1PERF-RECORD(1) perf Manual PERF-RECORD(1)
2
6 perf-record - Run a command and record its profile into perf.data
7
9 perf record [-e <EVENT> | --event=EVENT] [-a] <command>
10 perf record [-e <EVENT> | --event=EVENT] [-a] — <command> [<options>]
11
13 This command runs a command and gathers a performance counter profile
14 from it, into perf.data - without displaying anything.
15 This file can then be inspected later on, using perf report.
17
19 <command>...
20 Any command you can specify in a shell.
21 -e, --event=
23 Select the PMU event. Selection can be:
24 · a symbolic event name (use perf list to list all events)
26 · a raw PMU event (eventsel+umask) in the form of rNNN where NNN
28 is a hexadecimal event descriptor.
29 · a symbolic or raw PMU event followed by an optional colon and a
31 list of event modifiers, e.g., cpu-cycles:p. See the perf-
32 list(1) man page for details on event modifiers.
33 · a symbolically formed PMU event like pmu/param1=0x3,param2/
35 where param1, param2, etc are defined as formats for the PMU in
36 /sys/bus/event_source/devices/<pmu>/format/*.
37 · a symbolically formed event like
39 pmu/config=M,config1=N,config3=K/
40 where M, N, K are numbers (in decimal, hex, octal format). Acceptable
42 values for each of 'config', 'config1' and 'config2' are defined by
43 corresponding entries in /sys/bus/event_source/devices/<pmu>/format/*
44 param1 and param2 are defined as formats for the PMU in:
45 /sys/bus/event_source/devices/<pmu>/format/*
46 There are also some parameters which are not defined in .../<pmu>/format/*.
48 These params can be used to overload default config values per event.
49 Here are some common parameters:
50 - 'period': Set event sampling period
51 - 'freq': Set event sampling frequency
52 - 'time': Disable/enable time stamping. Acceptable values are 1 for
53 enabling time stamping. 0 for disabling time stamping.
54 The default is 1.
55 - 'call-graph': Disable/enable callgraph. Acceptable str are "fp" for
56 FP mode, "dwarf" for DWARF mode, "lbr" for LBR mode and
57 "no" for disable callgraph.
58 - 'stack-size': user stack size for dwarf mode
59 - 'name' : User defined event name. Single quotes (') may be used to
60 escape symbols in the name from parsing by shell and tool
61 like this: name=\'CPU_CLK_UNHALTED.THREAD:cmask=0x1\'.
62 - 'aux-output': Generate AUX records instead of events. This requires
63 that an AUX area event is also provided.
64 - 'aux-sample-size': Set sample size for AUX area sampling. If the
65 '--aux-sample' option has been used, set aux-sample-size=0 to disable
66 AUX area sampling for the event.
67 See the linkperf:perf-list[1] man page for more parameters.
69 Note: If user explicitly sets options which conflict with the params,
71 the value set by the parameters will be overridden.
72 Also not defined in .../<pmu>/format/* are PMU driver specific
74 configuration parameters. Any configuration parameter preceded by
75 the letter '@' is not interpreted in user space and sent down directly
76 to the PMU driver. For example:
77 perf record -e some_event/@cfg1,@cfg2=config/ ...
79 will see 'cfg1' and 'cfg2=config' pushed to the PMU driver associated
81 with the event for further processing. There is no restriction on
82 what the configuration parameters are, as long as their semantic is
83 understood and supported by the PMU driver.
84 · a hardware breakpoint event in the form of
86 \mem:addr[/len][:access] where addr is the address in memory
87 you want to break in. Access is the memory access type (read,
88 write, execute) it can be passed as follows:
89 \mem:addr[:[r][w][x]]. len is the range, number of bytes from
90 specified addr, which the breakpoint will cover. If you want to
91 profile read-write accesses in 0x1000, just set mem:0x1000:rw.
92 If you want to profile write accesses in [0x1000~1008), just
93 set mem:0x1000/8:w.
94 · a BPF source file (ending in .c) or a precompiled object file
96 (ending in .o) selects one or more BPF events. The BPF program
97 can attach to various perf events based on the ELF section
98 names.
