1PERF-TOP(1) perf Manual PERF-TOP(1)
2
6 perf-top - System profiling tool.
7
9 perf top [-e <EVENT> | --event=EVENT] [<options>]
10
12 This command generates and displays a performance counter profile in
13 real time.
14
16 -a, --all-cpus
17 System-wide collection. (default)
18 -c <count>, --count=<count>
20 Event period to sample.
21 -C <cpu-list>, --cpu=<cpu>
23 Monitor only on the list of CPUs provided. Multiple CPUs can be
24 provided as a comma-separated list with no space: 0,1. Ranges of
25 CPUs are specified with -: 0-2. Default is to monitor all CPUS.
26 -d <seconds>, --delay=<seconds>
28 Number of seconds to delay between refreshes.
29 -e <event>, --event=<event>
31 Select the PMU event. Selection can be a symbolic event name (use
32 perf list to list all events) or a raw PMU event (eventsel+umask)
33 in the form of rNNN where NNN is a hexadecimal event descriptor.
34 -E <entries>, --entries=<entries>
36 Display this many functions.
37 -f <count>, --count-filter=<count>
39 Only display functions with more events than this.
40 --group
42 Put the counters into a counter group.
43 --group-sort-idx
45 Sort the output by the event at the index n in group. If n is
46 invalid, sort by the first event. It can support multiple groups
47 with different amount of events. WARNING: This should be used on
48 grouped events.
49 -F <freq>, --freq=<freq>
51 Profile at this frequency. Use max to use the currently maximum
52 allowed frequency, i.e. the value in the
53 kernel.perf_event_max_sample_rate sysctl.
54 -i, --inherit
56 Child tasks do not inherit counters.
57 -k <path>, --vmlinux=<path>
59 Path to vmlinux. Required for annotation functionality.
60 --ignore-vmlinux
62 Ignore vmlinux files.
63 --kallsyms=<file>
65 kallsyms pathname
66 -m <pages>, --mmap-pages=<pages>
68 Number of mmap data pages (must be a power of two) or size
69 specification with appended unit character - B/K/M/G. The size is
70 rounded up to have nearest pages power of two value.
71 -p <pid>, --pid=<pid>
73 Profile events on existing Process ID (comma separated list).
74 -t <tid>, --tid=<tid>
76 Profile events on existing thread ID (comma separated list).
77 -u, --uid=
79 Record events in threads owned by uid. Name or number.
80 -r <priority>, --realtime=<priority>
82 Collect data with this RT SCHED_FIFO priority.
83 --sym-annotate=<symbol>
85 Annotate this symbol.
86 -K, --hide_kernel_symbols
88 Hide kernel symbols.
89 -U, --hide_user_symbols
91 Hide user symbols.
92 --demangle-kernel
94 Demangle kernel symbols.
95 -D, --dump-symtab
97 Dump the symbol table used for profiling.
98 -v, --verbose
100 Be more verbose (show counter open errors, etc).
101 -z, --zero
103 Zero history across display updates.
104 -s, --sort
106 Sort by key(s): pid, comm, dso, symbol, parent, srcline, weight,
107 local_weight, abort, in_tx, transaction, overhead, sample, period.
108 Please see description of --sort in the perf-report man page.
109 --fields=
111 Specify output field - multiple keys can be specified in CSV
112 format. Following fields are available: overhead, overhead_sys,
113 overhead_us, overhead_children, sample and period. Also it can
114 contain any sort key(s).
115 By default, every sort keys not specified in --field will be appended
117 automatically.
118 -n, --show-nr-samples
120 Show a column with the number of samples.
121 --show-total-period
123 Show a column with the sum of periods.
124 --dsos
126 Only consider symbols in these dsos. This option will affect the
127 percentage of the overhead column. See --percentage for more info.
128 --comms
130 Only consider symbols in these comms. This option will affect the
131 percentage of the overhead column. See --percentage for more info.
132 --symbols
134 Only consider these symbols. This option will affect the percentage
135 of the overhead column. See --percentage for more info.
136 -M, --disassembler-style=
138 Set disassembler style for objdump.
139 --prefix=PREFIX, --prefix-strip=N
141 Remove first N entries from source file path names in executables
142 and add PREFIX. This allows to display source code compiled on
143 systems with different file system layout.
144 --source
146 Interleave source code with assembly code. Enabled by default,
147 disable with --no-source.
148 --asm-raw
150 Show raw instruction encoding of assembly instructions.
151 -g
153 Enables call-graph (stack chain/backtrace) recording.
154 --call-graph [mode,type,min[,limit],order[,key][,branch]]
156 Setup and enable call-graph (stack chain/backtrace) recording,
157 implies -g. See --call-graph section in perf-record and perf-report
158 man pages for details.
159 --children
161 Accumulate callchain of children to parent entry so that then can
162 show up in the output. The output will have a new "Children" column
163 and will be sorted on the data. It requires -g/--call-graph option
164 enabled. See the ‘overhead calculation’ section for more details.
