Loading...
1perf-record(1)
2==============
3
4NAME
5----
6perf-record - Run a command and record its profile into perf.data
7
8SYNOPSIS
9--------
10[verse]
11'perf record' [-e <EVENT> | --event=EVENT] [-a] <command>
12'perf record' [-e <EVENT> | --event=EVENT] [-a] \-- <command> [<options>]
13
14DESCRIPTION
15-----------
16This command runs a command and gathers a performance counter profile
17from it, into perf.data - without displaying anything.
18
19This file can then be inspected later on, using 'perf report'.
20
21
22OPTIONS
23-------
24<command>...::
25 Any command you can specify in a shell.
26
27-e::
28--event=::
29 Select the PMU event. Selection can be:
30
31 - a symbolic event name (use 'perf list' to list all events)
32
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_source/devices/cpu/format/*.
37
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.
41
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/*.
45
46 - a symbolically formed event like 'pmu/config=M,config1=N,config3=K/'
47
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/*
53
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.
61 The default is 1.
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.
74
75 See the linkperf:perf-list[1] man page for more parameters.
76
77 Note: If user explicitly sets options which conflict with the params,
78 the value set by the parameters will be overridden.
79
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:
84
85 perf record -e some_event/@cfg1,@cfg2=config/ ...
86
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.
91
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
98 'mem:0x1000:rw'.
99 If you want to profile write accesses in [0x1000~1008), just set
100 'mem:0x1000/8:w'.
101
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
105 names.
106
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.:
110
111 perf record --clang-opt "-DLINUX_VERSION_CODE=0x50000" \
112 -e tests/bpf-script-example.c
113
114 Note: '--clang-opt' must be placed before '--event/-e'.
115
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.
120
121--filter=<filter>::
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).
125
126 - tracepoint filters
127
128 In the case of tracepoints, multiple '--filter' options are combined
129 using '&&'.
130
131 - address filters
132
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.
136
137 Address filters have the format:
138
139 filter|start|stop|tracestop <start> [/ <size>] [@<file name>]
140
141 Where:
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.
146
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>.
150
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.
153
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
161 of that symbol.
162
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
165 file.
166
167 If symbol names (or '*') are provided, they must be surrounded by white
168 space.
169
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.
172
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.
176
177 Multiple filters can be separated with space or comma.
178
179--exclude-perf::
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
184 them by '&&'.
185
186-a::
187--all-cpus::
188 System-wide collection from all CPUs (default if no target is specified).
189
190-p::
191--pid=::
192 Record events on existing process ID (comma separated list).
193
194-t::
195--tid=::
196 Record events on existing thread ID (comma separated list).
197 This option also disables inheritance by default. Enable it by adding
198 --inherit.
199
200-u::
201--uid=::
202 Record events in threads owned by uid. Name or number.
203
204-r::
205--realtime=::
206 Collect data with this RT SCHED_FIFO priority.
207
208--no-buffering::
209 Collect data without buffering.
210
211-c::
212--count=::
213 Event period to sample.
214
215-o::
216--output=::
217 Output file name.
218
219-i::
220--no-inherit::
221 Child tasks do not inherit counters.
222
223-F::
224--freq=::
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.
228 See --strict-freq.
229
230--strict-freq::
231 Fail if the specified frequency can't be used.
232
233-m::
234--mmap-pages=::
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.
240
241-g::
242 Enables call-graph (stack chain/backtrace) recording for both
243 kernel space and user space.
244
245--call-graph::
246 Setup and enable call-graph (stack chain/backtrace) recording,
247 implies -g. Default is "fp" (for user space).
248
249 The unwinding method used for kernel space is dependent on the
250 unwinder used by the active kernel configuration, i.e
251 CONFIG_UNWINDER_FRAME_POINTER (fp) or CONFIG_UNWINDER_ORC (orc)
252
253 Any option specified here controls the method used for user space.
254
255 Valid options are "fp" (frame pointer), "dwarf" (DWARF's CFI -
256 Call Frame Information) or "lbr" (Hardware Last Branch Record
257 facility).
258
259 In some systems, where binaries are build with gcc
260 --fomit-frame-pointer, using the "fp" method will produce bogus
261 call graphs, using "dwarf", if available (perf tools linked to
262 the libunwind or libdw library) should be used instead.
263 Using the "lbr" method doesn't require any compiler options. It
264 will produce call graphs from the hardware LBR registers. The
265 main limitation is that it is only available on new Intel
266 platforms, such as Haswell. It can only get user call chain. It
267 doesn't work with branch stack sampling at the same time.
268
269 When "dwarf" recording is used, perf also records (user) stack dump
270 when sampled. Default size of the stack dump is 8192 (bytes).
271 User can change the size by passing the size after comma like
272 "--call-graph dwarf,4096".
273
274 When "fp" recording is used, perf tries to save stack enties
275 up to the number specified in sysctl.kernel.perf_event_max_stack
276 by default. User can change the number by passing it after comma
277 like "--call-graph fp,32".
278
279-q::
280--quiet::
281 Don't print any warnings or messages, useful for scripting.
282
283-v::
284--verbose::
285 Be more verbose (show counter open errors, etc).
286
287-s::
288--stat::
289 Record per-thread event counts. Use it with 'perf report -T' to see
290 the values.
