1 .\" Copyright (c) 2012, Vincent Weaver
3 .\" %%%LICENSE_START(GPLv2+_DOC_FULL)
4 .\" This is free documentation; you can redistribute it and/or
5 .\" modify it under the terms of the GNU General Public License as
6 .\" published by the Free Software Foundation; either version 2 of
7 .\" the License, or (at your option) any later version.
9 .\" The GNU General Public License's references to "object code"
10 .\" and "executables" are to be interpreted as the output of any
11 .\" document formatting or typesetting system, including
12 .\" intermediate and printed output.
14 .\" This manual is distributed in the hope that it will be useful,
15 .\" but WITHOUT ANY WARRANTY; without even the implied warranty of
16 .\" MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 .\" GNU General Public License for more details.
19 .\" You should have received a copy of the GNU General Public
20 .\" License along with this manual; if not, see
21 .\" <http://www.gnu.org/licenses/>.
24 .\" This document is based on the perf_event.h header file, the
25 .\" tools/perf/design.txt file, and a lot of bitter experience.
27 .TH PERF_EVENT_OPEN 2 2015-01-10 "Linux" "Linux Programmer's Manual"
29 perf_event_open \- set up performance monitoring
32 .B #include <linux/perf_event.h>
33 .B #include <linux/hw_breakpoint.h>
35 .BI "int perf_event_open(struct perf_event_attr *" attr ,
36 .BI " pid_t " pid ", int " cpu ", int " group_fd ,
37 .BI " unsigned long " flags );
41 There is no glibc wrapper for this system call; see NOTES.
43 Given a list of parameters,
44 .BR perf_event_open ()
45 returns a file descriptor, for use in subsequent system calls
46 .RB ( read "(2), " mmap "(2), " prctl "(2), " fcntl "(2), etc.)."
49 .BR perf_event_open ()
50 creates a file descriptor that allows measuring performance
52 Each file descriptor corresponds to one
53 event that is measured; these can be grouped together
54 to measure multiple events simultaneously.
56 Events can be enabled and disabled in two ways: via
60 When an event is disabled it does not count or generate overflows but does
61 continue to exist and maintain its count value.
63 Events come in two flavors: counting and sampled.
66 event is one that is used for counting the aggregate number of events
68 In general, counting event results are gathered with a
73 event periodically writes measurements to a buffer that can then
82 arguments allow specifying which process and CPU to monitor:
84 .BR "pid == 0" " and " "cpu == \-1"
85 This measures the calling process/thread on any CPU.
87 .BR "pid == 0" " and " "cpu >= 0"
88 This measures the calling process/thread only
89 when running on the specified CPU.
91 .BR "pid > 0" " and " "cpu == \-1"
92 This measures the specified process/thread on any CPU.
94 .BR "pid > 0" " and " "cpu >= 0"
95 This measures the specified process/thread only
96 when running on the specified CPU.
98 .BR "pid == \-1" " and " "cpu >= 0"
99 This measures all processes/threads on the specified CPU.
103 .I /proc/sys/kernel/perf_event_paranoid
104 value of less than 1.
106 .BR "pid == \-1" " and " "cpu == \-1"
107 This setting is invalid and will return an error.
111 argument allows event groups to be created.
112 An event group has one event which is the group leader.
113 The leader is created first, with
114 .IR group_fd " = \-1."
115 The rest of the group members are created with subsequent
116 .BR perf_event_open ()
119 being set to the file descriptor of the group leader.
120 (A single event on its own is created with
121 .IR group_fd " = \-1"
122 and is considered to be a group with only 1 member.)
123 An event group is scheduled onto the CPU as a unit: it will
124 be put onto the CPU only if all of the events in the group can be put onto
126 This means that the values of the member events can be
127 meaningfully compared\(emadded, divided (to get ratios), and so on\(emwith each
128 other, since they have counted events for the same set of executed
133 argument is formed by ORing together zero or more of the following values:
135 .BR PERF_FLAG_FD_CLOEXEC " (since Linux 3.14)"
136 .\" commit a21b0b354d4ac39be691f51c53562e2c24443d9e
137 This flag enables the close-on-exec flag for the created
138 event file descriptor,
139 so that the file descriptor is automatically closed on
141 Setting the close-on-exec flags at creation time, rather than later with
143 avoids potential race conditions where the calling thread invokes
144 .BR perf_event_open ()
147 at the same time as another thread calls
152 .BR PERF_FLAG_FD_NO_GROUP
153 This flag tells the event to ignore the
155 parameter except for the purpose of setting up output redirection
157 .B PERF_FLAG_FD_OUTPUT
160 .BR PERF_FLAG_FD_OUTPUT " (broken since Linux 2.6.35)"
161 .\" commit ac9721f3f54b27a16c7e1afb2481e7ee95a70318
162 This flag re-routes the event's sampled output to instead
163 be included in the mmap buffer of the event specified by
166 .BR PERF_FLAG_PID_CGROUP " (since Linux 2.6.39)"
167 .\" commit e5d1367f17ba6a6fed5fd8b74e4d5720923e0c25
168 This flag activates per-container system-wide monitoring.
170 is an abstraction that isolates a set of resources for finer-grained
171 control (CPUs, memory, etc.).
172 In this mode, the event is measured
173 only if the thread running on the monitored CPU belongs to the designated
175 The cgroup is identified by passing a file descriptor
176 opened on its directory in the cgroupfs filesystem.
178 cgroup to monitor is called
180 then a file descriptor opened on
182 (assuming cgroupfs is mounted on
184 must be passed as the
187 cgroup monitoring is available only
188 for system-wide events and may therefore require extra permissions.
192 structure provides detailed configuration information
193 for the event being created.