165 Enabled by default, disable with --no-children.
166 --max-stack
168 Set the stack depth limit when parsing the callchain, anything
169 beyond the specified depth will be ignored. This is a trade-off
170 between information loss and faster processing especially for
171 workloads that can have a very long callchain stack.
172 Default: /proc/sys/kernel/perf_event_max_stack when present, 127 otherwise.
174 --ignore-callees=<regex>
176 Ignore callees of the function(s) matching the given regex. This
177 has the effect of collecting the callers of each such function into
178 one place in the call-graph tree.
179 --percent-limit
181 Do not show entries which have an overhead under that percent.
182 (Default: 0).
183 --percentage
185 Determine how to display the overhead percentage of filtered
186 entries. Filters can be applied by --comms, --dsos and/or --symbols
187 options and Zoom operations on the TUI (thread, dso, etc).
188 "relative" means it's relative to filtered entries only so that the
190 sum of shown entries will be always 100%. "absolute" means it retains
191 the original value before and after the filter is applied.
192 -w, --column-widths=<width[,width...]>
194 Force each column width to the provided list, for large terminal
195 readability. 0 means no limit (default behavior).
196 --proc-map-timeout
198 When processing pre-existing threads /proc/XXX/mmap, it may take a
199 long time, because the file may be huge. A time out is needed in
200 such cases. This option sets the time out limit. The default value
201 is 500 ms.
202 -b, --branch-any
204 Enable taken branch stack sampling. Any type of taken branch may be
205 sampled. This is a shortcut for --branch-filter any. See
206 --branch-filter for more infos.
207 -j, --branch-filter
209 Enable taken branch stack sampling. Each sample captures a series
210 of consecutive taken branches. The number of branches captured with
211 each sample depends on the underlying hardware, the type of
212 branches of interest, and the executed code. It is possible to
213 select the types of branches captured by enabling filters. For a
214 full list of modifiers please see the perf record manpage.
215 The option requires at least one branch type among any, any_call, any_ret, ind_call, cond.
217 The privilege levels may be omitted, in which case, the privilege levels of the associated
218 event are applied to the branch filter. Both kernel (k) and hypervisor (hv) privilege
219 levels are subject to permissions. When sampling on multiple events, branch stack sampling
220 is enabled for all the sampling events. The sampled branch type is the same for all events.
221 The various filters must be specified as a comma separated list: --branch-filter any_ret,u,k
222 Note that this feature may not be available on all processors.
223 --raw-trace
225 When displaying traceevent output, do not use print fmt or plugins.
226 --hierarchy
228 Enable hierarchy output.
229 --overwrite
231 Enable this to use just the most recent records, which helps in
232 high core count machines such as Knights Landing/Mill, but right
233 now is disabled by default as the pausing used in this technique is
234 leading to loss of metadata events such as PERF_RECORD_MMAP which
235 makes perf top unable to resolve samples, leading to lots of
236 unknown samples appearing on the UI. Enable this if you are in such
237 machines and profiling a workload that doesn’t creates short lived
238 threads and/or doesn’t uses many executable mmap operations. Work
239 is being planed to solve this situation, till then, this will
240 remain disabled by default.
241 --force
243 Don’t do ownership validation.
244 --num-thread-synthesize
246 The number of threads to run when synthesizing events for existing
247 processes. By default, the number of threads equals to the number
248 of online CPUs.
249 --namespaces
251 Record events of type PERF_RECORD_NAMESPACES and display it with
252 the cgroup_id sort key.
253 -G name, --cgroup name
255 monitor only in the container (cgroup) called "name". This option
256 is available only in per-cpu mode. The cgroup filesystem must be
257 mounted. All threads belonging to container "name" are monitored
258 when they run on the monitored CPUs. Multiple cgroups can be
259 provided. Each cgroup is applied to the corresponding event, i.e.,
260 first cgroup to first event, second cgroup to second event and so
261 on. It is possible to provide an empty cgroup (monitor all the
262 time) using, e.g., -G foo,,bar. Cgroups must have corresponding
263 events, i.e., they always refer to events defined earlier on the
264 command line. If the user wants to track multiple events for a
265 specific cgroup, the user can use -e e1 -e e2 -G foo,foo or just
266 use -e e1 -e e2 -G foo.
267 --all-cgroups
269 Record events of type PERF_RECORD_CGROUP and display it with the
270 cgroup sort key.
271 --switch-on EVENT_NAME
273 Only consider events after this event is found.
274 E.g.:
276 Find out where broadcast packets are handled
278 perf probe -L icmp_rcv
280 Insert a probe there:
282 perf probe icmp_rcv:59
284 Start perf top and ask it to only consider the cycles events when a
286 broadcast packet arrives This will show a menu with two entries and
287 will start counting when a broadcast packet arrives:
288 perf top -e cycles,probe:icmp_rcv --switch-on=probe:icmp_rcv
290 Alternatively one can ask for --group and then two overhead columns
292 will appear, the first for cycles and the second for the switch-on event.