291
292-d::
293--data::
294 Record the sample virtual addresses.
295
296--phys-data::
297 Record the sample physical addresses.
298
299--data-page-size::
300 Record the sampled data address data page size.
301
302--code-page-size::
303 Record the sampled code address (ip) page size
304
305-T::
306--timestamp::
307 Record the sample timestamps. Use it with 'perf report -D' to see the
308 timestamps, for instance.
309
310-P::
311--period::
312 Record the sample period.
313
314--sample-cpu::
315 Record the sample cpu.
316
317--sample-identifier::
318 Record the sample identifier i.e. PERF_SAMPLE_IDENTIFIER bit set in
319 the sample_type member of the struct perf_event_attr argument to the
320 perf_event_open system call.
321
322-n::
323--no-samples::
324 Don't sample.
325
326-R::
327--raw-samples::
328Collect raw sample records from all opened counters (default for tracepoint counters).
329
330-C::
331--cpu::
332Collect samples only on the list of CPUs provided. Multiple CPUs can be provided as a
333comma-separated list with no space: 0,1. Ranges of CPUs are specified with -: 0-2.
334In per-thread mode with inheritance mode on (default), samples are captured only when
335the thread executes on the designated CPUs. Default is to monitor all CPUs.
336
337-B::
338--no-buildid::
339Do not save the build ids of binaries in the perf.data files. This skips
340post processing after recording, which sometimes makes the final step in
341the recording process to take a long time, as it needs to process all
342events looking for mmap records. The downside is that it can misresolve
343symbols if the workload binaries used when recording get locally rebuilt
344or upgraded, because the only key available in this case is the
345pathname. You can also set the "record.build-id" config variable to
346'skip to have this behaviour permanently.
347
348-N::
349--no-buildid-cache::
350Do not update the buildid cache. This saves some overhead in situations
351where the information in the perf.data file (which includes buildids)
352is sufficient. You can also set the "record.build-id" config variable to
353'no-cache' to have the same effect.
354
355-G name,...::
356--cgroup name,...::
357monitor only in the container (cgroup) called "name". This option is available only
358in per-cpu mode. The cgroup filesystem must be mounted. All threads belonging to
359container "name" are monitored when they run on the monitored CPUs. Multiple cgroups
360can be provided. Each cgroup is applied to the corresponding event, i.e., first cgroup
361to first event, second cgroup to second event and so on. It is possible to provide
362an empty cgroup (monitor all the time) using, e.g., -G foo,,bar. Cgroups must have
363corresponding events, i.e., they always refer to events defined earlier on the command
364line. If the user wants to track multiple events for a specific cgroup, the user can
365use '-e e1 -e e2 -G foo,foo' or just use '-e e1 -e e2 -G foo'.
366
367If wanting to monitor, say, 'cycles' for a cgroup and also for system wide, this
368command line can be used: 'perf stat -e cycles -G cgroup_name -a -e cycles'.
369
370-b::
371--branch-any::
372Enable taken branch stack sampling. Any type of taken branch may be sampled.
373This is a shortcut for --branch-filter any. See --branch-filter for more infos.
374
375-j::
376--branch-filter::
377Enable taken branch stack sampling. Each sample captures a series of consecutive
378taken branches. The number of branches captured with each sample depends on the
379underlying hardware, the type of branches of interest, and the executed code.
380It is possible to select the types of branches captured by enabling filters. The
381following filters are defined:
382
383 - any: any type of branches
384 - any_call: any function call or system call
385 - any_ret: any function return or system call return
386 - ind_call: any indirect branch
387 - ind_jmp: any indirect jump
388 - call: direct calls, including far (to/from kernel) calls
389 - u: only when the branch target is at the user level
390 - k: only when the branch target is in the kernel
391 - hv: only when the target is at the hypervisor level
392 - in_tx: only when the target is in a hardware transaction
393 - no_tx: only when the target is not in a hardware transaction
394 - abort_tx: only when the target is a hardware transaction abort
395 - cond: conditional branches
396 - call_stack: save call stack
397 - no_flags: don't save branch flags e.g prediction, misprediction etc
398 - no_cycles: don't save branch cycles
399 - hw_index: save branch hardware index
400 - save_type: save branch type during sampling in case binary is not available later
401 For the platforms with Intel Arch LBR support (12th-Gen+ client or
402 4th-Gen Xeon+ server), the save branch type is unconditionally enabled
403 when the taken branch stack sampling is enabled.
404 - priv: save privilege state during sampling in case binary is not available later
405
406+
407The option requires at least one branch type among any, any_call, any_ret, ind_call, cond.
408The privilege levels may be omitted, in which case, the privilege levels of the associated
409event are applied to the branch filter. Both kernel (k) and hypervisor (hv) privilege
410levels are subject to permissions. When sampling on multiple events, branch stack sampling
411is enabled for all the sampling events. The sampled branch type is the same for all events.
412The various filters must be specified as a comma separated list: --branch-filter any_ret,u,k
413Note that this feature may not be available on all processors.
414
415-W::
416--weight::
417Enable weightened sampling. An additional weight is recorded per sample and can be
418displayed with the weight and local_weight sort keys. This currently works for TSX
419abort events and some memory events in precise mode on modern Intel CPUs.
420
421--namespaces::
422Record events of type PERF_RECORD_NAMESPACES. This enables 'cgroup_id' sort key.
