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 in the form of rN
33 where N is a hexadecimal value that represents the raw register
34 encoding with the layout of the event control registers as
35 described by entries in /sys/bus/event_source/devices/cpu/format/*.
36 -E <entries>, --entries=<entries>
38 Display this many functions.
39 -f <count>, --count-filter=<count>
41 Only display functions with more events than this.
42 --group-sort-idx
44 Sort the output by the event at the index n in group. If n is
45 invalid, sort by the first event. It can support multiple groups
46 with different amount of events. WARNING: This should be used on
47 grouped events.
48 -F <freq>, --freq=<freq>
50 Profile at this frequency. Use max to use the currently maximum
51 allowed frequency, i.e. the value in the
52 kernel.perf_event_max_sample_rate sysctl.
53 -i, --inherit
55 Child tasks do not inherit counters.
56 -k <path>, --vmlinux=<path>
58 Path to vmlinux. Required for annotation functionality.
59 --ignore-vmlinux
61 Ignore vmlinux files.
62 --kallsyms=<file>
64 kallsyms pathname
65 -m <pages>, --mmap-pages=<pages>
67 Number of mmap data pages (must be a power of two) or size
68 specification with appended unit character - B/K/M/G. The size is
69 rounded up to have nearest pages power of two value.
70 -p <pid>, --pid=<pid>
72 Profile events on existing Process ID (comma separated list).
73 -t <tid>, --tid=<tid>
75 Profile events on existing thread ID (comma separated list).
76 -u, --uid=
78 Record events in threads owned by uid. Name or number.
79 -r <priority>, --realtime=<priority>
81 Collect data with this RT SCHED_FIFO priority.
82 --sym-annotate=<symbol>
84 Annotate this symbol.
85 -K, --hide_kernel_symbols
87 Hide kernel symbols.
88 -U, --hide_user_symbols
90 Hide user symbols.
91 --demangle-kernel
93 Demangle kernel symbols.
94 -D, --dump-symtab
96 Dump the symbol table used for profiling.
97 -v, --verbose
99 Be more verbose (show counter open errors, etc).
100 -z, --zero
102 Zero history across display updates.
103 -s, --sort
105 Sort by key(s): pid, comm, dso, symbol, parent, srcline, weight,
106 local_weight, abort, in_tx, transaction, overhead, sample, period.
107 Please see description of --sort in the perf-report man page.
108 --fields=
110 Specify output field - multiple keys can be specified in CSV
111 format. Following fields are available: overhead, overhead_sys,
112 overhead_us, overhead_children, sample and period. Also it can
113 contain any sort key(s).
114 By default, every sort keys not specified in --field will be appended
116 automatically.
117 -n, --show-nr-samples
119 Show a column with the number of samples.
120 --show-total-period
122 Show a column with the sum of periods.
123 --dsos
125 Only consider symbols in these dsos. This option will affect the
126 percentage of the overhead column. See --percentage for more info.
127 --comms
129 Only consider symbols in these comms. This option will affect the
130 percentage of the overhead column. See --percentage for more info.
131 --symbols
133 Only consider these symbols. This option will affect the percentage
134 of the overhead column. See --percentage for more info.
135 -M, --disassembler-style=
137 Set disassembler style for objdump.
138 --addr2line=<path>
140 Path to addr2line binary.
141 --objdump=<path>
143 Path to objdump binary.
144 --prefix=PREFIX, --prefix-strip=N
146 Remove first N entries from source file path names in executables
147 and add PREFIX. This allows to display source code compiled on
148 systems with different file system layout.
149 --source
151 Interleave source code with assembly code. Enabled by default,
152 disable with --no-source.
153 --asm-raw
155 Show raw instruction encoding of assembly instructions.
156 -g
158 Enables call-graph (stack chain/backtrace) recording.
159 --call-graph [mode,type,min[,limit],order[,key][,branch]]
161 Setup and enable call-graph (stack chain/backtrace) recording,
162 implies -g. See --call-graph section in perf-record and perf-report
163 man pages for details.
164 --children
166 Accumulate callchain of children to parent entry so that then can
167 show up in the output. The output will have a new "Children" column
168 and will be sorted on the data. It requires -g/--call-graph option
169 enabled. See the ‘overhead calculation’ section for more details.
170 Enabled by default, disable with --no-children.
171 --max-stack
173 Set the stack depth limit when parsing the callchain, anything
174 beyond the specified depth will be ignored. This is a trade-off
175 between information loss and faster processing especially for
176 workloads that can have a very long callchain stack.
177 Default: /proc/sys/kernel/perf_event_max_stack when present, 127 otherwise.
179 --ignore-callees=<regex>
181 Ignore callees of the function(s) matching the given regex. This
182 has the effect of collecting the callers of each such function into
183 one place in the call-graph tree.