99 When processing a '.c' file, perf searches an installed LLVM to compile it
101 into an object file first. Optional clang options can be passed via the
102 '--clang-opt' command line option, e.g.:
103 perf record --clang-opt "-DLINUX_VERSION_CODE=0x50000" \
105 -e tests/bpf-script-example.c
106 Note: '--clang-opt' must be placed before '--event/-e'.
108 · a group of events surrounded by a pair of brace
110 ("{event1,event2,...}"). Each event is separated by commas and
111 the group should be quoted to prevent the shell interpretation.
112 You also need to use --group on "perf report" to view group
113 events together.
114 --filter=<filter>
116 Event filter. This option should follow an event selector (-e)
117 which selects either tracepoint event(s) or a hardware trace PMU
118 (e.g. Intel PT or CoreSight).
119 · tracepoint filters
121 In the case of tracepoints, multiple '--filter' options are combined
123 using '&&'.
124 · address filters
126 A hardware trace PMU advertises its ability to accept a number of
128 address filters by specifying a non-zero value in
129 /sys/bus/event_source/devices/<pmu>/nr_addr_filters.
130 Address filters have the format:
132 filter|start|stop|tracestop <start> [/ <size>] [@<file name>]
134 Where:
136 - 'filter': defines a region that will be traced.
137 - 'start': defines an address at which tracing will begin.
138 - 'stop': defines an address at which tracing will stop.
139 - 'tracestop': defines a region in which tracing will stop.
140 <file name> is the name of the object file, <start> is the offset to the
142 code to trace in that file, and <size> is the size of the region to
143 trace. 'start' and 'stop' filters need not specify a <size>.
144 If no object file is specified then the kernel is assumed, in which case
146 the start address must be a current kernel memory address.
147 <start> can also be specified by providing the name of a symbol. If the
149 symbol name is not unique, it can be disambiguated by inserting #n where
150 'n' selects the n'th symbol in address order. Alternately #0, #g or #G
151 select only a global symbol. <size> can also be specified by providing
152 the name of a symbol, in which case the size is calculated to the end
153 of that symbol. For 'filter' and 'tracestop' filters, if <size> is
154 omitted and <start> is a symbol, then the size is calculated to the end
155 of that symbol.
156 If <size> is omitted and <start> is '*', then the start and size will
158 be calculated from the first and last symbols, i.e. to trace the whole
159 file.
160 If symbol names (or '*') are provided, they must be surrounded by white
162 space.
163 The filter passed to the kernel is not necessarily the same as entered.
165 To see the filter that is passed, use the -v option.
166 The kernel may not be able to configure a trace region if it is not
168 within a single mapping. MMAP events (or /proc/<pid>/maps) can be
169 examined to determine if that is a possibility.
170 Multiple filters can be separated with space or comma.
172 --exclude-perf
174 Don’t record events issued by perf itself. This option should
175 follow an event selector (-e) which selects tracepoint event(s). It
176 adds a filter expression common_pid != $PERFPID to filters. If
177 other --filter exists, the new filter expression will be combined
178 with them by &&.
179 -a, --all-cpus
181 System-wide collection from all CPUs (default if no target is
182 specified).
183 -p, --pid=
185 Record events on existing process ID (comma separated list).
186 -t, --tid=
188 Record events on existing thread ID (comma separated list). This
189 option also disables inheritance by default. Enable it by adding
190 --inherit.
191 -u, --uid=
193 Record events in threads owned by uid. Name or number.
194 -r, --realtime=
196 Collect data with this RT SCHED_FIFO priority.
197 --no-buffering
199 Collect data without buffering.
200 -c, --count=
202 Event period to sample.
203 -o, --output=
205 Output file name.
206 -i, --no-inherit
208 Child tasks do not inherit counters.
209 -F, --freq=
211 Profile at this frequency. Use max to use the currently maximum
212 allowed frequency, i.e. the value in the
213 kernel.perf_event_max_sample_rate sysctl. Will throttle down to the
214 currently maximum allowed frequency. See --strict-freq.
215 --strict-freq
217 Fail if the specified frequency can’t be used.
218 -m, --mmap-pages=
220 Number of mmap data pages (must be a power of two) or size
221 specification with appended unit character - B/K/M/G. The size is
222 rounded up to have nearest pages power of two value. Also, by
223 adding a comma, the number of mmap pages for AUX area tracing can
224 be specified.
