6 perf-list - List all symbolic event types
11 'perf list' [--no-desc] [--long-desc]
12 [hw|sw|cache|tracepoint|pmu|sdt|metric|metricgroup|event_glob]
16 This command displays the symbolic event types which can be selected in the
17 various perf commands with the -e option.
23 Print extra event descriptions. (default)
26 Don't print descriptions.
30 Print longer event descriptions.
33 Enable debugging output.
36 Print how named events are resolved internally into perf events, and also
37 any extra expressions computed by perf stat.
40 Print deprecated events. By default the deprecated events are hidden.
43 Print PMU events and metrics limited to the specific PMU name.
44 (e.g. --unit cpu, --unit msr, --unit cpu_core, --unit cpu_atom)
48 Output in JSON format.
54 Events can optionally have a modifier by appending a colon and one or
55 more modifiers. Modifiers allow the user to restrict the events to be
56 counted. The following modifiers exist:
58 u - user-space counting
60 h - hypervisor counting
62 G - guest counting (in KVM guests)
63 H - host counting (not in KVM guests)
65 P - use maximum detected precise level
66 S - read sample value (PERF_SAMPLE_READ)
67 D - pin the event to the PMU
68 W - group is weak and will fallback to non-group if not schedulable,
69 e - group or event are exclusive and do not share the PMU
71 The 'p' modifier can be used for specifying how precise the instruction
72 address should be. The 'p' modifier can be specified multiple times:
74 0 - SAMPLE_IP can have arbitrary skid
75 1 - SAMPLE_IP must have constant skid
76 2 - SAMPLE_IP requested to have 0 skid
77 3 - SAMPLE_IP must have 0 skid, or uses randomization to avoid
78 sample shadowing effects.
80 For Intel systems precise event sampling is implemented with PEBS
81 which supports up to precise-level 2, and precise level 3 for
84 On AMD systems it is implemented using IBS (up to precise-level 2).
85 The precise modifier works with event types 0x76 (cpu-cycles, CPU
86 clocks not halted) and 0xC1 (micro-ops retired). Both events map to
87 IBS execution sampling (IBS op) with the IBS Op Counter Control bit
88 (IbsOpCntCtl) set respectively (see the
89 Core Complex (CCX) -> Processor x86 Core -> Instruction Based Sampling (IBS)
90 section of the [AMD Processor Programming Reference (PPR)] relevant to the
91 family, model and stepping of the processor being used).
93 Manual Volume 2: System Programming, 13.3 Instruction-Based
94 Sampling). Examples to use IBS:
96 perf record -a -e cpu-cycles:p ... # use ibs op counting cycles
97 perf record -a -e r076:p ... # same as -e cpu-cycles:p
98 perf record -a -e r0C1:p ... # use ibs op counting micro-ops
100 RAW HARDWARE EVENT DESCRIPTOR
101 -----------------------------
102 Even when an event is not available in a symbolic form within perf right now,
103 it can be encoded in a per processor specific way.
105 For instance on x86 CPUs, N is a hexadecimal value that represents the raw register encoding with the
106 layout of IA32_PERFEVTSELx MSRs (see [IntelĀ® 64 and IA-32 Architectures Software Developer's Manual Volume 3B: System Programming Guide] Figure 30-1 Layout
107 of IA32_PERFEVTSELx MSRs) or AMD's PERF_CTL MSRs (see the
108 Core Complex (CCX) -> Processor x86 Core -> MSR Registers section of the
109 [AMD Processor Programming Reference (PPR)] relevant to the family, model
110 and stepping of the processor being used).
