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-action': "pause" or "resume" to pause or resume an AUX 72 area event (the group leader) when this event occurs. 73 "start-paused" on an AUX area event itself, will 74 start in a paused state. 75 - 'aux-sample-size': Set sample size for AUX area sampling. If the 76 '--aux-sample' option has been used, set aux-sample-size=0 to disable 77 AUX area sampling for the event. 78 79 See the linkperf:perf-list[1] man page for more parameters. 80 81 Note: If user explicitly sets options which conflict with the params, 82 the value set by the parameters will be overridden. 83 84 Also not defined in .../<pmu>/format/* are PMU driver specific 85 configuration parameters. Any configuration parameter preceded by 86 the letter '@' is not interpreted in user space and sent down directly 87 to the PMU driver. For example: 88 89 perf record -e some_event/@cfg1,@cfg2=config/ ... 90 91 will see 'cfg1' and 'cfg2=config' pushed to the PMU driver associated 92 with the event for further processing. There is no restriction on 93 what the configuration parameters are, as long as their semantic is 94 understood and supported by the PMU driver. 95 96 - a hardware breakpoint event in the form of '\mem:addr[/len][:access]' 97 where addr is the address in memory you want to break in. 98 Access is the memory access type (read, write, execute) it can 99 be passed as follows: '\mem:addr[:[r][w][x]]'. len is the range, 100 number of bytes from specified addr, which the breakpoint will cover. 101 If you want to profile read-write accesses in 0x1000, just set 102 'mem:0x1000:rw'. 103 If you want to profile write accesses in [0x1000~1008), just set 104 'mem:0x1000/8:w'. 105 106 - a group of events surrounded by a pair of brace ("{event1,event2,...}"). 107 Each event is separated by commas and the group should be quoted to 108 prevent the shell interpretation. You also need to use --group on 109 "perf report" to view group events together. 110 111--filter=<filter>:: 112 Event filter. This option should follow an event selector (-e). 113 If the event is a tracepoint, the filter string will be parsed by 114 the kernel. If the event is a hardware trace PMU (e.g. Intel PT 115 or CoreSight), it'll be processed as an address filter. Otherwise 116 it means a general filter using BPF which can be applied for any 117 kind of event. 118 119 - tracepoint filters 120 121 In the case of tracepoints, multiple '--filter' options are combined 122 using '&&'. 123 124 - address filters 125 126 A hardware trace PMU advertises its ability to accept a number of 127 address filters by specifying a non-zero value in 128 /sys/bus/event_source/devices/<pmu>/nr_addr_filters. 129 130 Address filters have the format: 131 132 filter|start|stop|tracestop <start> [/ <size>] [@<file name>] 133 134 Where: 135 - 'filter': defines a region that will be traced. 136 - 'start': defines an address at which tracing will begin. 137 - 'stop': defines an address at which tracing will stop. 138 - 'tracestop': defines a region in which tracing will stop. 139 140 <file name> is the name of the object file, <start> is the offset to the 141 code to trace in that file, and <size> is the size of the region to 142 trace. 'start' and 'stop' filters need not specify a <size>. 143 144 If no object file is specified then the kernel is assumed, in which case 145 the start address must be a current kernel memory address. 146 147 <start> can also be specified by providing the name of a symbol. If the 148 symbol name is not unique, it can be disambiguated by inserting #n where 149 'n' selects the n'th symbol in address order. Alternately #0, #g or #G 150 select only a global symbol. <size> can also be specified by providing 151 the name of a symbol, in which case the size is calculated to the end 152 of that symbol. For 'filter' and 'tracestop' filters, if <size> is 153 omitted and <start> is a symbol, then the size is calculated to the end 154 of that symbol. 155 156 If <size> is omitted and <start> is '*', then the start and size will 157 be calculated from the first and last symbols, i.e. to trace the whole 158 file. 159 160 If symbol names (or '*') are provided, they must be surrounded by white 161 space. 