6 perf-record - Run a command and record its profile into perf.data
11 'perf record' [-e <EVENT> | --event=EVENT] [-a] <command>
12 'perf record' [-e <EVENT> | --event=EVENT] [-a] \-- <command> [<options>]
16 This command runs a command and gathers a performance counter profile
17 from it, into perf.data - without displaying anything.
19 This file can then be inspected later on, using 'perf report'.
25 Any command you can specify in a shell.
29 Select the PMU event. Selection can be:
31 - a symbolic event name (use 'perf list' to list all events)
33 - a raw PMU event in the form of rN where N is a hexadecimal value
34 that represents the raw register encoding with the layout of the
35 event control registers as described by entries in
36 /sys/bus/event_sources/devices/cpu/format/*.
38 - a symbolic or raw PMU event followed by an optional colon
39 and a list of event modifiers, e.g., cpu-cycles:p. See the
40 linkperf:perf-list[1] man page for details on event modifiers.
42 - a symbolically formed PMU event like 'pmu/param1=0x3,param2/' where
43 'param1', 'param2', etc are defined as formats for the PMU in
44 /sys/bus/event_source/devices/<pmu>/format/*.
46 - a symbolically formed event like 'pmu/config=M,config1=N,config3=K/'
48 where M, N, K are numbers (in decimal, hex, octal format). Acceptable
49 values for each of 'config', 'config1' and 'config2' are defined by
50 corresponding entries in /sys/bus/event_source/devices/<pmu>/format/*
51 param1 and param2 are defined as formats for the PMU in:
52 /sys/bus/event_source/devices/<pmu>/format/*
54 There are also some parameters which are not defined in .../<pmu>/format/*.
55 These params can be used to overload default config values per event.
56 Here are some common parameters:
57 - 'period': Set event sampling period
58 - 'freq': Set event sampling frequency
59 - 'time': Disable/enable time stamping. Acceptable values are 1 for
60 enabling time stamping. 0 for disabling time stamping.
62 - 'call-graph': Disable/enable callgraph. Acceptable str are "fp" for
63 FP mode, "dwarf" for DWARF mode, "lbr" for LBR mode and
64 "no" for disable callgraph.
65 - 'stack-size': user stack size for dwarf mode
66 - 'name' : User defined event name. Single quotes (') may be used to
67 escape symbols in the name from parsing by shell and tool
68 like this: name=\'CPU_CLK_UNHALTED.THREAD:cmask=0x1\'.
69 - 'aux-output': Generate AUX records instead of events. This requires
70 that an AUX area event is also provided.
71 - 'aux-sample-size': Set sample size for AUX area sampling. If the
72 '--aux-sample' option has been used, set aux-sample-size=0 to disable
73 AUX area sampling for the event.
75 See the linkperf:perf-list[1] man page for more parameters.
77 Note: If user explicitly sets options which conflict with the params,
78 the value set by the parameters will be overridden.
80 Also not defined in .../<pmu>/format/* are PMU driver specific
81 configuration parameters. Any configuration parameter preceded by
82 the letter '@' is not interpreted in user space and sent down directly
83 to the PMU driver. For example:
85 perf record -e some_event/@cfg1,@cfg2=config/ ...
87 will see 'cfg1' and 'cfg2=config' pushed to the PMU driver associated
88 with the event for further processing. There is no restriction on
89 what the configuration parameters are, as long as their semantic is
90 understood and supported by the PMU driver.
92 - a hardware breakpoint event in the form of '\mem:addr[/len][:access]'
93 where addr is the address in memory you want to break in.
94 Access is the memory access type (read, write, execute) it can
95 be passed as follows: '\mem:addr[:[r][w][x]]'. len is the range,
96 number of bytes from specified addr, which the breakpoint will cover.
97 If you want to profile read-write accesses in 0x1000, just set
99 If you want to profile write accesses in [0x1000~1008), just set
102 - a BPF source file (ending in .c) or a precompiled object file (ending
103 in .o) selects one or more BPF events.