197 struct perf_event_attr {
198 __u32 type; /* Type of event */
199 __u32 size; /* Size of attribute structure */
200 __u64 config; /* Type-specific configuration */
203 __u64 sample_period; /* Period of sampling */
204 __u64 sample_freq; /* Frequency of sampling */
207 __u64 sample_type; /* Specifies values included in sample */
208 __u64 read_format; /* Specifies values returned in read */
210 __u64 disabled : 1, /* off by default */
211 inherit : 1, /* children inherit it */
212 pinned : 1, /* must always be on PMU */
213 exclusive : 1, /* only group on PMU */
214 exclude_user : 1, /* don't count user */
215 exclude_kernel : 1, /* don't count kernel */
216 exclude_hv : 1, /* don't count hypervisor */
217 exclude_idle : 1, /* don't count when idle */
218 mmap : 1, /* include mmap data */
219 comm : 1, /* include comm data */
220 freq : 1, /* use freq, not period */
221 inherit_stat : 1, /* per task counts */
222 enable_on_exec : 1, /* next exec enables */
223 task : 1, /* trace fork/exit */
224 watermark : 1, /* wakeup_watermark */
225 precise_ip : 2, /* skid constraint */
226 mmap_data : 1, /* non-exec mmap data */
227 sample_id_all : 1, /* sample_type all events */
228 exclude_host : 1, /* don't count in host */
229 exclude_guest : 1, /* don't count in guest */
230 exclude_callchain_kernel : 1,
231 /* exclude kernel callchains */
232 exclude_callchain_user : 1,
233 /* exclude user callchains */
234 mmap2 : 1, /* include mmap with inode data */
235 comm_exec : 1, /* flag comm events that are due to exec */
239 __u32 wakeup_events; /* wakeup every n events */
240 __u32 wakeup_watermark; /* bytes before wakeup */
243 __u32 bp_type; /* breakpoint type */
246 __u64 bp_addr; /* breakpoint address */
247 __u64 config1; /* extension of config */
251 __u64 bp_len; /* breakpoint length */
252 __u64 config2; /* extension of config1 */
254 __u64 branch_sample_type; /* enum perf_branch_sample_type */
255 __u64 sample_regs_user; /* user regs to dump on samples */
256 __u32 sample_stack_user; /* size of stack to dump on
258 __u32 __reserved_2; /* Align to u64 */
266 structure are described in more detail below:
269 This field specifies the overall event type.
270 It has one of the following values:
273 .B PERF_TYPE_HARDWARE
274 This indicates one of the "generalized" hardware events provided
278 field definition for more details.
280 .B PERF_TYPE_SOFTWARE
281 This indicates one of the software-defined events provided by the kernel
282 (even if no hardware support is available).
284 .B PERF_TYPE_TRACEPOINT
285 This indicates a tracepoint
286 provided by the kernel tracepoint infrastructure.
288 .B PERF_TYPE_HW_CACHE
289 This indicates a hardware cache event.
290 This has a special encoding, described in the
295 This indicates a "raw" implementation-specific event in the
298 .BR PERF_TYPE_BREAKPOINT " (since Linux 2.6.33)"
299 .\" commit 24f1e32c60c45c89a997c73395b69c8af6f0a84e
300 This indicates a hardware breakpoint as provided by the CPU.
301 Breakpoints can be read/write accesses to an address as well as
302 execution of an instruction address.
306 .\" commit 2e80a82a49c4c7eca4e35734380f28298ba5db19
307 .BR perf_event_open ()
308 can support multiple PMUs.
309 To enable this, a value exported by the kernel can be used in the
311 field to indicate which PMU to use.
312 The value to use can be found in the sysfs filesystem:
313 there is a subdirectory per PMU instance under
314 .IR /sys/bus/event_source/devices .
315 In each subdirectory there is a
317 file whose content is an integer that can be used in the
321 .I /sys/bus/event_source/devices/cpu/type
322 contains the value for the core CPU PMU, which is usually 4.
328 structure for forward/backward compatibility.
330 .I sizeof(struct perf_event_attr)
331 to allow the kernel to see
332 the struct size at the time of compilation.
335 .B PERF_ATTR_SIZE_VER0
336 is set to 64; this was the size of the first published struct.
337 .B PERF_ATTR_SIZE_VER1
338 is 72, corresponding to the addition of breakpoints in Linux 2.6.33.
339 .\" commit cb5d76999029ae7a517cb07dfa732c1b5a934fc2
340 .\" this was added much later when PERF_ATTR_SIZE_VER2 happened
341 .\" but the actual attr_size had increased in 2.6.33
342 .B PERF_ATTR_SIZE_VER2
343 is 80 corresponding to the addition of branch sampling in Linux 3.4.
344 .\" commit cb5d76999029ae7a517cb07dfa732c1b5a934fc2
345 .B PERF_ATTR_SIZE_VER3
346 is 96 corresponding to the addition
352 .\" commit 1659d129ed014b715b0b2120e6fd929bdd33ed03
355 This specifies which event you want, in conjunction with
360 .IR config1 " and " config2
361 fields are also taken into account in cases where 64 bits is not
362 enough to fully specify the event.
363 The encoding of these fields are event dependent.
365 There are various ways to set the
367 field that are dependent on the value of the previously
371 What follows are various possible settings for
379 .BR PERF_TYPE_HARDWARE ,
380 we are measuring one of the generalized hardware CPU events.
381 Not all of these are available on all platforms.
384 to one of the following:
387 .B PERF_COUNT_HW_CPU_CYCLES
389 Be wary of what happens during CPU frequency scaling.
391 .B PERF_COUNT_HW_INSTRUCTIONS
392 Retired instructions.
393 Be careful, these can be affected by various
394 issues, most notably hardware interrupt counts.
396 .B PERF_COUNT_HW_CACHE_REFERENCES
398 Usually this indicates Last Level Cache accesses but this may
399 vary depending on your CPU.
400 This may include prefetches and coherency messages; again this
401 depends on the design of your CPU.
403 .B PERF_COUNT_HW_CACHE_MISSES
405 Usually this indicates Last Level Cache misses; this is intended to be
406 used in conjunction with the
407 .B PERF_COUNT_HW_CACHE_REFERENCES
408 event to calculate cache miss rates.
410 .B PERF_COUNT_HW_BRANCH_INSTRUCTIONS
411 Retired branch instructions.
412 Prior to Linux 2.6.35, this used
413 the wrong event on AMD processors.
