1PERF-REPORT(1) perf Manual PERF-REPORT(1)
2
6 perf-report - Read perf.data (created by perf record) and display the
7 profile
8
10 perf report [-i <file> | --input=file]
11
13 This command displays the performance counter profile information
14 recorded via perf record.
15
17 -i, --input=
18 Input file name. (default: perf.data unless stdin is a fifo)
19 -v, --verbose
21 Be more verbose. (show symbol address, etc)
22 -q, --quiet
24 Do not show any message. (Suppress -v)
25 -n, --show-nr-samples
27 Show the number of samples for each symbol
28 --show-cpu-utilization
30 Show sample percentage for different cpu modes.
31 -T, --threads
33 Show per-thread event counters. The input data file should be
34 recorded with -s option.
35 -c, --comms=
37 Only consider symbols in these comms. CSV that understands
38 file://filename entries. This option will affect the percentage of
39 the overhead column. See --percentage for more info.
40 --pid=
42 Only show events for given process ID (comma separated list).
43 --tid=
45 Only show events for given thread ID (comma separated list).
46 -d, --dsos=
48 Only consider symbols in these dsos. CSV that understands
49 file://filename entries. This option will affect the percentage of
50 the overhead column. See --percentage for more info.
51 -S, --symbols=
53 Only consider these symbols. CSV that understands file://filename
54 entries. This option will affect the percentage of the overhead
55 column. See --percentage for more info.
56 --symbol-filter=
58 Only show symbols that match (partially) with this filter.
59 -U, --hide-unresolved
61 Only display entries resolved to a symbol.
62 -s, --sort=
64 Sort histogram entries by given key(s) - multiple keys can be
65 specified in CSV format. Following sort keys are available: pid,
66 comm, dso, symbol, parent, cpu, socket, srcline, weight,
67 local_weight, cgroup_id.
68 Each key has following meaning:
70 • comm: command (name) of the task which can be read via
72 /proc/<pid>/comm
73 • pid: command and tid of the task
75 • dso: name of library or module executed at the time of sample
77 • dso_size: size of library or module executed at the time of
79 sample
80 • symbol: name of function executed at the time of sample
82 • symbol_size: size of function executed at the time of sample
84 • parent: name of function matched to the parent regex filter.
86 Unmatched entries are displayed as "[other]".
87 • cpu: cpu number the task ran at the time of sample
89 • socket: processor socket number the task ran at the time of
91 sample
92 • srcline: filename and line number executed at the time of
94 sample. The DWARF debugging info must be provided.
95 • srcfile: file name of the source file of the samples. Requires
97 dwarf information.
98 • weight: Event specific weight, e.g. memory latency or
100 transaction abort cost. This is the global weight.
101 • local_weight: Local weight version of the weight above.
103 • cgroup_id: ID derived from cgroup namespace device and inode
105 numbers.
106 • cgroup: cgroup pathname in the cgroupfs.
108 • transaction: Transaction abort flags.
110 • overhead: Overhead percentage of sample
112 • overhead_sys: Overhead percentage of sample running in system
114 mode
115 • overhead_us: Overhead percentage of sample running in user mode
117 • overhead_guest_sys: Overhead percentage of sample running in
119 system mode on guest machine
120 • overhead_guest_us: Overhead percentage of sample running in
122 user mode on guest machine
123 • sample: Number of sample
125 • period: Raw number of event count of sample
127 • time: Separate the samples by time stamp with the resolution
129 specified by --time-quantum (default 100ms). Specify with
130 overhead and before it.
131 • code_page_size: the code page size of sampled code address (ip)
133 • ins_lat: Instruction latency in core cycles. This is the global
135 instruction latency
136 • local_ins_lat: Local instruction latency version
138 • p_stage_cyc: On powerpc, this presents the number of cycles
140 spent in a pipeline stage. And currently supported only on
141 powerpc.
142 By default, comm, dso and symbol keys are used.
144 (i.e. --sort comm,dso,symbol)
145 If --branch-stack option is used, following sort keys are also
147 available:
148 • dso_from: name of library or module branched from
150 • dso_to: name of library or module branched to
152 • symbol_from: name of function branched from
154 • symbol_to: name of function branched to
156 • srcline_from: source file and line branched from
158 • srcline_to: source file and line branched to
160 • mispredict: "N" for predicted branch, "Y" for mispredicted
162 branch
163 • in_tx: branch in TSX transaction
165 • abort: TSX transaction abort.
