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 symbolically formed PMU event like pmu/param1=0x3,param2/
31 where param1, param2, etc are defined as formats for the PMU in
32 /sys/bus/event_source/devices/<pmu>/format/*.
33 · a symbolically formed event like
35 pmu/config=M,config1=N,config3=K/
36 where M, N, K are numbers (in decimal, hex, octal format). Acceptable
38 values for each of 'config', 'config1' and 'config2' are defined by
39 corresponding entries in /sys/bus/event_source/devices/<pmu>/format/*
40 param1 and param2 are defined as formats for the PMU in:
41 /sys/bus/event_source/devices/<pmu>/format/*
42 There are also some parameters which are not defined in .../<pmu>/format/*.
44 These params can be used to overload default config values per event.
45 Here are some common parameters:
46 - 'period': Set event sampling period
47 - 'freq': Set event sampling frequency
48 - 'time': Disable/enable time stamping. Acceptable values are 1 for
49 enabling time stamping. 0 for disabling time stamping.
50 The default is 1.
51 - 'call-graph': Disable/enable callgraph. Acceptable str are "fp" for
52 FP mode, "dwarf" for DWARF mode, "lbr" for LBR mode and
53 "no" for disable callgraph.
54 - 'stack-size': user stack size for dwarf mode
55 - 'name' : User defined event name. Single quotes (') may be used to
56 escape symbols in the name from parsing by shell and tool
57 like this: name=\'CPU_CLK_UNHALTED.THREAD:cmask=0x1\'.
58 See the linkperf:perf-list[1] man page for more parameters.
60 Note: If user explicitly sets options which conflict with the params,
62 the value set by the parameters will be overridden.
63 Also not defined in .../<pmu>/format/* are PMU driver specific
65 configuration parameters. Any configuration parameter preceded by
66 the letter '@' is not interpreted in user space and sent down directly
67 to the PMU driver. For example:
68 perf record -e some_event/@cfg1,@cfg2=config/ ...
70 will see 'cfg1' and 'cfg2=config' pushed to the PMU driver associated
72 with the event for further processing. There is no restriction on
73 what the configuration parameters are, as long as their semantic is
74 understood and supported by the PMU driver.
75 · a hardware breakpoint event in the form of
77 \mem:addr[/len][:access] where addr is the address in memory
78 you want to break in. Access is the memory access type (read,
79 write, execute) it can be passed as follows:
80 \mem:addr[:[r][w][x]]. len is the range, number of bytes from
81 specified addr, which the breakpoint will cover. If you want to
82 profile read-write accesses in 0x1000, just set mem:0x1000:rw.
83 If you want to profile write accesses in [0x1000~1008), just
84 set mem:0x1000/8:w.
85 · a BPF source file (ending in .c) or a precompiled object file
87 (ending in .o) selects one or more BPF events. The BPF program
88 can attach to various perf events based on the ELF section
89 names.
90 When processing a '.c' file, perf searches an installed LLVM to compile it
92 into an object file first. Optional clang options can be passed via the
93 '--clang-opt' command line option, e.g.:
94 perf record --clang-opt "-DLINUX_VERSION_CODE=0x50000" \
96 -e tests/bpf-script-example.c
97 Note: '--clang-opt' must be placed before '--event/-e'.
99 · a group of events surrounded by a pair of brace
101 ("{event1,event2,...}"). Each event is separated by commas and
102 the group should be quoted to prevent the shell interpretation.
103 You also need to use --group on "perf report" to view group
104 events together.
105 --filter=<filter>
107 Event filter. This option should follow an event selector (-e)
108 which selects either tracepoint event(s) or a hardware trace PMU
109 (e.g. Intel PT or CoreSight).
110 · tracepoint filters
112 In the case of tracepoints, multiple '--filter' options are combined
114 using '&&'.
115 · address filters
117 A hardware trace PMU advertises its ability to accept a number of
119 address filters by specifying a non-zero value in
120 /sys/bus/event_source/devices/<pmu>/nr_addr_filters.