104 The BPF program can attach to various perf events based on the ELF section
107 When processing a '.c' file, perf searches an installed LLVM to compile it
108 into an object file first. Optional clang options can be passed via the
109 '--clang-opt' command line option, e.g.:
111 perf record --clang-opt "-DLINUX_VERSION_CODE=0x50000" \
112 -e tests/bpf-script-example.c
114 Note: '--clang-opt' must be placed before '--event/-e'.
116 - a group of events surrounded by a pair of brace ("{event1,event2,...}").
117 Each event is separated by commas and the group should be quoted to
118 prevent the shell interpretation. You also need to use --group on
119 "perf report" to view group events together.
122 Event filter. This option should follow an event selector (-e) which
123 selects either tracepoint event(s) or a hardware trace PMU
124 (e.g. Intel PT or CoreSight).
128 In the case of tracepoints, multiple '--filter' options are combined
133 A hardware trace PMU advertises its ability to accept a number of
134 address filters by specifying a non-zero value in
135 /sys/bus/event_source/devices/<pmu>/nr_addr_filters.
137 Address filters have the format:
139 filter|start|stop|tracestop <start> [/ <size>] [@<file name>]
142 - 'filter': defines a region that will be traced.
143 - 'start': defines an address at which tracing will begin.
144 - 'stop': defines an address at which tracing will stop.
145 - 'tracestop': defines a region in which tracing will stop.
147 <file name> is the name of the object file, <start> is the offset to the
148 code to trace in that file, and <size> is the size of the region to
149 trace. 'start' and 'stop' filters need not specify a <size>.
151 If no object file is specified then the kernel is assumed, in which case
152 the start address must be a current kernel memory address.
154 <start> can also be specified by providing the name of a symbol. If the
155 symbol name is not unique, it can be disambiguated by inserting #n where
156 'n' selects the n'th symbol in address order. Alternately #0, #g or #G
157 select only a global symbol. <size> can also be specified by providing
158 the name of a symbol, in which case the size is calculated to the end
159 of that symbol. For 'filter' and 'tracestop' filters, if <size> is
160 omitted and <start> is a symbol, then the size is calculated to the end
163 If <size> is omitted and <start> is '*', then the start and size will
164 be calculated from the first and last symbols, i.e. to trace the whole
167 If symbol names (or '*') are provided, they must be surrounded by white
170 The filter passed to the kernel is not necessarily the same as entered.
171 To see the filter that is passed, use the -v option.
173 The kernel may not be able to configure a trace region if it is not
174 within a single mapping. MMAP events (or /proc/<pid>/maps) can be
175 examined to determine if that is a possibility.
177 Multiple filters can be separated with space or comma.
180 Don't record events issued by perf itself. This option should follow
181 an event selector (-e) which selects tracepoint event(s). It adds a
182 filter expression 'common_pid != $PERFPID' to filters. If other
183 '--filter' exists, the new filter expression will be combined with
188 System-wide collection from all CPUs (default if no target is specified).
192 Record events on existing process ID (comma separated list).
196 Record events on existing thread ID (comma separated list).
197 This option also disables inheritance by default. Enable it by adding
202 Record events in threads owned by uid. Name or number.
206 Collect data with this RT SCHED_FIFO priority.
209 Collect data without buffering.
213 Event period to sample.
221 Child tasks do not inherit counters.
225 Profile at this frequency. Use 'max' to use the currently maximum
226 allowed frequency, i.e. the value in the kernel.perf_event_max_sample_rate
227 sysctl. Will throttle down to the currently maximum allowed frequency.
231 Fail if the specified frequency can't be used.
235 Number of mmap data pages (must be a power of two) or size
236 specification with appended unit character - B/K/M/G. The
237 size is rounded up to have nearest pages power of two value.
238 Also, by adding a comma, the number of mmap pages for AUX
239 area tracing can be specified.
242 Put all events in a single event group. This precedes the --event
243 option and remains only for backward compatibility. See --event.
246 Enables call-graph (stack chain/backtrace) recording for both
247 kernel space and user space.
250 Setup and enable call-graph (stack chain/backtrace) recording,
251 implies -g. Default is "fp" (for user space).