423
424--all-cgroups::
425Record events of type PERF_RECORD_CGROUP. This enables 'cgroup' sort key.
426
427--transaction::
428Record transaction flags for transaction related events.
429
430--per-thread::
431Use per-thread mmaps. By default per-cpu mmaps are created. This option
432overrides that and uses per-thread mmaps. A side-effect of that is that
433inheritance is automatically disabled. --per-thread is ignored with a warning
434if combined with -a or -C options.
435
436-D::
437--delay=::
438After starting the program, wait msecs before measuring (-1: start with events
439disabled), or enable events only for specified ranges of msecs (e.g.
440-D 10-20,30-40 means wait 10 msecs, enable for 10 msecs, wait 10 msecs, enable
441for 10 msecs, then stop). Note, delaying enabling of events is useful to filter
442out the startup phase of the program, which is often very different.
443
444-I::
445--intr-regs::
446Capture machine state (registers) at interrupt, i.e., on counter overflows for
447each sample. List of captured registers depends on the architecture. This option
448is off by default. It is possible to select the registers to sample using their
449symbolic names, e.g. on x86, ax, si. To list the available registers use
450--intr-regs=\?. To name registers, pass a comma separated list such as
451--intr-regs=ax,bx. The list of register is architecture dependent.
452
453--user-regs::
454Similar to -I, but capture user registers at sample time. To list the available
455user registers use --user-regs=\?.
456
457--running-time::
458Record running and enabled time for read events (:S)
459
460-k::
461--clockid::
462Sets the clock id to use for the various time fields in the perf_event_type
463records. See clock_gettime(). In particular CLOCK_MONOTONIC and
464CLOCK_MONOTONIC_RAW are supported, some events might also allow
465CLOCK_BOOTTIME, CLOCK_REALTIME and CLOCK_TAI.
466
467-S::
468--snapshot::
469Select AUX area tracing Snapshot Mode. This option is valid only with an
470AUX area tracing event. Optionally, certain snapshot capturing parameters
471can be specified in a string that follows this option:
472 'e': take one last snapshot on exit; guarantees that there is at least one
473 snapshot in the output file;
474 <size>: if the PMU supports this, specify the desired snapshot size.
475
476In Snapshot Mode trace data is captured only when signal SIGUSR2 is received
477and on exit if the above 'e' option is given.
478
479--aux-sample[=OPTIONS]::
480Select AUX area sampling. At least one of the events selected by the -e option
481must be an AUX area event. Samples on other events will be created containing
482data from the AUX area. Optionally sample size may be specified, otherwise it
483defaults to 4KiB.
484
485--proc-map-timeout::
486When processing pre-existing threads /proc/XXX/mmap, it may take a long time,
487because the file may be huge. A time out is needed in such cases.
488This option sets the time out limit. The default value is 500 ms.
489
490--switch-events::
491Record context switch events i.e. events of type PERF_RECORD_SWITCH or
492PERF_RECORD_SWITCH_CPU_WIDE. In some cases (e.g. Intel PT, CoreSight or Arm SPE)
493switch events will be enabled automatically, which can be suppressed by
494by the option --no-switch-events.
495
496--clang-path=PATH::
497Path to clang binary to use for compiling BPF scriptlets.
498(enabled when BPF support is on)
499
500--clang-opt=OPTIONS::
501Options passed to clang when compiling BPF scriptlets.
502(enabled when BPF support is on)
503
504--vmlinux=PATH::
505Specify vmlinux path which has debuginfo.
506(enabled when BPF prologue is on)
507
508--buildid-all::
509Record build-id of all DSOs regardless whether it's actually hit or not.
510
511--buildid-mmap::
512Record build ids in mmap2 events, disables build id cache (implies --no-buildid).
513
514--aio[=n]::
515Use <n> control blocks in asynchronous (Posix AIO) trace writing mode (default: 1, max: 4).
516Asynchronous mode is supported only when linking Perf tool with libc library
517providing implementation for Posix AIO API.
518
519--affinity=mode::
520Set affinity mask of trace reading thread according to the policy defined by 'mode' value:
521 node - thread affinity mask is set to NUMA node cpu mask of the processed mmap buffer
522 cpu - thread affinity mask is set to cpu of the processed mmap buffer
523
524--mmap-flush=number::
525
526Specify minimal number of bytes that is extracted from mmap data pages and
527processed for output. One can specify the number using B/K/M/G suffixes.
528
529The maximal allowed value is a quarter of the size of mmaped data pages.
530
531The default option value is 1 byte which means that every time that the output
532writing thread finds some new data in the mmaped buffer the data is extracted,
533possibly compressed (-z) and written to the output, perf.data or pipe.
534
535Larger data chunks are compressed more effectively in comparison to smaller
536chunks so extraction of larger chunks from the mmap data pages is preferable
537from the perspective of output size reduction.
538
539Also at some cases executing less output write syscalls with bigger data size
540can take less time than executing more output write syscalls with smaller data
541size thus lowering runtime profiling overhead.
542
543-z::
544--compression-level[=n]::
545Produce compressed trace using specified level n (default: 1 - fastest compression,
54622 - smallest trace)
547
548--all-kernel::
549Configure all used events to run in kernel space.
550
551--all-user::
552Configure all used events to run in user space.
553
554--kernel-callchains::
555Collect callchains only from kernel space. I.e. this option sets
556perf_event_attr.exclude_callchain_user to 1.