162 163 The filter passed to the kernel is not necessarily the same as entered. 164 To see the filter that is passed, use the -v option. 165 166 The kernel may not be able to configure a trace region if it is not 167 within a single mapping. MMAP events (or /proc/<pid>/maps) can be 168 examined to determine if that is a possibility. 169 170 Multiple filters can be separated with space or comma. 171 172 - bpf filters 173 174 A BPF filter can access the sample data and make a decision based on the 175 data. Users need to set an appropriate sample type to use the BPF 176 filter. BPF filters need root privilege. 177 178 The sample data field can be specified in lower case letter. Multiple 179 filters can be separated with comma. For example, 180 181 --filter 'period > 1000, cpu == 1' 182 or 183 --filter 'mem_op == load || mem_op == store, mem_lvl > l1' 184 185 The former filter only accept samples with period greater than 1000 AND 186 CPU number is 1. The latter one accepts either load and store memory 187 operations but it should have memory level above the L1. Since the 188 mem_op and mem_lvl fields come from the (memory) data_source, it'd only 189 work with some events which set the data_source field. 190 191 Also user should request to collect that information (with -d option in 192 the above case). Otherwise, the following message will be shown. 193 194 $ sudo perf record -e cycles --filter 'mem_op == load' 195 Error: cycles event does not have PERF_SAMPLE_DATA_SRC 196 Hint: please add -d option to perf record. 197 failed to set filter "BPF" on event cycles with 22 (Invalid argument) 198 199 Essentially the BPF filter expression is: 200 201 <term> <operator> <value> (("," | "||") <term> <operator> <value>)* 202 203 The <term> can be one of: 204 ip, id, tid, pid, cpu, time, addr, period, txn, weight, phys_addr, 205 code_pgsz, data_pgsz, weight1, weight2, weight3, ins_lat, retire_lat, 206 p_stage_cyc, mem_op, mem_lvl, mem_snoop, mem_remote, mem_lock, 207 mem_dtlb, mem_blk, mem_hops, uid, gid 208 209 The <operator> can be one of: 210 ==, !=, >, >=, <, <=, & 211 212 The <value> can be one of: 213 <number> (for any term) 214 na, load, store, pfetch, exec (for mem_op) 215 l1, l2, l3, l4, cxl, io, any_cache, lfb, ram, pmem (for mem_lvl) 216 na, none, hit, miss, hitm, fwd, peer (for mem_snoop) 217 remote (for mem_remote) 218 na, locked (for mem_locked) 219 na, l1_hit, l1_miss, l2_hit, l2_miss, any_hit, any_miss, walk, fault (for mem_dtlb) 220 na, by_data, by_addr (for mem_blk) 221 hops0, hops1, hops2, hops3 (for mem_hops) 222 223--exclude-perf:: 224 Don't record events issued by perf itself. This option should follow 225 an event selector (-e) which selects tracepoint event(s). It adds a 226 filter expression 'common_pid != $PERFPID' to filters. If other 227 '--filter' exists, the new filter expression will be combined with 228 them by '&&'. 229 230--latency:: 231 Enable data collection for latency profiling. 232 Use perf report --latency for latency-centric profile. 233 234-a:: 235--all-cpus:: 236 System-wide collection from all CPUs (default if no target is specified). 237 238-p:: 239--pid=:: 240 Record events on existing process ID (comma separated list). 241 242-t:: 243--tid=:: 244 Record events on existing thread ID (comma separated list). 245 This option also disables inheritance by default. Enable it by adding 246 --inherit. 247 248-u:: 249--uid=:: 250 Record events in threads owned by uid. Name or number. 251 252-r:: 253--realtime=:: 254 Collect data with this RT SCHED_FIFO priority. 255 256--no-buffering:: 257 Collect data without buffering. 258 259-c:: 260--count=:: 261 Event period to sample. 262 263-o:: 264--output=:: 265 Output file name. 266 267-i:: 268--no-inherit:: 269 Child tasks do not inherit counters. 270 271-F:: 272--freq=:: 273 Profile at this frequency. Use 'max' to use the currently maximum 274 allowed frequency, i.e. the value in the kernel.perf_event_max_sample_rate 275 sysctl. Will throttle down to the currently maximum allowed frequency. 