perf-top.txt 13 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398
  1. perf-top(1)
  2. ===========
  3. NAME
  4. ----
  5. perf-top - System profiling tool.
  6. SYNOPSIS
  7. --------
  8. [verse]
  9. 'perf top' [-e <EVENT> | --event=EVENT] [<options>]
  10. DESCRIPTION
  11. -----------
  12. This command generates and displays a performance counter profile in real time.
  13. OPTIONS
  14. -------
  15. -a::
  16. --all-cpus::
  17. System-wide collection. (default)
  18. -c <count>::
  19. --count=<count>::
  20. Event period to sample.
  21. -C <cpu-list>::
  22. --cpu=<cpu>::
  23. Monitor only on the list of CPUs provided. Multiple CPUs can be provided as a
  24. comma-separated list with no space: 0,1. Ranges of CPUs are specified with -: 0-2.
  25. Default is to monitor all CPUS.
  26. -d <seconds>::
  27. --delay=<seconds>::
  28. Number of seconds to delay between refreshes.
  29. -e <event>::
  30. --event=<event>::
  31. Select the PMU event. Selection can be a symbolic event name
  32. (use 'perf list' to list all events) or a raw PMU event in the form
  33. of rN where N is a hexadecimal value that represents the raw register
  34. encoding with the layout of the event control registers as described
  35. by entries in /sys/bus/event_source/devices/cpu/format/*.
  36. -E <entries>::
  37. --entries=<entries>::
  38. Display this many functions.
  39. -f <count>::
  40. --count-filter=<count>::
  41. Only display functions with more events than this.
  42. --group::
  43. Put the counters into a counter group.
  44. --group-sort-idx::
  45. Sort the output by the event at the index n in group. If n is invalid,
  46. sort by the first event. It can support multiple groups with different
  47. amount of events. WARNING: This should be used on grouped events.
  48. -F <freq>::
  49. --freq=<freq>::
  50. Profile at this frequency. Use 'max' to use the currently maximum
  51. allowed frequency, i.e. the value in the kernel.perf_event_max_sample_rate
  52. sysctl.
  53. -i::
  54. --inherit::
  55. Child tasks do not inherit counters.
  56. -k <path>::
  57. --vmlinux=<path>::
  58. Path to vmlinux. Required for annotation functionality.
  59. --ignore-vmlinux::
  60. Ignore vmlinux files.
  61. --kallsyms=<file>::
  62. kallsyms pathname
  63. -m <pages>::
  64. --mmap-pages=<pages>::
  65. Number of mmap data pages (must be a power of two) or size
  66. specification with appended unit character - B/K/M/G. The
  67. size is rounded up to have nearest pages power of two value.
  68. -p <pid>::
  69. --pid=<pid>::
  70. Profile events on existing Process ID (comma separated list).
  71. -t <tid>::
  72. --tid=<tid>::
  73. Profile events on existing thread ID (comma separated list).
  74. -u::
  75. --uid=::
  76. Record events in threads owned by uid. Name or number.
  77. -r <priority>::
  78. --realtime=<priority>::
  79. Collect data with this RT SCHED_FIFO priority.
  80. --sym-annotate=<symbol>::
  81. Annotate this symbol.
  82. -K::
  83. --hide_kernel_symbols::
  84. Hide kernel symbols.
  85. -U::
  86. --hide_user_symbols::
  87. Hide user symbols.
  88. --demangle-kernel::
  89. Demangle kernel symbols.
  90. -D::
  91. --dump-symtab::
  92. Dump the symbol table used for profiling.
  93. -v::
  94. --verbose::
  95. Be more verbose (show counter open errors, etc).
  96. -z::
  97. --zero::
  98. Zero history across display updates.
  99. -s::
  100. --sort::
  101. Sort by key(s): pid, comm, dso, symbol, parent, srcline, weight,
  102. local_weight, abort, in_tx, transaction, overhead, sample, period.
  103. Please see description of --sort in the perf-report man page.
  104. --fields=::
  105. Specify output field - multiple keys can be specified in CSV format.
  106. Following fields are available:
  107. overhead, overhead_sys, overhead_us, overhead_children, sample and period.
  108. Also it can contain any sort key(s).
  109. By default, every sort keys not specified in --field will be appended
  110. automatically.
  111. -n::
  112. --show-nr-samples::
  113. Show a column with the number of samples.
  114. --show-total-period::
  115. Show a column with the sum of periods.
  116. --dsos::
  117. Only consider symbols in these dsos. This option will affect the
  118. percentage of the overhead column. See --percentage for more info.
  119. --comms::
  120. Only consider symbols in these comms. This option will affect the
  121. percentage of the overhead column. See --percentage for more info.