557
558--user-callchains::
559Collect callchains only from user space. I.e. this option sets
560perf_event_attr.exclude_callchain_kernel to 1.
561
562Don't use both --kernel-callchains and --user-callchains at the same time or no
563callchains will be collected.
564
565--timestamp-filename
566Append timestamp to output file name.
567
568--timestamp-boundary::
569Record timestamp boundary (time of first/last samples).
570
571--switch-output[=mode]::
572Generate multiple perf.data files, timestamp prefixed, switching to a new one
573based on 'mode' value:
574 "signal" - when receiving a SIGUSR2 (default value) or
575 <size> - when reaching the size threshold, size is expected to
576 be a number with appended unit character - B/K/M/G
577 <time> - when reaching the time threshold, size is expected to
578 be a number with appended unit character - s/m/h/d
579
580 Note: the precision of the size threshold hugely depends
581 on your configuration - the number and size of your ring
582 buffers (-m). It is generally more precise for higher sizes
583 (like >5M), for lower values expect different sizes.
584
585A possible use case is to, given an external event, slice the perf.data file
586that gets then processed, possibly via a perf script, to decide if that
587particular perf.data snapshot should be kept or not.
588
589Implies --timestamp-filename, --no-buildid and --no-buildid-cache.
590The reason for the latter two is to reduce the data file switching
591overhead. You can still switch them on with:
592
593 --switch-output --no-no-buildid --no-no-buildid-cache
594
595--switch-output-event::
596Events that will cause the switch of the perf.data file, auto-selecting
597--switch-output=signal, the results are similar as internally the side band
598thread will also send a SIGUSR2 to the main one.
599
600Uses the same syntax as --event, it will just not be recorded, serving only to
601switch the perf.data file as soon as the --switch-output event is processed by
602a separate sideband thread.
603
604This sideband thread is also used to other purposes, like processing the
605PERF_RECORD_BPF_EVENT records as they happen, asking the kernel for extra BPF
606information, etc.
607
608--switch-max-files=N::
609
610When rotating perf.data with --switch-output, only keep N files.
611
612--dry-run::
613Parse options then exit. --dry-run can be used to detect errors in cmdline
614options.
615
616'perf record --dry-run -e' can act as a BPF script compiler if llvm.dump-obj
617in config file is set to true.
618
619--synth=TYPE::
620Collect and synthesize given type of events (comma separated). Note that
621this option controls the synthesis from the /proc filesystem which represent
622task status for pre-existing threads.
623
624Kernel (and some other) events are recorded regardless of the
625choice in this option. For example, --synth=no would have MMAP events for
626kernel and modules.
627
628Available types are:
629 'task' - synthesize FORK and COMM events for each task
630 'mmap' - synthesize MMAP events for each process (implies 'task')
631 'cgroup' - synthesize CGROUP events for each cgroup
632 'all' - synthesize all events (default)
633 'no' - do not synthesize any of the above events
634
635--tail-synthesize::
636Instead of collecting non-sample events (for example, fork, comm, mmap) at
637the beginning of record, collect them during finalizing an output file.
638The collected non-sample events reflects the status of the system when
639record is finished.
640
641--overwrite::
642Makes all events use an overwritable ring buffer. An overwritable ring
643buffer works like a flight recorder: when it gets full, the kernel will
644overwrite the oldest records, that thus will never make it to the
645perf.data file.
646
647When '--overwrite' and '--switch-output' are used perf records and drops
648events until it receives a signal, meaning that something unusual was
649detected that warrants taking a snapshot of the most current events,
650those fitting in the ring buffer at that moment.
651
652'overwrite' attribute can also be set or canceled for an event using
653config terms. For example: 'cycles/overwrite/' and 'instructions/no-overwrite/'.
654
655Implies --tail-synthesize.
656
657--kcore::
658Make a copy of /proc/kcore and place it into a directory with the perf data file.
659
660--max-size=<size>::
661Limit the sample data max size, <size> is expected to be a number with
662appended unit character - B/K/M/G
663
664--num-thread-synthesize::
665 The number of threads to run when synthesizing events for existing processes.
666 By default, the number of threads equals 1.
667
668ifdef::HAVE_LIBPFM[]
669--pfm-events events::
670Select a PMU event using libpfm4 syntax (see http://perfmon2.sf.net)
671including support for event filters. For example '--pfm-events
672inst_retired:any_p:u:c=1:i'. More than one event can be passed to the
673option using the comma separator. Hardware events and generic hardware
674events cannot be mixed together. The latter must be used with the -e
675option. The -e option and this one can be mixed and matched. Events
676can be grouped using the {} notation.
677endif::HAVE_LIBPFM[]
678
679--control=fifo:ctl-fifo[,ack-fifo]::
680--control=fd:ctl-fd[,ack-fd]::
681ctl-fifo / ack-fifo are opened and used as ctl-fd / ack-fd as follows.
682Listen on ctl-fd descriptor for command to control measurement.
683
684Available commands:
685 'enable' : enable events
686 'disable' : disable events
687 'enable name' : enable event 'name'
688 'disable name' : disable event 'name'
689 'snapshot' : AUX area tracing snapshot).
690 'stop' : stop perf record
691 'ping' : ping
692
693 'evlist [-v|-g|-F] : display all events
694 -F Show just the sample frequency used for each event.
