0001 perf-top(1)
0002 ===========
0003
0004 NAME
0005 ----
0006 perf-top - System profiling tool.
0007
0008 SYNOPSIS
0009 --------
0010 [verse]
0011 'perf top' [-e <EVENT> | --event=EVENT] [<options>]
0012
0013 DESCRIPTION
0014 -----------
0015 This command generates and displays a performance counter profile in real time.
0016
0017
0018 OPTIONS
0019 -------
0020 -a::
0021 --all-cpus::
0022 System-wide collection. (default)
0023
0024 -c <count>::
0025 --count=<count>::
0026 Event period to sample.
0027
0028 -C <cpu-list>::
0029 --cpu=<cpu>::
0030 Monitor only on the list of CPUs provided. Multiple CPUs can be provided as a
0031 comma-separated list with no space: 0,1. Ranges of CPUs are specified with -: 0-2.
0032 Default is to monitor all CPUS.
0033
0034 -d <seconds>::
0035 --delay=<seconds>::
0036 Number of seconds to delay between refreshes.
0037
0038 -e <event>::
0039 --event=<event>::
0040 Select the PMU event. Selection can be a symbolic event name
0041 (use 'perf list' to list all events) or a raw PMU event in the form
0042 of rN where N is a hexadecimal value that represents the raw register
0043 encoding with the layout of the event control registers as described
0044 by entries in /sys/bus/event_source/devices/cpu/format/*.
0045
0046 -E <entries>::
0047 --entries=<entries>::
0048 Display this many functions.
0049
0050 -f <count>::
0051 --count-filter=<count>::
0052 Only display functions with more events than this.
0053
0054 --group::
0055 Put the counters into a counter group.
0056
0057 --group-sort-idx::
0058 Sort the output by the event at the index n in group. If n is invalid,
0059 sort by the first event. It can support multiple groups with different
0060 amount of events. WARNING: This should be used on grouped events.
0061
0062 -F <freq>::
0063 --freq=<freq>::
0064 Profile at this frequency. Use 'max' to use the currently maximum
0065 allowed frequency, i.e. the value in the kernel.perf_event_max_sample_rate
0066 sysctl.
0067
0068 -i::
0069 --inherit::
0070 Child tasks do not inherit counters.
0071
0072 -k <path>::
0073 --vmlinux=<path>::
0074 Path to vmlinux. Required for annotation functionality.
0075
0076 --ignore-vmlinux::
0077 Ignore vmlinux files.
0078
0079 --kallsyms=<file>::
0080 kallsyms pathname
0081
0082 -m <pages>::
0083 --mmap-pages=<pages>::
0084 Number of mmap data pages (must be a power of two) or size
0085 specification with appended unit character - B/K/M/G. The
0086 size is rounded up to have nearest pages power of two value.
0087
0088 -p <pid>::
0089 --pid=<pid>::
0090 Profile events on existing Process ID (comma separated list).
0091
0092 -t <tid>::
0093 --tid=<tid>::
0094 Profile events on existing thread ID (comma separated list).
0095
0096 -u::
0097 --uid=::
0098 Record events in threads owned by uid. Name or number.
0099
0100 -r <priority>::
0101 --realtime=<priority>::
0102 Collect data with this RT SCHED_FIFO priority.
0103
0104 --sym-annotate=<symbol>::
0105 Annotate this symbol.
0106
0107 -K::
0108 --hide_kernel_symbols::
0109 Hide kernel symbols.
0110
0111 -U::
0112 --hide_user_symbols::
0113 Hide user symbols.
0114
0115 --demangle-kernel::
0116 Demangle kernel symbols.
0117
0118 -D::
0119 --dump-symtab::
0120 Dump the symbol table used for profiling.
0121
0122 -v::
0123 --verbose::
0124 Be more verbose (show counter open errors, etc).
0125
0126 -z::
0127 --zero::
0128 Zero history across display updates.
0129
0130 -s::
0131 --sort::
0132 Sort by key(s): pid, comm, dso, symbol, parent, srcline, weight,
0133 local_weight, abort, in_tx, transaction, overhead, sample, period.
0134 Please see description of --sort in the perf-report man page.
0135
0136 --fields=::
0137 Specify output field - multiple keys can be specified in CSV format.
0138 Following fields are available:
0139 overhead, overhead_sys, overhead_us, overhead_children, sample and period.
0140 Also it can contain any sort key(s).
0141
0142 By default, every sort keys not specified in --field will be appended
0143 automatically.
0144
0145 -n::
0146 --show-nr-samples::
0147 Show a column with the number of samples.
0148
0149 --show-total-period::
0150 Show a column with the sum of periods.
0151
0152 --dsos::
0153 Only consider symbols in these dsos. This option will affect the
0154 percentage of the overhead column. See --percentage for more info.
