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
40 Only consider symbols in these comms. CSV that understands
41 file://filename entries. This option will affect the percentage of
42 the overhead column. See --percentage for more info.
45 Only consider symbols in these dsos. CSV that understands
46 file://filename entries. This option will affect the percentage of
47 the overhead column. See --percentage for more info.
50 Only consider these symbols. CSV that understands
51 file://filename entries. This option will affect the percentage of
52 the overhead column. See --percentage for more info.
55 Only show symbols that match (partially) with this filter.
59 Only display entries resolved to a symbol.
63 Sort histogram entries by given key(s) - multiple keys can be specified
64 in CSV format. Following sort keys are available:
65 pid, comm, dso, symbol, parent, cpu, srcline, weight, local_weight.
67 Each key has following meaning:
69 - comm: command (name) of the task which can be read via /proc/<pid>/comm
70 - pid: command and tid of the task
71 - dso: name of library or module executed at the time of sample
72 - symbol: name of function executed at the time of sample
73 - parent: name of function matched to the parent regex filter. Unmatched
74 entries are displayed as "[other]".
75 - cpu: cpu number the task ran at the time of sample
76 - srcline: filename and line number executed at the time of sample. The
77 DWARF debugging info must be provided.
78 - weight: Event specific weight, e.g. memory latency or transaction
79 abort cost. This is the global weight.
80 - local_weight: Local weight version of the weight above.
81 - transaction: Transaction abort flags.
82 - overhead: Overhead percentage of sample
83 - overhead_sys: Overhead percentage of sample running in system mode
84 - overhead_us: Overhead percentage of sample running in user mode
85 - overhead_guest_sys: Overhead percentage of sample running in system mode
87 - overhead_guest_us: Overhead percentage of sample running in user mode on
89 - sample: Number of sample
90 - period: Raw number of event count of sample
92 By default, comm, dso and symbol keys are used.
93 (i.e. --sort comm,dso,symbol)
95 If --branch-stack option is used, following sort keys are also
97 dso_from, dso_to, symbol_from, symbol_to, mispredict.
99 - dso_from: name of library or module branched from
100 - dso_to: name of library or module branched to
101 - symbol_from: name of function branched from
102 - symbol_to: name of function branched to
103 - mispredict: "N" for predicted branch, "Y" for mispredicted branch
104 - in_tx: branch in TSX transaction
105 - abort: TSX transaction abort.
107 And default sort keys are changed to comm, dso_from, symbol_from, dso_to
108 and symbol_to, see '--branch-stack'.
112 Specify output field - multiple keys can be specified in CSV format.
113 Following fields are available:
114 overhead, overhead_sys, overhead_us, overhead_children, sample and period.
115 Also it can contain any sort key(s).
117 By default, every sort keys not specified in -F will be appended
120 If --mem-mode option is used, following sort keys are also available
121 (incompatible with --branch-stack):
122 symbol_daddr, dso_daddr, locked, tlb, mem, snoop, dcacheline.
124 - symbol_daddr: name of data symbol being executed on at the time of sample
125 - dso_daddr: name of library or module containing the data being executed
126 on at the time of sample
127 - locked: whether the bus was locked at the time of sample
128 - tlb: type of tlb access for the data at the time of sample
129 - mem: type of memory access for the data at the time of sample
130 - snoop: type of snoop (if any) for the data at the time of sample
131 - dcacheline: the cacheline the data address is on at the time of sample
133 And default sort keys are changed to local_weight, mem, sym, dso,
134 symbol_daddr, dso_daddr, snoop, tlb, locked, see '--mem-mode'.
138 A regex filter to identify parent. The parent is a caller of this
139 function and searched through the callchain, thus it requires callchain
140 information recorded. The pattern is in the exteneded regex format and
141 defaults to "\^sys_|^do_page_fault", see '--sort parent'.
145 Only display entries with parent-match.
148 --column-widths=<width[,width...]>::
149 Force each column width to the provided list, for large terminal
150 readability. 0 means no limit (default behavior).
154 Use a special separator character and don't pad with spaces, replacing
155 all occurrences of this separator in symbol names (and other output)
156 with a '.' character, that thus it's the only non valid separator.