253 The unwinding method used for kernel space is dependent on the
254 unwinder used by the active kernel configuration, i.e
255 CONFIG_UNWINDER_FRAME_POINTER (fp) or CONFIG_UNWINDER_ORC (orc)
257 Any option specified here controls the method used for user space.
259 Valid options are "fp" (frame pointer), "dwarf" (DWARF's CFI -
260 Call Frame Information) or "lbr" (Hardware Last Branch Record
263 In some systems, where binaries are build with gcc
264 --fomit-frame-pointer, using the "fp" method will produce bogus
265 call graphs, using "dwarf", if available (perf tools linked to
266 the libunwind or libdw library) should be used instead.
267 Using the "lbr" method doesn't require any compiler options. It
268 will produce call graphs from the hardware LBR registers. The
269 main limitation is that it is only available on new Intel
270 platforms, such as Haswell. It can only get user call chain. It
271 doesn't work with branch stack sampling at the same time.
273 When "dwarf" recording is used, perf also records (user) stack dump
274 when sampled. Default size of the stack dump is 8192 (bytes).
275 User can change the size by passing the size after comma like
276 "--call-graph dwarf,4096".
280 Don't print any message, useful for scripting.
284 Be more verbose (show counter open errors, etc).
288 Record per-thread event counts. Use it with 'perf report -T' to see
293 Record the sample virtual addresses.
296 Record the sample physical addresses.
299 Record the sampled data address data page size.
302 Record the sampled code address (ip) page size
306 Record the sample timestamps. Use it with 'perf report -D' to see the
307 timestamps, for instance.
311 Record the sample period.
314 Record the sample cpu.
322 Collect raw sample records from all opened counters (default for tracepoint counters).
326 Collect samples only on the list of CPUs provided. Multiple CPUs can be provided as a
327 comma-separated list with no space: 0,1. Ranges of CPUs are specified with -: 0-2.
328 In per-thread mode with inheritance mode on (default), samples are captured only when
329 the thread executes on the designated CPUs. Default is to monitor all CPUs.
333 Do not save the build ids of binaries in the perf.data files. This skips
334 post processing after recording, which sometimes makes the final step in
335 the recording process to take a long time, as it needs to process all
336 events looking for mmap records. The downside is that it can misresolve
337 symbols if the workload binaries used when recording get locally rebuilt
338 or upgraded, because the only key available in this case is the
339 pathname. You can also set the "record.build-id" config variable to
340 'skip to have this behaviour permanently.
344 Do not update the buildid cache. This saves some overhead in situations
345 where the information in the perf.data file (which includes buildids)
346 is sufficient. You can also set the "record.build-id" config variable to
347 'no-cache' to have the same effect.
351 monitor only in the container (cgroup) called "name". This option is available only
352 in per-cpu mode. The cgroup filesystem must be mounted. All threads belonging to
353 container "name" are monitored when they run on the monitored CPUs. Multiple cgroups
354 can be provided. Each cgroup is applied to the corresponding event, i.e., first cgroup
355 to first event, second cgroup to second event and so on. It is possible to provide
356 an empty cgroup (monitor all the time) using, e.g., -G foo,,bar. Cgroups must have
357 corresponding events, i.e., they always refer to events defined earlier on the command
358 line. If the user wants to track multiple events for a specific cgroup, the user can
359 use '-e e1 -e e2 -G foo,foo' or just use '-e e1 -e e2 -G foo'.
361 If wanting to monitor, say, 'cycles' for a cgroup and also for system wide, this
362 command line can be used: 'perf stat -e cycles -G cgroup_name -a -e cycles'.
366 Enable taken branch stack sampling. Any type of taken branch may be sampled.
367 This is a shortcut for --branch-filter any. See --branch-filter for more infos.
371 Enable taken branch stack sampling. Each sample captures a series of consecutive
372 taken branches. The number of branches captured with each sample depends on the
373 underlying hardware, the type of branches of interest, and the executed code.