695 -v Show all fields.
696 -g Show event group information.
697
698Measurements can be started with events disabled using --delay=-1 option. Optionally
699send control command completion ('ack\n') to ack-fd descriptor to synchronize with the
700controlling process. Example of bash shell script to enable and disable events during
701measurements:
702
703 #!/bin/bash
704
705 ctl_dir=/tmp/
706
707 ctl_fifo=${ctl_dir}perf_ctl.fifo
708 test -p ${ctl_fifo} && unlink ${ctl_fifo}
709 mkfifo ${ctl_fifo}
710 exec {ctl_fd}<>${ctl_fifo}
711
712 ctl_ack_fifo=${ctl_dir}perf_ctl_ack.fifo
713 test -p ${ctl_ack_fifo} && unlink ${ctl_ack_fifo}
714 mkfifo ${ctl_ack_fifo}
715 exec {ctl_fd_ack}<>${ctl_ack_fifo}
716
717 perf record -D -1 -e cpu-cycles -a \
718 --control fd:${ctl_fd},${ctl_fd_ack} \
719 -- sleep 30 &
720 perf_pid=$!
721
722 sleep 5 && echo 'enable' >&${ctl_fd} && read -u ${ctl_fd_ack} e1 && echo "enabled(${e1})"
723 sleep 10 && echo 'disable' >&${ctl_fd} && read -u ${ctl_fd_ack} d1 && echo "disabled(${d1})"
724
725 exec {ctl_fd_ack}>&-
726 unlink ${ctl_ack_fifo}
727
728 exec {ctl_fd}>&-
729 unlink ${ctl_fifo}
730
731 wait -n ${perf_pid}
732 exit $?
733
734--threads=<spec>::
735Write collected trace data into several data files using parallel threads.
736<spec> value can be user defined list of masks. Masks separated by colon
737define CPUs to be monitored by a thread and affinity mask of that thread
738is separated by slash:
739
740 <cpus mask 1>/<affinity mask 1>:<cpus mask 2>/<affinity mask 2>:...
741
742CPUs or affinity masks must not overlap with other corresponding masks.
743Invalid CPUs are ignored, but masks containing only invalid CPUs are not
744allowed.
745
746For example user specification like the following:
747
748 0,2-4/2-4:1,5-7/5-7
749
750specifies parallel threads layout that consists of two threads,
751the first thread monitors CPUs 0 and 2-4 with the affinity mask 2-4,
752the second monitors CPUs 1 and 5-7 with the affinity mask 5-7.
753
754<spec> value can also be a string meaning predefined parallel threads
755layout:
756
757 cpu - create new data streaming thread for every monitored cpu
758 core - create new thread to monitor CPUs grouped by a core
759 package - create new thread to monitor CPUs grouped by a package
760 numa - create new threed to monitor CPUs grouped by a NUMA domain
761
762Predefined layouts can be used on systems with large number of CPUs in
763order not to spawn multiple per-cpu streaming threads but still avoid LOST
764events in data directory files. Option specified with no or empty value
765defaults to CPU layout. Masks defined or provided by the option value are
766filtered through the mask provided by -C option.
767
768--debuginfod[=URLs]::
769 Specify debuginfod URL to be used when cacheing perf.data binaries,
770 it follows the same syntax as the DEBUGINFOD_URLS variable, like:
771
772 http://192.168.122.174:8002
773
774 If the URLs is not specified, the value of DEBUGINFOD_URLS
775 system environment variable is used.
776
777--off-cpu::
778 Enable off-cpu profiling with BPF. The BPF program will collect
779 task scheduling information with (user) stacktrace and save them
780 as sample data of a software event named "offcpu-time". The
781 sample period will have the time the task slept in nanoseconds.
782
783 Note that BPF can collect stack traces using frame pointer ("fp")
784 only, as of now. So the applications built without the frame
785 pointer might see bogus addresses.
786
787include::intel-hybrid.txt[]
788
789SEE ALSO
790--------
791linkperf:perf-stat[1], linkperf:perf-list[1], linkperf:perf-intel-pt[1]
1perf-record(1)
2==============
3
4NAME
5----
6perf-record - Run a command and record its profile into perf.data
7
8SYNOPSIS
9--------
10[verse]
11'perf record' [-e <EVENT> | --event=EVENT] [-l] [-a] <command>
12'perf record' [-e <EVENT> | --event=EVENT] [-l] [-a] -- <command> [<options>]
13
14DESCRIPTION
15-----------
16This command runs a command and gathers a performance counter profile
17from it, into perf.data - without displaying anything.
18
19This file can then be inspected later on, using 'perf report'.
20
21
22OPTIONS
23-------
24<command>...::
25 Any command you can specify in a shell.
26
27-e::
28--event=::
29 Select the PMU event. Selection can be:
30
31 - a symbolic event name (use 'perf list' to list all events)
32
33 - a raw PMU event (eventsel+umask) in the form of rNNN where NNN is a
34 hexadecimal event descriptor.
35
36 - a symbolically formed PMU event like 'pmu/param1=0x3,param2/' where
37 'param1', 'param2', etc are defined as formats for the PMU in
38 /sys/bus/event_source/devices/<pmu>/format/*.