121 Address filters have the format:
123 filter|start|stop|tracestop <start> [/ <size>] [@<file name>]
125 Where:
127 - 'filter': defines a region that will be traced.
128 - 'start': defines an address at which tracing will begin.
129 - 'stop': defines an address at which tracing will stop.
130 - 'tracestop': defines a region in which tracing will stop.
131 <file name> is the name of the object file, <start> is the offset to the
133 code to trace in that file, and <size> is the size of the region to
134 trace. 'start' and 'stop' filters need not specify a <size>.
135 If no object file is specified then the kernel is assumed, in which case
137 the start address must be a current kernel memory address.
138 <start> can also be specified by providing the name of a symbol. If the
140 symbol name is not unique, it can be disambiguated by inserting #n where
141 'n' selects the n'th symbol in address order. Alternately #0, #g or #G
142 select only a global symbol. <size> can also be specified by providing
143 the name of a symbol, in which case the size is calculated to the end
144 of that symbol. For 'filter' and 'tracestop' filters, if <size> is
145 omitted and <start> is a symbol, then the size is calculated to the end
146 of that symbol.
147 If <size> is omitted and <start> is '*', then the start and size will
149 be calculated from the first and last symbols, i.e. to trace the whole
150 file.
151 If symbol names (or '*') are provided, they must be surrounded by white
153 space.
154 The filter passed to the kernel is not necessarily the same as entered.
156 To see the filter that is passed, use the -v option.
157 The kernel may not be able to configure a trace region if it is not
159 within a single mapping. MMAP events (or /proc/<pid>/maps) can be
160 examined to determine if that is a possibility.
161 Multiple filters can be separated with space or comma.
163 --exclude-perf
165 Don’t record events issued by perf itself. This option should
166 follow an event selector (-e) which selects tracepoint event(s). It
167 adds a filter expression common_pid != $PERFPID to filters. If
168 other --filter exists, the new filter expression will be combined
169 with them by &&.
170 -a, --all-cpus
172 System-wide collection from all CPUs (default if no target is
173 specified).
174 -p, --pid=
176 Record events on existing process ID (comma separated list).
177 -t, --tid=
179 Record events on existing thread ID (comma separated list). This
180 option also disables inheritance by default. Enable it by adding
181 --inherit.
182 -u, --uid=
184 Record events in threads owned by uid. Name or number.
185 -r, --realtime=
187 Collect data with this RT SCHED_FIFO priority.
188 --no-buffering
190 Collect data without buffering.
191 -c, --count=
193 Event period to sample.
194 -o, --output=
196 Output file name.
197 -i, --no-inherit
199 Child tasks do not inherit counters.
200 -F, --freq=
202 Profile at this frequency. Use max to use the currently maximum
203 allowed frequency, i.e. the value in the
204 kernel.perf_event_max_sample_rate sysctl. Will throttle down to the
205 currently maximum allowed frequency. See --strict-freq.
206 --strict-freq
208 Fail if the specified frequency can’t be used.
209 -m, --mmap-pages=
211 Number of mmap data pages (must be a power of two) or size
212 specification with appended unit character - B/K/M/G. The size is
213 rounded up to have nearest pages power of two value. Also, by
214 adding a comma, the number of mmap pages for AUX area tracing can
215 be specified.
216 --group
218 Put all events in a single event group. This precedes the --event
219 option and remains only for backward compatibility. See --event.
220 -g
222 Enables call-graph (stack chain/backtrace) recording.
223 --call-graph
225 Setup and enable call-graph (stack chain/backtrace) recording,
226 implies -g. Default is "fp".
227 Allows specifying "fp" (frame pointer) or "dwarf"
229 (DWARF's CFI - Call Frame Information) or "lbr"
230 (Hardware Last Branch Record facility) as the method to collect
231 the information used to show the call graphs.
232 In some systems, where binaries are build with gcc
234 --fomit-frame-pointer, using the "fp" method will produce bogus
235 call graphs, using "dwarf", if available (perf tools linked to
236 the libunwind or libdw library) should be used instead.