112 Note: Only the following bit fields can be set in x86 counter
113 registers: event, umask, edge, inv, cmask. Esp. guest/host only and
114 OS/user mode flags must be setup using <<EVENT_MODIFIERS, EVENT
119 If the Intel docs for a QM720 Core i7 describe an event as:
121 Event Umask Event Mask
122 Num. Value Mnemonic Description Comment
124 A8H 01H LSD.UOPS Counts the number of micro-ops Use cmask=1 and
125 delivered by loop stream detector invert to count
128 raw encoding of 0x1A8 can be used:
130 perf stat -e r1a8 -a sleep 1
131 perf record -e r1a8 ...
133 It's also possible to use pmu syntax:
135 perf record -e r1a8 -a sleep 1
136 perf record -e cpu/r1a8/ ...
137 perf record -e cpu/r0x1a8/ ...
139 Some processors, like those from AMD, support event codes and unit masks
140 larger than a byte. In such cases, the bits corresponding to the event
141 configuration parameters can be seen with:
143 cat /sys/bus/event_source/devices/<pmu>/format/<config>
147 If the AMD docs for an EPYC 7713 processor describe an event as:
149 Event Umask Event Mask
150 Num. Value Mnemonic Description
152 28FH 03H op_cache_hit_miss.op_cache_hit Counts Op Cache micro-tag
155 raw encoding of 0x0328F cannot be used since the upper nibble of the
156 EventSelect bits have to be specified via bits 32-35 as can be seen with:
158 cat /sys/bus/event_source/devices/cpu/format/event
160 raw encoding of 0x20000038F should be used instead:
162 perf stat -e r20000038f -a sleep 1
163 perf record -e r20000038f ...
165 It's also possible to use pmu syntax:
167 perf record -e r20000038f -a sleep 1
168 perf record -e cpu/r20000038f/ ...
169 perf record -e cpu/r0x20000038f/ ...
171 You should refer to the processor specific documentation for getting these
172 details. Some of them are referenced in the SEE ALSO section below.
177 perf also supports an extended syntax for specifying raw parameters
178 to PMUs. Using this typically requires looking up the specific event
179 in the CPU vendor specific documentation.
181 The available PMUs and their raw parameters can be listed with
183 ls /sys/devices/*/format
185 For example the raw event "LSD.UOPS" core pmu event above could
188 perf stat -e cpu/event=0xa8,umask=0x1,name=LSD.UOPS_CYCLES,cmask=0x1/ ...
190 or using extended name syntax
192 perf stat -e cpu/event=0xa8,umask=0x1,cmask=0x1,name=\'LSD.UOPS_CYCLES:cmask=0x1\'/ ...
197 Some PMUs are not associated with a core, but with a whole CPU socket.
198 Events on these PMUs generally cannot be sampled, but only counted globally
199 with perf stat -a. They can be bound to one logical CPU, but will measure
200 all the CPUs in the same socket.
202 This example measures memory bandwidth every second
203 on the first memory controller on socket 0 of a Intel Xeon system
205 perf stat -C 0 -a uncore_imc_0/cas_count_read/,uncore_imc_0/cas_count_write/ -I 1000 ...
207 Each memory controller has its own PMU. Measuring the complete system
208 bandwidth would require specifying all imc PMUs (see perf list output),
209 and adding the values together. To simplify creation of multiple events,
210 prefix and glob matching is supported in the PMU name, and the prefix
211 'uncore_' is also ignored when performing the match. So the command above
212 can be expanded to all memory controllers by using the syntaxes:
214 perf stat -C 0 -a imc/cas_count_read/,imc/cas_count_write/ -I 1000 ...
215 perf stat -C 0 -a *imc*/cas_count_read/,*imc*/cas_count_write/ -I 1000 ...
217 This example measures the combined core power every second
219 perf stat -I 1000 -e power/energy-cores/ -a
224 For non root users generally only context switched PMU events are available.
225 This is normally only the events in the cpu PMU, the predefined events
226 like cycles and instructions and some software events.
228 Other PMUs and global measurements are normally root only.
229 Some event qualifiers, such as "any", are also root only.
231 This can be overridden by setting the kernel.perf_event_paranoid
232 sysctl to -1, which allows non root to use these events.