39
40 - a symbolically formed event like 'pmu/config=M,config1=N,config3=K/'
41
42 where M, N, K are numbers (in decimal, hex, octal format). Acceptable
43 values for each of 'config', 'config1' and 'config2' are defined by
44 corresponding entries in /sys/bus/event_source/devices/<pmu>/format/*
45 param1 and param2 are defined as formats for the PMU in:
46 /sys/bus/event_source/devices/<pmu>/format/*
47
48 There are also some parameters which are not defined in .../<pmu>/format/*.
49 These params can be used to overload default config values per event.
50 Here are some common parameters:
51 - 'period': Set event sampling period
52 - 'freq': Set event sampling frequency
53 - 'time': Disable/enable time stamping. Acceptable values are 1 for
54 enabling time stamping. 0 for disabling time stamping.
55 The default is 1.
56 - 'call-graph': Disable/enable callgraph. Acceptable str are "fp" for
57 FP mode, "dwarf" for DWARF mode, "lbr" for LBR mode and
58 "no" for disable callgraph.
59 - 'stack-size': user stack size for dwarf mode
60
61 See the linkperf:perf-list[1] man page for more parameters.
62
63 Note: If user explicitly sets options which conflict with the params,
64 the value set by the parameters will be overridden.
65
66 Also not defined in .../<pmu>/format/* are PMU driver specific
67 configuration parameters. Any configuration parameter preceded by
68 the letter '@' is not interpreted in user space and sent down directly
69 to the PMU driver. For example:
70
71 perf record -e some_event/@cfg1,@cfg2=config/ ...
72
73 will see 'cfg1' and 'cfg2=config' pushed to the PMU driver associated
74 with the event for further processing. There is no restriction on
75 what the configuration parameters are, as long as their semantic is
76 understood and supported by the PMU driver.
77
78 - a hardware breakpoint event in the form of '\mem:addr[/len][:access]'
79 where addr is the address in memory you want to break in.
80 Access is the memory access type (read, write, execute) it can
81 be passed as follows: '\mem:addr[:[r][w][x]]'. len is the range,
82 number of bytes from specified addr, which the breakpoint will cover.
83 If you want to profile read-write accesses in 0x1000, just set
84 'mem:0x1000:rw'.
85 If you want to profile write accesses in [0x1000~1008), just set
86 'mem:0x1000/8:w'.
87
88 - a group of events surrounded by a pair of brace ("{event1,event2,...}").
89 Each event is separated by commas and the group should be quoted to
90 prevent the shell interpretation. You also need to use --group on
91 "perf report" to view group events together.
92
93--filter=<filter>::
94 Event filter. This option should follow a event selector (-e) which
95 selects either tracepoint event(s) or a hardware trace PMU
96 (e.g. Intel PT or CoreSight).
97
98 - tracepoint filters
99
100 In the case of tracepoints, multiple '--filter' options are combined
101 using '&&'.
102
103 - address filters
104
105 A hardware trace PMU advertises its ability to accept a number of
106 address filters by specifying a non-zero value in
107 /sys/bus/event_source/devices/<pmu>/nr_addr_filters.
108
109 Address filters have the format:
110
111 filter|start|stop|tracestop <start> [/ <size>] [@<file name>]
112
113 Where:
114 - 'filter': defines a region that will be traced.
115 - 'start': defines an address at which tracing will begin.
116 - 'stop': defines an address at which tracing will stop.
117 - 'tracestop': defines a region in which tracing will stop.
118
119 <file name> is the name of the object file, <start> is the offset to the
120 code to trace in that file, and <size> is the size of the region to
121 trace. 'start' and 'stop' filters need not specify a <size>.
122
123 If no object file is specified then the kernel is assumed, in which case
124 the start address must be a current kernel memory address.
125
126 <start> can also be specified by providing the name of a symbol. If the
127 symbol name is not unique, it can be disambiguated by inserting #n where
128 'n' selects the n'th symbol in address order. Alternately #0, #g or #G
129 select only a global symbol. <size> can also be specified by providing
130 the name of a symbol, in which case the size is calculated to the end
131 of that symbol. For 'filter' and 'tracestop' filters, if <size> is
132 omitted and <start> is a symbol, then the size is calculated to the end
133 of that symbol.
134
135 If <size> is omitted and <start> is '*', then the start and size will
136 be calculated from the first and last symbols, i.e. to trace the whole
137 file.
138
139 If symbol names (or '*') are provided, they must be surrounded by white
140 space.
141
142 The filter passed to the kernel is not necessarily the same as entered.
143 To see the filter that is passed, use the -v option.
144
145 The kernel may not be able to configure a trace region if it is not
146 within a single mapping. MMAP events (or /proc/<pid>/maps) can be
147 examined to determine if that is a possibility.
148
149 Multiple filters can be separated with space or comma.
150
151--exclude-perf::
152 Don't record events issued by perf itself. This option should follow
153 a event selector (-e) which selects tracepoint event(s). It adds a
154 filter expression 'common_pid != $PERFPID' to filters. If other
155 '--filter' exists, the new filter expression will be combined with
156 them by '&&'.
157
158-a::
159--all-cpus::
160 System-wide collection from all CPUs.
161
162-p::
163--pid=::
164 Record events on existing process ID (comma separated list).
165
166-t::
167--tid=::
168 Record events on existing thread ID (comma separated list).
169 This option also disables inheritance by default. Enable it by adding
170 --inherit.