225 --group
227 Put all events in a single event group. This precedes the --event
228 option and remains only for backward compatibility. See --event.
229 -g
231 Enables call-graph (stack chain/backtrace) recording for both
232 kernel space and user space.
233 --call-graph
235 Setup and enable call-graph (stack chain/backtrace) recording,
236 implies -g. Default is "fp" (for user space).
237 The unwinding method used for kernel space is dependent on the
239 unwinder used by the active kernel configuration, i.e
240 CONFIG_UNWINDER_FRAME_POINTER (fp) or CONFIG_UNWINDER_ORC (orc)
241 Any option specified here controls the method used for user space.
243 Valid options are "fp" (frame pointer), "dwarf" (DWARF's CFI -
245 Call Frame Information) or "lbr" (Hardware Last Branch Record
246 facility).
247 In some systems, where binaries are build with gcc
249 --fomit-frame-pointer, using the "fp" method will produce bogus
250 call graphs, using "dwarf", if available (perf tools linked to
251 the libunwind or libdw library) should be used instead.
252 Using the "lbr" method doesn't require any compiler options. It
253 will produce call graphs from the hardware LBR registers. The
254 main limitation is that it is only available on new Intel
255 platforms, such as Haswell. It can only get user call chain. It
256 doesn't work with branch stack sampling at the same time.
257 When "dwarf" recording is used, perf also records (user) stack dump
259 when sampled. Default size of the stack dump is 8192 (bytes).
260 User can change the size by passing the size after comma like
261 "--call-graph dwarf,4096".
262 -q, --quiet
264 Don’t print any message, useful for scripting.
265 -v, --verbose
267 Be more verbose (show counter open errors, etc).
268 -s, --stat
270 Record per-thread event counts. Use it with perf report -T to see
271 the values.
272 -d, --data
274 Record the sample virtual addresses.
275 --phys-data
277 Record the sample physical addresses.
278 --data-page-size
280 Record the sampled data address data page size.
281 -T, --timestamp
283 Record the sample timestamps. Use it with perf report -D to see the
284 timestamps, for instance.
285 -P, --period
287 Record the sample period.
288 --sample-cpu
290 Record the sample cpu.
291 -n, --no-samples
293 Don’t sample.
294 -R, --raw-samples
296 Collect raw sample records from all opened counters (default for
297 tracepoint counters).
298 -C, --cpu
300 Collect samples only on the list of CPUs provided. Multiple CPUs
301 can be provided as a comma-separated list with no space: 0,1.
302 Ranges of CPUs are specified with -: 0-2. In per-thread mode with
303 inheritance mode on (default), samples are captured only when the
304 thread executes on the designated CPUs. Default is to monitor all
305 CPUs.
306 -B, --no-buildid
308 Do not save the build ids of binaries in the perf.data files. This
309 skips post processing after recording, which sometimes makes the
310 final step in the recording process to take a long time, as it
311 needs to process all events looking for mmap records. The downside
312 is that it can misresolve symbols if the workload binaries used
313 when recording get locally rebuilt or upgraded, because the only
314 key available in this case is the pathname. You can also set the
315 "record.build-id" config variable to 'skip to have this behaviour
316 permanently.
317 -N, --no-buildid-cache
319 Do not update the buildid cache. This saves some overhead in
320 situations where the information in the perf.data file (which
321 includes buildids) is sufficient. You can also set the
322 "record.build-id" config variable to no-cache to have the same
323 effect.
324 -G name,..., --cgroup name,...
326 monitor only in the container (cgroup) called "name". This option
327 is available only in per-cpu mode. The cgroup filesystem must be
328 mounted. All threads belonging to container "name" are monitored
329 when they run on the monitored CPUs. Multiple cgroups can be
330 provided. Each cgroup is applied to the corresponding event, i.e.,
331 first cgroup to first event, second cgroup to second event and so
332 on. It is possible to provide an empty cgroup (monitor all the
333 time) using, e.g., -G foo,,bar. Cgroups must have corresponding
334 events, i.e., they always refer to events defined earlier on the
335 command line. If the user wants to track multiple events for a
336 specific cgroup, the user can use -e e1 -e e2 -G foo,foo or just
337 use -e e1 -e e2 -G foo.