237 Using the "lbr" method doesn't require any compiler options. It
238 will produce call graphs from the hardware LBR registers. The
239 main limitation is that it is only available on new Intel
240 platforms, such as Haswell. It can only get user call chain. It
241 doesn't work with branch stack sampling at the same time.
242 When "dwarf" recording is used, perf also records (user) stack dump
244 when sampled. Default size of the stack dump is 8192 (bytes).
245 User can change the size by passing the size after comma like
246 "--call-graph dwarf,4096".
247 -q, --quiet
249 Don’t print any message, useful for scripting.
250 -v, --verbose
252 Be more verbose (show counter open errors, etc).
253 -s, --stat
255 Record per-thread event counts. Use it with perf report -T to see
256 the values.
257 -d, --data
259 Record the sample virtual addresses.
260 --phys-data
262 Record the sample physical addresses.
263 -T, --timestamp
265 Record the sample timestamps. Use it with perf report -D to see the
266 timestamps, for instance.
267 -P, --period
269 Record the sample period.
270 --sample-cpu
272 Record the sample cpu.
273 -n, --no-samples
275 Don’t sample.
276 -R, --raw-samples
278 Collect raw sample records from all opened counters (default for
279 tracepoint counters).
280 -C, --cpu
282 Collect samples only on the list of CPUs provided. Multiple CPUs
283 can be provided as a comma-separated list with no space: 0,1.
284 Ranges of CPUs are specified with -: 0-2. In per-thread mode with
285 inheritance mode on (default), samples are captured only when the
286 thread executes on the designated CPUs. Default is to monitor all
287 CPUs.
288 -B, --no-buildid
290 Do not save the build ids of binaries in the perf.data files. This
291 skips post processing after recording, which sometimes makes the
292 final step in the recording process to take a long time, as it
293 needs to process all events looking for mmap records. The downside
294 is that it can misresolve symbols if the workload binaries used
295 when recording get locally rebuilt or upgraded, because the only
296 key available in this case is the pathname. You can also set the
297 "record.build-id" config variable to 'skip to have this behaviour
298 permanently.
299 -N, --no-buildid-cache
301 Do not update the buildid cache. This saves some overhead in
302 situations where the information in the perf.data file (which
303 includes buildids) is sufficient. You can also set the
304 "record.build-id" config variable to no-cache to have the same
305 effect.
306 -G name,..., --cgroup name,...
308 monitor only in the container (cgroup) called "name". This option
309 is available only in per-cpu mode. The cgroup filesystem must be
310 mounted. All threads belonging to container "name" are monitored
311 when they run on the monitored CPUs. Multiple cgroups can be
312 provided. Each cgroup is applied to the corresponding event, i.e.,
313 first cgroup to first event, second cgroup to second event and so
314 on. It is possible to provide an empty cgroup (monitor all the
315 time) using, e.g., -G foo,,bar. Cgroups must have corresponding
316 events, i.e., they always refer to events defined earlier on the
317 command line. If the user wants to track multiple events for a
318 specific cgroup, the user can use -e e1 -e e2 -G foo,foo or just
319 use -e e1 -e e2 -G foo.
320 If wanting to monitor, say, cycles for a cgroup and also for system
322 wide, this command line can be used: perf stat -e cycles -G cgroup_name
323 -a -e cycles.
324 -b, --branch-any
326 Enable taken branch stack sampling. Any type of taken branch may be
327 sampled. This is a shortcut for --branch-filter any. See
328 --branch-filter for more infos.