414 .\" commit f287d332ce835f77a4f5077d2c0ef1e3f9ea42d2
416 .B PERF_COUNT_HW_BRANCH_MISSES
417 Mispredicted branch instructions.
419 .B PERF_COUNT_HW_BUS_CYCLES
420 Bus cycles, which can be different from total cycles.
422 .BR PERF_COUNT_HW_STALLED_CYCLES_FRONTEND " (since Linux 3.0)"
423 .\" commit 8f62242246351b5a4bc0c1f00c0c7003edea128a
424 Stalled cycles during issue.
426 .BR PERF_COUNT_HW_STALLED_CYCLES_BACKEND " (since Linux 3.0)"
427 .\" commit 8f62242246351b5a4bc0c1f00c0c7003edea128a
428 Stalled cycles during retirement.
430 .BR PERF_COUNT_HW_REF_CPU_CYCLES " (since Linux 3.3)"
431 .\" commit c37e17497e01fc0f5d2d6feb5723b210b3ab8890
432 Total cycles; not affected by CPU frequency scaling.
438 .BR PERF_TYPE_SOFTWARE ,
439 we are measuring software events provided by the kernel.
442 to one of the following:
445 .B PERF_COUNT_SW_CPU_CLOCK
446 This reports the CPU clock, a high-resolution per-CPU timer.
448 .B PERF_COUNT_SW_TASK_CLOCK
449 This reports a clock count specific to the task that is running.
451 .B PERF_COUNT_SW_PAGE_FAULTS
452 This reports the number of page faults.
454 .B PERF_COUNT_SW_CONTEXT_SWITCHES
455 This counts context switches.
456 Until Linux 2.6.34, these were all reported as user-space
457 events, after that they are reported as happening in the kernel.
458 .\" commit e49a5bd38159dfb1928fd25b173bc9de4bbadb21
460 .B PERF_COUNT_SW_CPU_MIGRATIONS
461 This reports the number of times the process
462 has migrated to a new CPU.
464 .B PERF_COUNT_SW_PAGE_FAULTS_MIN
465 This counts the number of minor page faults.
466 These did not require disk I/O to handle.
468 .B PERF_COUNT_SW_PAGE_FAULTS_MAJ
469 This counts the number of major page faults.
470 These required disk I/O to handle.
472 .BR PERF_COUNT_SW_ALIGNMENT_FAULTS " (since Linux 2.6.33)"
473 .\" commit f7d7986060b2890fc26db6ab5203efbd33aa2497
474 This counts the number of alignment faults.
475 These happen when unaligned memory accesses happen; the kernel
476 can handle these but it reduces performance.
477 This happens only on some architectures (never on x86).
479 .BR PERF_COUNT_SW_EMULATION_FAULTS " (since Linux 2.6.33)"
480 .\" commit f7d7986060b2890fc26db6ab5203efbd33aa2497
481 This counts the number of emulation faults.
482 The kernel sometimes traps on unimplemented instructions
483 and emulates them for user space.
484 This can negatively impact performance.
486 .BR PERF_COUNT_SW_DUMMY " (since Linux 3.12)"
487 .\" commit fa0097ee690693006ab1aea6c01ad3c851b65c77
488 This is a placeholder event that counts nothing.
489 Informational sample record types such as mmap or comm
490 must be associated with an active event.
491 This dummy event allows gathering such records without requiring
499 .BR PERF_TYPE_TRACEPOINT ,
500 then we are measuring kernel tracepoints.
503 can be obtained from under debugfs
504 .I tracing/events/*/*/id
505 if ftrace is enabled in the kernel.
512 .BR PERF_TYPE_HW_CACHE ,
513 then we are measuring a hardware CPU cache event.
514 To calculate the appropriate
516 value use the following equation:
520 (perf_hw_cache_id) | (perf_hw_cache_op_id << 8) |
521 (perf_hw_cache_op_result_id << 16)
529 .B PERF_COUNT_HW_CACHE_L1D
530 for measuring Level 1 Data Cache
532 .B PERF_COUNT_HW_CACHE_L1I
533 for measuring Level 1 Instruction Cache
535 .B PERF_COUNT_HW_CACHE_LL
536 for measuring Last-Level Cache
538 .B PERF_COUNT_HW_CACHE_DTLB
539 for measuring the Data TLB
541 .B PERF_COUNT_HW_CACHE_ITLB
542 for measuring the Instruction TLB
544 .B PERF_COUNT_HW_CACHE_BPU
545 for measuring the branch prediction unit
547 .BR PERF_COUNT_HW_CACHE_NODE " (since Linux 3.1)"
548 .\" commit 89d6c0b5bdbb1927775584dcf532d98b3efe1477
549 for measuring local memory accesses
553 .I perf_hw_cache_op_id
557 .B PERF_COUNT_HW_CACHE_OP_READ
560 .B PERF_COUNT_HW_CACHE_OP_WRITE
563 .B PERF_COUNT_HW_CACHE_OP_PREFETCH
564 for prefetch accesses
568 .I perf_hw_cache_op_result_id
572 .B PERF_COUNT_HW_CACHE_RESULT_ACCESS
575 .B PERF_COUNT_HW_CACHE_RESULT_MISS
587 Most CPUs support events that are not covered by the "generalized" events.
588 These are implementation defined; see your CPU manual (for example
589 the Intel Volume 3B documentation or the AMD BIOS and Kernel Developer
591 The libpfm4 library can be used to translate from the name in the
592 architectural manuals to the raw hex value
593 .BR perf_event_open ()
594 expects in this field.
599 .BR PERF_TYPE_BREAKPOINT ,
603 Its parameters are set in other places.
606 .IR sample_period ", " sample_freq
607 A "sampling" event is one that generates an overflow notification
608 every N events, where N is given by
611 .IR sample_period " > 0."
612 When an overflow occurs, requested data is recorded
616 field controls what data is recorded on each overflow.
619 can be used if you wish to use frequency rather than period.
620 In this case, you set the
623 The kernel will adjust the sampling period
624 to try and achieve the desired rate.
625 The rate of adjustment is a
629 The various bits in this field specify which values to include
631 They will be recorded in a ring-buffer,
632 which is available to user space using
634 The order in which the values are saved in the
635 sample are documented in the MMAP Layout subsection below;
637 .I "enum perf_event_sample_format"
642 Records instruction pointer.
645 Records the process and thread IDs.
651 Records an address, if applicable.
654 Record counter values for all events in a group, not just the group leader.
656 .B PERF_SAMPLE_CALLCHAIN
657 Records the callchain (stack backtrace).
660 Records a unique ID for the opened event's group leader.
665 .B PERF_SAMPLE_PERIOD
666 Records the current sampling period.
668 .B PERF_SAMPLE_STREAM_ID
669 Records a unique ID for the opened event.
672 the actual ID is returned, not the group leader.