167 • cycles: Cycles in basic block
169 And default sort keys are changed to comm, dso_from, symbol_from, dso_to
171 and symbol_to, see '--branch-stack'.
172 When the sort key symbol is specified, columns "IPC" and "IPC Coverage"
174 are enabled automatically. Column "IPC" reports the average IPC per function
175 and column "IPC coverage" reports the percentage of instructions with
176 sampled IPC in this function. IPC means Instruction Per Cycle. If it's low,
177 it indicates there may be a performance bottleneck when the function is
178 executed, such as a memory access bottleneck. If a function has high overhead
179 and low IPC, it's worth further analyzing it to optimize its performance.
180 If the --mem-mode option is used, the following sort keys are also available
182 (incompatible with --branch-stack):
183 symbol_daddr, dso_daddr, locked, tlb, mem, snoop, dcacheline, blocked.
184 • symbol_daddr: name of data symbol being executed on at the time
186 of sample
187 • dso_daddr: name of library or module containing the data being
189 executed on at the time of the sample
190 • locked: whether the bus was locked at the time of the sample
192 • tlb: type of tlb access for the data at the time of the sample
194 • mem: type of memory access for the data at the time of the
196 sample
197 • snoop: type of snoop (if any) for the data at the time of the
199 sample
200 • dcacheline: the cacheline the data address is on at the time of
202 the sample
203 • phys_daddr: physical address of data being executed on at the
205 time of sample
206 • data_page_size: the data page size of data being executed on at
208 the time of sample
209 • blocked: reason of blocked load access for the data at the time
211 of the sample
212 And the default sort keys are changed to local_weight, mem, sym, dso,
214 symbol_daddr, dso_daddr, snoop, tlb, locked, blocked, local_ins_lat,
215 see '--mem-mode'.
216 If the data file has tracepoint event(s), following (dynamic) sort keys
218 are also available:
219 trace, trace_fields, [<event>.]<field>[/raw]
220 • trace: pretty printed trace output in a single column
222 • trace_fields: fields in tracepoints in separate columns
224 • <field name>: optional event and field name for a specific
226 field
227 The last form consists of event and field names. If event name is
229 omitted, it searches all events for matching field name. The matched
230 field will be shown only for the event has the field. The event name
231 supports substring match so user doesn't need to specify full subsystem
232 and event name everytime. For example, 'sched:sched_switch' event can
233 be shortened to 'switch' as long as it's not ambiguous. Also event can
234 be specified by its index (starting from 1) preceded by the '%'.
235 So '%1' is the first event, '%2' is the second, and so on.
236 The field name can have '/raw' suffix which disables pretty printing
238 and shows raw field value like hex numbers. The --raw-trace option
239 has the same effect for all dynamic sort keys.
240 The default sort keys are changed to 'trace' if all events in the data
242 file are tracepoint.
243 -F, --fields=
245 Specify output field - multiple keys can be specified in CSV
246 format. Following fields are available: overhead, overhead_sys,
247 overhead_us, overhead_children, sample and period. Also it can
248 contain any sort key(s).
249 By default, every sort keys not specified in -F will be appended
251 automatically.
252 If the keys starts with a prefix '+', then it will append the specified
254 field(s) to the default field order. For example: perf report -F +period,sample.
255 -p, --parent=<regex>
257 A regex filter to identify parent. The parent is a caller of this
258 function and searched through the callchain, thus it requires
259 callchain information recorded. The pattern is in the extended
260 regex format and defaults to "^sys_|^do_page_fault", see --sort
261 parent.
262 -x, --exclude-other
264 Only display entries with parent-match.
265 -w, --column-widths=<width[,width...]>
267 Force each column width to the provided list, for large terminal
268 readability. 0 means no limit (default behavior).
269 -t, --field-separator=
271 Use a special separator character and don’t pad with spaces,
272 replacing all occurrences of this separator in symbol names (and
273 other output) with a . character, that thus it’s the only non
274 valid separator.