160 Dump raw trace in ASCII.
162 -g [type,min[,limit],order[,key][,branch]]::
164 Display call chains using type, min percent threshold, optional print
167 - flat: single column, linear exposure of call chains.
168 - graph: use a graph tree, displaying absolute overhead rates.
169 - fractal: like graph, but displays relative rates. Each branch of
170 the tree is considered as a new profiled object. +
173 - callee: callee based call graph.
174 - caller: inverted caller based call graph.
177 - function: compare on functions
178 - address: compare on individual code addresses
181 - branch: include last branch information in callgraph
182 when available. Usually more convenient to use --branch-history
185 Default: fractal,0.5,callee,function.
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 callchains are recorded.
193 Set the stack depth limit when parsing the callchain, anything
194 beyond the specified depth will be ignored. This is a trade-off
195 between information loss and faster processing especially for
196 workloads that can have a very long callchain stack.
202 alias for inverted caller based call graph.
204 --ignore-callees=<regex>::
205 Ignore callees of the function(s) matching the given regex.
206 This has the effect of collecting the callers of each such
207 function into one place in the call-graph tree.
210 Pretty printing style. key: normal, raw
212 --stdio:: Use the stdio interface.
214 --tui:: Use the TUI interface, that is integrated with annotate and allows
215 zooming into DSOs or threads, among other features. Use of --tui
216 requires a tty, if one is not present, as when piping to other
217 commands, the stdio interface is used.
219 --gtk:: Use the GTK2 interface.
230 Load module symbols. WARNING: This should only be used with -k and
235 Don't complain, do it.
237 --symfs=<directory>::
238 Look for files with symbols relative to this directory.
241 --cpu:: Only report samples for the list of CPUs provided. Multiple CPUs can
242 be provided as a comma-separated list with no space: 0,1. Ranges of
243 CPUs are specified with -: 0-2. Default is to report samples on all
247 --disassembler-style=:: Set disassembler style for objdump.
250 Interleave source code with assembly code. Enabled by default,
251 disable with --no-source.
254 Show raw instruction encoding of assembly instructions.
256 --show-total-period:: Show a column with the sum of periods.
260 Display extended information about the perf.data file. This adds
261 information which may be very large and thus may clutter the display.
262 It currently includes: cpu and numa topology of the host system.
266 Use the addresses of sampled taken branches instead of the instruction
267 address to build the histograms. To generate meaningful output, the
268 perf.data file must have been obtained using perf record -b or
269 perf record --branch-filter xxx where xxx is a branch filter option.
270 perf report is able to auto-detect whether a perf.data file contains
271 branch stacks and it will automatically switch to the branch view mode,
272 unless --no-branch-stack is used.
275 Add the addresses of sampled taken branches to the callstack.
276 This allows to examine the path the program took to each sample.
277 The data collection must have used -b (or -j) and -g.
280 Path to objdump binary.
283 Show event group information together.
286 Demangle symbol names to human readable form. It's enabled by default,
287 disable with --no-demangle.
290 Demangle kernel symbol names to human readable form (for C++ kernels).
293 Use the data addresses of samples in addition to instruction addresses
294 to build the histograms. To generate meaningful output, the perf.data
295 file must have been obtained using perf record -d -W and using a
296 special event -e cpu/mem-loads/ or -e cpu/mem-stores/. See
297 'perf mem' for simpler access.
300 Do not show entries which have an overhead under that percent.
304 Determine how to display the overhead percentage of filtered entries.
305 Filters can be applied by --comms, --dsos and/or --symbols options and
306 Zoom operations on the TUI (thread, dso, etc).
308 "relative" means it's relative to filtered entries only so that the
309 sum of shown entries will be always 100%. "absolute" means it retains
310 the original value before and after the filter is applied.
313 Show header information in the perf.data file. This includes
314 various information like hostname, OS and perf version, cpu/mem
315 info, perf command line, event list and so on. Currently only
316 --stdio output supports this feature.
319 Show only perf.data header (forces --stdio).
323 linkperf:perf-stat[1], linkperf:perf-annotate[1]