673 This ID is the same as the one returned by
677 Records additional data, if applicable.
678 Usually returned by tracepoint events.
680 .BR PERF_SAMPLE_BRANCH_STACK " (since Linux 3.4)"
681 .\" commit bce38cd53e5ddba9cb6d708c4ef3d04a4016ec7e
682 This provides a record of recent branches, as provided
683 by CPU branch sampling hardware (such as Intel Last Branch Record).
684 Not all hardware supports this feature.
687 .I branch_sample_type
688 field for how to filter which branches are reported.
690 .BR PERF_SAMPLE_REGS_USER " (since Linux 3.7)"
691 .\" commit 4018994f3d8785275ef0e7391b75c3462c029e56
692 Records the current user-level CPU register state
693 (the values in the process before the kernel was called).
695 .BR PERF_SAMPLE_STACK_USER " (since Linux 3.7)"
696 .\" commit c5ebcedb566ef17bda7b02686e0d658a7bb42ee7
697 Records the user level stack, allowing stack unwinding.
699 .BR PERF_SAMPLE_WEIGHT " (since Linux 3.10)"
700 .\" commit c3feedf2aaf9ac8bad6f19f5d21e4ee0b4b87e9c
701 Records a hardware provided weight value that expresses how
702 costly the sampled event was.
703 This allows the hardware to highlight expensive events in
706 .BR PERF_SAMPLE_DATA_SRC " (since Linux 3.10)"
707 .\" commit d6be9ad6c960f43800a6f118932bc8a5a4eadcd1
708 Records the data source: where in the memory hierarchy
709 the data associated with the sampled instruction came from.
710 This is available only if the underlying hardware
711 supports this feature.
713 .BR PERF_SAMPLE_IDENTIFIER " (since Linux 3.12)"
714 .\" commit ff3d527cebc1fa3707c617bfe9e74f53fcfb0955
717 value in a fixed position in the record,
718 either at the beginning (for sample events) or at the end
719 (if a non-sample event).
721 This was necessary because a sample stream may have
722 records from various different event sources with different
725 Parsing the event stream properly was not possible because the
726 format of the record was needed to find
729 the format could not be found without knowing what
730 event the sample belonged to (causing a circular
734 .B PERF_SAMPLE_IDENTIFIER
735 setting makes the event stream always parsable
738 in a fixed location, even though
739 it means having duplicate
743 .BR PERF_SAMPLE_TRANSACTION " (since Linux 3.13)"
744 .\" commit fdfbbd07e91f8fe387140776f3fd94605f0c89e5
745 Records reasons for transactional memory abort events
746 (for example, from Intel TSX transactional memory support).
750 setting must be greater than 0 and a transactional memory abort
751 event must be measured or no values will be recorded.
752 Also note that some perf_event measurements, such as sampled
753 cycle counting, may cause extraneous aborts (by causing an
754 interrupt during a transaction).
758 This field specifies the format of the data returned by
761 .BR perf_event_open ()
765 .B PERF_FORMAT_TOTAL_TIME_ENABLED
769 This can be used to calculate estimated totals if
770 the PMU is overcommitted and multiplexing is happening.
772 .B PERF_FORMAT_TOTAL_TIME_RUNNING
776 This can be used to calculate estimated totals if
777 the PMU is overcommitted and multiplexing is happening.
780 Adds a 64-bit unique value that corresponds to the event group.
783 Allows all counter values in an event group to be read with one read.
789 bit specifies whether the counter starts out disabled or enabled.
790 If disabled, the event can later be enabled by
796 When creating an event group, typically the group leader is initialized
799 set to 1 and any child events are initialized with
804 being 0, the child events will not start until the group leader
810 bit specifies that this counter should count events of child
811 tasks as well as the task specified.
812 This applies only to new children, not to any existing children at
813 the time the counter is created (nor to any new children of
816 Inherit does not work for some combinations of
819 .BR PERF_FORMAT_GROUP .
824 bit specifies that the counter should always be on the CPU if at all
826 It applies only to hardware counters and only to group leaders.
827 If a pinned counter cannot be put onto the CPU (e.g., because there are
828 not enough hardware counters or because of a conflict with some other
829 event), then the counter goes into an 'error' state, where reads
830 return end-of-file (i.e.,
832 returns 0) until the counter is subsequently enabled or disabled.
837 bit specifies that when this counter's group is on the CPU,
838 it should be the only group using the CPU's counters.
839 In the future this may allow monitoring programs to
840 support PMU features that need to run alone so that they do not
841 disrupt other hardware counters.
843 Note that many unexpected situations may prevent events with the
845 bit set from ever running.
846 This includes any users running a system-wide
847 measurement as well as any kernel use of the performance counters
848 (including the commonly enabled NMI Watchdog Timer interface).
851 If this bit is set, the count excludes events that happen in user space.
854 If this bit is set, the count excludes events that happen in kernel-space.
857 If this bit is set, the count excludes events that happen in the
859 This is mainly for PMUs that have built-in support for handling this
861 Extra support is needed for handling hypervisor measurements on most
865 If set, don't count when the CPU is idle.
870 bit enables generation of
877 This allows tools to notice new executable code being mapped into
878 a program (dynamic shared libraries for example)
879 so that addresses can be mapped back to the original code.
884 bit enables tracking of process command name as modified by the
887 .BR prctl (PR_SET_NAME)
888 system calls as well as writing to
889 .IR /proc/self/comm .
892 flag is also successfully set (possible since Linux 3.16),
893 .\" commit 82b897782d10fcc4930c9d4a15b175348fdd2871
895 .B PERF_RECORD_MISC_COMM_EXEC
896 can be used to differentiate the
898 case from the others.
901 If this bit is set, then
905 is used when setting up the sampling interval.
908 This bit enables saving of event counts on context switch for
910 This is meaningful only if the
915 If this bit is set, a counter is automatically
916 enabled after a call to
920 If this bit is set, then
921 fork/exit notifications are included in the ring buffer.
924 If set, have an overflow notification happen when we cross the
927 Otherwise, overflow notifications happen after
931 .IR "precise_ip" " (since Linux 2.6.35)"
932 .\" commit ab608344bcbde4f55ec4cd911b686b0ce3eae076
933 This controls the amount of skid.
934 Skid is how many instructions
935 execute between an event of interest happening and the kernel
936 being able to stop and record the event.