234 For accessing trace point events perf needs to have read access to
235 /sys/kernel/tracing, even when perf_event_paranoid is in a relaxed
241 Some PMUs control advanced hardware tracing capabilities, such as Intel PT,
242 that allows low overhead execution tracing. These are described in a separate
243 intel-pt.txt document.
248 Some pmu events listed by 'perf-list' will be displayed with '?' in them. For
251 hv_gpci/dtbp_ptitc,phys_processor_idx=?/
253 This means that when provided as an event, a value for '?' must
254 also be supplied. For example:
256 perf stat -C 0 -e 'hv_gpci/dtbp_ptitc,phys_processor_idx=0x2/' ...
260 It is also possible to add extra qualifiers to an event:
264 Sums up the event counts for all hardware threads in a core, e.g.:
267 perf stat -e cpu/event=0,umask=0x3,percore=1/
273 Perf supports time based multiplexing of events, when the number of events
274 active exceeds the number of hardware performance counters. Multiplexing
275 can cause measurement errors when the workload changes its execution
278 When metrics are computed using formulas from event counts, it is useful to
279 ensure some events are always measured together as a group to minimize multiplexing
280 errors. Event groups can be specified using { }.
282 perf stat -e '{instructions,cycles}' ...
284 The number of available performance counters depend on the CPU. A group
285 cannot contain more events than available counters.
286 For example Intel Core CPUs typically have four generic performance counters
287 for the core, plus three fixed counters for instructions, cycles and
288 ref-cycles. Some special events have restrictions on which counter they
289 can schedule, and may not support multiple instances in a single group.
290 When too many events are specified in the group some of them will not
293 Globally pinned events can limit the number of counters available for
294 other groups. On x86 systems, the NMI watchdog pins a counter by default.
295 The nmi watchdog can be disabled as root with
297 echo 0 > /proc/sys/kernel/nmi_watchdog
299 Events from multiple different PMUs cannot be mixed in a group, with
300 some exceptions for software events.
305 perf also supports group leader sampling using the :S specifier.
307 perf record -e '{cycles,instructions}:S' ...
310 Normally all events in an event group sample, but with :S only
311 the first event (the leader) samples, and it only reads the values of the
312 other events in the group.
314 However, in the case AUX area events (e.g. Intel PT or CoreSight), the AUX
315 area event must be the leader, so then the second event samples, not the first.
320 Without options all known events will be listed.
322 To limit the list use:
324 . 'hw' or 'hardware' to list hardware events such as cache-misses, etc.
326 . 'sw' or 'software' to list software events such as context switches, etc.
328 . 'cache' or 'hwcache' to list hardware cache events such as L1-dcache-loads, etc.
330 . 'tracepoint' to list all tracepoint events, alternatively use
331 'subsys_glob:event_glob' to filter by tracepoint subsystems such as sched,
334 . 'pmu' to print the kernel supplied PMU events.
336 . 'sdt' to list all Statically Defined Tracepoint events.
338 . 'metric' to list metrics
340 . 'metricgroup' to list metricgroups with metrics.
342 . If none of the above is matched, it will apply the supplied glob to all
343 events, printing the ones that match.
345 . As a last resort, it will do a substring search in all event names.
347 One or more types can be used at the same time, listing the events for the
352 . '--raw-dump', shows the raw-dump of all the events.
353 . '--raw-dump [hw|sw|cache|tracepoint|pmu|event_glob]', shows the raw-dump of
354 a certain kind of events.
358 linkperf:perf-stat[1], linkperf:perf-top[1],
359 linkperf:perf-record[1],
360 http://www.intel.com/sdm/[IntelĀ® 64 and IA-32 Architectures Software Developer's Manual Volume 3B: System Programming Guide],
361 https://bugzilla.kernel.org/show_bug.cgi?id=206537[AMD Processor Programming Reference (PPR)]