171
172-u::
173--uid=::
174 Record events in threads owned by uid. Name or number.
175
176-r::
177--realtime=::
178 Collect data with this RT SCHED_FIFO priority.
179
180--no-buffering::
181 Collect data without buffering.
182
183-c::
184--count=::
185 Event period to sample.
186
187-o::
188--output=::
189 Output file name.
190
191-i::
192--no-inherit::
193 Child tasks do not inherit counters.
194-F::
195--freq=::
196 Profile at this frequency.
197
198-m::
199--mmap-pages=::
200 Number of mmap data pages (must be a power of two) or size
201 specification with appended unit character - B/K/M/G. The
202 size is rounded up to have nearest pages power of two value.
203 Also, by adding a comma, the number of mmap pages for AUX
204 area tracing can be specified.
205
206--group::
207 Put all events in a single event group. This precedes the --event
208 option and remains only for backward compatibility. See --event.
209
210-g::
211 Enables call-graph (stack chain/backtrace) recording.
212
213--call-graph::
214 Setup and enable call-graph (stack chain/backtrace) recording,
215 implies -g. Default is "fp".
216
217 Allows specifying "fp" (frame pointer) or "dwarf"
218 (DWARF's CFI - Call Frame Information) or "lbr"
219 (Hardware Last Branch Record facility) as the method to collect
220 the information used to show the call graphs.
221
222 In some systems, where binaries are build with gcc
223 --fomit-frame-pointer, using the "fp" method will produce bogus
224 call graphs, using "dwarf", if available (perf tools linked to
225 the libunwind or libdw library) should be used instead.
226 Using the "lbr" method doesn't require any compiler options. It
227 will produce call graphs from the hardware LBR registers. The
228 main limition is that it is only available on new Intel
229 platforms, such as Haswell. It can only get user call chain. It
230 doesn't work with branch stack sampling at the same time.
231
232 When "dwarf" recording is used, perf also records (user) stack dump
233 when sampled. Default size of the stack dump is 8192 (bytes).
234 User can change the size by passing the size after comma like
235 "--call-graph dwarf,4096".
236
237-q::
238--quiet::
239 Don't print any message, useful for scripting.
240
241-v::
242--verbose::
243 Be more verbose (show counter open errors, etc).
244
245-s::
246--stat::
247 Record per-thread event counts. Use it with 'perf report -T' to see
248 the values.
249
250-d::
251--data::
252 Record the sample addresses.
253
254-T::
255--timestamp::
256 Record the sample timestamps. Use it with 'perf report -D' to see the
257 timestamps, for instance.
258
259-P::
260--period::
261 Record the sample period.
262
263--sample-cpu::
264 Record the sample cpu.
265
266-n::
267--no-samples::
268 Don't sample.
269
270-R::
271--raw-samples::
272Collect raw sample records from all opened counters (default for tracepoint counters).
273
274-C::
275--cpu::
276Collect samples only on the list of CPUs provided. Multiple CPUs can be provided as a
277comma-separated list with no space: 0,1. Ranges of CPUs are specified with -: 0-2.
278In per-thread mode with inheritance mode on (default), samples are captured only when
279the thread executes on the designated CPUs. Default is to monitor all CPUs.
280
281-B::
282--no-buildid::
283Do not save the build ids of binaries in the perf.data files. This skips
284post processing after recording, which sometimes makes the final step in
285the recording process to take a long time, as it needs to process all
286events looking for mmap records. The downside is that it can misresolve
287symbols if the workload binaries used when recording get locally rebuilt
288or upgraded, because the only key available in this case is the
289pathname. You can also set the "record.build-id" config variable to
290'skip to have this behaviour permanently.
291
292-N::
293--no-buildid-cache::
294Do not update the buildid cache. This saves some overhead in situations
295where the information in the perf.data file (which includes buildids)
296is sufficient. You can also set the "record.build-id" config variable to
297'no-cache' to have the same effect.
298
299-G name,...::
300--cgroup name,...::
301monitor only in the container (cgroup) called "name". This option is available only
302in per-cpu mode. The cgroup filesystem must be mounted. All threads belonging to
303container "name" are monitored when they run on the monitored CPUs. Multiple cgroups
304can be provided. Each cgroup is applied to the corresponding event, i.e., first cgroup
305to first event, second cgroup to second event and so on. It is possible to provide
306an empty cgroup (monitor all the time) using, e.g., -G foo,,bar. Cgroups must have
307corresponding events, i.e., they always refer to events defined earlier on the command
308line.
309
310-b::
311--branch-any::
312Enable taken branch stack sampling. Any type of taken branch may be sampled.
313This is a shortcut for --branch-filter any. See --branch-filter for more infos.
314
315-j::
316--branch-filter::
317Enable taken branch stack sampling. Each sample captures a series of consecutive
318taken branches. The number of branches captured with each sample depends on the
319underlying hardware, the type of branches of interest, and the executed code.
320It is possible to select the types of branches captured by enabling filters. The
321following filters are defined:
322
323 - any: any type of branches
324 - any_call: any function call or system call
325 - any_ret: any function return or system call return
326 - ind_call: any indirect branch
327 - call: direct calls, including far (to/from kernel) calls
328 - u: only when the branch target is at the user level
329 - k: only when the branch target is in the kernel
330 - hv: only when the target is at the hypervisor level
331 - in_tx: only when the target is in a hardware transaction
332 - no_tx: only when the target is not in a hardware transaction
333 - abort_tx: only when the target is a hardware transaction abort
334 - cond: conditional branches
335
336+
337The option requires at least one branch type among any, any_call, any_ret, ind_call, cond.