293 perf top --group -e cycles,probe:icmp_rcv --switch-on=probe:icmp_rcv
295 This may be interesting to measure a workload only after some initialization
297 phase is over, i.e. insert a perf probe at that point and use the above
298 examples replacing probe:icmp_rcv with the just-after-init probe.
299 --switch-off EVENT_NAME
301 Stop considering events after this event is found.
302 --show-on-off-events
304 Show the --switch-on/off events too. This has no effect in perf top
305 now but probably we’ll make the default not to show the
306 switch-on/off events on the --group mode and if there is only one
307 event besides the off/on ones, go straight to the histogram
308 browser, just like perf top with no events explicitly specified
309 does.
310 --stitch-lbr
312 Show callgraph with stitched LBRs, which may have more complete
313 callgraph. The option must be used with --call-graph lbr recording.
314 Disabled by default. In common cases with call stack overflows, it
315 can recreate better call stacks than the default lbr call stack
316 output. But this approach is not full proof. There can be cases
317 where it creates incorrect call stacks from incorrect matches. The
318 known limitations include exception handing such as setjmp/longjmp
319 will have calls/returns not match.
320
322 [d]
323 Display refresh delay.
324 [e]
326 Number of entries to display.
327 [E]
329 Event to display when multiple counters are active.
330 [f]
332 Profile display filter (>= hit count).
333 [F]
335 Annotation display filter (>= % of total).
336 [s]
338 Annotate symbol.
339 [S]
341 Stop annotation, return to full profile display.
342 [K]
344 Hide kernel symbols.
345 [U]
347 Hide user symbols.
348 [z]
350 Toggle event count zeroing across display updates.
351 [qQ]
353 Quit.
354 Pressing any unmapped key displays a menu, and prompts for input.
356
358 The overhead can be shown in two columns as Children and Self when perf
359 collects callchains. The self overhead is simply calculated by adding
360 all period values of the entry - usually a function (symbol). This is
361 the value that perf shows traditionally and sum of all the self
362 overhead values should be 100%.
363 The children overhead is calculated by adding all period values of the
365 child functions so that it can show the total overhead of the higher
366 level functions even if they don’t directly execute much. Children here
367 means functions that are called from another (parent) function.
368 It might be confusing that the sum of all the children overhead values
370 exceeds 100% since each of them is already an accumulation of self
371 overhead of its child functions. But with this enabled, users can find
372 which function has the most overhead even if samples are spread over
373 the children.
374 Consider the following example; there are three functions like below.
376 .ft C
379 void foo(void) {
380 /* do something */
381 }
382 void bar(void) {
384 /* do something */
385 foo();
386 }
387 int main(void) {
389 bar()
390 return 0;
391 }
392 .ft
393 In this case foo is a child of bar, and bar is an immediate child of
396 main so foo also is a child of main. In other words, main is a parent
397 of foo and bar, and bar is a parent of foo.
398 Suppose all samples are recorded in foo and bar only. When it’s
400 recorded with callchains the output will show something like below in
401 the usual (self-overhead-only) output of perf report:
402 .ft C
405 Overhead Symbol
406 ........ .....................
407 60.00% foo
408 |
409 --- foo
410 bar
411 main
412 __libc_start_main
413 40.00% bar
415 |
416 --- bar
417 main
418 __libc_start_main
419 .ft
420 When the --children option is enabled, the self overhead values of
423 child functions (i.e. foo and bar) are added to the parents to
424 calculate the children overhead. In this case the report could be
425 displayed as:
426 .ft C
429 Children Self Symbol
430 ........ ........ ....................
431 100.00% 0.00% __libc_start_main
432 |
433 --- __libc_start_main
434 100.00% 0.00% main
436 |
437 --- main
438 __libc_start_main
439 100.00% 40.00% bar
441 |
442 --- bar
443 main
444 __libc_start_main
445 60.00% 60.00% foo
447 |
448 --- foo
449 bar
450 main
451 __libc_start_main
452 .ft
453 In the above output, the self overhead of foo (60%) was add to the
456 children overhead of bar, main and __libc_start_main. Likewise, the
457 self overhead of bar (40%) was added to the children overhead of main
458 and \_\_libc_start_main.
459 So \_\_libc_start_main and main are shown first since they have same
461 (100%) children overhead (even though they have zero self overhead) and
462 they are the parents of foo and bar.
463 Since v3.16 the children overhead is shown by default and the output is
465 sorted by its values. The children overhead is disabled by specifying
466 --no-children option on the command line or by adding report.children =
467 false or top.children = false in the perf config file.
468
470 perf-stat(1), perf-list(1), perf-report(1)
471perf 11/22/2021 PERF-TOP(1)