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 By default, comm, dso and symbol keys are used.
140 (i.e. --sort comm,dso,symbol)
141 If --branch-stack option is used, following sort keys are also
143 available:
144 • dso_from: name of library or module branched from
146 • dso_to: name of library or module branched to
148 • symbol_from: name of function branched from
150 • symbol_to: name of function branched to
152 • srcline_from: source file and line branched from
154 • srcline_to: source file and line branched to
156 • mispredict: "N" for predicted branch, "Y" for mispredicted
158 branch
159 • in_tx: branch in TSX transaction
161 • abort: TSX transaction abort.
163 • cycles: Cycles in basic block
165 And default sort keys are changed to comm, dso_from, symbol_from, dso_to
167 and symbol_to, see '--branch-stack'.
168 When the sort key symbol is specified, columns "IPC" and "IPC Coverage"
170 are enabled automatically. Column "IPC" reports the average IPC per function
171 and column "IPC coverage" reports the percentage of instructions with
172 sampled IPC in this function. IPC means Instruction Per Cycle. If it's low,
173 it indicates there may be a performance bottleneck when the function is
174 executed, such as a memory access bottleneck. If a function has high overhead
175 and low IPC, it's worth further analyzing it to optimize its performance.
176 If the --mem-mode option is used, the following sort keys are also available
178 (incompatible with --branch-stack):
179 symbol_daddr, dso_daddr, locked, tlb, mem, snoop, dcacheline, blocked.
180 • symbol_daddr: name of data symbol being executed on at the time
182 of sample
183 • dso_daddr: name of library or module containing the data being
185 executed on at the time of the sample
186 • locked: whether the bus was locked at the time of the sample
188 • tlb: type of tlb access for the data at the time of the sample
190 • mem: type of memory access for the data at the time of the
192 sample
193 • snoop: type of snoop (if any) for the data at the time of the
195 sample
196 • dcacheline: the cacheline the data address is on at the time of
198 the sample
199 • phys_daddr: physical address of data being executed on at the
201 time of sample
202 • data_page_size: the data page size of data being executed on at
204 the time of sample
205 • blocked: reason of blocked load access for the data at the time
207 of the sample
208 And the default sort keys are changed to local_weight, mem, sym, dso,
210 symbol_daddr, dso_daddr, snoop, tlb, locked, blocked, local_ins_lat,
211 see '--mem-mode'.
212 If the data file has tracepoint event(s), following (dynamic) sort keys
214 are also available:
215 trace, trace_fields, [<event>.]<field>[/raw]
216 • trace: pretty printed trace output in a single column
218 • trace_fields: fields in tracepoints in separate columns
220 • <field name>: optional event and field name for a specific
222 field
223 The last form consists of event and field names. If event name is
225 omitted, it searches all events for matching field name. The matched
226 field will be shown only for the event has the field. The event name
227 supports substring match so user doesn't need to specify full subsystem
228 and event name everytime. For example, 'sched:sched_switch' event can
229 be shortened to 'switch' as long as it's not ambiguous. Also event can
230 be specified by its index (starting from 1) preceded by the '%'.
231 So '%1' is the first event, '%2' is the second, and so on.
232 The field name can have '/raw' suffix which disables pretty printing
234 and shows raw field value like hex numbers. The --raw-trace option
235 has the same effect for all dynamic sort keys.
236 The default sort keys are changed to 'trace' if all events in the data
238 file are tracepoint.
239 -F, --fields=
241 Specify output field - multiple keys can be specified in CSV
242 format. Following fields are available: overhead, overhead_sys,
243 overhead_us, overhead_children, sample and period. Also it can
244 contain any sort key(s).
245 By default, every sort keys not specified in -F will be appended
247 automatically.
248 If the keys starts with a prefix '+', then it will append the specified
250 field(s) to the default field order. For example: perf report -F +period,sample.
251 -p, --parent=<regex>
253 A regex filter to identify parent. The parent is a caller of this
254 function and searched through the callchain, thus it requires
255 callchain information recorded. The pattern is in the extended
256 regex format and defaults to "^sys_|^do_page_fault", see --sort
257 parent.
258 -x, --exclude-other
260 Only display entries with parent-match.
261 -w, --column-widths=<width[,width...]>
263 Force each column width to the provided list, for large terminal
264 readability. 0 means no limit (default behavior).