276 See --strict-freq. 277 278--strict-freq:: 279 Fail if the specified frequency can't be used. 280 281-m:: 282--mmap-pages=:: 283 Number of mmap data pages (must be a power of two) or size 284 specification in bytes with appended unit character - B/K/M/G. 285 The size is rounded up to the nearest power-of-two page value. 286 By adding a comma, an additional parameter with the same 287 semantics used for the normal mmap areas can be specified for 288 AUX tracing area. 289 290-g:: 291 Enables call-graph (stack chain/backtrace) recording for both 292 kernel space and user space. 293 294--call-graph:: 295 Setup and enable call-graph (stack chain/backtrace) recording, 296 implies -g. Default is "fp" (for user space). 297 298 The unwinding method used for kernel space is dependent on the 299 unwinder used by the active kernel configuration, i.e 300 CONFIG_UNWINDER_FRAME_POINTER (fp) or CONFIG_UNWINDER_ORC (orc) 301 302 Any option specified here controls the method used for user space. 303 304 Valid options are "fp" (frame pointer), "dwarf" (DWARF's CFI - 305 Call Frame Information) or "lbr" (Hardware Last Branch Record 306 facility). 307 308 In some systems, where binaries are build with gcc 309 --fomit-frame-pointer, using the "fp" method will produce bogus 310 call graphs, using "dwarf", if available (perf tools linked to 311 the libunwind or libdw library) should be used instead. 312 Using the "lbr" method doesn't require any compiler options. It 313 will produce call graphs from the hardware LBR registers. The 314 main limitation is that it is only available on new Intel 315 platforms, such as Haswell. It can only get user call chain. It 316 doesn't work with branch stack sampling at the same time. 317 318 When "dwarf" recording is used, perf also records (user) stack dump 319 when sampled. Default size of the stack dump is 8192 (bytes). 320 User can change the size by passing the size after comma like 321 "--call-graph dwarf,4096". 322 323 When "fp" recording is used, perf tries to save stack entries 324 up to the number specified in sysctl.kernel.perf_event_max_stack 325 by default. User can change the number by passing it after comma 326 like "--call-graph fp,32". 327 328 Also "defer" can be used with "fp" (like "--call-graph fp,defer") to 329 enable deferred user callchain which will collect user-space callchains 330 when the thread returns to the user space. 331 332-q:: 333--quiet:: 334 Don't print any warnings or messages, useful for scripting. 335 336-v:: 337--verbose:: 338 Be more verbose (show counter open errors, etc). 339 340-s:: 341--stat:: 342 Record per-thread event counts. Use it with 'perf report -T' to see 343 the values. 344 345-d:: 346--data:: 347 Record the sample virtual addresses. Implies --sample-mem-info. 348 349--phys-data:: 350 Record the sample physical addresses. 351 352--data-page-size:: 353 Record the sampled data address data page size. 354 355--code-page-size:: 356 Record the sampled code address (ip) page size 357 358-T:: 359--timestamp:: 360 Record the sample timestamps. Use it with 'perf report -D' to see the 361 timestamps, for instance. 362 363-P:: 364--period:: 365 Record the sample period. 366 367--sample-cpu:: 368 Record the sample cpu. 369 370--sample-identifier:: 371 Record the sample identifier i.e. PERF_SAMPLE_IDENTIFIER bit set in 372 the sample_type member of the struct perf_event_attr argument to the 373 perf_event_open system call. 374 375--sample-mem-info:: 376 Record the sample data source information for memory operations. 377 It requires hardware supports and may work on specific events only. 378 Please consider using 'perf mem record' instead if you're not sure. 379 380-n:: 381--no-samples:: 382 Don't sample. 383 384-R:: 385--raw-samples:: 386Collect raw sample records from all opened counters (default for tracepoint counters). 387 388-C:: 389--cpu:: 390Collect samples only on the list of CPUs provided. Multiple CPUs can be provided as a 391comma-separated list with no space: 0,1. Ranges of CPUs are specified with -: 0-2. 