338 If wanting to monitor, say, cycles for a cgroup and also for system
340 wide, this command line can be used: perf stat -e cycles -G cgroup_name
341 -a -e cycles.
342 -b, --branch-any
344 Enable taken branch stack sampling. Any type of taken branch may be
345 sampled. This is a shortcut for --branch-filter any. See
346 --branch-filter for more infos.
347 -j, --branch-filter
349 Enable taken branch stack sampling. Each sample captures a series
350 of consecutive taken branches. The number of branches captured with
351 each sample depends on the underlying hardware, the type of
352 branches of interest, and the executed code. It is possible to
353 select the types of branches captured by enabling filters. The
354 following filters are defined:
355 · any: any type of branches
357 · any_call: any function call or system call
359 · any_ret: any function return or system call return
361 · ind_call: any indirect branch
363 · call: direct calls, including far (to/from kernel) calls
365 · u: only when the branch target is at the user level
367 · k: only when the branch target is in the kernel
369 · hv: only when the target is at the hypervisor level
371 · in_tx: only when the target is in a hardware transaction
373 · no_tx: only when the target is not in a hardware transaction
375 · abort_tx: only when the target is a hardware transaction abort
377 · cond: conditional branches
379 · save_type: save branch type during sampling in case binary is
381 not available later
382 The option requires at least one branch type among any, any_call,
384 any_ret, ind_call, cond. The privilege levels may be omitted, in
385 which case, the privilege levels of the associated event are
386 applied to the branch filter. Both kernel (k) and hypervisor (hv)
387 privilege levels are subject to permissions. When sampling on
388 multiple events, branch stack sampling is enabled for all the
389 sampling events. The sampled branch type is the same for all
390 events. The various filters must be specified as a comma separated
391 list: --branch-filter any_ret,u,k Note that this feature may not be
392 available on all processors.
393 --weight
395 Enable weightened sampling. An additional weight is recorded per
396 sample and can be displayed with the weight and local_weight sort
397 keys. This currently works for TSX abort events and some memory
398 events in precise mode on modern Intel CPUs.
399 --namespaces
401 Record events of type PERF_RECORD_NAMESPACES. This enables
402 cgroup_id sort key.
403 --all-cgroups
405 Record events of type PERF_RECORD_CGROUP. This enables cgroup sort
406 key.
407 --transaction
409 Record transaction flags for transaction related events.
410 --per-thread
412 Use per-thread mmaps. By default per-cpu mmaps are created. This
413 option overrides that and uses per-thread mmaps. A side-effect of
414 that is that inheritance is automatically disabled. --per-thread is
415 ignored with a warning if combined with -a or -C options.
416 -D, --delay=
418 After starting the program, wait msecs before measuring (-1: start
419 with events disabled). This is useful to filter out the startup
420 phase of the program, which is often very different.
421 -I, --intr-regs
423 Capture machine state (registers) at interrupt, i.e., on counter
424 overflows for each sample. List of captured registers depends on
425 the architecture. This option is off by default. It is possible to
426 select the registers to sample using their symbolic names, e.g. on
427 x86, ax, si. To list the available registers use --intr-regs=\?. To
428 name registers, pass a comma separated list such as
429 --intr-regs=ax,bx. The list of register is architecture dependent.
430 --user-regs
432 Similar to -I, but capture user registers at sample time. To list
433 the available user registers use --user-regs=\?.
434 --running-time
436 Record running and enabled time for read events (:S)
437 -k, --clockid
439 Sets the clock id to use for the various time fields in the
440 perf_event_type records. See clock_gettime(). In particular
441 CLOCK_MONOTONIC and CLOCK_MONOTONIC_RAW are supported, some events
442 might also allow CLOCK_BOOTTIME, CLOCK_REALTIME and CLOCK_TAI.
443 -S, --snapshot
445 Select AUX area tracing Snapshot Mode. This option is valid only
446 with an AUX area tracing event. Optionally, certain snapshot
447 capturing parameters can be specified in a string that follows this
448 option: e: take one last snapshot on exit; guarantees that there is
449 at least one snapshot in the output file; <size>: if the PMU
450 supports this, specify the desired snapshot size.
451 In Snapshot Mode trace data is captured only when signal SIGUSR2 is
453 received and on exit if the above e option is given.
