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 -F <freq>, --freq=<freq>
45 Profile at this frequency. Use max to use the currently maximum
46 allowed frequency, i.e. the value in the
47 kernel.perf_event_max_sample_rate sysctl.
48 -i, --inherit
50 Child tasks do not inherit counters.
51 -k <path>, --vmlinux=<path>
53 Path to vmlinux. Required for annotation functionality.
54 --ignore-vmlinux
56 Ignore vmlinux files.
57 --kallsyms=<file>
59 kallsyms pathname
60 -m <pages>, --mmap-pages=<pages>
62 Number of mmap data pages (must be a power of two) or size
63 specification with appended unit character - B/K/M/G. The size is
64 rounded up to have nearest pages power of two value.
65 -p <pid>, --pid=<pid>
67 Profile events on existing Process ID (comma separated list).
68 -t <tid>, --tid=<tid>
70 Profile events on existing thread ID (comma separated list).
71 -u, --uid=
73 Record events in threads owned by uid. Name or number.
74 -r <priority>, --realtime=<priority>
76 Collect data with this RT SCHED_FIFO priority.
77 --sym-annotate=<symbol>
79 Annotate this symbol.
80 -K, --hide_kernel_symbols
82 Hide kernel symbols.
83 -U, --hide_user_symbols
85 Hide user symbols.
86 --demangle-kernel
88 Demangle kernel symbols.
89 -D, --dump-symtab
91 Dump the symbol table used for profiling.
92 -v, --verbose
94 Be more verbose (show counter open errors, etc).
95 -z, --zero
97 Zero history across display updates.
98 -s, --sort
100 Sort by key(s): pid, comm, dso, symbol, parent, srcline, weight,
101 local_weight, abort, in_tx, transaction, overhead, sample, period.
102 Please see description of --sort in the perf-report man page.
103 --fields=
105 Specify output field - multiple keys can be specified in CSV
106 format. Following fields are available: overhead, overhead_sys,
107 overhead_us, overhead_children, sample and period. Also it can
108 contain any sort key(s).
109 By default, every sort keys not specified in --field will be appended
111 automatically.
112 -n, --show-nr-samples
114 Show a column with the number of samples.
115 --show-total-period
117 Show a column with the sum of periods.
118 --dsos
120 Only consider symbols in these dsos. This option will affect the
121 percentage of the overhead column. See --percentage for more info.
122 --comms
124 Only consider symbols in these comms. This option will affect the
125 percentage of the overhead column. See --percentage for more info.
126 --symbols
128 Only consider these symbols. This option will affect the percentage
129 of the overhead column. See --percentage for more info.
130 -M, --disassembler-style=
132 Set disassembler style for objdump.
133 --prefix=PREFIX, --prefix-strip=N
135 Remove first N entries from source file path names in executables
136 and add PREFIX. This allows to display source code compiled on
137 systems with different file system layout.
138 --source
140 Interleave source code with assembly code. Enabled by default,
141 disable with --no-source.
142 --asm-raw
144 Show raw instruction encoding of assembly instructions.
145 -g
147 Enables call-graph (stack chain/backtrace) recording.
148 --call-graph [mode,type,min[,limit],order[,key][,branch]]
150 Setup and enable call-graph (stack chain/backtrace) recording,
151 implies -g. See --call-graph section in perf-record and perf-report
152 man pages for details.
153 --children
155 Accumulate callchain of children to parent entry so that then can
156 show up in the output. The output will have a new "Children" column
157 and will be sorted on the data. It requires -g/--call-graph option
158 enabled. See the ‘overhead calculation’ section for more details.
159 Enabled by default, disable with --no-children.
160 --max-stack
162 Set the stack depth limit when parsing the callchain, anything
163 beyond the specified depth will be ignored. This is a trade-off
164 between information loss and faster processing especially for
165 workloads that can have a very long callchain stack.