392In per-thread mode with inheritance mode on (default), samples are captured only when 393the thread executes on the designated CPUs. Default is to monitor all CPUs. 394 395User space tasks can migrate between CPUs, so when tracing selected CPUs, 396a dummy event is created to track sideband for all CPUs. 397 398-B:: 399--no-buildid:: 400Do not save the build ids of binaries in the perf.data files. This skips 401post processing after recording, which sometimes makes the final step in 402the recording process to take a long time, as it needs to process all 403events looking for mmap records. The downside is that it can misresolve 404symbols if the workload binaries used when recording get locally rebuilt 405or upgraded, because the only key available in this case is the 406pathname. You can also set the "record.build-id" config variable to 407'skip to have this behaviour permanently. 408 409-N:: 410--no-buildid-cache:: 411Do not update the buildid cache. This saves some overhead in situations 412where the information in the perf.data file (which includes buildids) 413is sufficient. You can also set the "record.build-id" config variable to 414'no-cache' to have the same effect. 415 416-G name,...:: 417--cgroup name,...:: 418monitor only in the container (cgroup) called "name". This option is available only 419in per-cpu mode. The cgroup filesystem must be mounted. All threads belonging to 420container "name" are monitored when they run on the monitored CPUs. Multiple cgroups 421can be provided. Each cgroup is applied to the corresponding event, i.e., first cgroup 422to first event, second cgroup to second event and so on. It is possible to provide 423an empty cgroup (monitor all the time) using, e.g., -G foo,,bar. Cgroups must have 424corresponding events, i.e., they always refer to events defined earlier on the command 425line. If the user wants to track multiple events for a specific cgroup, the user can 426use '-e e1 -e e2 -G foo,foo' or just use '-e e1 -e e2 -G foo'. 427 428If wanting to monitor, say, 'cycles' for a cgroup and also for system wide, this 429command line can be used: 'perf stat -e cycles -G cgroup_name -a -e cycles'. 430 431-b:: 432--branch-any:: 433Enable taken branch stack sampling. Any type of taken branch may be sampled. 434This is a shortcut for --branch-filter any. See --branch-filter for more infos. 435 436-j:: 437--branch-filter:: 438Enable taken branch stack sampling. Each sample captures a series of consecutive 439taken branches. The number of branches captured with each sample depends on the 440underlying hardware, the type of branches of interest, and the executed code. 441It is possible to select the types of branches captured by enabling filters. The 442following filters are defined: 443 444 - any: any type of branches 445 - any_call: any function call or system call 446 - any_ret: any function return or system call return 447 - ind_call: any indirect branch 448 - ind_jmp: any indirect jump 449 - call: direct calls, including far (to/from kernel) calls 450 - u: only when the branch target is at the user level 451 - k: only when the branch target is in the kernel 452 - hv: only when the target is at the hypervisor level 453 - in_tx: only when the target is in a hardware transaction 454 - no_tx: only when the target is not in a hardware transaction 455 - abort_tx: only when the target is a hardware transaction abort 456 - cond: conditional branches 457 - call_stack: save call stack 458 - no_flags: don't save branch flags e.g prediction, misprediction etc 459 - no_cycles: don't save branch cycles 460 - hw_index: save branch hardware index 461 - save_type: save branch type during sampling in case binary is not available later 462 For the platforms with Intel Arch LBR support (12th-Gen+ client or 463 4th-Gen Xeon+ server), the save branch type is unconditionally enabled 464 when the taken branch stack sampling is enabled. 