184 --percent-limit
186 Do not show entries which have an overhead under that percent.
187 (Default: 0).
188 --percentage
190 Determine how to display the overhead percentage of filtered
191 entries. Filters can be applied by --comms, --dsos and/or --symbols
192 options and Zoom operations on the TUI (thread, dso, etc).
193 "relative" means it's relative to filtered entries only so that the
195 sum of shown entries will be always 100%. "absolute" means it retains
196 the original value before and after the filter is applied.
197 -w, --column-widths=<width[,width...]>
199 Force each column width to the provided list, for large terminal
200 readability. 0 means no limit (default behavior).
201 --proc-map-timeout
203 When processing pre-existing threads /proc/XXX/mmap, it may take a
204 long time, because the file may be huge. A time out is needed in
205 such cases. This option sets the time out limit. The default value
206 is 500 ms.
207 -b, --branch-any
209 Enable taken branch stack sampling. Any type of taken branch may be
210 sampled. This is a shortcut for --branch-filter any. See
211 --branch-filter for more infos.
212 -j, --branch-filter
214 Enable taken branch stack sampling. Each sample captures a series
215 of consecutive taken branches. The number of branches captured with
216 each sample depends on the underlying hardware, the type of
217 branches of interest, and the executed code. It is possible to
218 select the types of branches captured by enabling filters. For a
219 full list of modifiers please see the perf record manpage.
220 The option requires at least one branch type among any, any_call, any_ret, ind_call, cond.
222 The privilege levels may be omitted, in which case, the privilege levels of the associated
223 event are applied to the branch filter. Both kernel (k) and hypervisor (hv) privilege
224 levels are subject to permissions. When sampling on multiple events, branch stack sampling
225 is enabled for all the sampling events. The sampled branch type is the same for all events.
226 The various filters must be specified as a comma separated list: --branch-filter any_ret,u,k
227 Note that this feature may not be available on all processors.
228 --branch-history
230 Add the addresses of sampled taken branches to the callstack. This
231 allows to examine the path the program took to each sample.
232 --raw-trace
234 When displaying traceevent output, do not use print fmt or plugins.
235 --hierarchy
237 Enable hierarchy output.
238 --overwrite
240 Enable this to use just the most recent records, which helps in
241 high core count machines such as Knights Landing/Mill, but right
242 now is disabled by default as the pausing used in this technique is
243 leading to loss of metadata events such as PERF_RECORD_MMAP which
244 makes perf top unable to resolve samples, leading to lots of
245 unknown samples appearing on the UI. Enable this if you are in such
246 machines and profiling a workload that doesn’t creates short lived
247 threads and/or doesn’t uses many executable mmap operations. Work
248 is being planed to solve this situation, till then, this will
249 remain disabled by default.
250 --force
252 Don’t do ownership validation.
253 --num-thread-synthesize
255 The number of threads to run when synthesizing events for existing
256 processes. By default, the number of threads equals to the number
257 of online CPUs.
258 --namespaces
260 Record events of type PERF_RECORD_NAMESPACES and display it with
261 the cgroup_id sort key.
262 -G name, --cgroup name
264 monitor only in the container (cgroup) called "name". This option
265 is available only in per-cpu mode. The cgroup filesystem must be
266 mounted. All threads belonging to container "name" are monitored
267 when they run on the monitored CPUs. Multiple cgroups can be
268 provided. Each cgroup is applied to the corresponding event, i.e.,
269 first cgroup to first event, second cgroup to second event and so
270 on. It is possible to provide an empty cgroup (monitor all the
271 time) using, e.g., -G foo,,bar. Cgroups must have corresponding
272 events, i.e., they always refer to events defined earlier on the
273 command line. If the user wants to track multiple events for a
274 specific cgroup, the user can use -e e1 -e e2 -G foo,foo or just
275 use -e e1 -e e2 -G foo.
276 --all-cgroups
278 Record events of type PERF_RECORD_CGROUP and display it with the
279 cgroup sort key.
280 --switch-on EVENT_NAME
282 Only consider events after this event is found.
283 E.g.:
285 Find out where broadcast packets are handled
287 perf probe -L icmp_rcv
289 Insert a probe there:
291 perf probe icmp_rcv:59
293 Start perf top and ask it to only consider the cycles events when a
295 broadcast packet arrives This will show a menu with two entries and
296 will start counting when a broadcast packet arrives:
297 perf top -e cycles,probe:icmp_rcv --switch-on=probe:icmp_rcv
299 Alternatively one can ask for a group and then two overhead columns
301 will appear, the first for cycles and the second for the switch-on event.