329 -j, --branch-filter
331 Enable taken branch stack sampling. Each sample captures a series
332 of consecutive taken branches. The number of branches captured with
333 each sample depends on the underlying hardware, the type of
334 branches of interest, and the executed code. It is possible to
335 select the types of branches captured by enabling filters. The
336 following filters are defined:
337 · any: any type of branches
339 · any_call: any function call or system call
341 · any_ret: any function return or system call return
343 · ind_call: any indirect branch
345 · call: direct calls, including far (to/from kernel) calls
347 · u: only when the branch target is at the user level
349 · k: only when the branch target is in the kernel
351 · hv: only when the target is at the hypervisor level
353 · in_tx: only when the target is in a hardware transaction
355 · no_tx: only when the target is not in a hardware transaction
357 · abort_tx: only when the target is a hardware transaction abort
359 · cond: conditional branches
361 · save_type: save branch type during sampling in case binary is
363 not available later
364 The option requires at least one branch type among any, any_call,
366 any_ret, ind_call, cond. The privilege levels may be omitted, in
367 which case, the privilege levels of the associated event are
368 applied to the branch filter. Both kernel (k) and hypervisor (hv)
369 privilege levels are subject to permissions. When sampling on
370 multiple events, branch stack sampling is enabled for all the
371 sampling events. The sampled branch type is the same for all
372 events. The various filters must be specified as a comma separated
373 list: --branch-filter any_ret,u,k Note that this feature may not be
374 available on all processors.
375 --weight
377 Enable weightened sampling. An additional weight is recorded per
378 sample and can be displayed with the weight and local_weight sort
379 keys. This currently works for TSX abort events and some memory
380 events in precise mode on modern Intel CPUs.
381 --namespaces
383 Record events of type PERF_RECORD_NAMESPACES.
384 --transaction
386 Record transaction flags for transaction related events.
387 --per-thread
389 Use per-thread mmaps. By default per-cpu mmaps are created. This
390 option overrides that and uses per-thread mmaps. A side-effect of
391 that is that inheritance is automatically disabled. --per-thread is
392 ignored with a warning if combined with -a or -C options.
393 -D, --delay=
395 After starting the program, wait msecs before measuring. This is
396 useful to filter out the startup phase of the program, which is
397 often very different.
398 -I, --intr-regs
400 Capture machine state (registers) at interrupt, i.e., on counter
401 overflows for each sample. List of captured registers depends on
402 the architecture. This option is off by default. It is possible to
403 select the registers to sample using their symbolic names, e.g. on
404 x86, ax, si. To list the available registers use --intr-regs=\?. To
405 name registers, pass a comma separated list such as
406 --intr-regs=ax,bx. The list of register is architecture dependent.
407 --user-regs
409 Similar to -I, but capture user registers at sample time. To list
410 the available user registers use --user-regs=\?.
411 --running-time
413 Record running and enabled time for read events (:S)
414 -k, --clockid
416 Sets the clock id to use for the various time fields in the
417 perf_event_type records. See clock_gettime(). In particular
418 CLOCK_MONOTONIC and CLOCK_MONOTONIC_RAW are supported, some events
419 might also allow CLOCK_BOOTTIME, CLOCK_REALTIME and CLOCK_TAI.
420 -S, --snapshot
422 Select AUX area tracing Snapshot Mode. This option is valid only
423 with an AUX area tracing event. Optionally the number of bytes to
424 capture per snapshot can be specified. In Snapshot Mode, trace data
425 is captured only when signal SIGUSR2 is received.
426 --proc-map-timeout
428 When processing pre-existing threads /proc/XXX/mmap, it may take a
429 long time, because the file may be huge. A time out is needed in
430 such cases. This option sets the time out limit. The default value
431 is 500 ms.
432 --switch-events
434 Record context switch events i.e. events of type PERF_RECORD_SWITCH
435 or PERF_RECORD_SWITCH_CPU_WIDE.
436 --clang-path=PATH
438 Path to clang binary to use for compiling BPF scriptlets. (enabled
439 when BPF support is on)
440 --clang-opt=OPTIONS
442 Options passed to clang when compiling BPF scriptlets. (enabled
443 when BPF support is on)
444 --vmlinux=PATH
446 Specify vmlinux path which has debuginfo. (enabled when BPF
447 prologue is on)
448 --buildid-all
450 Record build-id of all DSOs regardless whether it’s actually hit or
451 not.