938 better and allows more accurate reporting of which events
939 correspond to which instructions, but hardware is often limited
940 with how small this can be.
942 The values of this are the following:
947 can have arbitrary skid.
951 must have constant skid.
955 requested to have 0 skid.
961 .BR PERF_RECORD_MISC_EXACT_IP .
964 .IR "mmap_data" " (since Linux 2.6.36)"
965 .\" commit 3af9e859281bda7eb7c20b51879cf43aa788ac2e
966 The counterpart of the
969 This enables generation of
973 calls that do not have
975 set (for example data and SysV shared memory).
977 .IR "sample_id_all" " (since Linux 2.6.38)"
978 .\" commit c980d1091810df13f21aabbce545fd98f545bbf7
979 If set, then TID, TIME, ID, STREAM_ID, and CPU can
980 additionally be included in
981 .RB non- PERF_RECORD_SAMPLE s
987 .B PERF_SAMPLE_IDENTIFIER
988 is specified, then an additional ID value is included
989 as the last value to ease parsing the record stream.
992 value appearing twice.
994 The layout is described by this pseudo-structure:
998 { u32 pid, tid; } /* if PERF_SAMPLE_TID set */
999 { u64 time; } /* if PERF_SAMPLE_TIME set */
1000 { u64 id; } /* if PERF_SAMPLE_ID set */
1001 { u64 stream_id;} /* if PERF_SAMPLE_STREAM_ID set */
1002 { u32 cpu, res; } /* if PERF_SAMPLE_CPU set */
1003 { u64 id; } /* if PERF_SAMPLE_IDENTIFIER set */
1007 .IR "exclude_host" " (since Linux 3.2)"
1008 .\" commit a240f76165e6255384d4bdb8139895fac7988799
1009 Do not measure time spent in VM host.
1011 .IR "exclude_guest" " (since Linux 3.2)"
1012 .\" commit a240f76165e6255384d4bdb8139895fac7988799
1013 Do not measure time spent in VM guest.
1015 .IR "exclude_callchain_kernel" " (since Linux 3.7)"
1016 .\" commit d077526485d5c9b12fe85d0b2b3b7041e6bc5f91
1017 Do not include kernel callchains.
1019 .IR "exclude_callchain_user" " (since Linux 3.7)"
1020 .\" commit d077526485d5c9b12fe85d0b2b3b7041e6bc5f91
1021 Do not include user callchains.
1023 .IR "mmap2" " (since Linux 3.16)"
1024 .\" commit 13d7a2410fa637f450a29ecb515ac318ee40c741
1025 .\" This is tricky; was committed during 3.12 development
1026 .\" but right before release was disabled.
1027 .\" So while you could select mmap2 starting with 3.12
1028 .\" it did not work until 3.16
1029 .\" commit a5a5ba72843dd05f991184d6cb9a4471acce1005
1030 Generate an extended executable mmap record that contains enough
1031 additional information to uniquely identify shared mappings.
1034 flag must also be set for this to work.
1036 .IR "comm_exec" " (since Linux 3.16)"
1037 .\" commit 82b897782d10fcc4930c9d4a15b175348fdd2871
1038 This is purely a feature-detection flag, it does not change
1040 If this flag can successfully be set, then, when
1043 .B PERF_RECORD_MISC_COMM_EXEC
1044 flag will be set in the
1046 field of a comm record header if the rename event being
1047 reported was caused by a call to
1049 This allows tools to distinguish between the various
1050 types of process renaming.
1052 .IR "wakeup_events" ", " "wakeup_watermark"
1053 This union sets how many samples
1054 .RI ( wakeup_events )
1056 .RI ( wakeup_watermark )
1057 happen before an overflow notification happens.
1058 Which one is used is selected by the
1064 .B PERF_RECORD_SAMPLE
1066 To receive overflow notification for all
1068 types choose watermark and set
1072 Prior to Linux 3.0 setting
1073 .\" commit f506b3dc0ec454a16d40cab9ee5d75435b39dc50
1075 to 0 resulted in no overflow notifications;
1076 more recent kernels treat 0 the same as 1.
1078 .IR "bp_type" " (since Linux 2.6.33)"
1079 .\" commit 24f1e32c60c45c89a997c73395b69c8af6f0a84e
1080 This chooses the breakpoint type.
1084 .BR HW_BREAKPOINT_EMPTY
1088 Count when we read the memory location.
1091 Count when we write the memory location.
1093 .BR HW_BREAKPOINT_RW
1094 Count when we read or write the memory location.
1097 Count when we execute code at the memory location.
1099 The values can be combined via a bitwise or, but the
1109 .IR "bp_addr" " (since Linux 2.6.33)"
1110 .\" commit 24f1e32c60c45c89a997c73395b69c8af6f0a84e
1112 address of the breakpoint.
1113 For execution breakpoints this is the memory address of the instruction
1114 of interest; for read and write breakpoints it is the memory address
1115 of the memory location of interest.
1117 .IR "config1" " (since Linux 2.6.39)"
1118 .\" commit a7e3ed1e470116c9d12c2f778431a481a6be8ab6
1120 is used for setting events that need an extra register or otherwise
1121 do not fit in the regular config field.
1122 Raw OFFCORE_EVENTS on Nehalem/Westmere/SandyBridge use this field
1123 on 3.3 and later kernels.
1125 .IR "bp_len" " (since Linux 2.6.33)"
1126 .\" commit 24f1e32c60c45c89a997c73395b69c8af6f0a84e
1128 is the length of the breakpoint being measured if
1131 .BR PERF_TYPE_BREAKPOINT .
1133 .BR HW_BREAKPOINT_LEN_1 ,
1134 .BR HW_BREAKPOINT_LEN_2 ,
1135 .BR HW_BREAKPOINT_LEN_4 ,
1136 .BR HW_BREAKPOINT_LEN_8 .
1137 For an execution breakpoint, set this to
1140 .IR "config2" " (since Linux 2.6.39)"
1141 .\" commit a7e3ed1e470116c9d12c2f778431a481a6be8ab6
1144 is a further extension of the
1148 .IR "branch_sample_type" " (since Linux 3.4)"
1149 .\" commit bce38cd53e5ddba9cb6d708c4ef3d04a4016ec7e
1151 .B PERF_SAMPLE_BRANCH_STACK
1152 is enabled, then this specifies what branches to include
1153 in the branch record.