465 - priv: save privilege state during sampling in case binary is not available later 466 - counter: save occurrences of the event since the last branch entry. Currently, the 467 feature is only supported by a newer CPU, e.g., Intel Sierra Forest and 468 later platforms. An error out is expected if it's used on the unsupported 469 kernel or CPUs. 470 471+ 472The option requires at least one branch type among any, any_call, any_ret, ind_call, cond. 473The privilege levels may be omitted, in which case, the privilege levels of the associated 474event are applied to the branch filter. Both kernel (k) and hypervisor (hv) privilege 475levels are subject to permissions. When sampling on multiple events, branch stack sampling 476is enabled for all the sampling events. The sampled branch type is the same for all events. 477The various filters must be specified as a comma separated list: --branch-filter any_ret,u,k 478Note that this feature may not be available on all processors. 479 480-W:: 481--weight:: 482Enable weightened sampling. An additional weight is recorded per sample and can be 483displayed with the weight and local_weight sort keys. This currently works for TSX 484abort events and some memory events in precise mode on modern Intel CPUs. 485 486--namespaces:: 487Record events of type PERF_RECORD_NAMESPACES. This enables 'cgroup_id' sort key. 488 489--all-cgroups:: 490Record events of type PERF_RECORD_CGROUP. This enables 'cgroup' sort key. 491 492--transaction:: 493Record transaction flags for transaction related events. 494 495--per-thread:: 496Use per-thread mmaps. By default per-cpu mmaps are created. This option 497overrides that and uses per-thread mmaps. A side-effect of that is that 498inheritance is automatically disabled. --per-thread is ignored with a warning 499if combined with -a or -C options. 500 501-D:: 502--delay=:: 503After starting the program, wait msecs before measuring (-1: start with events 504disabled), or enable events only for specified ranges of msecs (e.g. 505-D 10-20,30-40 means wait 10 msecs, enable for 10 msecs, wait 10 msecs, enable 506for 10 msecs, then stop). Note, delaying enabling of events is useful to filter 507out the startup phase of the program, which is often very different. 508 509-I:: 510--intr-regs:: 511Capture machine state (registers) at interrupt, i.e., on counter overflows for 512each sample. List of captured registers depends on the architecture. This option 513is off by default. It is possible to select the registers to sample using their 514symbolic names, e.g. on x86, ax, si. To list the available registers use 515--intr-regs=\?. To name registers, pass a comma separated list such as 516--intr-regs=ax,bx. The list of register is architecture dependent. 517 518--user-regs:: 519Similar to -I, but capture user registers at sample time. To list the available 520user registers use --user-regs=\?. 521 522--running-time:: 523Record running and enabled time for read events (:S) 524 525-k:: 526--clockid:: 527Sets the clock id to use for the various time fields in the perf_event_type 528records. See clock_gettime(). In particular CLOCK_MONOTONIC and 529CLOCK_MONOTONIC_RAW are supported, some events might also allow 530CLOCK_BOOTTIME, CLOCK_REALTIME and CLOCK_TAI. 531 532-S:: 533--snapshot:: 534Select AUX area tracing Snapshot Mode. This option is valid only with an 535AUX area tracing event. Optionally, certain snapshot capturing parameters 536can be specified in a string that follows this option: 537 538 - 'e': take one last snapshot on exit; guarantees that there is at least one 539 snapshot in the output file; 540 - <size>: if the PMU supports this, specify the desired snapshot size. 541 542In Snapshot Mode trace data is captured only when signal SIGUSR2 is received 543and on exit if the above 'e' option is given. 544 545--aux-sample[=OPTIONS]:: 546Select AUX area sampling. At least one of the events selected by the -e option 547must be an AUX area event. Samples on other events will be created containing 548data from the AUX area. Optionally sample size may be specified, otherwise it 549defaults to 4KiB. 550 551--proc-map-timeout:: 552When processing pre-existing threads /proc/XXX/mmap, it may take a long time, 553because the file may be huge. A time out is needed in such cases. 554This option sets the time out limit. The default value is 500 ms. 555 556--switch-events:: 557Record context switch events i.e. events of type PERF_RECORD_SWITCH or 558PERF_RECORD_SWITCH_CPU_WIDE. In some cases (e.g. Intel PT, CoreSight or Arm SPE) 559switch events will be enabled automatically, which can be suppressed by 560by the option --no-switch-events. 