * Approximate RIP event size: 24 bytes.
768 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
769 (double)bytes_written
/ 1024.0 / 1024.0,
776 static const char * const record_usage
[] = {
777 "perf record [<options>] [<command>]",
778 "perf record [<options>] -- <command> [<options>]",
782 static bool force
, append_file
;
784 static const struct option options
[] = {
785 OPT_CALLBACK('e', "event", NULL
, "event",
786 "event selector. use 'perf list' to list available events",
788 OPT_CALLBACK(0, "filter", NULL
, "filter",
789 "event filter", parse_filter
),
790 OPT_INTEGER('p', "pid", &target_pid
,
791 "record events on existing process id"),
792 OPT_INTEGER('t', "tid", &target_tid
,
793 "record events on existing thread id"),
794 OPT_INTEGER('r', "realtime", &realtime_prio
,
795 "collect data with this RT SCHED_FIFO priority"),
796 OPT_BOOLEAN('R', "raw-samples", &raw_samples
,
797 "collect raw sample records from all opened counters"),
798 OPT_BOOLEAN('a', "all-cpus", &system_wide
,
799 "system-wide collection from all CPUs"),
800 OPT_BOOLEAN('A', "append", &append_file
,
801 "append to the output file to do incremental profiling"),
802 OPT_STRING('C', "cpu", &cpu_list
, "cpu",
803 "list of cpus to monitor"),
804 OPT_BOOLEAN('f', "force", &force
,
805 "overwrite existing data file (deprecated)"),
806 OPT_U64('c', "count", &user_interval
, "event period to sample"),
807 OPT_STRING('o', "output", &output_name
, "file",
809 OPT_BOOLEAN('i', "no-inherit", &no_inherit
,
810 "child tasks do not inherit counters"),
811 OPT_UINTEGER('F', "freq", &user_freq
, "profile at this frequency"),
812 OPT_UINTEGER('m', "mmap-pages", &mmap_pages
, "number of mmap data pages"),
813 OPT_BOOLEAN('g', "call-graph", &call_graph
,
814 "do call-graph (stack chain/backtrace) recording"),
815 OPT_INCR('v', "verbose", &verbose
,
816 "be more verbose (show counter open errors, etc)"),
817 OPT_BOOLEAN('s', "stat", &inherit_stat
,
818 "per thread counts"),
819 OPT_BOOLEAN('d', "data", &sample_address
,
821 OPT_BOOLEAN('n', "no-samples", &no_samples
,
826 int cmd_record(int argc
, const char **argv
, const char *prefix __used
)
830 argc
= parse_options(argc
, argv
, options
, record_usage
,
831 PARSE_OPT_STOP_AT_NON_OPTION
);
832 if (!argc
&& target_pid
== -1 && target_tid
== -1 &&
833 !system_wide
&& !cpu_list
)
834 usage_with_options(record_usage
, options
);
836 if (force
&& append_file
) {
837 fprintf(stderr
, "Can't overwrite and append at the same time."
838 " You need to choose between -f and -A");
839 usage_with_options(record_usage
, options
);
840 } else if (append_file
) {
841 write_mode
= WRITE_APPEND
;
843 write_mode
= WRITE_FORCE
;
850 attrs
[0].type
= PERF_TYPE_HARDWARE
;
851 attrs
[0].config
= PERF_COUNT_HW_CPU_CYCLES
;
854 if (target_pid
!= -1) {
855 target_tid
= target_pid
;
856 thread_num
= find_all_tid(target_pid
, &all_tids
);
857 if (thread_num
<= 0) {
858 fprintf(stderr
, "Can't find all threads of pid %d\n",
860 usage_with_options(record_usage
, options
);
863 all_tids
=malloc(sizeof(pid_t
));
867 all_tids
[0] = target_tid
;
871 for (i
= 0; i
< MAX_NR_CPUS
; i
++) {
872 for (j
= 0; j
< MAX_COUNTERS
; j
++) {
873 fd
[i
][j
] = malloc(sizeof(int)*thread_num
);
878 event_array
= malloc(
879 sizeof(struct pollfd
)*MAX_NR_CPUS
*MAX_COUNTERS
*thread_num
);
883 if (user_interval
!= ULLONG_MAX
)
884 default_interval
= user_interval
;
885 if (user_freq
!= UINT_MAX
)
889 * User specified count overrides default frequency.
891 if (default_interval
)
894 default_interval
= freq
;
896 fprintf(stderr
, "frequency and count are zero, aborting\n");
900 return __cmd_record(argc
, argv
);