6 perf-top - System profiling tool.
11 'perf top' [-e <EVENT> | --event=EVENT] [<options>]
15 This command generates and displays a performance counter profile in real time.
22 System-wide collection. (default)
26 Event period to sample.
30 Monitor only on the list of CPUs provided. Multiple CPUs can be provided as a
31 comma-separated list with no space: 0,1. Ranges of CPUs are specified with -: 0-2.
32 Default is to monitor all CPUS.
36 Number of seconds to delay between refreshes.
40 Select the PMU event. Selection can be a symbolic event name
41 (use 'perf list' to list all events) or a raw PMU
42 event (eventsel+umask) in the form of rNNN where NNN is a
43 hexadecimal event descriptor.
47 Display this many functions.
50 --count-filter=<count>::
51 Only display functions with more events than this.
54 Put the counters into a counter group.
57 Sort the output by the event at the index n in group. If n is invalid,
58 sort by the first event. It can support multiple groups with different
59 amount of events. WARNING: This should be used on grouped events.
63 Profile at this frequency. Use 'max' to use the currently maximum
64 allowed frequency, i.e. the value in the kernel.perf_event_max_sample_rate
69 Child tasks do not inherit counters.
73 Path to vmlinux. Required for annotation functionality.
82 --mmap-pages=<pages>::
83 Number of mmap data pages (must be a power of two) or size
84 specification with appended unit character - B/K/M/G. The
85 size is rounded up to have nearest pages power of two value.
89 Profile events on existing Process ID (comma separated list).
93 Profile events on existing thread ID (comma separated list).
97 Record events in threads owned by uid. Name or number.
100 --realtime=<priority>::
101 Collect data with this RT SCHED_FIFO priority.
103 --sym-annotate=<symbol>::
104 Annotate this symbol.
107 --hide_kernel_symbols::
111 --hide_user_symbols::
115 Demangle kernel symbols.
119 Dump the symbol table used for profiling.
123 Be more verbose (show counter open errors, etc).
127 Zero history across display updates.
131 Sort by key(s): pid, comm, dso, symbol, parent, srcline, weight,
132 local_weight, abort, in_tx, transaction, overhead, sample, period.
133 Please see description of --sort in the perf-report man page.
136 Specify output field - multiple keys can be specified in CSV format.
137 Following fields are available:
138 overhead, overhead_sys, overhead_us, overhead_children, sample and period.
139 Also it can contain any sort key(s).
141 By default, every sort keys not specified in --field will be appended
146 Show a column with the number of samples.
148 --show-total-period::
149 Show a column with the sum of periods.
152 Only consider symbols in these dsos. This option will affect the
153 percentage of the overhead column. See --percentage for more info.
156 Only consider symbols in these comms. This option will affect the
157 percentage of the overhead column. See --percentage for more info.
160 Only consider these symbols. This option will affect the
161 percentage of the overhead column. See --percentage for more info.
164 --disassembler-style=:: Set disassembler style for objdump.
168 Remove first N entries from source file path names in executables
169 and add PREFIX. This allows to display source code compiled on systems
170 with different file system layout.
173 Interleave source code with assembly code. Enabled by default,
174 disable with --no-source.
177 Show raw instruction encoding of assembly instructions.
180 Enables call-graph (stack chain/backtrace) recording.
182 --call-graph [mode,type,min[,limit],order[,key][,branch]]::
183 Setup and enable call-graph (stack chain/backtrace) recording,
184 implies -g. See `--call-graph` section in perf-record and
185 perf-report man pages for details.
188 Accumulate callchain of children to parent entry so that then can
189 show up in the output. The output will have a new "Children" column
190 and will be sorted on the data. It requires -g/--call-graph option
191 enabled. See the `overhead calculation' section for more details.
192 Enabled by default, disable with --no-children.
195 Set the stack depth limit when parsing the callchain, anything
196 beyond the specified depth will be ignored. This is a trade-off
197 between information loss and faster processing especially for
198 workloads that can have a very long callchain stack.
200 Default: /proc/sys/kernel/perf_event_max_stack when present, 127 otherwise.
202 --ignore-callees=<regex>::
203 Ignore callees of the function(s) matching the given regex.
204 This has the effect of collecting the callers of each such
205 function into one place in the call-graph tree.
208 Do not show entries which have an overhead under that percent.
212 Determine how to display the overhead percentage of filtered entries.
213 Filters can be applied by --comms, --dsos and/or --symbols options and
214 Zoom operations on the TUI (thread, dso, etc).
216 "relative" means it's relative to filtered entries only so that the
217 sum of shown entries will be always 100%. "absolute" means it retains
218 the original value before and after the filter is applied.
221 --column-widths=<width[,width...]>::
222 Force each column width to the provided list, for large terminal
223 readability. 0 means no limit (default behavior).
226 When processing pre-existing threads /proc/XXX/mmap, it may take
227 a long time, because the file may be huge. A time out is needed
229 This option sets the time out limit. The default value is 500 ms.
234 Enable taken branch stack sampling. Any type of taken branch may be sampled.
235 This is a shortcut for --branch-filter any. See --branch-filter for more infos.
239 Enable taken branch stack sampling. Each sample captures a series of consecutive
240 taken branches. The number of branches captured with each sample depends on the
241 underlying hardware, the type of branches of interest, and the executed code.
