1perf-list(1)
2============
3
4NAME
5----
6perf-list - List all symbolic event types
7
8SYNOPSIS
9--------
10[verse]
11'perf list' [--no-desc] [--long-desc]
12            [hw|sw|cache|tracepoint|pmu|sdt|metric|metricgroup|event_glob]
13
14DESCRIPTION
15-----------
16This command displays the symbolic event types which can be selected in the
17various perf commands with the -e option.
18
19OPTIONS
20-------
21-d::
22--desc::
23Print extra event descriptions. (default)
24
25--no-desc::
26Don't print descriptions.
27
28-v::
29--long-desc::
30Print longer event descriptions.
31
32--debug::
33Enable debugging output.
34
35--details::
36Print how named events are resolved internally into perf events, and also
37any extra expressions computed by perf stat.
38
39--deprecated::
40Print deprecated events. By default the deprecated events are hidden.
41
42--cputype::
43Print events applying cpu with this type for hybrid platform
44(e.g. --cputype core or --cputype atom)
45
46[[EVENT_MODIFIERS]]
47EVENT MODIFIERS
48---------------
49
50Events can optionally have a modifier by appending a colon and one or
51more modifiers. Modifiers allow the user to restrict the events to be
52counted. The following modifiers exist:
53
54 u - user-space counting
55 k - kernel counting
56 h - hypervisor counting
57 I - non idle counting
58 G - guest counting (in KVM guests)
59 H - host counting (not in KVM guests)
60 p - precise level
61 P - use maximum detected precise level
62 S - read sample value (PERF_SAMPLE_READ)
63 D - pin the event to the PMU
64 W - group is weak and will fallback to non-group if not schedulable,
65 e - group or event are exclusive and do not share the PMU
66
67The 'p' modifier can be used for specifying how precise the instruction
68address should be. The 'p' modifier can be specified multiple times:
69
70 0 - SAMPLE_IP can have arbitrary skid
71 1 - SAMPLE_IP must have constant skid
72 2 - SAMPLE_IP requested to have 0 skid
73 3 - SAMPLE_IP must have 0 skid, or uses randomization to avoid
74     sample shadowing effects.
75
76For Intel systems precise event sampling is implemented with PEBS
77which supports up to precise-level 2, and precise level 3 for
78some special cases
79
80On AMD systems it is implemented using IBS (up to precise-level 2).
81The precise modifier works with event types 0x76 (cpu-cycles, CPU
82clocks not halted) and 0xC1 (micro-ops retired). Both events map to
83IBS execution sampling (IBS op) with the IBS Op Counter Control bit
84(IbsOpCntCtl) set respectively (see AMD64 Architecture Programmer’s
85Manual Volume 2: System Programming, 13.3 Instruction-Based
86Sampling). Examples to use IBS:
87
88 perf record -a -e cpu-cycles:p ...    # use ibs op counting cycles
89 perf record -a -e r076:p ...          # same as -e cpu-cycles:p
90 perf record -a -e r0C1:p ...          # use ibs op counting micro-ops
91
92RAW HARDWARE EVENT DESCRIPTOR
93-----------------------------
94Even when an event is not available in a symbolic form within perf right now,
95it can be encoded in a per processor specific way.
96
97For instance For x86 CPUs NNN represents the raw register encoding with the
98layout of IA32_PERFEVTSELx MSRs (see [Intel® 64 and IA-32 Architectures Software Developer's Manual Volume 3B: System Programming Guide] Figure 30-1 Layout
99of IA32_PERFEVTSELx MSRs) or AMD's PerfEvtSeln (see [AMD64 Architecture Programmer’s Manual Volume 2: System Programming], Page 344,
100Figure 13-7 Performance Event-Select Register (PerfEvtSeln)).
101
102Note: Only the following bit fields can be set in x86 counter
103registers: event, umask, edge, inv, cmask. Esp. guest/host only and
104OS/user mode flags must be setup using <<EVENT_MODIFIERS, EVENT
105MODIFIERS>>.