454 --aux-sample[=OPTIONS]
456 Select AUX area sampling. At least one of the events selected by
457 the -e option must be an AUX area event. Samples on other events
458 will be created containing data from the AUX area. Optionally
459 sample size may be specified, otherwise it defaults to 4KiB.
460 --proc-map-timeout
462 When processing pre-existing threads /proc/XXX/mmap, it may take a
463 long time, because the file may be huge. A time out is needed in
464 such cases. This option sets the time out limit. The default value
465 is 500 ms.
466 --switch-events
468 Record context switch events i.e. events of type PERF_RECORD_SWITCH
469 or PERF_RECORD_SWITCH_CPU_WIDE. In some cases (e.g. Intel PT or
470 CoreSight) switch events will be enabled automatically, which can
471 be suppressed by by the option --no-switch-events.
472 --clang-path=PATH
474 Path to clang binary to use for compiling BPF scriptlets. (enabled
475 when BPF support is on)
476 --clang-opt=OPTIONS
478 Options passed to clang when compiling BPF scriptlets. (enabled
479 when BPF support is on)
480 --vmlinux=PATH
482 Specify vmlinux path which has debuginfo. (enabled when BPF
483 prologue is on)
484 --buildid-all
486 Record build-id of all DSOs regardless whether it’s actually hit or
487 not.
488 --aio[=n]
490 Use <n> control blocks in asynchronous (Posix AIO) trace writing
491 mode (default: 1, max: 4). Asynchronous mode is supported only when
492 linking Perf tool with libc library providing implementation for
493 Posix AIO API.
494 --affinity=mode
496 Set affinity mask of trace reading thread according to the policy
497 defined by mode value: node - thread affinity mask is set to NUMA
498 node cpu mask of the processed mmap buffer cpu - thread affinity
499 mask is set to cpu of the processed mmap buffer
500 --mmap-flush=number
502 Specify minimal number of bytes that is extracted from mmap data
503 pages and processed for output. One can specify the number using
504 B/K/M/G suffixes.
505 The maximal allowed value is a quarter of the size of mmaped data
507 pages.
508 The default option value is 1 byte which means that every time that the
510 output writing thread finds some new data in the mmaped buffer the data
511 is extracted, possibly compressed (-z) and written to the output,
512 perf.data or pipe.
513 Larger data chunks are compressed more effectively in comparison to
515 smaller chunks so extraction of larger chunks from the mmap data pages
516 is preferable from the perspective of output size reduction.
517 Also at some cases executing less output write syscalls with bigger
519 data size can take less time than executing more output write syscalls
520 with smaller data size thus lowering runtime profiling overhead.
521 -z, --compression-level[=n]
523 Produce compressed trace using specified level n (default: 1 -
524 fastest compression, 22 - smallest trace)
525 --all-kernel
527 Configure all used events to run in kernel space.
528 --all-user
530 Configure all used events to run in user space.
531 --kernel-callchains
533 Collect callchains only from kernel space. I.e. this option sets
534 perf_event_attr.exclude_callchain_user to 1.
535 --user-callchains
537 Collect callchains only from user space. I.e. this option sets
538 perf_event_attr.exclude_callchain_kernel to 1.
539 Don’t use both --kernel-callchains and --user-callchains at the same
541 time or no callchains will be collected.
542 --timestamp-filename Append timestamp to output file name.
544 --timestamp-boundary
546 Record timestamp boundary (time of first/last samples).
547 --switch-output[=mode]
549 Generate multiple perf.data files, timestamp prefixed, switching to
550 a new one based on mode value: "signal" - when receiving a SIGUSR2
551 (default value) or <size> - when reaching the size threshold, size
552 is expected to be a number with appended unit character - B/K/M/G
553 <time> - when reaching the time threshold, size is expected to be a
554 number with appended unit character - s/m/h/d
555 Note: the precision of the size threshold hugely depends
557 on your configuration - the number and size of your ring
558 buffers (-m). It is generally more precise for higher sizes
559 (like >5M), for lower values expect different sizes.
560 A possible use case is to, given an external event, slice the perf.data
562 file that gets then processed, possibly via a perf script, to decide if
563 that particular perf.data snapshot should be kept or not.