561 562--vmlinux=PATH:: 563Specify vmlinux path which has debuginfo. 564(enabled when BPF prologue is on) 565 566--buildid-all:: 567Record build-id of all DSOs regardless whether it's actually hit or not. 568 569--buildid-mmap:: 570Legacy record build-id in map events option which is now the 571default. Behaves indentically to --no-buildid. Disable with 572--no-buildid-mmap. 573 574--aio[=n]:: 575Use <n> control blocks in asynchronous (Posix AIO) trace writing mode (default: 1, max: 4). 576Asynchronous mode is supported only when linking Perf tool with libc library 577providing implementation for Posix AIO API. 578 579--affinity=mode:: 580Set affinity mask of trace reading thread according to the policy defined by 'mode' value: 581 582 - node - thread affinity mask is set to NUMA node cpu mask of the processed mmap buffer 583 - cpu - thread affinity mask is set to cpu of the processed mmap buffer 584 585--mmap-flush=number:: 586 587Specify minimal number of bytes that is extracted from mmap data pages and 588processed for output. One can specify the number using B/K/M/G suffixes. 589 590The maximal allowed value is a quarter of the size of mmaped data pages. 591 592The default option value is 1 byte which means that every time that the output 593writing thread finds some new data in the mmaped buffer the data is extracted, 594possibly compressed (-z) and written to the output, perf.data or pipe. 595 596Larger data chunks are compressed more effectively in comparison to smaller 597chunks so extraction of larger chunks from the mmap data pages is preferable 598from the perspective of output size reduction. 599 600Also at some cases executing less output write syscalls with bigger data size 601can take less time than executing more output write syscalls with smaller data 602size thus lowering runtime profiling overhead. 603 604-z:: 605--compression-level[=n]:: 606Produce compressed trace using specified level n (default: 1 - fastest compression, 60722 - smallest trace) 608 609--all-kernel:: 610Configure all used events to run in kernel space. 611 612--all-user:: 613Configure all used events to run in user space. 614 615--kernel-callchains:: 616Collect callchains only from kernel space. I.e. this option sets 617perf_event_attr.exclude_callchain_user to 1. 618 619--user-callchains:: 620Collect callchains only from user space. I.e. this option sets 621perf_event_attr.exclude_callchain_kernel to 1. 622 623Don't use both --kernel-callchains and --user-callchains at the same time or no 624callchains will be collected. 625 626--timestamp-filename 627Append timestamp to output file name. 628 629--timestamp-boundary:: 630Record timestamp boundary (time of first/last samples). 631 632--switch-output[=mode]:: 633Generate multiple perf.data files, timestamp prefixed, switching to a new one 634based on 'mode' value: 635 636 - "signal" - when receiving a SIGUSR2 (default value) or 637 - <size> - when reaching the size threshold, size is expected to 638 be a number with appended unit character - B/K/M/G 639 - <time> - when reaching the time threshold, size is expected to 640 be a number with appended unit character - s/m/h/d 641 642 Note: the precision of the size threshold hugely depends 643 on your configuration - the number and size of your ring 644 buffers (-m). It is generally more precise for higher sizes 645 (like >5M), for lower values expect different sizes. 646 647A possible use case is to, given an external event, slice the perf.data file 648that gets then processed, possibly via a perf script, to decide if that 649particular perf.data snapshot should be kept or not. 650 651Implies --timestamp-filename, --no-buildid and --no-buildid-cache. 