275 -D, --dump-raw-trace
277 Dump raw trace in ASCII.
278 --disable-order
280 Disable raw trace ordering.
281 -g,
283 --call-graph=<print_type,threshold[,print_limit],order,sort_key[,branch],value>
284 Display call chains using type, min percent threshold, print limit,
285 call order, sort key, optional branch and value. Note that ordering
286 is not fixed so any parameter can be given in an arbitrary order.
287 One exception is the print_limit which should be preceded by
288 threshold.
289 print_type can be either:
291 - flat: single column, linear exposure of call chains.
292 - graph: use a graph tree, displaying absolute overhead rates. (default)
293 - fractal: like graph, but displays relative rates. Each branch of
294 the tree is considered as a new profiled object.
295 - folded: call chains are displayed in a line, separated by semicolons
296 - none: disable call chain display.
297 threshold is a percentage value which specifies a minimum percent to be
299 included in the output call graph. Default is 0.5 (%).
300 print_limit is only applied when stdio interface is used. It's to limit
302 number of call graph entries in a single hist entry. Note that it needs
303 to be given after threshold (but not necessarily consecutive).
304 Default is 0 (unlimited).
305 order can be either:
307 - callee: callee based call graph.
308 - caller: inverted caller based call graph.
309 Default is 'caller' when --children is used, otherwise 'callee'.
310 sort_key can be:
312 - function: compare on functions (default)
313 - address: compare on individual code addresses
314 - srcline: compare on source filename and line number
315 branch can be:
317 - branch: include last branch information in callgraph when available.
318 Usually more convenient to use --branch-history for this.
319 value can be:
321 - percent: display overhead percent (default)
322 - period: display event period
323 - count: display event count
324 --children
326 Accumulate callchain of children to parent entry so that then can
327 show up in the output. The output will have a new "Children" column
328 and will be sorted on the data. It requires callchains are
329 recorded. See the ‘overhead calculation’ section for more details.
330 Enabled by default, disable with --no-children.
331 --max-stack
333 Set the stack depth limit when parsing the callchain, anything
334 beyond the specified depth will be ignored. This is a trade-off
335 between information loss and faster processing especially for
336 workloads that can have a very long callchain stack. Note that when
337 using the --itrace option the synthesized callchain size will
338 override this value if the synthesized callchain size is bigger.
339 Default: 127
341 -G, --inverted
343 alias for inverted caller based call graph.
344 --ignore-callees=<regex>
346 Ignore callees of the function(s) matching the given regex. This
347 has the effect of collecting the callers of each such function into
348 one place in the call-graph tree.
349 --pretty=<key>
351 Pretty printing style. key: normal, raw
352 --stdio
354 Use the stdio interface.
355 --stdio-color
357 always, never or auto, allowing configuring color output via the
358 command line, in addition to via "color.ui" .perfconfig. Use
359 --stdio-color always to generate color even when redirecting to a
360 pipe or file. Using just --stdio-color is equivalent to using
361 always.
362 --tui
364 Use the TUI interface, that is integrated with annotate and allows
365 zooming into DSOs or threads, among other features. Use of --tui
366 requires a tty, if one is not present, as when piping to other
367 commands, the stdio interface is used.
368 --gtk
370 Use the GTK2 interface.
371 -k, --vmlinux=<file>
373 vmlinux pathname
374 --ignore-vmlinux
376 Ignore vmlinux files.
377 --kallsyms=<file>
379 kallsyms pathname
380 -m, --modules
382 Load module symbols. WARNING: This should only be used with -k and
383 a LIVE kernel.
384 -f, --force
386 Don’t do ownership validation.
387 --symfs=<directory>
389 Look for files with symbols relative to this directory.
390 -C, --cpu
392 Only report samples for the list of CPUs provided. Multiple CPUs
393 can be provided as a comma-separated list with no space: 0,1.
394 Ranges of CPUs are specified with -: 0-2. Default is to report
395 samples on all CPUs.
396 -M, --disassembler-style=
398 Set disassembler style for objdump.
399 --source
401 Interleave source code with assembly code. Enabled by default,
402 disable with --no-source.
403 --asm-raw
405 Show raw instruction encoding of assembly instructions.
406 --show-total-period
408 Show a column with the sum of periods.
409 -I, --show-info
411 Display extended information about the perf.data file. This adds
412 information which may be very large and thus may clutter the
413 display. It currently includes: cpu and numa topology of the host
414 system.
