6 perf-report - Read perf.data (created by perf record) and display the profile
11 'perf report' [-i <file> | --input=file]
15 This command displays the performance counter profile information recorded
22 Input file name. (default: perf.data unless stdin is a fifo)
26 Be more verbose. (show symbol address, etc)
30 Do not show any warnings or messages. (Suppress -v)
34 Show the number of samples for each symbol
36 --show-cpu-utilization::
37 Show sample percentage for different cpu modes.
41 Show per-thread event counters. The input data file should be recorded
45 Only consider symbols in these comms. CSV that understands
46 file://filename entries. This option will affect the percentage of
47 the overhead column. See --percentage for more info.
49 Only show events for given process ID (comma separated list).
52 Only show events for given thread ID (comma separated list).
55 Only consider symbols in these dsos. CSV that understands
56 file://filename entries. This option will affect the percentage of
57 the overhead column. See --percentage for more info.
60 Only consider these symbols. CSV that understands
61 file://filename entries. This option will affect the percentage of
62 the overhead column. See --percentage for more info.
65 Only show symbols that match (partially) with this filter.
69 Only display entries resolved to a symbol.
73 Sort histogram entries by given key(s) - multiple keys can be specified
74 in CSV format. Following sort keys are available:
75 pid, comm, dso, symbol, parent, cpu, socket, srcline, weight,
76 local_weight, cgroup_id, addr.
78 Each key has following meaning:
80 - comm: command (name) of the task which can be read via /proc/<pid>/comm
81 - pid: command and tid of the task
82 - dso: name of library or module executed at the time of sample
83 - dso_size: size of library or module executed at the time of sample
84 - symbol: name of function executed at the time of sample
85 - symbol_size: size of function executed at the time of sample
86 - parent: name of function matched to the parent regex filter. Unmatched
87 entries are displayed as "[other]".
88 - cpu: cpu number the task ran at the time of sample
89 - socket: processor socket number the task ran at the time of sample
90 - srcline: filename and line number executed at the time of sample. The
91 DWARF debugging info must be provided.
92 - srcfile: file name of the source file of the samples. Requires dwarf
94 - weight: Event specific weight, e.g. memory latency or transaction
95 abort cost. This is the global weight.
96 - local_weight: Local weight version of the weight above.
97 - cgroup_id: ID derived from cgroup namespace device and inode numbers.
98 - cgroup: cgroup pathname in the cgroupfs.
99 - transaction: Transaction abort flags.
100 - overhead: Overhead percentage of sample
101 - overhead_sys: Overhead percentage of sample running in system mode
102 - overhead_us: Overhead percentage of sample running in user mode
103 - overhead_guest_sys: Overhead percentage of sample running in system mode
105 - overhead_guest_us: Overhead percentage of sample running in user mode on
107 - sample: Number of sample
108 - period: Raw number of event count of sample
109 - time: Separate the samples by time stamp with the resolution specified by
110 --time-quantum (default 100ms). Specify with overhead and before it.
111 - code_page_size: the code page size of sampled code address (ip)
112 - ins_lat: Instruction latency in core cycles. This is the global instruction
114 - local_ins_lat: Local instruction latency version
115 - p_stage_cyc: On powerpc, this presents the number of cycles spent in a
116 pipeline stage. And currently supported only on powerpc.
117 - addr: (Full) virtual address of the sampled instruction
119 By default, comm, dso and symbol keys are used.
120 (i.e. --sort comm,dso,symbol)
122 If --branch-stack option is used, following sort keys are also
125 - dso_from: name of library or module branched from
126 - dso_to: name of library or module branched to
127 - symbol_from: name of function branched from
128 - symbol_to: name of function branched to
129 - srcline_from: source file and line branched from
130 - srcline_to: source file and line branched to
131 - mispredict: "N" for predicted branch, "Y" for mispredicted branch
132 - in_tx: branch in TSX transaction
133 - abort: TSX transaction abort.
134 - cycles: Cycles in basic block
136 And default sort keys are changed to comm, dso_from, symbol_from, dso_to
137 and symbol_to, see '--branch-stack'.
139 When the sort key symbol is specified, columns "IPC" and "IPC Coverage"
140 are enabled automatically. Column "IPC" reports the average IPC per function
141 and column "IPC coverage" reports the percentage of instructions with
142 sampled IPC in this function. IPC means Instruction Per Cycle. If it's low,
143 it indicates there may be a performance bottleneck when the function is
144 executed, such as a memory access bottleneck. If a function has high overhead
145 and low IPC, it's worth further analyzing it to optimize its performance.