564 Implies --timestamp-filename, --no-buildid and --no-buildid-cache. The
566 reason for the latter two is to reduce the data file switching
567 overhead. You can still switch them on with:
568 --switch-output --no-no-buildid --no-no-buildid-cache
570 --switch-output-event
572 Events that will cause the switch of the perf.data file,
573 auto-selecting --switch-output=signal, the results are similar as
574 internally the side band thread will also send a SIGUSR2 to the
575 main one.
576 Uses the same syntax as --event, it will just not be recorded, serving
578 only to switch the perf.data file as soon as the --switch-output event
579 is processed by a separate sideband thread.
580 This sideband thread is also used to other purposes, like processing
582 the PERF_RECORD_BPF_EVENT records as they happen, asking the kernel for
583 extra BPF information, etc.
584 --switch-max-files=N
586 When rotating perf.data with --switch-output, only keep N files.
587 --dry-run
589 Parse options then exit. --dry-run can be used to detect errors in
590 cmdline options.
591 perf record --dry-run -e can act as a BPF script compiler if
593 llvm.dump-obj in config file is set to true.
594 --tail-synthesize
596 Instead of collecting non-sample events (for example, fork, comm,
597 mmap) at the beginning of record, collect them during finalizing an
598 output file. The collected non-sample events reflects the status of
599 the system when record is finished.
600 --overwrite
602 Makes all events use an overwritable ring buffer. An overwritable
603 ring buffer works like a flight recorder: when it gets full, the
604 kernel will overwrite the oldest records, that thus will never make
605 it to the perf.data file.
606 When --overwrite and --switch-output are used perf records and drops
608 events until it receives a signal, meaning that something unusual was
609 detected that warrants taking a snapshot of the most current events,
610 those fitting in the ring buffer at that moment.
611 overwrite attribute can also be set or canceled for an event using
613 config terms. For example: cycles/overwrite/ and
614 instructions/no-overwrite/.
615 Implies --tail-synthesize.
617 --kcore
619 Make a copy of /proc/kcore and place it into a directory with the
620 perf data file.
621 --max-size=<size>
623 Limit the sample data max size, <size> is expected to be a number
624 with appended unit character - B/K/M/G
625 --num-thread-synthesize
627 The number of threads to run when synthesizing events for existing
628 processes. By default, the number of threads equals 1.
629 --control=fifo:ctl-fifo[,ack-fifo], --control=fd:ctl-fd[,ack-fd]
631 ctl-fifo / ack-fifo are opened and used as ctl-fd / ack-fd as
632 follows. Listen on ctl-fd descriptor for command to control
633 measurement.
634 Available commands: enable : enable events disable : disable events
636 snapshot: AUX area tracing snapshot).
637 Measurements can be started with events disabled using --delay=-1
639 option. Optionally send control command completion (ack\n) to ack-fd
640 descriptor to synchronize with the controlling process. Example of bash
641 shell script to enable and disable events during measurements:
642 #!/bin/bash
644 ctl_dir=/tmp/
646 ctl_fifo=${ctl_dir}perf_ctl.fifo
648 test -p ${ctl_fifo} && unlink ${ctl_fifo}
649 mkfifo ${ctl_fifo}
650 exec {ctl_fd}<>${ctl_fifo}
651 ctl_ack_fifo=${ctl_dir}perf_ctl_ack.fifo
653 test -p ${ctl_ack_fifo} && unlink ${ctl_ack_fifo}
654 mkfifo ${ctl_ack_fifo}
655 exec {ctl_fd_ack}<>${ctl_ack_fifo}
656 perf record -D -1 -e cpu-cycles -a \
658 --control fd:${ctl_fd},${ctl_fd_ack} \
659 -- sleep 30 &
660 perf_pid=$!
661 sleep 5 && echo 'enable' >&${ctl_fd} && read -u ${ctl_fd_ack} e1 && echo "enabled(${e1})"
663 sleep 10 && echo 'disable' >&${ctl_fd} && read -u ${ctl_fd_ack} d1 && echo "disabled(${d1})"
664 exec {ctl_fd_ack}>&-
666 unlink ${ctl_ack_fifo}
667 exec {ctl_fd}>&-
669 unlink ${ctl_fifo}
670 wait -n ${perf_pid}
672 exit $?
673
675 perf-stat(1), perf-list(1), perf-intel-pt(1)
676perf 03/30/2021 PERF-RECORD(1)