415 -b, --branch-stack
417 Use the addresses of sampled taken branches instead of the
418 instruction address to build the histograms. To generate meaningful
419 output, the perf.data file must have been obtained using perf
420 record -b or perf record --branch-filter xxx where xxx is a branch
421 filter option. perf report is able to auto-detect whether a
422 perf.data file contains branch stacks and it will automatically
423 switch to the branch view mode, unless --no-branch-stack is used.
424 --branch-history
426 Add the addresses of sampled taken branches to the callstack. This
427 allows to examine the path the program took to each sample. The
428 data collection must have used -b (or -j) and -g.
429 --objdump=<path>
431 Path to objdump binary.
432 --prefix=PREFIX, --prefix-strip=N
434 Remove first N entries from source file path names in executables
435 and add PREFIX. This allows to display source code compiled on
436 systems with different file system layout.
437 --group
439 Show event group information together. It forces group output also
440 if there are no groups defined in data file.
441 --group-sort-idx
443 Sort the output by the event at the index n in group. If n is
444 invalid, sort by the first event. It can support multiple groups
445 with different amount of events. WARNING: This should be used on
446 grouped events.
447 --demangle
449 Demangle symbol names to human readable form. It’s enabled by
450 default, disable with --no-demangle.
451 --demangle-kernel
453 Demangle kernel symbol names to human readable form (for C++
454 kernels).
455 --mem-mode
457 Use the data addresses of samples in addition to instruction
458 addresses to build the histograms. To generate meaningful output,
459 the perf.data file must have been obtained using perf record -d -W
460 and using a special event -e cpu/mem-loads/p or -e
461 cpu/mem-stores/p. See perf mem for simpler access.
462 --percent-limit
464 Do not show entries which have an overhead under that percent.
465 (Default: 0). Note that this option also sets the percent limit
466 (threshold) of callchains. However the default value of callchain
467 threshold is different than the default value of hist entries.
468 Please see the --call-graph option for details.
469 --percentage
471 Determine how to display the overhead percentage of filtered
472 entries. Filters can be applied by --comms, --dsos and/or --symbols
473 options and Zoom operations on the TUI (thread, dso, etc).
474 "relative" means it's relative to filtered entries only so that the
476 sum of shown entries will be always 100%. "absolute" means it retains
477 the original value before and after the filter is applied.
478 --header
480 Show header information in the perf.data file. This includes
481 various information like hostname, OS and perf version, cpu/mem
482 info, perf command line, event list and so on. Currently only
483 --stdio output supports this feature.
484 --header-only
486 Show only perf.data header (forces --stdio).
487 --time
489 Only analyze samples within given time window: <start>,<stop>.
490 Times have the format seconds.nanoseconds. If start is not given
491 (i.e. time string is ,x.y) then analysis starts at the beginning of
492 the file. If stop time is not given (i.e. time string is x.y,) then
493 analysis goes to end of file. Multiple ranges can be separated by
494 spaces, which requires the argument to be quoted e.g. --time
495 "1234.567,1234.789 1235,"
496 Also support time percent with multiple time ranges. Time string is
498 'a%/n,b%/m,...' or 'a%-b%,c%-%d,...'.
499 For example:
501 Select the second 10% time slice:
502 perf report --time 10%/2
504 Select from 0% to 10% time slice:
506 perf report --time 0%-10%
508 Select the first and second 10% time slices:
510 perf report --time 10%/1,10%/2
512 Select from 0% to 10% and 30% to 40% slices:
514 perf report --time 0%-10%,30%-40%
516 --switch-on EVENT_NAME
518 Only consider events after this event is found.
519 This may be interesting to measure a workload only after some initialization
521 phase is over, i.e. insert a perf probe at that point and then using this
522 option with that probe.
523 --switch-off EVENT_NAME
525 Stop considering events after this event is found.
526 --show-on-off-events
528 Show the --switch-on/off events too. This has no effect in perf
529 report now but probably we’ll make the default not to show the
530 switch-on/off events on the --group mode and if there is only one
531 event besides the off/on ones, go straight to the histogram
532 browser, just like perf report with no events explicitly specified
533 does.