106
107Example:
108
109If the Intel docs for a QM720 Core i7 describe an event as:
110
111  Event  Umask  Event Mask
112  Num.   Value  Mnemonic    Description                        Comment
113
114  A8H      01H  LSD.UOPS    Counts the number of micro-ops     Use cmask=1 and
115                            delivered by loop stream detector  invert to count
116                                                               cycles
117
118raw encoding of 0x1A8 can be used:
119
120 perf stat -e r1a8 -a sleep 1
121 perf record -e r1a8 ...
122
123It's also possible to use pmu syntax:
124
125 perf record -e r1a8 -a sleep 1
126 perf record -e cpu/r1a8/ ...
127 perf record -e cpu/r0x1a8/ ...
128
129You should refer to the processor specific documentation for getting these
130details. Some of them are referenced in the SEE ALSO section below.
131
132ARBITRARY PMUS
133--------------
134
135perf also supports an extended syntax for specifying raw parameters
136to PMUs. Using this typically requires looking up the specific event
137in the CPU vendor specific documentation.
138
139The available PMUs and their raw parameters can be listed with
140
141  ls /sys/devices/*/format
142
143For example the raw event "LSD.UOPS" core pmu event above could
144be specified as
145
146  perf stat -e cpu/event=0xa8,umask=0x1,name=LSD.UOPS_CYCLES,cmask=0x1/ ...
147
148  or using extended name syntax
149
150  perf stat -e cpu/event=0xa8,umask=0x1,cmask=0x1,name=\'LSD.UOPS_CYCLES:cmask=0x1\'/ ...
151
152PER SOCKET PMUS
153---------------
154
155Some PMUs are not associated with a core, but with a whole CPU socket.
156Events on these PMUs generally cannot be sampled, but only counted globally
157with perf stat -a. They can be bound to one logical CPU, but will measure
158all the CPUs in the same socket.
159
160This example measures memory bandwidth every second
161on the first memory controller on socket 0 of a Intel Xeon system
162
163  perf stat -C 0 -a uncore_imc_0/cas_count_read/,uncore_imc_0/cas_count_write/ -I 1000 ...
164
165Each memory controller has its own PMU.  Measuring the complete system
166bandwidth would require specifying all imc PMUs (see perf list output),
167and adding the values together. To simplify creation of multiple events,
168prefix and glob matching is supported in the PMU name, and the prefix
169'uncore_' is also ignored when performing the match. So the command above
170can be expanded to all memory controllers by using the syntaxes:
171
172  perf stat -C 0 -a imc/cas_count_read/,imc/cas_count_write/ -I 1000 ...
173  perf stat -C 0 -a *imc*/cas_count_read/,*imc*/cas_count_write/ -I 1000 ...
174
175This example measures the combined core power every second
176
177  perf stat -I 1000 -e power/energy-cores/  -a
178
179ACCESS RESTRICTIONS
180-------------------
181
182For non root users generally only context switched PMU events are available.
183This is normally only the events in the cpu PMU, the predefined events
184like cycles and instructions and some software events.
185
186Other PMUs and global measurements are normally root only.
187Some event qualifiers, such as "any", are also root only.
188
189This can be overridden by setting the kernel.perf_event_paranoid
190sysctl to -1, which allows non root to use these events.
191
192For accessing trace point events perf needs to have read access to
193/sys/kernel/debug/tracing, even when perf_event_paranoid is in a relaxed
194setting.
195
196TRACING
197-------
198
199Some PMUs control advanced hardware tracing capabilities, such as Intel PT,
200that allows low overhead execution tracing.  These are described in a separate
201intel-pt.txt document.