652The reason for the latter two is to reduce the data file switching 653overhead. You can still switch them on with: 654 655 --switch-output --no-no-buildid --no-no-buildid-cache 656 657--switch-output-event:: 658Events that will cause the switch of the perf.data file, auto-selecting 659--switch-output=signal, the results are similar as internally the side band 660thread will also send a SIGUSR2 to the main one. 661 662Uses the same syntax as --event, it will just not be recorded, serving only to 663switch the perf.data file as soon as the --switch-output event is processed by 664a separate sideband thread. 665 666This sideband thread is also used to other purposes, like processing the 667PERF_RECORD_BPF_EVENT records as they happen, asking the kernel for extra BPF 668information, etc. 669 670--switch-max-files=N:: 671 672When rotating perf.data with --switch-output, only keep N files. 673 674--dry-run:: 675Parse options then exit. --dry-run can be used to detect errors in cmdline 676options. 677 678'perf record --dry-run -e' can act as a BPF script compiler if llvm.dump-obj 679in config file is set to true. 680 681--synth=TYPE:: 682Collect and synthesize given type of events (comma separated). Note that 683this option controls the synthesis from the /proc filesystem which represent 684task status for pre-existing threads. 685 686Kernel (and some other) events are recorded regardless of the 687choice in this option. For example, --synth=no would have MMAP events for 688kernel and modules. 689 690Available types are: 691 692 - 'task' - synthesize FORK and COMM events for each task 693 - 'mmap' - synthesize MMAP events for each process (implies 'task') 694 - 'cgroup' - synthesize CGROUP events for each cgroup 695 - 'all' - synthesize all events (default) 696 - 'no' - do not synthesize any of the above events 697 698--tail-synthesize:: 699Instead of collecting non-sample events (for example, fork, comm, mmap) at 700the beginning of record, collect them during finalizing an output file. 701The collected non-sample events reflects the status of the system when 702record is finished. 703 704--overwrite:: 705Makes all events use an overwritable ring buffer. An overwritable ring 706buffer works like a flight recorder: when it gets full, the kernel will 707overwrite the oldest records, that thus will never make it to the 708perf.data file. 709 710When '--overwrite' and '--switch-output' are used perf records and drops 711events until it receives a signal, meaning that something unusual was 712detected that warrants taking a snapshot of the most current events, 713those fitting in the ring buffer at that moment. 714 715'overwrite' attribute can also be set or canceled for an event using 716config terms. For example: 'cycles/overwrite/' and 'instructions/no-overwrite/'. 717 718Implies --tail-synthesize. 719 720--kcore:: 721Make a copy of /proc/kcore and place it into a directory with the perf data file. 722 723--max-size=<size>:: 724Limit the sample data max size, <size> is expected to be a number with 725appended unit character - B/K/M/G 726 727--num-thread-synthesize:: 728 The number of threads to run when synthesizing events for existing processes. 729 By default, the number of threads equals 1. 730 731ifdef::HAVE_LIBPFM[] 732--pfm-events events:: 733Select a PMU event using libpfm4 syntax (see http://perfmon2.sf.net) 734including support for event filters. For example '--pfm-events 735inst_retired:any_p:u:c=1:i'. More than one event can be passed to the 736option using the comma separator. Hardware events and generic hardware 737events cannot be mixed together. The latter must be used with the -e 738option. The -e option and this one can be mixed and matched. Events 739can be grouped using the {} notation. 740endif::HAVE_LIBPFM[] 741 742--control=fifo:ctl-fifo[,ack-fifo]:: 743--control=fd:ctl-fd[,ack-fd]:: 744ctl-fifo / ack-fifo are opened and used as ctl-fd / ack-fd as follows. 745Listen on ctl-fd descriptor for command to control measurement. 746 747Available commands: 748 749 - 'enable' : enable events 750 - 'disable' : disable events 751 - 'enable name' : enable event 'name' 752 - 'disable name' : disable event 'name' 753 - 'snapshot' : AUX area tracing snapshot). 754 - 'stop' : stop perf record 755 - 'ping' : ping 756 - 'evlist [-v|-g|-F] : display all events 757 758 -F Show just the sample frequency used for each event. 759 -v Show all fields. 760 -g Show event group information. 