265 -t, --field-separator=
267 Use a special separator character and don’t pad with spaces,
268 replacing all occurrences of this separator in symbol names (and
269 other output) with a . character, that thus it’s the only non
270 valid separator.
271 -D, --dump-raw-trace
273 Dump raw trace in ASCII.
274 -g,
276 --call-graph=<print_type,threshold[,print_limit],order,sort_key[,branch],value>
277 Display call chains using type, min percent threshold, print limit,
278 call order, sort key, optional branch and value. Note that ordering
279 is not fixed so any parameter can be given in an arbitrary order.
280 One exception is the print_limit which should be preceded by
281 threshold.
282 print_type can be either:
284 - flat: single column, linear exposure of call chains.
285 - graph: use a graph tree, displaying absolute overhead rates. (default)
286 - fractal: like graph, but displays relative rates. Each branch of
287 the tree is considered as a new profiled object.
288 - folded: call chains are displayed in a line, separated by semicolons
289 - none: disable call chain display.
290 threshold is a percentage value which specifies a minimum percent to be
292 included in the output call graph. Default is 0.5 (%).
293 print_limit is only applied when stdio interface is used. It's to limit
295 number of call graph entries in a single hist entry. Note that it needs
296 to be given after threshold (but not necessarily consecutive).
297 Default is 0 (unlimited).
298 order can be either:
300 - callee: callee based call graph.
301 - caller: inverted caller based call graph.
302 Default is 'caller' when --children is used, otherwise 'callee'.
303 sort_key can be:
305 - function: compare on functions (default)
306 - address: compare on individual code addresses
307 - srcline: compare on source filename and line number
308 branch can be:
310 - branch: include last branch information in callgraph when available.
311 Usually more convenient to use --branch-history for this.
312 value can be:
314 - percent: display overhead percent (default)
315 - period: display event period
316 - count: display event count
317 --children
319 Accumulate callchain of children to parent entry so that then can
320 show up in the output. The output will have a new "Children" column
321 and will be sorted on the data. It requires callchains are
322 recorded. See the ‘overhead calculation’ section for more details.
323 Enabled by default, disable with --no-children.
324 --max-stack
326 Set the stack depth limit when parsing the callchain, anything
327 beyond the specified depth will be ignored. This is a trade-off
328 between information loss and faster processing especially for
329 workloads that can have a very long callchain stack. Note that when
330 using the --itrace option the synthesized callchain size will
331 override this value if the synthesized callchain size is bigger.
332 Default: 127
334 -G, --inverted
336 alias for inverted caller based call graph.
337 --ignore-callees=<regex>
339 Ignore callees of the function(s) matching the given regex. This
340 has the effect of collecting the callers of each such function into
341 one place in the call-graph tree.
342 --pretty=<key>
344 Pretty printing style. key: normal, raw
345 --stdio
347 Use the stdio interface.
348 --stdio-color
350 always, never or auto, allowing configuring color output via the
351 command line, in addition to via "color.ui" .perfconfig. Use
352 --stdio-color always to generate color even when redirecting to a
353 pipe or file. Using just --stdio-color is equivalent to using
354 always.
355 --tui
357 Use the TUI interface, that is integrated with annotate and allows
358 zooming into DSOs or threads, among other features. Use of --tui
359 requires a tty, if one is not present, as when piping to other
360 commands, the stdio interface is used.
361 --gtk
363 Use the GTK2 interface.
364 -k, --vmlinux=<file>
366 vmlinux pathname
367 --ignore-vmlinux
369 Ignore vmlinux files.
370 --kallsyms=<file>
372 kallsyms pathname
373 -m, --modules
375 Load module symbols. WARNING: This should only be used with -k and
376 a LIVE kernel.
377 -f, --force
379 Don’t do ownership validation.
380 --symfs=<directory>
382 Look for files with symbols relative to this directory.
383 -C, --cpu
385 Only report samples for the list of CPUs provided. Multiple CPUs
386 can be provided as a comma-separated list with no space: 0,1.
387 Ranges of CPUs are specified with -: 0-2. Default is to report
388 samples on all CPUs.
389 -M, --disassembler-style=
391 Set disassembler style for objdump.
392 --source
394 Interleave source code with assembly code. Enabled by default,
395 disable with --no-source.