0155
0156 --comms::
0157 Only consider symbols in these comms. This option will affect the
0158 percentage of the overhead column. See --percentage for more info.
0159
0160 --symbols::
0161 Only consider these symbols. This option will affect the
0162 percentage of the overhead column. See --percentage for more info.
0163
0164 -M::
0165 --disassembler-style=:: Set disassembler style for objdump.
0166
0167 --prefix=PREFIX::
0168 --prefix-strip=N::
0169 Remove first N entries from source file path names in executables
0170 and add PREFIX. This allows to display source code compiled on systems
0171 with different file system layout.
0172
0173 --source::
0174 Interleave source code with assembly code. Enabled by default,
0175 disable with --no-source.
0176
0177 --asm-raw::
0178 Show raw instruction encoding of assembly instructions.
0179
0180 -g::
0181 Enables call-graph (stack chain/backtrace) recording.
0182
0183 --call-graph [mode,type,min[,limit],order[,key][,branch]]::
0184 Setup and enable call-graph (stack chain/backtrace) recording,
0185 implies -g. See `--call-graph` section in perf-record and
0186 perf-report man pages for details.
0187
0188 --children::
0189 Accumulate callchain of children to parent entry so that then can
0190 show up in the output. The output will have a new "Children" column
0191 and will be sorted on the data. It requires -g/--call-graph option
0192 enabled. See the `overhead calculation' section for more details.
0193 Enabled by default, disable with --no-children.
0194
0195 --max-stack::
0196 Set the stack depth limit when parsing the callchain, anything
0197 beyond the specified depth will be ignored. This is a trade-off
0198 between information loss and faster processing especially for
0199 workloads that can have a very long callchain stack.
0200
0201 Default: /proc/sys/kernel/perf_event_max_stack when present, 127 otherwise.
0202
0203 --ignore-callees=<regex>::
0204 Ignore callees of the function(s) matching the given regex.
0205 This has the effect of collecting the callers of each such
0206 function into one place in the call-graph tree.
0207
0208 --percent-limit::
0209 Do not show entries which have an overhead under that percent.
0210 (Default: 0).
0211
0212 --percentage::
0213 Determine how to display the overhead percentage of filtered entries.
0214 Filters can be applied by --comms, --dsos and/or --symbols options and
0215 Zoom operations on the TUI (thread, dso, etc).
0216
0217 "relative" means it's relative to filtered entries only so that the
0218 sum of shown entries will be always 100%. "absolute" means it retains
0219 the original value before and after the filter is applied.
0220
0221 -w::
0222 --column-widths=<width[,width...]>::
0223 Force each column width to the provided list, for large terminal
0224 readability. 0 means no limit (default behavior).
0225
0226 --proc-map-timeout::
0227 When processing pre-existing threads /proc/XXX/mmap, it may take
0228 a long time, because the file may be huge. A time out is needed
0229 in such cases.
0230 This option sets the time out limit. The default value is 500 ms.
0231
0232
0233 -b::
0234 --branch-any::
0235 Enable taken branch stack sampling. Any type of taken branch may be sampled.
0236 This is a shortcut for --branch-filter any. See --branch-filter for more infos.
0237
0238 -j::
0239 --branch-filter::
0240 Enable taken branch stack sampling. Each sample captures a series of consecutive
0241 taken branches. The number of branches captured with each sample depends on the
0242 underlying hardware, the type of branches of interest, and the executed code.
0243 It is possible to select the types of branches captured by enabling filters.
0244 For a full list of modifiers please see the perf record manpage.
0245
0246 The option requires at least one branch type among any, any_call, any_ret, ind_call, cond.
0247 The privilege levels may be omitted, in which case, the privilege levels of the associated
0248 event are applied to the branch filter. Both kernel (k) and hypervisor (hv) privilege
0249 levels are subject to permissions. When sampling on multiple events, branch stack sampling
0250 is enabled for all the sampling events. The sampled branch type is the same for all events.
0251 The various filters must be specified as a comma separated list: --branch-filter any_ret,u,k
0252 Note that this feature may not be available on all processors.
0253
0254 --raw-trace::
0255 When displaying traceevent output, do not use print fmt or plugins.
0256
0257 --hierarchy::
0258 Enable hierarchy output.
0259
0260 --overwrite::
0261 Enable this to use just the most recent records, which helps in high core count
0262 machines such as Knights Landing/Mill, but right now is disabled by default as
0263 the pausing used in this technique is leading to loss of metadata events such
0264 as PERF_RECORD_MMAP which makes 'perf top' unable to resolve samples, leading
0265 to lots of unknown samples appearing on the UI. Enable this if you are in such
0266 machines and profiling a workload that doesn't creates short lived threads and/or
0267 doesn't uses many executable mmap operations. Work is being planed to solve
0268 this situation, till then, this will remain disabled by default.