338The privilege levels may be omitted, in which case, the privilege levels of the associated
339event are applied to the branch filter. Both kernel (k) and hypervisor (hv) privilege
340levels are subject to permissions. When sampling on multiple events, branch stack sampling
341is enabled for all the sampling events. The sampled branch type is the same for all events.
342The various filters must be specified as a comma separated list: --branch-filter any_ret,u,k
343Note that this feature may not be available on all processors.
344
345--weight::
346Enable weightened sampling. An additional weight is recorded per sample and can be
347displayed with the weight and local_weight sort keys. This currently works for TSX
348abort events and some memory events in precise mode on modern Intel CPUs.
349
350--transaction::
351Record transaction flags for transaction related events.
352
353--per-thread::
354Use per-thread mmaps. By default per-cpu mmaps are created. This option
355overrides that and uses per-thread mmaps. A side-effect of that is that
356inheritance is automatically disabled. --per-thread is ignored with a warning
357if combined with -a or -C options.
358
359-D::
360--delay=::
361After starting the program, wait msecs before measuring. This is useful to
362filter out the startup phase of the program, which is often very different.
363
364-I::
365--intr-regs::
366Capture machine state (registers) at interrupt, i.e., on counter overflows for
367each sample. List of captured registers depends on the architecture. This option
368is off by default. It is possible to select the registers to sample using their
369symbolic names, e.g. on x86, ax, si. To list the available registers use
370--intr-regs=\?. To name registers, pass a comma separated list such as
371--intr-regs=ax,bx. The list of register is architecture dependent.
372
373
374--running-time::
375Record running and enabled time for read events (:S)
376
377-k::
378--clockid::
379Sets the clock id to use for the various time fields in the perf_event_type
380records. See clock_gettime(). In particular CLOCK_MONOTONIC and
381CLOCK_MONOTONIC_RAW are supported, some events might also allow
382CLOCK_BOOTTIME, CLOCK_REALTIME and CLOCK_TAI.
383
384-S::
385--snapshot::
386Select AUX area tracing Snapshot Mode. This option is valid only with an
387AUX area tracing event. Optionally the number of bytes to capture per
388snapshot can be specified. In Snapshot Mode, trace data is captured only when
389signal SIGUSR2 is received.
390
391--proc-map-timeout::
392When processing pre-existing threads /proc/XXX/mmap, it may take a long time,
393because the file may be huge. A time out is needed in such cases.
394This option sets the time out limit. The default value is 500 ms.
395
396--switch-events::
397Record context switch events i.e. events of type PERF_RECORD_SWITCH or
398PERF_RECORD_SWITCH_CPU_WIDE.
399
400--clang-path=PATH::
401Path to clang binary to use for compiling BPF scriptlets.
402(enabled when BPF support is on)
403
404--clang-opt=OPTIONS::
405Options passed to clang when compiling BPF scriptlets.
406(enabled when BPF support is on)
407
408--vmlinux=PATH::
409Specify vmlinux path which has debuginfo.
410(enabled when BPF prologue is on)
411
412--buildid-all::
413Record build-id of all DSOs regardless whether it's actually hit or not.
414
415--all-kernel::
416Configure all used events to run in kernel space.
417
418--all-user::
419Configure all used events to run in user space.
420
421--timestamp-filename
422Append timestamp to output file name.
423
424--switch-output::
425Generate multiple perf.data files, timestamp prefixed, switching to a new one
426when receiving a SIGUSR2.
427
428A possible use case is to, given an external event, slice the perf.data file
429that gets then processed, possibly via a perf script, to decide if that
430particular perf.data snapshot should be kept or not.
431
432Implies --timestamp-filename, --no-buildid and --no-buildid-cache.
433The reason for the latter two is to reduce the data file switching
434overhead. You can still switch them on with:
435
436 --switch-output --no-no-buildid --no-no-buildid-cache
437
438--dry-run::
439Parse options then exit. --dry-run can be used to detect errors in cmdline
440options.
441
442'perf record --dry-run -e' can act as a BPF script compiler if llvm.dump-obj
443in config file is set to true.
444
445--tail-synthesize::
446Instead of collecting non-sample events (for example, fork, comm, mmap) at
447the beginning of record, collect them during finalizing an output file.
448The collected non-sample events reflects the status of the system when
449record is finished.
450
451--overwrite::
452Makes all events use an overwritable ring buffer. An overwritable ring
453buffer works like a flight recorder: when it gets full, the kernel will
454overwrite the oldest records, that thus will never make it to the
455perf.data file.
456
457When '--overwrite' and '--switch-output' are used perf records and drops
458events until it receives a signal, meaning that something unusual was
459detected that warrants taking a snapshot of the most current events,
460those fitting in the ring buffer at that moment.
461
462'overwrite' attribute can also be set or canceled for an event using
463config terms. For example: 'cycles/overwrite/' and 'instructions/no-overwrite/'.
464
465Implies --tail-synthesize.
466
467SEE ALSO
468--------
469linkperf:perf-stat[1], linkperf:perf-list[1]