1155 The first part of the value is the privilege level, which
1156 is a combination of one of the following values.
1157 If the user does not set privilege level explicitly, the kernel
1158 will use the event's privilege level.
1159 Event and branch privilege levels do not have to match.
1162 .B PERF_SAMPLE_BRANCH_USER
1163 Branch target is in user space.
1165 .B PERF_SAMPLE_BRANCH_KERNEL
1166 Branch target is in kernel space.
1168 .B PERF_SAMPLE_BRANCH_HV
1169 Branch target is in hypervisor.
1171 .B PERF_SAMPLE_BRANCH_PLM_ALL
1172 A convenience value that is the three preceding values ORed together.
1175 In addition to the privilege value, at least one or more of the
1176 following bits must be set.
1179 .B PERF_SAMPLE_BRANCH_ANY
1182 .B PERF_SAMPLE_BRANCH_ANY_CALL
1185 .B PERF_SAMPLE_BRANCH_ANY_RETURN
1188 .B PERF_SAMPLE_BRANCH_IND_CALL
1191 .BR PERF_SAMPLE_BRANCH_COND " (since Linux 3.16)"
1192 .\" commit bac52139f0b7ab31330e98fd87fc5a2664951050
1193 Conditional branches.
1195 .BR PERF_SAMPLE_BRANCH_ABORT_TX " (since Linux 3.11)"
1196 .\" commit 135c5612c460f89657c4698fe2ea753f6f667963
1197 Transactional memory aborts.
1199 .BR PERF_SAMPLE_BRANCH_IN_TX " (since Linux 3.11)"
1200 .\" commit 135c5612c460f89657c4698fe2ea753f6f667963
1201 Branch in transactional memory transaction.
1203 .BR PERF_SAMPLE_BRANCH_NO_TX " (since Linux 3.11)"
1204 .\" commit 135c5612c460f89657c4698fe2ea753f6f667963
1205 Branch not in transactional memory transaction.
1209 .IR "sample_regs_user" " (since Linux 3.7)"
1210 .\" commit 4018994f3d8785275ef0e7391b75c3462c029e56
1211 This bit mask defines the set of user CPU registers to dump on samples.
1212 The layout of the register mask is architecture-specific and
1213 described in the kernel header
1214 .IR arch/ARCH/include/uapi/asm/perf_regs.h .
1216 .IR "sample_stack_user" " (since Linux 3.7)"
1217 .\" commit c5ebcedb566ef17bda7b02686e0d658a7bb42ee7
1218 This defines the size of the user stack to dump if
1219 .B PERF_SAMPLE_STACK_USER
1223 .BR perf_event_open ()
1224 file descriptor has been opened, the values
1225 of the events can be read from the file descriptor.
1226 The values that are there are specified by the
1230 structure at open time.
1232 If you attempt to read into a buffer that is not big enough to hold the
1237 Here is the layout of the data returned by a read:
1240 .B PERF_FORMAT_GROUP
1241 was specified to allow reading all events in a group at once:
1245 struct read_format {
1246 u64 nr; /* The number of events */
1247 u64 time_enabled; /* if PERF_FORMAT_TOTAL_TIME_ENABLED */
1248 u64 time_running; /* if PERF_FORMAT_TOTAL_TIME_RUNNING */
1250 u64 value; /* The value of the event */
1251 u64 id; /* if PERF_FORMAT_ID */
1258 .B PERF_FORMAT_GROUP
1265 struct read_format {
1266 u64 value; /* The value of the event */
1267 u64 time_enabled; /* if PERF_FORMAT_TOTAL_TIME_ENABLED */
1268 u64 time_running; /* if PERF_FORMAT_TOTAL_TIME_RUNNING */
1269 u64 id; /* if PERF_FORMAT_ID */
1274 The values read are as follows:
1277 The number of events in this file descriptor.
1279 .B PERF_FORMAT_GROUP
1282 .IR time_enabled ", " time_running
1283 Total time the event was enabled and running.
1284 Normally these are the same.
1285 If more events are started,
1286 then available counter slots on the PMU, then multiplexing
1287 happens and events run only part of the time.
1292 values can be used to scale an estimated value for the count.
1295 An unsigned 64-bit value containing the counter result.
1298 A globally unique value for this particular event, only present if
1304 .BR perf_event_open ()
1305 in sampled mode, asynchronous events
1306 (like counter overflow or
1309 are logged into a ring-buffer.
1310 This ring-buffer is created and accessed through
1313 The mmap size should be 1+2^n pages, where the first page is a
1315 .RI ( "struct perf_event_mmap_page" )
1316 that contains various
1317 bits of information such as where the ring-buffer head is.
1319 Before kernel 2.6.39, there is a bug that means you must allocate a mmap
1320 ring buffer when sampling even if you do not plan to access it.
1322 The structure of the first metadata mmap page is as follows:
1326 struct perf_event_mmap_page {
1327 __u32 version; /* version number of this structure */
1328 __u32 compat_version; /* lowest version this is compat with */
1329 __u32 lock; /* seqlock for synchronization */
1330 __u32 index; /* hardware counter identifier */
1331 __s64 offset; /* add to hardware counter value */
1332 __u64 time_enabled; /* time event active */
1333 __u64 time_running; /* time event on CPU */
1337 __u64 cap_usr_time / cap_usr_rdpmc / cap_bit0 : 1,
1338 cap_bit0_is_deprecated : 1,
1341 cap_user_time_zero : 1,
1348 __u64 __reserved[120]; /* Pad to 1k */
1349 __u64 data_head; /* head in the data section */
1350 __u64 data_tail; /* user-space written tail */
1355 The following list describes the fields in the
1356 .I perf_event_mmap_page
1357 structure in more detail:
1360 Version number of this structure.
1363 The lowest version this is compatible with.
1366 A seqlock for synchronization.
1369 A unique hardware counter identifier.
1372 When using rdpmc for reads this offset value
1373 must be added to the one returned by rdpmc to get
1374 the current total event count.
1377 Time the event was active.
1380 Time the event was running.
1382 .IR cap_usr_time " / " cap_usr_rdpmc " / " cap_bit0 " (since Linux 3.4)"
1383 .\" commit c7206205d00ab375839bd6c7ddb247d600693c09
1384 There was a bug in the definition of
1388 from Linux 3.4 until Linux 3.11.