534 --itrace
536 Options for decoding instruction tracing data. The options are:
537 i synthesize instructions events
539 b synthesize branches events (branch misses for Arm SPE)
540 c synthesize branches events (calls only)
541 r synthesize branches events (returns only)
542 x synthesize transactions events
543 w synthesize ptwrite events
544 p synthesize power events (incl. PSB events for Intel PT)
545 o synthesize other events recorded due to the use
546 of aux-output (refer to perf record)
547 e synthesize error events
548 d create a debug log
549 f synthesize first level cache events
550 m synthesize last level cache events
551 M synthesize memory events
552 t synthesize TLB events
553 a synthesize remote access events
554 g synthesize a call chain (use with i or x)
555 G synthesize a call chain on existing event records
556 l synthesize last branch entries (use with i or x)
557 L synthesize last branch entries on existing event records
558 s skip initial number of events
559 q quicker (less detailed) decoding
560 Z prefer to ignore timestamps (so-called "timeless" decoding)
561 The default is all events i.e. the same as --itrace=ibxwpe,
563 except for perf script where it is --itrace=ce
564 In addition, the period (default 100000, except for perf script where it is 1)
566 for instructions events can be specified in units of:
567 i instructions
569 t ticks
570 ms milliseconds
571 us microseconds
572 ns nanoseconds (default)
573 Also the call chain size (default 16, max. 1024) for instructions or
575 transactions events can be specified.
576 Also the number of last branch entries (default 64, max. 1024) for
578 instructions or transactions events can be specified.
579 Similar to options g and l, size may also be specified for options G and L.
581 On x86, note that G and L work poorly when data has been recorded with
582 large PEBS. Refer linkperf:perf-intel-pt[1] man page for details.
583 It is also possible to skip events generated (instructions, branches, transactions,
585 ptwrite, power) at the beginning. This is useful to ignore initialization code.
586 --itrace=i0nss1000000
588 skips the first million instructions.
590 The 'e' option may be followed by flags which affect what errors will or
592 will not be reported. Each flag must be preceded by either '+' or '-'.
593 The flags are:
594 o overflow
595 l trace data lost
596 If supported, the 'd' option may be followed by flags which affect what
598 debug messages will or will not be logged. Each flag must be preceded
599 by either '+' or '-'. The flags are:
600 a all perf events
601 If supported, the 'q' option may be repeated to increase the effect.
603 To disable decoding entirely, use --no-itrace.
605 --full-source-path
607 Show the full path for source files for srcline output.
608 --show-ref-call-graph
610 When multiple events are sampled, it may not be needed to collect
611 callgraphs for all of them. The sample sites are usually nearby,
612 and it’s enough to collect the callgraphs on a reference event. So
613 user can use "call-graph=no" event modifier to disable callgraph
614 for other events to reduce the overhead. However, perf report
615 cannot show callgraphs for the event which disable the callgraph.
616 This option extends the perf report to show reference callgraphs,
617 which collected by reference event, in no callgraph event.
618 --stitch-lbr
620 Show callgraph with stitched LBRs, which may have more complete
621 callgraph. The perf.data file must have been obtained using perf
622 record --call-graph lbr. Disabled by default. In common cases with
623 call stack overflows, it can recreate better call stacks than the
624 default lbr call stack output. But this approach is not full proof.
625 There can be cases where it creates incorrect call stacks from
626 incorrect matches. The known limitations include exception handing
627 such as setjmp/longjmp will have calls/returns not match.
628 --socket-filter
630 Only report the samples on the processor socket that match with
631 this filter
632 --samples=N
634 Save N individual samples for each histogram entry to show context
635 in perf report tui browser.
636 --raw-trace
638 When displaying traceevent output, do not use print fmt or plugins.
639 --hierarchy
641 Enable hierarchical output.
642 --inline
644 If a callgraph address belongs to an inlined function, the inline
645 stack will be printed. Each entry is function name or file/line.
646 Enabled by default, disable with --no-inline.
647 --mmaps
649 Show --tasks output plus mmap information in a format similar to
650 /proc/<PID>/maps.
651 Please note that not all mmaps are stored, options affecting which ones
653 are include 'perf record --data', for instance.