242 It is possible to select the types of branches captured by enabling filters.
243 For a full list of modifiers please see the perf record manpage.
245 The option requires at least one branch type among any, any_call, any_ret, ind_call, cond.
246 The privilege levels may be omitted, in which case, the privilege levels of the associated
247 event are applied to the branch filter. Both kernel (k) and hypervisor (hv) privilege
248 levels are subject to permissions. When sampling on multiple events, branch stack sampling
249 is enabled for all the sampling events. The sampled branch type is the same for all events.
250 The various filters must be specified as a comma separated list: --branch-filter any_ret,u,k
251 Note that this feature may not be available on all processors.
254 When displaying traceevent output, do not use print fmt or plugins.
257 Enable hierarchy output.
260 Enable this to use just the most recent records, which helps in high core count
261 machines such as Knights Landing/Mill, but right now is disabled by default as
262 the pausing used in this technique is leading to loss of metadata events such
263 as PERF_RECORD_MMAP which makes 'perf top' unable to resolve samples, leading
264 to lots of unknown samples appearing on the UI. Enable this if you are in such
265 machines and profiling a workload that doesn't creates short lived threads and/or
266 doesn't uses many executable mmap operations. Work is being planed to solve
267 this situation, till then, this will remain disabled by default.
270 Don't do ownership validation.
272 --num-thread-synthesize::
273 The number of threads to run when synthesizing events for existing processes.
274 By default, the number of threads equals to the number of online CPUs.
277 Record events of type PERF_RECORD_NAMESPACES and display it with the
278 'cgroup_id' sort key.
282 monitor only in the container (cgroup) called "name". This option is available only
283 in per-cpu mode. The cgroup filesystem must be mounted. All threads belonging to
284 container "name" are monitored when they run on the monitored CPUs. Multiple cgroups
285 can be provided. Each cgroup is applied to the corresponding event, i.e., first cgroup
286 to first event, second cgroup to second event and so on. It is possible to provide
287 an empty cgroup (monitor all the time) using, e.g., -G foo,,bar. Cgroups must have
288 corresponding events, i.e., they always refer to events defined earlier on the command
289 line. If the user wants to track multiple events for a specific cgroup, the user can
290 use '-e e1 -e e2 -G foo,foo' or just use '-e e1 -e e2 -G foo'.
293 Record events of type PERF_RECORD_CGROUP and display it with the
296 --switch-on EVENT_NAME::
297 Only consider events after this event is found.
301 Find out where broadcast packets are handled
303 perf probe -L icmp_rcv
305 Insert a probe there:
307 perf probe icmp_rcv:59
309 Start perf top and ask it to only consider the cycles events when a
310 broadcast packet arrives This will show a menu with two entries and
311 will start counting when a broadcast packet arrives:
313 perf top -e cycles,probe:icmp_rcv --switch-on=probe:icmp_rcv
315 Alternatively one can ask for --group and then two overhead columns
316 will appear, the first for cycles and the second for the switch-on event.
318 perf top --group -e cycles,probe:icmp_rcv --switch-on=probe:icmp_rcv
320 This may be interesting to measure a workload only after some initialization
321 phase is over, i.e. insert a perf probe at that point and use the above
322 examples replacing probe:icmp_rcv with the just-after-init probe.
324 --switch-off EVENT_NAME::
325 Stop considering events after this event is found.
327 --show-on-off-events::
328 Show the --switch-on/off events too. This has no effect in 'perf top' now
329 but probably we'll make the default not to show the switch-on/off events
330 on the --group mode and if there is only one event besides the off/on ones,
331 go straight to the histogram browser, just like 'perf top' with no events
332 explicitly specified does.
335 Show callgraph with stitched LBRs, which may have more complete
336 callgraph. The option must be used with --call-graph lbr recording.
337 Disabled by default. In common cases with call stack overflows,
338 it can recreate better call stacks than the default lbr call stack
339 output. But this approach is not full proof. There can be cases
340 where it creates incorrect call stacks from incorrect matches.
341 The known limitations include exception handing such as
342 setjmp/longjmp will have calls/returns not match.
345 --pfm-events events::
346 Select a PMU event using libpfm4 syntax (see http://perfmon2.sf.net)
347 including support for event filters. For example '--pfm-events
348 inst_retired:any_p:u:c=1:i'. More than one event can be passed to the
349 option using the comma separator. Hardware events and generic hardware
350 events cannot be mixed together. The latter must be used with the -e
351 option. The -e option and this one can be mixed and matched. Events
352 can be grouped using the {} notation.
355 INTERACTIVE PROMPTING KEYS
356 --------------------------
359 Display refresh delay.
362 Number of entries to display.
365 Event to display when multiple counters are active.
368 Profile display filter (>= hit count).
371 Annotation display filter (>= % of total).
377 Stop annotation, return to full profile display.
386 Toggle event count zeroing across display updates.
391 Pressing any unmapped key displays a menu, and prompts for input.
393 include::callchain-overhead-calculation.txt[]
397 linkperf:perf-stat[1], linkperf:perf-list[1], linkperf:perf-report[1]