  122. --symbols::
  123. Only consider these symbols. This option will affect the
  124. percentage of the overhead column. See --percentage for more info.
  125. -M::
  126. --disassembler-style=:: Set disassembler style for objdump.
  127. --prefix=PREFIX::
  128. --prefix-strip=N::
  129. Remove first N entries from source file path names in executables
  130. and add PREFIX. This allows to display source code compiled on systems
  131. with different file system layout.
  132. --source::
  133. Interleave source code with assembly code. Enabled by default,
  134. disable with --no-source.
  135. --asm-raw::
  136. Show raw instruction encoding of assembly instructions.
  137. -g::
  138. Enables call-graph (stack chain/backtrace) recording.
  139. --call-graph [mode,type,min[,limit],order[,key][,branch]]::
  140. Setup and enable call-graph (stack chain/backtrace) recording,
  141. implies -g. See `--call-graph` section in perf-record and
  142. perf-report man pages for details.
  143. --children::
  144. Accumulate callchain of children to parent entry so that then can
  145. show up in the output. The output will have a new "Children" column
  146. and will be sorted on the data. It requires -g/--call-graph option
  147. enabled. See the `overhead calculation' section for more details.
  148. Enabled by default, disable with --no-children.
  149. --max-stack::
  150. Set the stack depth limit when parsing the callchain, anything
  151. beyond the specified depth will be ignored. This is a trade-off
  152. between information loss and faster processing especially for
  153. workloads that can have a very long callchain stack.
  154. Default: /proc/sys/kernel/perf_event_max_stack when present, 127 otherwise.
  155. --ignore-callees=<regex>::
  156. Ignore callees of the function(s) matching the given regex.
  157. This has the effect of collecting the callers of each such
  158. function into one place in the call-graph tree.
  159. --percent-limit::
  160. Do not show entries which have an overhead under that percent.
  161. (Default: 0).
  162. --percentage::
  163. Determine how to display the overhead percentage of filtered entries.
  164. Filters can be applied by --comms, --dsos and/or --symbols options and
  165. Zoom operations on the TUI (thread, dso, etc).
  166. "relative" means it's relative to filtered entries only so that the
  167. sum of shown entries will be always 100%. "absolute" means it retains
  168. the original value before and after the filter is applied.
  169. -w::
  170. --column-widths=<width[,width...]>::
  171. Force each column width to the provided list, for large terminal
  172. readability. 0 means no limit (default behavior).
  173. --proc-map-timeout::
  174. When processing pre-existing threads /proc/XXX/mmap, it may take
  175. a long time, because the file may be huge. A time out is needed
  176. in such cases.
  177. This option sets the time out limit. The default value is 500 ms.
  178. -b::
  179. --branch-any::
  180. Enable taken branch stack sampling. Any type of taken branch may be sampled.
  181. This is a shortcut for --branch-filter any. See --branch-filter for more infos.
  182. -j::
  183. --branch-filter::
  184. Enable taken branch stack sampling. Each sample captures a series of consecutive
  185. taken branches. The number of branches captured with each sample depends on the
  186. underlying hardware, the type of branches of interest, and the executed code.
  187. It is possible to select the types of branches captured by enabling filters.
  188. For a full list of modifiers please see the perf record manpage.
  189. The option requires at least one branch type among any, any_call, any_ret, ind_call, cond.
  190. The privilege levels may be omitted, in which case, the privilege levels of the associated
  191. event are applied to the branch filter. Both kernel (k) and hypervisor (hv) privilege
  192. levels are subject to permissions. When sampling on multiple events, branch stack sampling
  193. is enabled for all the sampling events. The sampled branch type is the same for all events.
  194. The various filters must be specified as a comma separated list: --branch-filter any_ret,u,k
  195. Note that this feature may not be available on all processors.
  196. --raw-trace::
  197. When displaying traceevent output, do not use print fmt or plugins.
  198. --hierarchy::
  199. Enable hierarchy output.
  200. --overwrite::
  201. Enable this to use just the most recent records, which helps in high core count
  202. machines such as Knights Landing/Mill, but right now is disabled by default as
  203. the pausing used in this technique is leading to loss of metadata events such
  204. as PERF_RECORD_MMAP which makes 'perf top' unable to resolve samples, leading
  205. to lots of unknown samples appearing on the UI. Enable this if you are in such
  206. machines and profiling a workload that doesn't creates short lived threads and/or
  207. doesn't uses many executable mmap operations. Work is being planed to solve
  208. this situation, till then, this will remain disabled by default.