166 Default: /proc/sys/kernel/perf_event_max_stack when present, 127 otherwise.
168 --ignore-callees=<regex>
170 Ignore callees of the function(s) matching the given regex. This
171 has the effect of collecting the callers of each such function into
172 one place in the call-graph tree.
173 --percent-limit
175 Do not show entries which have an overhead under that percent.
176 (Default: 0).
177 --percentage
179 Determine how to display the overhead percentage of filtered
180 entries. Filters can be applied by --comms, --dsos and/or --symbols
181 options and Zoom operations on the TUI (thread, dso, etc).
182 "relative" means it's relative to filtered entries only so that the
184 sum of shown entries will be always 100%. "absolute" means it retains
185 the original value before and after the filter is applied.
186 -w, --column-widths=<width[,width...]>
188 Force each column width to the provided list, for large terminal
189 readability. 0 means no limit (default behavior).
190 --proc-map-timeout
192 When processing pre-existing threads /proc/XXX/mmap, it may take a
193 long time, because the file may be huge. A time out is needed in
194 such cases. This option sets the time out limit. The default value
195 is 500 ms.
196 -b, --branch-any
198 Enable taken branch stack sampling. Any type of taken branch may be
199 sampled. This is a shortcut for --branch-filter any. See
200 --branch-filter for more infos.
201 -j, --branch-filter
203 Enable taken branch stack sampling. Each sample captures a series
204 of consecutive taken branches. The number of branches captured with
205 each sample depends on the underlying hardware, the type of
206 branches of interest, and the executed code. It is possible to
207 select the types of branches captured by enabling filters. For a
208 full list of modifiers please see the perf record manpage.
209 The option requires at least one branch type among any, any_call, any_ret, ind_call, cond.
211 The privilege levels may be omitted, in which case, the privilege levels of the associated
212 event are applied to the branch filter. Both kernel (k) and hypervisor (hv) privilege
213 levels are subject to permissions. When sampling on multiple events, branch stack sampling
214 is enabled for all the sampling events. The sampled branch type is the same for all events.
215 The various filters must be specified as a comma separated list: --branch-filter any_ret,u,k
216 Note that this feature may not be available on all processors.
217 --raw-trace
219 When displaying traceevent output, do not use print fmt or plugins.
220 --hierarchy
222 Enable hierarchy output.
223 --overwrite
225 Enable this to use just the most recent records, which helps in
226 high core count machines such as Knights Landing/Mill, but right
227 now is disabled by default as the pausing used in this technique is
228 leading to loss of metadata events such as PERF_RECORD_MMAP which
229 makes perf top unable to resolve samples, leading to lots of
230 unknown samples appearing on the UI. Enable this if you are in such
231 machines and profiling a workload that doesn’t creates short lived
232 threads and/or doesn’t uses many executable mmap operations. Work
233 is being planed to solve this situation, till then, this will
234 remain disabled by default.
235 --force
237 Don’t do ownership validation.
238 --num-thread-synthesize
240 The number of threads to run when synthesizing events for existing
241 processes. By default, the number of threads equals to the number
242 of online CPUs.
243 --namespaces
245 Record events of type PERF_RECORD_NAMESPACES and display it with
246 the cgroup_id sort key.
247 --switch-on EVENT_NAME
249 Only consider events after this event is found.
250 E.g.:
252 Find out where broadcast packets are handled
254 perf probe -L icmp_rcv
256 Insert a probe there:
258 perf probe icmp_rcv:59
260 Start perf top and ask it to only consider the cycles events when a
262 broadcast packet arrives This will show a menu with two entries and
263 will start counting when a broadcast packet arrives:
264 perf top -e cycles,probe:icmp_rcv --switch-on=probe:icmp_rcv
266 Alternatively one can ask for --group and then two overhead columns
268 will appear, the first for cycles and the second for the switch-on event.
269 perf top --group -e cycles,probe:icmp_rcv --switch-on=probe:icmp_rcv
271 This may be interesting to measure a workload only after some initialization
273 phase is over, i.e. insert a perf probe at that point and use the above
274 examples replacing probe:icmp_rcv with the just-after-init probe.