396 --asm-raw
398 Show raw instruction encoding of assembly instructions.
399 --show-total-period
401 Show a column with the sum of periods.
402 -I, --show-info
404 Display extended information about the perf.data file. This adds
405 information which may be very large and thus may clutter the
406 display. It currently includes: cpu and numa topology of the host
407 system.
408 -b, --branch-stack
410 Use the addresses of sampled taken branches instead of the
411 instruction address to build the histograms. To generate meaningful
412 output, the perf.data file must have been obtained using perf
413 record -b or perf record --branch-filter xxx where xxx is a branch
414 filter option. perf report is able to auto-detect whether a
415 perf.data file contains branch stacks and it will automatically
416 switch to the branch view mode, unless --no-branch-stack is used.
417 --branch-history
419 Add the addresses of sampled taken branches to the callstack. This
420 allows to examine the path the program took to each sample. The
421 data collection must have used -b (or -j) and -g.
422 --objdump=<path>
424 Path to objdump binary.
425 --prefix=PREFIX, --prefix-strip=N
427 Remove first N entries from source file path names in executables
428 and add PREFIX. This allows to display source code compiled on
429 systems with different file system layout.
430 --group
432 Show event group information together. It forces group output also
433 if there are no groups defined in data file.
434 --group-sort-idx
436 Sort the output by the event at the index n in group. If n is
437 invalid, sort by the first event. It can support multiple groups
438 with different amount of events. WARNING: This should be used on
439 grouped events.
440 --demangle
442 Demangle symbol names to human readable form. It’s enabled by
443 default, disable with --no-demangle.
444 --demangle-kernel
446 Demangle kernel symbol names to human readable form (for C++
447 kernels).
448 --mem-mode
450 Use the data addresses of samples in addition to instruction
451 addresses to build the histograms. To generate meaningful output,
452 the perf.data file must have been obtained using perf record -d -W
453 and using a special event -e cpu/mem-loads/p or -e
454 cpu/mem-stores/p. See perf mem for simpler access.
455 --percent-limit
457 Do not show entries which have an overhead under that percent.
458 (Default: 0). Note that this option also sets the percent limit
459 (threshold) of callchains. However the default value of callchain
460 threshold is different than the default value of hist entries.
461 Please see the --call-graph option for details.
462 --percentage
464 Determine how to display the overhead percentage of filtered
465 entries. Filters can be applied by --comms, --dsos and/or --symbols
466 options and Zoom operations on the TUI (thread, dso, etc).
467 "relative" means it's relative to filtered entries only so that the
469 sum of shown entries will be always 100%. "absolute" means it retains
470 the original value before and after the filter is applied.
471 --header
473 Show header information in the perf.data file. This includes
474 various information like hostname, OS and perf version, cpu/mem
475 info, perf command line, event list and so on. Currently only
476 --stdio output supports this feature.
477 --header-only
479 Show only perf.data header (forces --stdio).
480 --time
482 Only analyze samples within given time window: <start>,<stop>.
483 Times have the format seconds.nanoseconds. If start is not given
484 (i.e. time string is ,x.y) then analysis starts at the beginning of
485 the file. If stop time is not given (i.e. time string is x.y,) then
486 analysis goes to end of file. Multiple ranges can be separated by
487 spaces, which requires the argument to be quoted e.g. --time
488 "1234.567,1234.789 1235,"
489 Also support time percent with multiple time ranges. Time string is
491 'a%/n,b%/m,...' or 'a%-b%,c%-%d,...'.
492 For example:
494 Select the second 10% time slice:
495 perf report --time 10%/2
497 Select from 0% to 10% time slice:
499 perf report --time 0%-10%
501 Select the first and second 10% time slices:
503 perf report --time 10%/1,10%/2
505 Select from 0% to 10% and 30% to 40% slices:
507 perf report --time 0%-10%,30%-40%
509 --switch-on EVENT_NAME
511 Only consider events after this event is found.
512 This may be interesting to measure a workload only after some initialization
514 phase is over, i.e. insert a perf probe at that point and then using this
515 option with that probe.