1389 Both bits were defined to point to the same location, so it was
1390 impossible to know if
1396 Starting with Linux 3.12, these are renamed to
1397 .\" commit fa7315871046b9a4c48627905691dbde57e51033
1399 and you should use the
1406 .IR cap_bit0_is_deprecated " (since Linux 3.12)"
1407 .\" commit fa7315871046b9a4c48627905691dbde57e51033
1408 If set, this bit indicates that the kernel supports
1409 the properly separated
1415 If not-set, it indicates an older kernel where
1419 map to the same bit and thus both features should
1420 be used with caution.
1423 .IR cap_user_rdpmc " (since Linux 3.12)"
1424 .\" commit fa7315871046b9a4c48627905691dbde57e51033
1425 If the hardware supports user-space read of performance counters
1426 without syscall (this is the "rdpmc" instruction on x86), then
1427 the following code can be used to do a read:
1431 u32 seq, time_mult, time_shift, idx, width;
1432 u64 count, enabled, running;
1433 u64 cyc, time_offset;
1438 enabled = pc\->time_enabled;
1439 running = pc\->time_running;
1441 if (pc\->cap_usr_time && enabled != running) {
1443 time_offset = pc\->time_offset;
1444 time_mult = pc\->time_mult;
1445 time_shift = pc\->time_shift;
1449 count = pc\->offset;
1451 if (pc\->cap_usr_rdpmc && idx) {
1452 width = pc\->pmc_width;
1453 count += rdpmc(idx \- 1);
1457 } while (pc\->lock != seq);
1461 .IR cap_user_time " (since Linux 3.12)"
1462 .\" commit fa7315871046b9a4c48627905691dbde57e51033
1463 This bit indicates the hardware has a constant, nonstop
1464 timestamp counter (TSC on x86).
1466 .IR cap_user_time_zero " (since Linux 3.12)"
1467 .\" commit fa7315871046b9a4c48627905691dbde57e51033
1468 Indicates the presence of
1470 which allows mapping timestamp values to
1476 this field provides the bit-width of the value
1477 read using the rdpmc or equivalent instruction.
1478 This can be used to sign extend the result like:
1482 pmc <<= 64 \- pmc_width;
1483 pmc >>= 64 \- pmc_width; // signed shift right
1488 .IR time_shift ", " time_mult ", " time_offset
1492 these fields can be used to compute the time
1493 delta since time_enabled (in nanoseconds) using rdtsc or similar.
1498 quot = (cyc >> time_shift);
1499 rem = cyc & ((1 << time_shift) \- 1);
1500 delta = time_offset + quot * time_mult +
1501 ((rem * time_mult) >> time_shift);
1511 seqcount loop described above.
1512 This delta can then be added to
1513 enabled and possible running (if idx), improving the scaling:
1519 quot = count / running;
1520 rem = count % running;
1521 count = quot * enabled + (rem * enabled) / running;
1524 .IR time_zero " (since Linux 3.12)"
1525 .\" commit fa7315871046b9a4c48627905691dbde57e51033
1528 .I cap_usr_time_zero
1529 is set, then the hardware clock (the TSC timestamp counter on x86)
1530 can be calculated from the
1531 .IR time_zero ", " time_mult ", and " time_shift " values:"
1534 time = timestamp - time_zero;
1535 quot = time / time_mult;
1536 rem = time % time_mult;
1537 cyc = (quot << time_shift) + (rem << time_shift) / time_mult;
1543 quot = cyc >> time_shift;
1544 rem = cyc & ((1 << time_shift) - 1);
1545 timestamp = time_zero + quot * time_mult +
1546 ((rem * time_mult) >> time_shift);
1550 This points to the head of the data section.
1551 The value continuously increases, it does not wrap.
1552 The value needs to be manually wrapped by the size of the mmap buffer
1553 before accessing the samples.
1555 On SMP-capable platforms, after reading the
1558 user space should issue an rmb().
1565 value should be written by user space to reflect the last read data.
1566 In this case, the kernel will not overwrite unread data.
1568 The following 2^n ring-buffer pages have the layout described below.
1571 .I perf_event_attr.sample_id_all
1572 is set, then all event types will
1573 have the sample_type selected fields related to where/when (identity)
1574 an event took place (TID, TIME, ID, CPU, STREAM_ID) described in
1575 .B PERF_RECORD_SAMPLE
1576 below, it will be stashed just after the
1577 .I perf_event_header
1578 and the fields already present for the existing
1579 fields, that is, at the end of the payload.
1580 That way a newer perf.data
1581 file will be supported by older perf tools, with these new optional
1582 fields being ignored.
1584 The mmap values start with a header:
1588 struct perf_event_header {
1596 Below, we describe the
1597 .I perf_event_header
1598 fields in more detail.
1599 For ease of reading,
1600 the fields with shorter descriptions are presented first.
1603 This indicates the size of the record.
1608 field contains additional information about the sample.
1610 The CPU mode can be determined from this value by masking with
1611 .B PERF_RECORD_MISC_CPUMODE_MASK
1612 and looking for one of the following (note these are not
1613 bit masks, only one can be set at a time):
1616 .B PERF_RECORD_MISC_CPUMODE_UNKNOWN
1619 .B PERF_RECORD_MISC_KERNEL
1620 Sample happened in the kernel.
1622 .B PERF_RECORD_MISC_USER
1623 Sample happened in user code.
1625 .B PERF_RECORD_MISC_HYPERVISOR
1626 Sample happened in the hypervisor.
1628 .BR PERF_RECORD_MISC_GUEST_KERNEL " (since Linux 2.6.35)"
1629 .\" commit 39447b386c846bbf1c56f6403c5282837486200f
1630 Sample happened in the guest kernel.
1632 .B PERF_RECORD_MISC_GUEST_USER " (since Linux 2.6.35)"
1633 .\" commit 39447b386c846bbf1c56f6403c5282837486200f
1634 Sample happened in guest user code.
1638 In addition, one of the following bits can be set:
1640 .BR PERF_RECORD_MISC_MMAP_DATA " (since Linux 3.10)"
1641 .\" commit 2fe85427e3bf65d791700d065132772fc26e4d75
1642 This is set when the mapping is not executable;
1643 otherwise the mapping is executable.