374 It is possible to select the types of branches captured by enabling filters. The
375 following filters are defined:
377 - any: any type of branches
378 - any_call: any function call or system call
379 - any_ret: any function return or system call return
380 - ind_call: any indirect branch
381 - call: direct calls, including far (to/from kernel) calls
382 - u: only when the branch target is at the user level
383 - k: only when the branch target is in the kernel
384 - hv: only when the target is at the hypervisor level
385 - in_tx: only when the target is in a hardware transaction
386 - no_tx: only when the target is not in a hardware transaction
387 - abort_tx: only when the target is a hardware transaction abort
388 - cond: conditional branches
389 - save_type: save branch type during sampling in case binary is not available later
392 The option requires at least one branch type among any, any_call, any_ret, ind_call, cond.
393 The privilege levels may be omitted, in which case, the privilege levels of the associated
394 event are applied to the branch filter. Both kernel (k) and hypervisor (hv) privilege
395 levels are subject to permissions. When sampling on multiple events, branch stack sampling
396 is enabled for all the sampling events. The sampled branch type is the same for all events.
397 The various filters must be specified as a comma separated list: --branch-filter any_ret,u,k
398 Note that this feature may not be available on all processors.
401 Enable weightened sampling. An additional weight is recorded per sample and can be
402 displayed with the weight and local_weight sort keys. This currently works for TSX
403 abort events and some memory events in precise mode on modern Intel CPUs.
406 Record events of type PERF_RECORD_NAMESPACES. This enables 'cgroup_id' sort key.
409 Record events of type PERF_RECORD_CGROUP. This enables 'cgroup' sort key.
412 Record transaction flags for transaction related events.
415 Use per-thread mmaps. By default per-cpu mmaps are created. This option
416 overrides that and uses per-thread mmaps. A side-effect of that is that
417 inheritance is automatically disabled. --per-thread is ignored with a warning
418 if combined with -a or -C options.
422 After starting the program, wait msecs before measuring (-1: start with events
423 disabled). This is useful to filter out the startup phase of the program, which
424 is often very different.
428 Capture machine state (registers) at interrupt, i.e., on counter overflows for
429 each sample. List of captured registers depends on the architecture. This option
430 is off by default. It is possible to select the registers to sample using their
431 symbolic names, e.g. on x86, ax, si. To list the available registers use
432 --intr-regs=\?. To name registers, pass a comma separated list such as
433 --intr-regs=ax,bx. The list of register is architecture dependent.
436 Similar to -I, but capture user registers at sample time. To list the available
437 user registers use --user-regs=\?.
440 Record running and enabled time for read events (:S)
444 Sets the clock id to use for the various time fields in the perf_event_type
445 records. See clock_gettime(). In particular CLOCK_MONOTONIC and
446 CLOCK_MONOTONIC_RAW are supported, some events might also allow
447 CLOCK_BOOTTIME, CLOCK_REALTIME and CLOCK_TAI.
451 Select AUX area tracing Snapshot Mode. This option is valid only with an
452 AUX area tracing event. Optionally, certain snapshot capturing parameters
453 can be specified in a string that follows this option:
454 'e': take one last snapshot on exit; guarantees that there is at least one
455 snapshot in the output file;
456 <size>: if the PMU supports this, specify the desired snapshot size.
458 In Snapshot Mode trace data is captured only when signal SIGUSR2 is received
459 and on exit if the above 'e' option is given.
461 --aux-sample[=OPTIONS]::
462 Select AUX area sampling. At least one of the events selected by the -e option
463 must be an AUX area event. Samples on other events will be created containing
464 data from the AUX area. Optionally sample size may be specified, otherwise it
468 When processing pre-existing threads /proc/XXX/mmap, it may take a long time,
469 because the file may be huge. A time out is needed in such cases.
470 This option sets the time out limit. The default value is 500 ms.
473 Record context switch events i.e. events of type PERF_RECORD_SWITCH or
474 PERF_RECORD_SWITCH_CPU_WIDE. In some cases (e.g. Intel PT, CoreSight or Arm SPE)
475 switch events will be enabled automatically, which can be suppressed by
476 by the option --no-switch-events.
479 Path to clang binary to use for compiling BPF scriptlets.
480 (enabled when BPF support is on)
482 --clang-opt=OPTIONS::
483 Options passed to clang when compiling BPF scriptlets.
484 (enabled when BPF support is on)
487 Specify vmlinux path which has debuginfo.
488 (enabled when BPF prologue is on)
491 Record build-id of all DSOs regardless whether it's actually hit or not.
494 Record build ids in mmap2 events, disables build id cache (implies --no-buildid).
497 Use <n> control blocks in asynchronous (Posix AIO) trace writing mode (default: 1, max: 4).
498 Asynchronous mode is supported only when linking Perf tool with libc library
499 providing implementation for Posix AIO API.