516 --switch-off EVENT_NAME
518 Stop considering events after this event is found.
519 --show-on-off-events
521 Show the --switch-on/off events too. This has no effect in perf
522 report now but probably we’ll make the default not to show the
523 switch-on/off events on the --group mode and if there is only one
524 event besides the off/on ones, go straight to the histogram
525 browser, just like perf report with no events explicitely specified
526 does.
527 --itrace
529 Options for decoding instruction tracing data. The options are:
530 i synthesize instructions events
532 b synthesize branches events (branch misses for Arm SPE)
533 c synthesize branches events (calls only)
534 r synthesize branches events (returns only)
535 x synthesize transactions events
536 w synthesize ptwrite events
537 p synthesize power events (incl. PSB events for Intel PT)
538 o synthesize other events recorded due to the use
539 of aux-output (refer to perf record)
540 e synthesize error events
541 d create a debug log
542 f synthesize first level cache events
543 m synthesize last level cache events
544 M synthesize memory events
545 t synthesize TLB events
546 a synthesize remote access events
547 g synthesize a call chain (use with i or x)
548 G synthesize a call chain on existing event records
549 l synthesize last branch entries (use with i or x)
550 L synthesize last branch entries on existing event records
551 s skip initial number of events
552 q quicker (less detailed) decoding
553 The default is all events i.e. the same as --itrace=ibxwpe,
555 except for perf script where it is --itrace=ce
556 In addition, the period (default 100000, except for perf script where it is 1)
558 for instructions events can be specified in units of:
559 i instructions
561 t ticks
562 ms milliseconds
563 us microseconds
564 ns nanoseconds (default)
565 Also the call chain size (default 16, max. 1024) for instructions or
567 transactions events can be specified.
568 Also the number of last branch entries (default 64, max. 1024) for
570 instructions or transactions events can be specified.
571 Similar to options g and l, size may also be specified for options G and L.
573 On x86, note that G and L work poorly when data has been recorded with
574 large PEBS. Refer linkperf:perf-intel-pt[1] man page for details.
575 It is also possible to skip events generated (instructions, branches, transactions,
577 ptwrite, power) at the beginning. This is useful to ignore initialization code.
578 --itrace=i0nss1000000
580 skips the first million instructions.
582 The 'e' option may be followed by flags which affect what errors will or
584 will not be reported. Each flag must be preceded by either '+' or '-'.
585 The flags are:
586 o overflow
587 l trace data lost
588 If supported, the 'd' option may be followed by flags which affect what
590 debug messages will or will not be logged. Each flag must be preceded
591 by either '+' or '-'. The flags are:
592 a all perf events
593 If supported, the 'q' option may be repeated to increase the effect.
595 To disable decoding entirely, use --no-itrace.
597 --full-source-path
599 Show the full path for source files for srcline output.
600 --show-ref-call-graph
602 When multiple events are sampled, it may not be needed to collect
603 callgraphs for all of them. The sample sites are usually nearby,
604 and it’s enough to collect the callgraphs on a reference event. So
605 user can use "call-graph=no" event modifier to disable callgraph
606 for other events to reduce the overhead. However, perf report
607 cannot show callgraphs for the event which disable the callgraph.
608 This option extends the perf report to show reference callgraphs,
609 which collected by reference event, in no callgraph event.
610 --stitch-lbr
612 Show callgraph with stitched LBRs, which may have more complete
613 callgraph. The perf.data file must have been obtained using perf
614 record --call-graph lbr. Disabled by default. In common cases with
615 call stack overflows, it can recreate better call stacks than the
616 default lbr call stack output. But this approach is not full proof.
617 There can be cases where it creates incorrect call stacks from
618 incorrect matches. The known limitations include exception handing
619 such as setjmp/longjmp will have calls/returns not match.
620 --socket-filter
622 Only report the samples on the processor socket that match with
623 this filter
624 --samples=N
626 Save N individual samples for each histogram entry to show context
627 in perf report tui browser.
628 --raw-trace
630 When displaying traceevent output, do not use print fmt or plugins.
631 --hierarchy
633 Enable hierarchical output.
634 --inline
636 If a callgraph address belongs to an inlined function, the inline
637 stack will be printed. Each entry is function name or file/line.
638 Enabled by default, disable with --no-inline.
639 --mmaps
641 Show --tasks output plus mmap information in a format similar to
642 /proc/<PID>/maps.
643 Please note that not all mmaps are stored, options affecting which ones
645 are include 'perf record --data', for instance.