147 If the --mem-mode option is used, the following sort keys are also available
148 (incompatible with --branch-stack):
149 symbol_daddr, dso_daddr, locked, tlb, mem, snoop, dcacheline, blocked.
151 - symbol_daddr: name of data symbol being executed on at the time of sample
152 - dso_daddr: name of library or module containing the data being executed
153 on at the time of the sample
154 - locked: whether the bus was locked at the time of the sample
155 - tlb: type of tlb access for the data at the time of the sample
156 - mem: type of memory access for the data at the time of the sample
157 - snoop: type of snoop (if any) for the data at the time of the sample
158 - dcacheline: the cacheline the data address is on at the time of the sample
159 - phys_daddr: physical address of data being executed on at the time of sample
160 - data_page_size: the data page size of data being executed on at the time of sample
161 - blocked: reason of blocked load access for the data at the time of the sample
163 And the default sort keys are changed to local_weight, mem, sym, dso,
164 symbol_daddr, dso_daddr, snoop, tlb, locked, blocked, local_ins_lat,
167 If the data file has tracepoint event(s), following (dynamic) sort keys
169 trace, trace_fields, [<event>.]<field>[/raw]
171 - trace: pretty printed trace output in a single column
172 - trace_fields: fields in tracepoints in separate columns
173 - <field name>: optional event and field name for a specific field
175 The last form consists of event and field names. If event name is
176 omitted, it searches all events for matching field name. The matched
177 field will be shown only for the event has the field. The event name
178 supports substring match so user doesn't need to specify full subsystem
179 and event name everytime. For example, 'sched:sched_switch' event can
180 be shortened to 'switch' as long as it's not ambiguous. Also event can
181 be specified by its index (starting from 1) preceded by the '%'.
182 So '%1' is the first event, '%2' is the second, and so on.
184 The field name can have '/raw' suffix which disables pretty printing
185 and shows raw field value like hex numbers. The --raw-trace option
186 has the same effect for all dynamic sort keys.
188 The default sort keys are changed to 'trace' if all events in the data
193 Specify output field - multiple keys can be specified in CSV format.
194 Following fields are available:
195 overhead, overhead_sys, overhead_us, overhead_children, sample and period.
196 Also it can contain any sort key(s).
198 By default, every sort keys not specified in -F will be appended
201 If the keys starts with a prefix '+', then it will append the specified
202 field(s) to the default field order. For example: perf report -F +period,sample.
206 A regex filter to identify parent. The parent is a caller of this
207 function and searched through the callchain, thus it requires callchain
208 information recorded. The pattern is in the extended regex format and
209 defaults to "\^sys_|^do_page_fault", see '--sort parent'.
213 Only display entries with parent-match.
216 --column-widths=<width[,width...]>::
217 Force each column width to the provided list, for large terminal
218 readability. 0 means no limit (default behavior).
222 Use a special separator character and don't pad with spaces, replacing
223 all occurrences of this separator in symbol names (and other output)
224 with a '.' character, that thus it's the only non valid separator.
228 Dump raw trace in ASCII.
231 Disable raw trace ordering.
234 --call-graph=<print_type,threshold[,print_limit],order,sort_key[,branch],value>::
235 Display call chains using type, min percent threshold, print limit,
236 call order, sort key, optional branch and value. Note that ordering
237 is not fixed so any parameter can be given in an arbitrary order.
238 One exception is the print_limit which should be preceded by threshold.
240 print_type can be either:
241 - flat: single column, linear exposure of call chains.
242 - graph: use a graph tree, displaying absolute overhead rates. (default)
243 - fractal: like graph, but displays relative rates. Each branch of
244 the tree is considered as a new profiled object.
245 - folded: call chains are displayed in a line, separated by semicolons
246 - none: disable call chain display.
248 threshold is a percentage value which specifies a minimum percent to be
249 included in the output call graph. Default is 0.5 (%).
251 print_limit is only applied when stdio interface is used. It's to limit
252 number of call graph entries in a single hist entry. Note that it needs
253 to be given after threshold (but not necessarily consecutive).
254 Default is 0 (unlimited).
257 - callee: callee based call graph.
258 - caller: inverted caller based call graph.
259 Default is 'caller' when --children is used, otherwise 'callee'.
262 - function: compare on functions (default)
263 - address: compare on individual code addresses
264 - srcline: compare on source filename and line number
267 - branch: include last branch information in callgraph when available.