302 perf top -e '{cycles,probe:icmp_rcv}' --switch-on=probe:icmp_rcv
304 This may be interesting to measure a workload only after some initialization
306 phase is over, i.e. insert a perf probe at that point and use the above
307 examples replacing probe:icmp_rcv with the just-after-init probe.
308 --switch-off EVENT_NAME
310 Stop considering events after this event is found.
311 --show-on-off-events
313 Show the --switch-on/off events too. This has no effect in perf top
314 now but probably we’ll make the default not to show the
315 switch-on/off events on the --group mode and if there is only one
316 event besides the off/on ones, go straight to the histogram
317 browser, just like perf top with no events explicitly specified
318 does.
319 --stitch-lbr
321 Show callgraph with stitched LBRs, which may have more complete
322 callgraph. The option must be used with --call-graph lbr recording.
323 Disabled by default. In common cases with call stack overflows, it
324 can recreate better call stacks than the default lbr call stack
325 output. But this approach is not foolproof. There can be cases
326 where it creates incorrect call stacks from incorrect matches. The
327 known limitations include exception handing such as setjmp/longjmp
328 will have calls/returns not match.
329
331 [d]
332 Display refresh delay.
333 [e]
335 Number of entries to display.
336 [E]
338 Event to display when multiple counters are active.
339 [f]
341 Profile display filter (>= hit count).
342 [F]
344 Annotation display filter (>= % of total).
345 [s]
347 Annotate symbol.
348 [S]
350 Stop annotation, return to full profile display.
351 [K]
353 Hide kernel symbols.
354 [U]
356 Hide user symbols.
357 [z]
359 Toggle event count zeroing across display updates.
360 [qQ]
362 Quit.
363 Pressing any unmapped key displays a menu, and prompts for input.
365
367 The overhead can be shown in two columns as Children and Self when perf
368 collects callchains. The self overhead is simply calculated by adding
369 all period values of the entry - usually a function (symbol). This is
370 the value that perf shows traditionally and sum of all the self
371 overhead values should be 100%.
372 The children overhead is calculated by adding all period values of the
374 child functions so that it can show the total overhead of the higher
375 level functions even if they don’t directly execute much. Children here
376 means functions that are called from another (parent) function.
377 It might be confusing that the sum of all the children overhead values
379 exceeds 100% since each of them is already an accumulation of self
380 overhead of its child functions. But with this enabled, users can find
381 which function has the most overhead even if samples are spread over
382 the children.
383 Consider the following example; there are three functions like below.
385 .ft C
388 void foo(void) {
389 /* do something */
390 }
391 void bar(void) {
393 /* do something */
394 foo();
395 }
396 int main(void) {
398 bar()
399 return 0;
400 }
401 .ft
402 In this case foo is a child of bar, and bar is an immediate child of
405 main so foo also is a child of main. In other words, main is a parent
406 of foo and bar, and bar is a parent of foo.
407 Suppose all samples are recorded in foo and bar only. When it’s
409 recorded with callchains the output will show something like below in
410 the usual (self-overhead-only) output of perf report:
411 .ft C
414 Overhead Symbol
415 ........ .....................
416 60.00% foo
417 |
418 --- foo
419 bar
420 main
421 __libc_start_main
422 40.00% bar
424 |
425 --- bar
426 main
427 __libc_start_main
428 .ft
429 When the --children option is enabled, the self overhead values of
432 child functions (i.e. foo and bar) are added to the parents to
433 calculate the children overhead. In this case the report could be
434 displayed as:
435 .ft C
438 Children Self Symbol
439 ........ ........ ....................
440 100.00% 0.00% __libc_start_main
441 |
442 --- __libc_start_main
443 100.00% 0.00% main
445 |
446 --- main
447 __libc_start_main
448 100.00% 40.00% bar
450 |
451 --- bar
452 main
453 __libc_start_main
454 60.00% 60.00% foo
456 |
457 --- foo
458 bar
459 main
460 __libc_start_main
461 .ft
462 In the above output, the self overhead of foo (60%) was add to the
465 children overhead of bar, main and __libc_start_main. Likewise, the
466 self overhead of bar (40%) was added to the children overhead of main
467 and \_\_libc_start_main.
468 So \_\_libc_start_main and main are shown first since they have same
470 (100%) children overhead (even though they have zero self overhead) and
471 they are the parents of foo and bar.
472 Since v3.16 the children overhead is shown by default and the output is
474 sorted by its values. The children overhead is disabled by specifying
475 --no-children option on the command line or by adding report.children =
476 false or top.children = false in the perf config file.
477
479 perf-stat(1), perf-list(1), perf-report(1)
480perf 11/28/2023 PERF-TOP(1)