761 762Measurements can be started with events disabled using --delay=-1 option. Optionally 763send control command completion ('ack\n') to ack-fd descriptor to synchronize with the 764controlling process. Example of bash shell script to enable and disable events during 765measurements: 766 767 #!/bin/bash 768 769 ctl_dir=/tmp/ 770 771 ctl_fifo=${ctl_dir}perf_ctl.fifo 772 test -p ${ctl_fifo} && unlink ${ctl_fifo} 773 mkfifo ${ctl_fifo} 774 exec {ctl_fd}<>${ctl_fifo} 775 776 ctl_ack_fifo=${ctl_dir}perf_ctl_ack.fifo 777 test -p ${ctl_ack_fifo} && unlink ${ctl_ack_fifo} 778 mkfifo ${ctl_ack_fifo} 779 exec {ctl_fd_ack}<>${ctl_ack_fifo} 780 781 perf record -D -1 -e cpu-cycles -a \ 782 --control fd:${ctl_fd},${ctl_fd_ack} \ 783 -- sleep 30 & 784 perf_pid=$! 785 786 sleep 5 && echo 'enable' >&${ctl_fd} && read -u ${ctl_fd_ack} e1 && echo "enabled(${e1})" 787 sleep 10 && echo 'disable' >&${ctl_fd} && read -u ${ctl_fd_ack} d1 && echo "disabled(${d1})" 788 789 exec {ctl_fd_ack}>&- 790 unlink ${ctl_ack_fifo} 791 792 exec {ctl_fd}>&- 793 unlink ${ctl_fifo} 794 795 wait -n ${perf_pid} 796 exit $? 797 798--threads=<spec>:: 799Write collected trace data into several data files using parallel threads. 800<spec> value can be user defined list of masks. Masks separated by colon 801define CPUs to be monitored by a thread and affinity mask of that thread 802is separated by slash: 803 804 <cpus mask 1>/<affinity mask 1>:<cpus mask 2>/<affinity mask 2>:... 805 806CPUs or affinity masks must not overlap with other corresponding masks. 807Invalid CPUs are ignored, but masks containing only invalid CPUs are not 808allowed. 809 810For example user specification like the following: 811 812 0,2-4/2-4:1,5-7/5-7 813 814specifies parallel threads layout that consists of two threads, 815the first thread monitors CPUs 0 and 2-4 with the affinity mask 2-4, 816the second monitors CPUs 1 and 5-7 with the affinity mask 5-7. 817 818<spec> value can also be a string meaning predefined parallel threads 819layout: 820 821 - cpu - create new data streaming thread for every monitored cpu 822 - core - create new thread to monitor CPUs grouped by a core 823 - package - create new thread to monitor CPUs grouped by a package 824 - numa - create new threed to monitor CPUs grouped by a NUMA domain 825 826Predefined layouts can be used on systems with large number of CPUs in 827order not to spawn multiple per-cpu streaming threads but still avoid LOST 828events in data directory files. Option specified with no or empty value 829defaults to CPU layout. Masks defined or provided by the option value are 830filtered through the mask provided by -C option. 831 832--debuginfod[=URLs]:: 833 Specify debuginfod URL to be used when cacheing perf.data binaries, 834 it follows the same syntax as the DEBUGINFOD_URLS variable, like: 835 836 http://192.168.122.174:8002 837 838 If the URLs is not specified, the value of DEBUGINFOD_URLS 839 system environment variable is used. 840 841--off-cpu:: 842 Enable off-cpu profiling with BPF. The BPF program will collect 843 task scheduling information with (user) stacktrace and save them 844 as sample data of a software event named "offcpu-time". The 845 sample period will have the time the task slept in nanoseconds. 846 847 Note that BPF can collect stack traces using frame pointer ("fp") 848 only, as of now. So the applications built without the frame 849 pointer might see bogus addresses. 850 851 off-cpu profiling consists two types of samples: direct samples, which 852 share the same behavior as regular samples, and the accumulated 853 samples, stored in BPF stack trace map, presented after all the regular 854 samples. 855 856--off-cpu-thresh:: 857 Once a task's off-cpu time reaches this threshold (in milliseconds), it 858 generates a direct off-cpu sample. The default is 500ms. 859 860--setup-filter=<action>:: 861 Prepare BPF filter to be used by regular users. The action should be 862 either "pin" or "unpin". The filter can be used after it's pinned. 863 864 865include::intel-hybrid.txt[] 866 867SEE ALSO 868-------- 869linkperf:perf-stat[1], linkperf:perf-list[1], linkperf:perf-intel-pt[1] 870