268 Usually more convenient to use --branch-history for this.
271 - percent: display overhead percent (default)
272 - period: display event period
273 - count: display event count
276 Accumulate callchain of children to parent entry so that then can
277 show up in the output. The output will have a new "Children" column
278 and will be sorted on the data. It requires callchains are recorded.
279 See the `overhead calculation' section for more details. Enabled by
280 default, disable with --no-children.
283 Set the stack depth limit when parsing the callchain, anything
284 beyond the specified depth will be ignored. This is a trade-off
285 between information loss and faster processing especially for
286 workloads that can have a very long callchain stack.
287 Note that when using the --itrace option the synthesized callchain size
288 will override this value if the synthesized callchain size is bigger.
294 alias for inverted caller based call graph.
296 --ignore-callees=<regex>::
297 Ignore callees of the function(s) matching the given regex.
298 This has the effect of collecting the callers of each such
299 function into one place in the call-graph tree.
302 Pretty printing style. key: normal, raw
304 --stdio:: Use the stdio interface.
307 'always', 'never' or 'auto', allowing configuring color output
308 via the command line, in addition to via "color.ui" .perfconfig.
309 Use '--stdio-color always' to generate color even when redirecting
310 to a pipe or file. Using just '--stdio-color' is equivalent to
313 --tui:: Use the TUI interface, that is integrated with annotate and allows
314 zooming into DSOs or threads, among other features. Use of --tui
315 requires a tty, if one is not present, as when piping to other
316 commands, the stdio interface is used.
318 --gtk:: Use the GTK2 interface.
325 Ignore vmlinux files.
332 Load module symbols. WARNING: This should only be used with -k and
337 Don't do ownership validation.
339 --symfs=<directory>::
340 Look for files with symbols relative to this directory.
343 --cpu:: Only report samples for the list of CPUs provided. Multiple CPUs can
344 be provided as a comma-separated list with no space: 0,1. Ranges of
345 CPUs are specified with -: 0-2. Default is to report samples on all
349 --disassembler-style=:: Set disassembler style for objdump.
352 Interleave source code with assembly code. Enabled by default,
353 disable with --no-source.
356 Show raw instruction encoding of assembly instructions.
358 --show-total-period:: Show a column with the sum of periods.
362 Display extended information about the perf.data file. This adds
363 information which may be very large and thus may clutter the display.
364 It currently includes: cpu and numa topology of the host system.
368 Use the addresses of sampled taken branches instead of the instruction
369 address to build the histograms. To generate meaningful output, the
370 perf.data file must have been obtained using perf record -b or
371 perf record --branch-filter xxx where xxx is a branch filter option.
372 perf report is able to auto-detect whether a perf.data file contains
373 branch stacks and it will automatically switch to the branch view mode,
374 unless --no-branch-stack is used.
377 Add the addresses of sampled taken branches to the callstack.
378 This allows to examine the path the program took to each sample.
379 The data collection must have used -b (or -j) and -g.
382 Path to objdump binary.
386 Remove first N entries from source file path names in executables
387 and add PREFIX. This allows to display source code compiled on systems
388 with different file system layout.
391 Show event group information together. It forces group output also
392 if there are no groups defined in data file.
395 Sort the output by the event at the index n in group. If n is invalid,
396 sort by the first event. It can support multiple groups with different
397 amount of events. WARNING: This should be used on grouped events.
400 Demangle symbol names to human readable form. It's enabled by default,
401 disable with --no-demangle.
404 Demangle kernel symbol names to human readable form (for C++ kernels).
407 Use the data addresses of samples in addition to instruction addresses
408 to build the histograms. To generate meaningful output, the perf.data
409 file must have been obtained using perf record -d -W and using a
410 special event -e cpu/mem-loads/p or -e cpu/mem-stores/p. See
411 'perf mem' for simpler access.
414 Do not show entries which have an overhead under that percent.
415 (Default: 0). Note that this option also sets the percent limit (threshold)
416 of callchains. However the default value of callchain threshold is
417 different than the default value of hist entries. Please see the
418 --call-graph option for details.
421 Determine how to display the overhead percentage of filtered entries.
422 Filters can be applied by --comms, --dsos and/or --symbols options and
423 Zoom operations on the TUI (thread, dso, etc).
425 "relative" means it's relative to filtered entries only so that the
426 sum of shown entries will be always 100%. "absolute" means it retains
427 the original value before and after the filter is applied.
430 Show header information in the perf.data file. This includes
431 various information like hostname, OS and perf version, cpu/mem
432 info, perf command line, event list and so on. Currently only
433 --stdio output supports this feature.