275 --switch-off EVENT_NAME
277 Stop considering events after this event is found.
278 --show-on-off-events
280 Show the --switch-on/off events too. This has no effect in perf top
281 now but probably we’ll make the default not to show the
282 switch-on/off events on the --group mode and if there is only one
283 event besides the off/on ones, go straight to the histogram
284 browser, just like perf top with no events explicitely specified
285 does.
286
288 [d]
289 Display refresh delay.
290 [e]
292 Number of entries to display.
293 [E]
295 Event to display when multiple counters are active.
296 [f]
298 Profile display filter (>= hit count).
299 [F]
301 Annotation display filter (>= % of total).
302 [s]
304 Annotate symbol.
305 [S]
307 Stop annotation, return to full profile display.
308 [K]
310 Hide kernel symbols.
311 [U]
313 Hide user symbols.
314 [z]
316 Toggle event count zeroing across display updates.
317 [qQ]
319 Quit.
320 Pressing any unmapped key displays a menu, and prompts for input.
322
324 The overhead can be shown in two columns as Children and Self when perf
325 collects callchains. The self overhead is simply calculated by adding
326 all period values of the entry - usually a function (symbol). This is
327 the value that perf shows traditionally and sum of all the self
328 overhead values should be 100%.
329 The children overhead is calculated by adding all period values of the
331 child functions so that it can show the total overhead of the higher
332 level functions even if they don’t directly execute much. Children here
333 means functions that are called from another (parent) function.
334 It might be confusing that the sum of all the children overhead values
336 exceeds 100% since each of them is already an accumulation of self
337 overhead of its child functions. But with this enabled, users can find
338 which function has the most overhead even if samples are spread over
339 the children.
340 Consider the following example; there are three functions like below.
342 .ft C
345 void foo(void) {
346 /* do something */
347 }
348 void bar(void) {
350 /* do something */
351 foo();
352 }
353 int main(void) {
355 bar()
356 return 0;
357 }
358 .ft
359 In this case foo is a child of bar, and bar is an immediate child of
362 main so foo also is a child of main. In other words, main is a parent
363 of foo and bar, and bar is a parent of foo.
364 Suppose all samples are recorded in foo and bar only. When it’s
366 recorded with callchains the output will show something like below in
367 the usual (self-overhead-only) output of perf report:
368 .ft C
371 Overhead Symbol
372 ........ .....................
373 60.00% foo
374 |
375 --- foo
376 bar
377 main
378 __libc_start_main
379 40.00% bar
381 |
382 --- bar
383 main
384 __libc_start_main
385 .ft
386 When the --children option is enabled, the self overhead values of
389 child functions (i.e. foo and bar) are added to the parents to
390 calculate the children overhead. In this case the report could be
391 displayed as:
392 .ft C
395 Children Self Symbol
396 ........ ........ ....................
397 100.00% 0.00% __libc_start_main
398 |
399 --- __libc_start_main
400 100.00% 0.00% main
402 |
403 --- main
404 __libc_start_main
405 100.00% 40.00% bar
407 |
408 --- bar
409 main
410 __libc_start_main
411 60.00% 60.00% foo
413 |
414 --- foo
415 bar
416 main
417 __libc_start_main
418 .ft
419 In the above output, the self overhead of foo (60%) was add to the
422 children overhead of bar, main and __libc_start_main. Likewise, the
423 self overhead of bar (40%) was added to the children overhead of main
424 and \_\_libc_start_main.
425 So \_\_libc_start_main and main are shown first since they have same
427 (100%) children overhead (even though they have zero self overhead) and
428 they are the parents of foo and bar.
429 Since v3.16 the children overhead is shown by default and the output is
431 sorted by its values. The children overhead is disabled by specifying
432 --no-children option on the command line or by adding report.children =
433 false or top.children = false in the perf config file.
434
436 perf-stat(1), perf-list(1), perf-report(1)
437perf 04/23/2020 PERF-TOP(1)