646 --ns
648 Show time stamps in nanoseconds.
649 --stats
651 Display overall events statistics without any further processing.
652 (like the one at the end of the perf report -D command)
653 --tasks
655 Display monitored tasks stored in perf data. Displaying
656 pid/tid/ppid plus the command string aligned to distinguish parent
657 and child tasks.
658 --percent-type
660 Set annotation percent type from following choices: global-period,
661 local-period, global-hits, local-hits
662 The local/global keywords set if the percentage is computed
664 in the scope of the function (local) or the whole data (global).
665 The period/hits keywords set the base the percentage is computed
666 on - the samples period or the number of samples (hits).
667 --time-quantum
669 Configure time quantum for time sort key. Default 100ms. Accepts s,
670 us, ms, ns units.
671 --total-cycles
673 When --total-cycles is specified, it supports sorting for all
674 blocks by Sampled Cycles%. This is useful to concentrate on the
675 globally hottest blocks. In output, there are some new columns:
676 'Sampled Cycles%' - block sampled cycles aggregation / total sampled cycles
678 'Sampled Cycles' - block sampled cycles aggregation
679 'Avg Cycles%' - block average sampled cycles / sum of total block average
680 sampled cycles
681 'Avg Cycles' - block average sampled cycles
682
684 The overhead can be shown in two columns as Children and Self when perf
685 collects callchains. The self overhead is simply calculated by adding
686 all period values of the entry - usually a function (symbol). This is
687 the value that perf shows traditionally and sum of all the self
688 overhead values should be 100%.
689 The children overhead is calculated by adding all period values of the
691 child functions so that it can show the total overhead of the higher
692 level functions even if they don’t directly execute much. Children here
693 means functions that are called from another (parent) function.
694 It might be confusing that the sum of all the children overhead values
696 exceeds 100% since each of them is already an accumulation of self
697 overhead of its child functions. But with this enabled, users can find
698 which function has the most overhead even if samples are spread over
699 the children.
700 Consider the following example; there are three functions like below.
702 .ft C
705 void foo(void) {
706 /* do something */
707 }
708 void bar(void) {
710 /* do something */
711 foo();
712 }
713 int main(void) {
715 bar()
716 return 0;
717 }
718 .ft
719 In this case foo is a child of bar, and bar is an immediate child of
722 main so foo also is a child of main. In other words, main is a parent
723 of foo and bar, and bar is a parent of foo.
724 Suppose all samples are recorded in foo and bar only. When it’s
726 recorded with callchains the output will show something like below in
727 the usual (self-overhead-only) output of perf report:
728 .ft C
731 Overhead Symbol
732 ........ .....................
733 60.00% foo
734 |
735 --- foo
736 bar
737 main
738 __libc_start_main
739 40.00% bar
741 |
742 --- bar
743 main
744 __libc_start_main
745 .ft
746 When the --children option is enabled, the self overhead values of
749 child functions (i.e. foo and bar) are added to the parents to
750 calculate the children overhead. In this case the report could be
751 displayed as:
752 .ft C
755 Children Self Symbol
756 ........ ........ ....................
757 100.00% 0.00% __libc_start_main
758 |
759 --- __libc_start_main
760 100.00% 0.00% main
762 |
763 --- main
764 __libc_start_main
765 100.00% 40.00% bar
767 |
768 --- bar
769 main
770 __libc_start_main
771 60.00% 60.00% foo
773 |
774 --- foo
775 bar
776 main
777 __libc_start_main
778 .ft
779 In the above output, the self overhead of foo (60%) was add to the
782 children overhead of bar, main and __libc_start_main. Likewise, the
783 self overhead of bar (40%) was added to the children overhead of main
784 and \_\_libc_start_main.
785 So \_\_libc_start_main and main are shown first since they have same
787 (100%) children overhead (even though they have zero self overhead) and
788 they are the parents of foo and bar.
789 Since v3.16 the children overhead is shown by default and the output is
791 sorted by its values. The children overhead is disabled by specifying
792 --no-children option on the command line or by adding report.children =
793 false or top.children = false in the perf config file.
794
796 perf-stat(1), perf-annotate(1), perf-record(1), perf-intel-pt(1)
797perf 06/03/2021 PERF-REPORT(1)