436 Show only perf.data header (forces --stdio).
439 Only analyze samples within given time window: <start>,<stop>. Times
440 have the format seconds.nanoseconds. If start is not given (i.e. time
441 string is ',x.y') then analysis starts at the beginning of the file. If
442 stop time is not given (i.e. time string is 'x.y,') then analysis goes
443 to end of file. Multiple ranges can be separated by spaces, which
444 requires the argument to be quoted e.g. --time "1234.567,1234.789 1235,"
446 Also support time percent with multiple time ranges. Time string is
447 'a%/n,b%/m,...' or 'a%-b%,c%-%d,...'.
450 Select the second 10% time slice:
452 perf report --time 10%/2
454 Select from 0% to 10% time slice:
456 perf report --time 0%-10%
458 Select the first and second 10% time slices:
460 perf report --time 10%/1,10%/2
462 Select from 0% to 10% and 30% to 40% slices:
464 perf report --time 0%-10%,30%-40%
466 --switch-on EVENT_NAME::
467 Only consider events after this event is found.
469 This may be interesting to measure a workload only after some initialization
470 phase is over, i.e. insert a perf probe at that point and then using this
471 option with that probe.
473 --switch-off EVENT_NAME::
474 Stop considering events after this event is found.
476 --show-on-off-events::
477 Show the --switch-on/off events too. This has no effect in 'perf report' now
478 but probably we'll make the default not to show the switch-on/off events
479 on the --group mode and if there is only one event besides the off/on ones,
480 go straight to the histogram browser, just like 'perf report' with no events
481 explicitly specified does.
484 Options for decoding instruction tracing data. The options are:
486 include::itrace.txt[]
488 To disable decoding entirely, use --no-itrace.
491 Show the full path for source files for srcline output.
493 --show-ref-call-graph::
494 When multiple events are sampled, it may not be needed to collect
495 callgraphs for all of them. The sample sites are usually nearby,
496 and it's enough to collect the callgraphs on a reference event.
497 So user can use "call-graph=no" event modifier to disable callgraph
498 for other events to reduce the overhead.
499 However, perf report cannot show callgraphs for the event which
500 disable the callgraph.
501 This option extends the perf report to show reference callgraphs,
502 which collected by reference event, in no callgraph event.
505 Show callgraph with stitched LBRs, which may have more complete
506 callgraph. The perf.data file must have been obtained using
507 perf record --call-graph lbr.
508 Disabled by default. In common cases with call stack overflows,
509 it can recreate better call stacks than the default lbr call stack
510 output. But this approach is not foolproof. There can be cases
511 where it creates incorrect call stacks from incorrect matches.
512 The known limitations include exception handing such as
513 setjmp/longjmp will have calls/returns not match.
516 Only report the samples on the processor socket that match with this filter
519 Save N individual samples for each histogram entry to show context in perf
523 When displaying traceevent output, do not use print fmt or plugins.
526 Enable hierarchical output.
529 If a callgraph address belongs to an inlined function, the inline stack
530 will be printed. Each entry is function name or file/line. Enabled by
531 default, disable with --no-inline.
534 Show --tasks output plus mmap information in a format similar to
537 Please note that not all mmaps are stored, options affecting which ones
538 are include 'perf record --data', for instance.
541 Show time stamps in nanoseconds.
544 Display overall events statistics without any further processing.
545 (like the one at the end of the perf report -D command)
548 Display monitored tasks stored in perf data. Displaying pid/tid/ppid
549 plus the command string aligned to distinguish parent and child tasks.
552 Set annotation percent type from following choices:
553 global-period, local-period, global-hits, local-hits
555 The local/global keywords set if the percentage is computed
556 in the scope of the function (local) or the whole data (global).
557 The period/hits keywords set the base the percentage is computed
558 on - the samples period or the number of samples (hits).
561 Configure time quantum for time sort key. Default 100ms.
562 Accepts s, us, ms, ns units.
565 When --total-cycles is specified, it supports sorting for all blocks by
566 'Sampled Cycles%'. This is useful to concentrate on the globally hottest
567 blocks. In output, there are some new columns:
569 'Sampled Cycles%' - block sampled cycles aggregation / total sampled cycles
570 'Sampled Cycles' - block sampled cycles aggregation
571 'Avg Cycles%' - block average sampled cycles / sum of total block average
573 'Avg Cycles' - block average sampled cycles
576 Do not print 0 results in the --stat output.
578 include::callchain-overhead-calculation.txt[]
582 linkperf:perf-stat[1], linkperf:perf-annotate[1], linkperf:perf-record[1],
583 linkperf:perf-intel-pt[1]