6 perf-report - Read perf.data (created by perf record) and display the profile
11 'perf report' [-i <file> | --input=file]
15 This command displays the performance counter profile information recorded
22 Input file name. (default: perf.data unless stdin is a fifo)
26 Be more verbose. (show symbol address, etc)
30 Show the number of samples for each symbol
32 --showcpuutilization::
33 Show sample percentage for different cpu modes.
37 Show per-thread event counters. The input data file should be recorded
41 Only consider symbols in these comms. CSV that understands
42 file://filename entries. This option will affect the percentage of
43 the overhead column. See --percentage for more info.
45 Only show events for given process ID (comma separated list).
48 Only show events for given thread ID (comma separated list).
51 Only consider symbols in these dsos. CSV that understands
52 file://filename entries. This option will affect the percentage of
53 the overhead column. See --percentage for more info.
56 Only consider these symbols. CSV that understands
57 file://filename entries. This option will affect the percentage of
58 the overhead column. See --percentage for more info.
61 Only show symbols that match (partially) with this filter.
65 Only display entries resolved to a symbol.
69 Sort histogram entries by given key(s) - multiple keys can be specified
70 in CSV format. Following sort keys are available:
71 pid, comm, dso, symbol, parent, cpu, srcline, weight, local_weight.
73 Each key has following meaning:
75 - comm: command (name) of the task which can be read via /proc/<pid>/comm
76 - pid: command and tid of the task
77 - dso: name of library or module executed at the time of sample
78 - symbol: name of function executed at the time of sample
79 - parent: name of function matched to the parent regex filter. Unmatched
80 entries are displayed as "[other]".
81 - cpu: cpu number the task ran at the time of sample
82 - srcline: filename and line number executed at the time of sample. The
83 DWARF debugging info must be provided.
84 - weight: Event specific weight, e.g. memory latency or transaction
85 abort cost. This is the global weight.
86 - local_weight: Local weight version of the weight above.
87 - transaction: Transaction abort flags.
88 - overhead: Overhead percentage of sample
89 - overhead_sys: Overhead percentage of sample running in system mode
90 - overhead_us: Overhead percentage of sample running in user mode
91 - overhead_guest_sys: Overhead percentage of sample running in system mode
93 - overhead_guest_us: Overhead percentage of sample running in user mode on
95 - sample: Number of sample
96 - period: Raw number of event count of sample
98 By default, comm, dso and symbol keys are used.
99 (i.e. --sort comm,dso,symbol)
101 If --branch-stack option is used, following sort keys are also
103 dso_from, dso_to, symbol_from, symbol_to, mispredict.
105 - dso_from: name of library or module branched from
106 - dso_to: name of library or module branched to
107 - symbol_from: name of function branched from
108 - symbol_to: name of function branched to
109 - mispredict: "N" for predicted branch, "Y" for mispredicted branch
110 - in_tx: branch in TSX transaction
111 - abort: TSX transaction abort.
113 And default sort keys are changed to comm, dso_from, symbol_from, dso_to
114 and symbol_to, see '--branch-stack'.
118 Specify output field - multiple keys can be specified in CSV format.
119 Following fields are available:
120 overhead, overhead_sys, overhead_us, overhead_children, sample and period.
121 Also it can contain any sort key(s).
123 By default, every sort keys not specified in -F will be appended
126 If --mem-mode option is used, following sort keys are also available
127 (incompatible with --branch-stack):
128 symbol_daddr, dso_daddr, locked, tlb, mem, snoop, dcacheline.
130 - symbol_daddr: name of data symbol being executed on at the time of sample
131 - dso_daddr: name of library or module containing the data being executed
132 on at the time of sample
133 - locked: whether the bus was locked at the time of sample
134 - tlb: type of tlb access for the data at the time of sample
135 - mem: type of memory access for the data at the time of sample
136 - snoop: type of snoop (if any) for the data at the time of sample
137 - dcacheline: the cacheline the data address is on at the time of sample
139 And default sort keys are changed to local_weight, mem, sym, dso,
140 symbol_daddr, dso_daddr, snoop, tlb, locked, see '--mem-mode'.
144 A regex filter to identify parent. The parent is a caller of this
145 function and searched through the callchain, thus it requires callchain
146 information recorded. The pattern is in the exteneded regex format and
147 defaults to "\^sys_|^do_page_fault", see '--sort parent'.
151 Only display entries with parent-match.
154 --column-widths=<width[,width...]>::
155 Force each column width to the provided list, for large terminal
156 readability. 0 means no limit (default behavior).
160 Use a special separator character and don't pad with spaces, replacing
161 all occurrences of this separator in symbol names (and other output)
162 with a '.' character, that thus it's the only non valid separator.