654 --ns
656 Show time stamps in nanoseconds.
657 --stats
659 Display overall events statistics without any further processing.
660 (like the one at the end of the perf report -D command)
661 --tasks
663 Display monitored tasks stored in perf data. Displaying
664 pid/tid/ppid plus the command string aligned to distinguish parent
665 and child tasks.
666 --percent-type
668 Set annotation percent type from following choices: global-period,
669 local-period, global-hits, local-hits
670 The local/global keywords set if the percentage is computed
672 in the scope of the function (local) or the whole data (global).
673 The period/hits keywords set the base the percentage is computed
674 on - the samples period or the number of samples (hits).
675 --time-quantum
677 Configure time quantum for time sort key. Default 100ms. Accepts s,
678 us, ms, ns units.
679 --total-cycles
681 When --total-cycles is specified, it supports sorting for all
682 blocks by Sampled Cycles%. This is useful to concentrate on the
683 globally hottest blocks. In output, there are some new columns:
684 'Sampled Cycles%' - block sampled cycles aggregation / total sampled cycles
686 'Sampled Cycles' - block sampled cycles aggregation
687 'Avg Cycles%' - block average sampled cycles / sum of total block average
688 sampled cycles
689 'Avg Cycles' - block average sampled cycles
690 --skip-empty
692 Do not print 0 results in the --stat output.
693
695 The overhead can be shown in two columns as Children and Self when perf
696 collects callchains. The self overhead is simply calculated by adding
697 all period values of the entry - usually a function (symbol). This is
698 the value that perf shows traditionally and sum of all the self
699 overhead values should be 100%.
700 The children overhead is calculated by adding all period values of the
702 child functions so that it can show the total overhead of the higher
703 level functions even if they don’t directly execute much. Children here
704 means functions that are called from another (parent) function.
705 It might be confusing that the sum of all the children overhead values
707 exceeds 100% since each of them is already an accumulation of self
708 overhead of its child functions. But with this enabled, users can find
709 which function has the most overhead even if samples are spread over
710 the children.
711 Consider the following example; there are three functions like below.
713 .ft C
716 void foo(void) {
717 /* do something */
718 }
719 void bar(void) {
721 /* do something */
722 foo();
723 }
724 int main(void) {
726 bar()
727 return 0;
728 }
729 .ft
730 In this case foo is a child of bar, and bar is an immediate child of
733 main so foo also is a child of main. In other words, main is a parent
734 of foo and bar, and bar is a parent of foo.
735 Suppose all samples are recorded in foo and bar only. When it’s
737 recorded with callchains the output will show something like below in
738 the usual (self-overhead-only) output of perf report:
739 .ft C
742 Overhead Symbol
743 ........ .....................
744 60.00% foo
745 |
746 --- foo
747 bar
748 main
749 __libc_start_main
750 40.00% bar
752 |
753 --- bar
754 main
755 __libc_start_main
756 .ft
757 When the --children option is enabled, the self overhead values of
760 child functions (i.e. foo and bar) are added to the parents to
761 calculate the children overhead. In this case the report could be
762 displayed as:
763 .ft C
766 Children Self Symbol
767 ........ ........ ....................
768 100.00% 0.00% __libc_start_main
769 |
770 --- __libc_start_main
771 100.00% 0.00% main
773 |
774 --- main
775 __libc_start_main
776 100.00% 40.00% bar
778 |
779 --- bar
780 main
781 __libc_start_main
782 60.00% 60.00% foo
784 |
785 --- foo
786 bar
787 main
788 __libc_start_main
789 .ft
790 In the above output, the self overhead of foo (60%) was add to the
793 children overhead of bar, main and __libc_start_main. Likewise, the
794 self overhead of bar (40%) was added to the children overhead of main
795 and \_\_libc_start_main.
796 So \_\_libc_start_main and main are shown first since they have same
798 (100%) children overhead (even though they have zero self overhead) and
799 they are the parents of foo and bar.
800 Since v3.16 the children overhead is shown by default and the output is
802 sorted by its values. The children overhead is disabled by specifying
803 --no-children option on the command line or by adding report.children =
804 false or top.children = false in the perf config file.
805
807 perf-stat(1), perf-annotate(1), perf-record(1), perf-intel-pt(1)
808perf 11/22/2021 PERF-REPORT(1)