502 Set affinity mask of trace reading thread according to the policy defined by 'mode' value:
503 node - thread affinity mask is set to NUMA node cpu mask of the processed mmap buffer
504 cpu - thread affinity mask is set to cpu of the processed mmap buffer
506 --mmap-flush=number::
508 Specify minimal number of bytes that is extracted from mmap data pages and
509 processed for output. One can specify the number using B/K/M/G suffixes.
511 The maximal allowed value is a quarter of the size of mmaped data pages.
513 The default option value is 1 byte which means that every time that the output
514 writing thread finds some new data in the mmaped buffer the data is extracted,
515 possibly compressed (-z) and written to the output, perf.data or pipe.
517 Larger data chunks are compressed more effectively in comparison to smaller
518 chunks so extraction of larger chunks from the mmap data pages is preferable
519 from the perspective of output size reduction.
521 Also at some cases executing less output write syscalls with bigger data size
522 can take less time than executing more output write syscalls with smaller data
523 size thus lowering runtime profiling overhead.
526 --compression-level[=n]::
527 Produce compressed trace using specified level n (default: 1 - fastest compression,
531 Configure all used events to run in kernel space.
534 Configure all used events to run in user space.
536 --kernel-callchains::
537 Collect callchains only from kernel space. I.e. this option sets
538 perf_event_attr.exclude_callchain_user to 1.
541 Collect callchains only from user space. I.e. this option sets
542 perf_event_attr.exclude_callchain_kernel to 1.
544 Don't use both --kernel-callchains and --user-callchains at the same time or no
545 callchains will be collected.
548 Append timestamp to output file name.
550 --timestamp-boundary::
551 Record timestamp boundary (time of first/last samples).
553 --switch-output[=mode]::
554 Generate multiple perf.data files, timestamp prefixed, switching to a new one
555 based on 'mode' value:
556 "signal" - when receiving a SIGUSR2 (default value) or
557 <size> - when reaching the size threshold, size is expected to
558 be a number with appended unit character - B/K/M/G
559 <time> - when reaching the time threshold, size is expected to
560 be a number with appended unit character - s/m/h/d
562 Note: the precision of the size threshold hugely depends
563 on your configuration - the number and size of your ring
564 buffers (-m). It is generally more precise for higher sizes
565 (like >5M), for lower values expect different sizes.
567 A possible use case is to, given an external event, slice the perf.data file
568 that gets then processed, possibly via a perf script, to decide if that
569 particular perf.data snapshot should be kept or not.
571 Implies --timestamp-filename, --no-buildid and --no-buildid-cache.
572 The reason for the latter two is to reduce the data file switching
573 overhead. You can still switch them on with:
575 --switch-output --no-no-buildid --no-no-buildid-cache
577 --switch-output-event::
578 Events that will cause the switch of the perf.data file, auto-selecting
579 --switch-output=signal, the results are similar as internally the side band
580 thread will also send a SIGUSR2 to the main one.
582 Uses the same syntax as --event, it will just not be recorded, serving only to
583 switch the perf.data file as soon as the --switch-output event is processed by
584 a separate sideband thread.
586 This sideband thread is also used to other purposes, like processing the
587 PERF_RECORD_BPF_EVENT records as they happen, asking the kernel for extra BPF
590 --switch-max-files=N::
592 When rotating perf.data with --switch-output, only keep N files.
595 Parse options then exit. --dry-run can be used to detect errors in cmdline
598 'perf record --dry-run -e' can act as a BPF script compiler if llvm.dump-obj
599 in config file is set to true.
602 Collect and synthesize given type of events (comma separated). Note that
603 this option controls the synthesis from the /proc filesystem which represent
604 task status for pre-existing threads.