166 Dump raw trace in ASCII.
168 -g [type,min[,limit],order[,key][,branch]]::
170 Display call chains using type, min percent threshold, optional print
173 - flat: single column, linear exposure of call chains.
174 - graph: use a graph tree, displaying absolute overhead rates.
175 - fractal: like graph, but displays relative rates. Each branch of
176 the tree is considered as a new profiled object. +
179 - callee: callee based call graph.
180 - caller: inverted caller based call graph.
183 - function: compare on functions
184 - address: compare on individual code addresses
187 - branch: include last branch information in callgraph
188 when available. Usually more convenient to use --branch-history
191 Default: fractal,0.5,callee,function.
194 Accumulate callchain of children to parent entry so that then can
195 show up in the output. The output will have a new "Children" column
196 and will be sorted on the data. It requires callchains are recorded.
197 See the `overhead calculation' section for more details.
200 Set the stack depth limit when parsing the callchain, anything
201 beyond the specified depth will be ignored. This is a trade-off
202 between information loss and faster processing especially for
203 workloads that can have a very long callchain stack.
209 alias for inverted caller based call graph.
211 --ignore-callees=<regex>::
212 Ignore callees of the function(s) matching the given regex.
213 This has the effect of collecting the callers of each such
214 function into one place in the call-graph tree.
217 Pretty printing style. key: normal, raw
219 --stdio:: Use the stdio interface.
221 --tui:: Use the TUI interface, that is integrated with annotate and allows
222 zooming into DSOs or threads, among other features. Use of --tui
223 requires a tty, if one is not present, as when piping to other
224 commands, the stdio interface is used.
226 --gtk:: Use the GTK2 interface.
237 Load module symbols. WARNING: This should only be used with -k and
242 Don't complain, do it.
244 --symfs=<directory>::
245 Look for files with symbols relative to this directory.
248 --cpu:: Only report samples for the list of CPUs provided. Multiple CPUs can
249 be provided as a comma-separated list with no space: 0,1. Ranges of
250 CPUs are specified with -: 0-2. Default is to report samples on all
254 --disassembler-style=:: Set disassembler style for objdump.
257 Interleave source code with assembly code. Enabled by default,
258 disable with --no-source.
261 Show raw instruction encoding of assembly instructions.
263 --show-total-period:: Show a column with the sum of periods.
267 Display extended information about the perf.data file. This adds
268 information which may be very large and thus may clutter the display.
269 It currently includes: cpu and numa topology of the host system.
273 Use the addresses of sampled taken branches instead of the instruction
274 address to build the histograms. To generate meaningful output, the
275 perf.data file must have been obtained using perf record -b or
276 perf record --branch-filter xxx where xxx is a branch filter option.
277 perf report is able to auto-detect whether a perf.data file contains
278 branch stacks and it will automatically switch to the branch view mode,
279 unless --no-branch-stack is used.
282 Add the addresses of sampled taken branches to the callstack.
283 This allows to examine the path the program took to each sample.
284 The data collection must have used -b (or -j) and -g.
287 Path to objdump binary.
290 Show event group information together.
293 Demangle symbol names to human readable form. It's enabled by default,
294 disable with --no-demangle.
297 Demangle kernel symbol names to human readable form (for C++ kernels).
300 Use the data addresses of samples in addition to instruction addresses
301 to build the histograms. To generate meaningful output, the perf.data
302 file must have been obtained using perf record -d -W and using a
303 special event -e cpu/mem-loads/ or -e cpu/mem-stores/. See
304 'perf mem' for simpler access.
307 Do not show entries which have an overhead under that percent.
311 Determine how to display the overhead percentage of filtered entries.
312 Filters can be applied by --comms, --dsos and/or --symbols options and
313 Zoom operations on the TUI (thread, dso, etc).
315 "relative" means it's relative to filtered entries only so that the
316 sum of shown entries will be always 100%. "absolute" means it retains
317 the original value before and after the filter is applied.
320 Show header information in the perf.data file. This includes
321 various information like hostname, OS and perf version, cpu/mem
322 info, perf command line, event list and so on. Currently only
323 --stdio output supports this feature.
326 Show only perf.data header (forces --stdio).
329 Options for decoding instruction tracing data. The options are:
331 i synthesize instructions events
332 b synthesize branches events
333 c synthesize branches events (calls only)
334 r synthesize branches events (returns only)
335 x synthesize transactions events
336 e synthesize error events
338 g synthesize a call chain (use with i or x)
340 The default is all events i.e. the same as --itrace=ibxe
342 In addition, the period (default 100000) for instructions events
343 can be specified in units of:
349 ns nanoseconds (default)
351 Also the call chain size (default 16, max. 1024) for instructions or
352 transactions events can be specified.
354 To disable decoding entirely, use --no-itrace.
357 include::callchain-overhead-calculation.txt[]
361 linkperf:perf-stat[1], linkperf:perf-annotate[1]