  209. --force::
  210. Don't do ownership validation.
  211. --num-thread-synthesize::
  212. The number of threads to run when synthesizing events for existing processes.
  213. By default, the number of threads equals to the number of online CPUs.
  214. --namespaces::
  215. Record events of type PERF_RECORD_NAMESPACES and display it with the
  216. 'cgroup_id' sort key.
  217. -G name::
  218. --cgroup name::
  219. monitor only in the container (cgroup) called "name". This option is available only
  220. in per-cpu mode. The cgroup filesystem must be mounted. All threads belonging to
  221. container "name" are monitored when they run on the monitored CPUs. Multiple cgroups
  222. can be provided. Each cgroup is applied to the corresponding event, i.e., first cgroup
  223. to first event, second cgroup to second event and so on. It is possible to provide
  224. an empty cgroup (monitor all the time) using, e.g., -G foo,,bar. Cgroups must have
  225. corresponding events, i.e., they always refer to events defined earlier on the command
  226. line. If the user wants to track multiple events for a specific cgroup, the user can
  227. use '-e e1 -e e2 -G foo,foo' or just use '-e e1 -e e2 -G foo'.
  228. --all-cgroups::
  229. Record events of type PERF_RECORD_CGROUP and display it with the
  230. 'cgroup' sort key.
  231. --switch-on EVENT_NAME::
  232. Only consider events after this event is found.
  233. E.g.:
  234. Find out where broadcast packets are handled
  235. perf probe -L icmp_rcv
  236. Insert a probe there:
  237. perf probe icmp_rcv:59
  238. Start perf top and ask it to only consider the cycles events when a
  239. broadcast packet arrives This will show a menu with two entries and
  240. will start counting when a broadcast packet arrives:
  241. perf top -e cycles,probe:icmp_rcv --switch-on=probe:icmp_rcv
  242. Alternatively one can ask for --group and then two overhead columns
  243. will appear, the first for cycles and the second for the switch-on event.
  244. perf top --group -e cycles,probe:icmp_rcv --switch-on=probe:icmp_rcv
  245. This may be interesting to measure a workload only after some initialization
  246. phase is over, i.e. insert a perf probe at that point and use the above
  247. examples replacing probe:icmp_rcv with the just-after-init probe.
  248. --switch-off EVENT_NAME::
  249. Stop considering events after this event is found.
  250. --show-on-off-events::
  251. Show the --switch-on/off events too. This has no effect in 'perf top' now
  252. but probably we'll make the default not to show the switch-on/off events
  253. on the --group mode and if there is only one event besides the off/on ones,
  254. go straight to the histogram browser, just like 'perf top' with no events
  255. explicitly specified does.
  256. --stitch-lbr::
  257. Show callgraph with stitched LBRs, which may have more complete
  258. callgraph. The option must be used with --call-graph lbr recording.
  259. Disabled by default. In common cases with call stack overflows,
  260. it can recreate better call stacks than the default lbr call stack
  261. output. But this approach is not full proof. There can be cases
  262. where it creates incorrect call stacks from incorrect matches.
  263. The known limitations include exception handing such as
  264. setjmp/longjmp will have calls/returns not match.
  265. ifdef::HAVE_LIBPFM[]
  266. --pfm-events events::
  267. Select a PMU event using libpfm4 syntax (see http://perfmon2.sf.net)
  268. including support for event filters. For example '--pfm-events
  269. inst_retired:any_p:u:c=1:i'. More than one event can be passed to the
  270. option using the comma separator. Hardware events and generic hardware
  271. events cannot be mixed together. The latter must be used with the -e
  272. option. The -e option and this one can be mixed and matched. Events
  273. can be grouped using the {} notation.
  274. endif::HAVE_LIBPFM[]
  275. INTERACTIVE PROMPTING KEYS
  276. --------------------------
  277. [d]::
  278. Display refresh delay.
  279. [e]::
  280. Number of entries to display.
  281. [E]::
  282. Event to display when multiple counters are active.
  283. [f]::
  284. Profile display filter (>= hit count).
  285. [F]::
  286. Annotation display filter (>= % of total).
  287. [s]::
  288. Annotate symbol.
  289. [S]::
  290. Stop annotation, return to full profile display.
  291. [K]::
  292. Hide kernel symbols.
  293. [U]::
  294. Hide user symbols.
  295. [z]::
  296. Toggle event count zeroing across display updates.
  297. [qQ]::
  298. Quit.
  299. Pressing any unmapped key displays a menu, and prompts for input.
  300. include::callchain-overhead-calculation.txt[]
  301. SEE ALSO
  302. --------
  303. linkperf:perf-stat[1], linkperf:perf-list[1], linkperf:perf-report[1]