606 Kernel (and some other) events are recorded regardless of the
607 choice in this option. For example, --synth=no would have MMAP events for
611 'task' - synthesize FORK and COMM events for each task
612 'mmap' - synthesize MMAP events for each process (implies 'task')
613 'cgroup' - synthesize CGROUP events for each cgroup
614 'all' - synthesize all events (default)
615 'no' - do not synthesize any of the above events
618 Instead of collecting non-sample events (for example, fork, comm, mmap) at
619 the beginning of record, collect them during finalizing an output file.
620 The collected non-sample events reflects the status of the system when
624 Makes all events use an overwritable ring buffer. An overwritable ring
625 buffer works like a flight recorder: when it gets full, the kernel will
626 overwrite the oldest records, that thus will never make it to the
629 When '--overwrite' and '--switch-output' are used perf records and drops
630 events until it receives a signal, meaning that something unusual was
631 detected that warrants taking a snapshot of the most current events,
632 those fitting in the ring buffer at that moment.
634 'overwrite' attribute can also be set or canceled for an event using
635 config terms. For example: 'cycles/overwrite/' and 'instructions/no-overwrite/'.
637 Implies --tail-synthesize.
640 Make a copy of /proc/kcore and place it into a directory with the perf data file.
643 Limit the sample data max size, <size> is expected to be a number with
644 appended unit character - B/K/M/G
646 --num-thread-synthesize::
647 The number of threads to run when synthesizing events for existing processes.
648 By default, the number of threads equals 1.
651 --pfm-events events::
652 Select a PMU event using libpfm4 syntax (see http://perfmon2.sf.net)
653 including support for event filters. For example '--pfm-events
654 inst_retired:any_p:u:c=1:i'. More than one event can be passed to the
655 option using the comma separator. Hardware events and generic hardware
656 events cannot be mixed together. The latter must be used with the -e
657 option. The -e option and this one can be mixed and matched. Events
658 can be grouped using the {} notation.
661 --control=fifo:ctl-fifo[,ack-fifo]::
662 --control=fd:ctl-fd[,ack-fd]::
663 ctl-fifo / ack-fifo are opened and used as ctl-fd / ack-fd as follows.
664 Listen on ctl-fd descriptor for command to control measurement.
667 'enable' : enable events
668 'disable' : disable events
669 'enable name' : enable event 'name'
670 'disable name' : disable event 'name'
671 'snapshot' : AUX area tracing snapshot).
672 'stop' : stop perf record
675 'evlist [-v|-g|-F] : display all events
676 -F Show just the sample frequency used for each event.
678 -g Show event group information.
680 Measurements can be started with events disabled using --delay=-1 option. Optionally
681 send control command completion ('ack\n') to ack-fd descriptor to synchronize with the
682 controlling process. Example of bash shell script to enable and disable events during
689 ctl_fifo=${ctl_dir}perf_ctl.fifo
690 test -p ${ctl_fifo} && unlink ${ctl_fifo}
692 exec {ctl_fd}<>${ctl_fifo}
694 ctl_ack_fifo=${ctl_dir}perf_ctl_ack.fifo
695 test -p ${ctl_ack_fifo} && unlink ${ctl_ack_fifo}
696 mkfifo ${ctl_ack_fifo}
697 exec {ctl_fd_ack}<>${ctl_ack_fifo}
699 perf record -D -1 -e cpu-cycles -a \
700 --control fd:${ctl_fd},${ctl_fd_ack} \
704 sleep 5 && echo 'enable' >&${ctl_fd} && read -u ${ctl_fd_ack} e1 && echo "enabled(${e1})"
705 sleep 10 && echo 'disable' >&${ctl_fd} && read -u ${ctl_fd_ack} d1 && echo "disabled(${d1})"
708 unlink ${ctl_ack_fifo}
716 include::intel-hybrid.txt[]
718 --debuginfod[=URLs]::
719 Specify debuginfod URL to be used when cacheing perf.data binaries,
720 it follows the same syntax as the DEBUGINFOD_URLS variable, like:
722 http://192.168.122.174:8002
724 If the URLs is not specified, the value of DEBUGINFOD_URLS
725 system environment variable is used.
729 linkperf:perf-stat[1], linkperf:perf-list[1], linkperf:perf-intel-pt[1]