0269
0270 --force::
0271 Don't do ownership validation.
0272
0273 --num-thread-synthesize::
0274 The number of threads to run when synthesizing events for existing processes.
0275 By default, the number of threads equals to the number of online CPUs.
0276
0277 --namespaces::
0278 Record events of type PERF_RECORD_NAMESPACES and display it with the
0279 'cgroup_id' sort key.
0280
0281 -G name::
0282 --cgroup name::
0283 monitor only in the container (cgroup) called "name". This option is available only
0284 in per-cpu mode. The cgroup filesystem must be mounted. All threads belonging to
0285 container "name" are monitored when they run on the monitored CPUs. Multiple cgroups
0286 can be provided. Each cgroup is applied to the corresponding event, i.e., first cgroup
0287 to first event, second cgroup to second event and so on. It is possible to provide
0288 an empty cgroup (monitor all the time) using, e.g., -G foo,,bar. Cgroups must have
0289 corresponding events, i.e., they always refer to events defined earlier on the command
0290 line. If the user wants to track multiple events for a specific cgroup, the user can
0291 use '-e e1 -e e2 -G foo,foo' or just use '-e e1 -e e2 -G foo'.
0292
0293 --all-cgroups::
0294 Record events of type PERF_RECORD_CGROUP and display it with the
0295 'cgroup' sort key.
0296
0297 --switch-on EVENT_NAME::
0298 Only consider events after this event is found.
0299
0300 E.g.:
0301
0302 Find out where broadcast packets are handled
0303
0304 perf probe -L icmp_rcv
0305
0306 Insert a probe there:
0307
0308 perf probe icmp_rcv:59
0309
0310 Start perf top and ask it to only consider the cycles events when a
0311 broadcast packet arrives This will show a menu with two entries and
0312 will start counting when a broadcast packet arrives:
0313
0314 perf top -e cycles,probe:icmp_rcv --switch-on=probe:icmp_rcv
0315
0316 Alternatively one can ask for --group and then two overhead columns
0317 will appear, the first for cycles and the second for the switch-on event.
0318
0319 perf top --group -e cycles,probe:icmp_rcv --switch-on=probe:icmp_rcv
0320
0321 This may be interesting to measure a workload only after some initialization
0322 phase is over, i.e. insert a perf probe at that point and use the above
0323 examples replacing probe:icmp_rcv with the just-after-init probe.
0324
0325 --switch-off EVENT_NAME::
0326 Stop considering events after this event is found.
0327
0328 --show-on-off-events::
0329 Show the --switch-on/off events too. This has no effect in 'perf top' now
0330 but probably we'll make the default not to show the switch-on/off events
0331 on the --group mode and if there is only one event besides the off/on ones,
0332 go straight to the histogram browser, just like 'perf top' with no events
0333 explicitly specified does.
0334
0335 --stitch-lbr::
0336 Show callgraph with stitched LBRs, which may have more complete
0337 callgraph. The option must be used with --call-graph lbr recording.
0338 Disabled by default. In common cases with call stack overflows,
0339 it can recreate better call stacks than the default lbr call stack
0340 output. But this approach is not full proof. There can be cases
0341 where it creates incorrect call stacks from incorrect matches.
0342 The known limitations include exception handing such as
0343 setjmp/longjmp will have calls/returns not match.
0344
0345 ifdef::HAVE_LIBPFM[]
0346 --pfm-events events::
0347 Select a PMU event using libpfm4 syntax (see http://perfmon2.sf.net)
0348 including support for event filters. For example '--pfm-events
0349 inst_retired:any_p:u:c=1:i'. More than one event can be passed to the
0350 option using the comma separator. Hardware events and generic hardware
0351 events cannot be mixed together. The latter must be used with the -e
0352 option. The -e option and this one can be mixed and matched. Events
0353 can be grouped using the {} notation.
0354 endif::HAVE_LIBPFM[]
0355
0356 INTERACTIVE PROMPTING KEYS
0357 --------------------------
0358
0359 [d]::
0360 Display refresh delay.
0361
0362 [e]::
0363 Number of entries to display.
0364
0365 [E]::
0366 Event to display when multiple counters are active.
0367
0368 [f]::
0369 Profile display filter (>= hit count).
0370
0371 [F]::
0372 Annotation display filter (>= % of total).
0373
0374 [s]::
0375 Annotate symbol.
0376
0377 [S]::
0378 Stop annotation, return to full profile display.
0379
0380 [K]::
0381 Hide kernel symbols.
0382
0383 [U]::
0384 Hide user symbols.
0385
0386 [z]::
0387 Toggle event count zeroing across display updates.
0388
0389 [qQ]::
0390 Quit.
0391
0392 Pressing any unmapped key displays a menu, and prompts for input.
0393
0394 include::callchain-overhead-calculation.txt[]
0395
0396 SEE ALSO
0397 --------
0398 linkperf:perf-stat[1], linkperf:perf-list[1], linkperf:perf-report[1]