452 --aio[=n]
454 Use <n> control blocks in asynchronous (Posix AIO) trace writing
455 mode (default: 1, max: 4). Asynchronous mode is supported only when
456 linking Perf tool with libc library providing implementation for
457 Posix AIO API.
458 --affinity=mode
460 Set affinity mask of trace reading thread according to the policy
461 defined by mode value: node - thread affinity mask is set to NUMA
462 node cpu mask of the processed mmap buffer cpu - thread affinity
463 mask is set to cpu of the processed mmap buffer
464 --mmap-flush=number
466 Specify minimal number of bytes that is extracted from mmap data
467 pages and processed for output. One can specify the number using
468 B/K/M/G suffixes.
469 The maximal allowed value is a quarter of the size of mmaped data
471 pages.
472 The default option value is 1 byte which means that every time that the
474 output writing thread finds some new data in the mmaped buffer the data
475 is extracted, possibly compressed (-z) and written to the output,
476 perf.data or pipe.
477 Larger data chunks are compressed more effectively in comparison to
479 smaller chunks so extraction of larger chunks from the mmap data pages
480 is preferable from the perspective of output size reduction.
481 Also at some cases executing less output write syscalls with bigger
483 data size can take less time than executing more output write syscalls
484 with smaller data size thus lowering runtime profiling overhead.
485 -z, --compression-level[=n]
487 Produce compressed trace using specified level n (default: 1 -
488 fastest compression, 22 - smallest trace)
489 --all-kernel
491 Configure all used events to run in kernel space.
492 --all-user
494 Configure all used events to run in user space.
495 --kernel-callchains
497 Collect callchains only from kernel space. I.e. this option sets
498 perf_event_attr.exclude_callchain_user to 1.
499 --user-callchains
501 Collect callchains only from user space. I.e. this option sets
502 perf_event_attr.exclude_callchain_kernel to 1.
503 Don’t use both --kernel-callchains and --user-callchains at the same
505 time or no callchains will be collected.
506 --timestamp-filename Append timestamp to output file name.
508 --timestamp-boundary
510 Record timestamp boundary (time of first/last samples).
511 --switch-output[=mode]
513 Generate multiple perf.data files, timestamp prefixed, switching to
514 a new one based on mode value: "signal" - when receiving a SIGUSR2
515 (default value) or <size> - when reaching the size threshold, size
516 is expected to be a number with appended unit character - B/K/M/G
517 <time> - when reaching the time threshold, size is expected to be a
518 number with appended unit character - s/m/h/d
519 Note: the precision of the size threshold hugely depends
521 on your configuration - the number and size of your ring
522 buffers (-m). It is generally more precise for higher sizes
523 (like >5M), for lower values expect different sizes.
524 A possible use case is to, given an external event, slice the perf.data
526 file that gets then processed, possibly via a perf script, to decide if
527 that particular perf.data snapshot should be kept or not.
528 Implies --timestamp-filename, --no-buildid and --no-buildid-cache. The
530 reason for the latter two is to reduce the data file switching
531 overhead. You can still switch them on with:
532 --switch-output --no-no-buildid --no-no-buildid-cache
534 --switch-max-files=N
536 When rotating perf.data with --switch-output, only keep N files.
537 --dry-run
539 Parse options then exit. --dry-run can be used to detect errors in
540 cmdline options.
541 perf record --dry-run -e can act as a BPF script compiler if
543 llvm.dump-obj in config file is set to true.
544 --tail-synthesize
546 Instead of collecting non-sample events (for example, fork, comm,
547 mmap) at the beginning of record, collect them during finalizing an
548 output file. The collected non-sample events reflects the status of
549 the system when record is finished.
550 --overwrite
552 Makes all events use an overwritable ring buffer. An overwritable
553 ring buffer works like a flight recorder: when it gets full, the
554 kernel will overwrite the oldest records, that thus will never make
555 it to the perf.data file.
556 When --overwrite and --switch-output are used perf records and drops
558 events until it receives a signal, meaning that something unusual was
559 detected that warrants taking a snapshot of the most current events,
560 those fitting in the ring buffer at that moment.
561 overwrite attribute can also be set or canceled for an event using
563 config terms. For example: cycles/overwrite/ and
564 instructions/no-overwrite/.
565 Implies --tail-synthesize.
567
569 perf-stat(1), perf-list(1)
570perf 11/12/2019 PERF-RECORD(1)