202
203PARAMETERIZED EVENTS
204--------------------
205
206Some pmu events listed by 'perf-list' will be displayed with '?' in them. For
207example:
208
209  hv_gpci/dtbp_ptitc,phys_processor_idx=?/
210
211This means that when provided as an event, a value for '?' must
212also be supplied. For example:
213
214  perf stat -C 0 -e 'hv_gpci/dtbp_ptitc,phys_processor_idx=0x2/' ...
215
216EVENT QUALIFIERS:
217
218It is also possible to add extra qualifiers to an event:
219
220percore:
221
222Sums up the event counts for all hardware threads in a core, e.g.:
223
224
225  perf stat -e cpu/event=0,umask=0x3,percore=1/
226
227
228EVENT GROUPS
229------------
230
231Perf supports time based multiplexing of events, when the number of events
232active exceeds the number of hardware performance counters. Multiplexing
233can cause measurement errors when the workload changes its execution
234profile.
235
236When metrics are computed using formulas from event counts, it is useful to
237ensure some events are always measured together as a group to minimize multiplexing
238errors. Event groups can be specified using { }.
239
240  perf stat -e '{instructions,cycles}' ...
241
242The number of available performance counters depend on the CPU. A group
243cannot contain more events than available counters.
244For example Intel Core CPUs typically have four generic performance counters
245for the core, plus three fixed counters for instructions, cycles and
246ref-cycles. Some special events have restrictions on which counter they
247can schedule, and may not support multiple instances in a single group.
248When too many events are specified in the group some of them will not
249be measured.
250
251Globally pinned events can limit the number of counters available for
252other groups. On x86 systems, the NMI watchdog pins a counter by default.
253The nmi watchdog can be disabled as root with
254
255	echo 0 > /proc/sys/kernel/nmi_watchdog
256
257Events from multiple different PMUs cannot be mixed in a group, with
258some exceptions for software events.
259
260LEADER SAMPLING
261---------------
262
263perf also supports group leader sampling using the :S specifier.
264
265  perf record -e '{cycles,instructions}:S' ...
266  perf report --group
267
268Normally all events in an event group sample, but with :S only
269the first event (the leader) samples, and it only reads the values of the
270other events in the group.
271
272However, in the case AUX area events (e.g. Intel PT or CoreSight), the AUX
273area event must be the leader, so then the second event samples, not the first.
274
275OPTIONS
276-------
277
278Without options all known events will be listed.
279
280To limit the list use:
281
282. 'hw' or 'hardware' to list hardware events such as cache-misses, etc.
283
284. 'sw' or 'software' to list software events such as context switches, etc.
285
286. 'cache' or 'hwcache' to list hardware cache events such as L1-dcache-loads, etc.
287
288. 'tracepoint' to list all tracepoint events, alternatively use
289  'subsys_glob:event_glob' to filter by tracepoint subsystems such as sched,
290  block, etc.
291
292. 'pmu' to print the kernel supplied PMU events.
293
294. 'sdt' to list all Statically Defined Tracepoint events.
295
296. 'metric' to list metrics
297
298. 'metricgroup' to list metricgroups with metrics.
299
300. If none of the above is matched, it will apply the supplied glob to all
301  events, printing the ones that match.
302
303. As a last resort, it will do a substring search in all event names.
304
305One or more types can be used at the same time, listing the events for the
306types specified.
307
308Support raw format:
309
310. '--raw-dump', shows the raw-dump of all the events.
311. '--raw-dump [hw|sw|cache|tracepoint|pmu|event_glob]', shows the raw-dump of
312  a certain kind of events.
313
314SEE ALSO
315--------
316linkperf:perf-stat[1], linkperf:perf-top[1],
317linkperf:perf-record[1],
318http://www.intel.com/sdm/[Intel® 64 and IA-32 Architectures Software Developer's Manual Volume 3B: System Programming Guide],
319http://support.amd.com/us/Processor_TechDocs/24593_APM_v2.pdf[AMD64 Architecture Programmer’s Manual Volume 2: System Programming]
320