1645 .BR PERF_RECORD_MISC_COMM_EXEC " (since Linux 3.16)"
1646 .\" commit 82b897782d10fcc4930c9d4a15b175348fdd2871
1649 record on kernels more recent than Linux 3.16
1650 if a process name change was caused by an
1654 .B PERF_RECORD_MISC_MMAP_DATA
1655 since the two values would not be set in the same record.
1657 .B PERF_RECORD_MISC_EXACT_IP
1658 This indicates that the content of
1661 to the actual instruction that triggered the event.
1663 .IR perf_event_attr.precise_ip .
1665 .BR PERF_RECORD_MISC_EXT_RESERVED " (since Linux 2.6.35)"
1666 .\" commit 1676b8a077c352085d52578fb4f29350b58b6e74
1667 This indicates there is extended data available (currently not used).
1673 value is one of the below.
1674 The values in the corresponding record (that follows the header)
1682 The MMAP events record the
1684 mappings so that we can correlate
1685 user-space IPs to code.
1686 They have the following structure:
1691 struct perf_event_header header;
1709 is the address of the allocated memory.
1711 is the length of the allocated memory.
1713 is the page offset of the allocated memory.
1715 is a string describing the backing of the allocated memory.
1719 This record indicates when events are lost.
1724 struct perf_event_header header;
1727 struct sample_id sample_id;
1734 is the unique event ID for the samples that were lost.
1737 is the number of events that were lost.
1741 This record indicates a change in the process name.
1746 struct perf_event_header header;
1750 struct sample_id sample_id;
1763 is a string containing the new name of the process.
1767 This record indicates a process exit event.
1772 struct perf_event_header header;
1776 struct sample_id sample_id;
1781 .BR PERF_RECORD_THROTTLE ", " PERF_RECORD_UNTHROTTLE
1782 This record indicates a throttle/unthrottle event.
1787 struct perf_event_header header;
1791 struct sample_id sample_id;
1797 This record indicates a fork event.
1802 struct perf_event_header header;
1806 struct sample_id sample_id;
1812 This record indicates a read event.
1817 struct perf_event_header header;
1819 struct read_format values;
1820 struct sample_id sample_id;
1825 .B PERF_RECORD_SAMPLE
1826 This record indicates a sample.
1831 struct perf_event_header header;
1832 u64 sample_id; /* if PERF_SAMPLE_IDENTIFIER */
1833 u64 ip; /* if PERF_SAMPLE_IP */
1834 u32 pid, tid; /* if PERF_SAMPLE_TID */
1835 u64 time; /* if PERF_SAMPLE_TIME */
1836 u64 addr; /* if PERF_SAMPLE_ADDR */
1837 u64 id; /* if PERF_SAMPLE_ID */
1838 u64 stream_id; /* if PERF_SAMPLE_STREAM_ID */
1839 u32 cpu, res; /* if PERF_SAMPLE_CPU */
1840 u64 period; /* if PERF_SAMPLE_PERIOD */
1841 struct read_format v; /* if PERF_SAMPLE_READ */
1842 u64 nr; /* if PERF_SAMPLE_CALLCHAIN */
1843 u64 ips[nr]; /* if PERF_SAMPLE_CALLCHAIN */
1844 u32 size; /* if PERF_SAMPLE_RAW */
1845 char data[size]; /* if PERF_SAMPLE_RAW */
1846 u64 bnr; /* if PERF_SAMPLE_BRANCH_STACK */
1847 struct perf_branch_entry lbr[bnr];
1848 /* if PERF_SAMPLE_BRANCH_STACK */
1849 u64 abi; /* if PERF_SAMPLE_REGS_USER */
1850 u64 regs[weight(mask)];
1851 /* if PERF_SAMPLE_REGS_USER */
1852 u64 size; /* if PERF_SAMPLE_STACK_USER */
1853 char data[size]; /* if PERF_SAMPLE_STACK_USER */
1854 u64 dyn_size; /* if PERF_SAMPLE_STACK_USER */
1855 u64 weight; /* if PERF_SAMPLE_WEIGHT */
1856 u64 data_src; /* if PERF_SAMPLE_DATA_SRC */
1857 u64 transaction;/* if PERF_SAMPLE_TRANSACTION */
1864 .B PERF_SAMPLE_IDENTIFIER
1865 is enabled, a 64-bit unique ID is included.
1866 This is a duplication of the
1869 value, but included at the beginning of the sample
1870 so parsers can easily obtain the value.
1875 is enabled, then a 64-bit instruction
1876 pointer value is included.
1881 is enabled, then a 32-bit process ID
1882 and 32-bit thread ID are included.
1887 is enabled, then a 64-bit timestamp
1889 This is obtained via local_clock() which is a hardware timestamp
1890 if available and the jiffies value if not.
1895 is enabled, then a 64-bit address is included.
1896 This is usually the address of a tracepoint,
1897 breakpoint, or software event; otherwise the value is 0.
1902 is enabled, a 64-bit unique ID is included.
1903 If the event is a member of an event group, the group leader ID is returned.
1904 This ID is the same as the one returned by
1905 .BR PERF_FORMAT_ID .
1909 .B PERF_SAMPLE_STREAM_ID
1910 is enabled, a 64-bit unique ID is included.
1913 the actual ID is returned, not the group leader.
1914 This ID is the same as the one returned by
1915 .BR PERF_FORMAT_ID .
1920 is enabled, this is a 32-bit value indicating
1921 which CPU was being used, in addition to a reserved (unused)
1926 .B PERF_SAMPLE_PERIOD
1927 is enabled, a 64-bit value indicating
1928 the current sampling period is written.
1933 is enabled, a structure of type read_format
1934 is included which has values for all events in the event group.
1935 The values included depend on the
1938 .BR perf_event_open ()
1943 .B PERF_SAMPLE_CALLCHAIN
1944 is enabled, then a 64-bit number is included
1945 which indicates how many following 64-bit instruction pointers will
1947 This is the current callchain.
1949 .IR size ", " data[size]
1952 is enabled, then a 32-bit value indicating size
1953 is included followed by an array of 8-bit values of length size.
1954 The values are padded with 0 to have 64-bit alignment.
1956 This RAW record data is opaque with respect to the ABI.
1957 The ABI doesn't make any promises with respect to the stability
1958 of its content, it may vary depending
1959 on event, hardware, and kernel version.
1961 .IR bnr ", " lbr[bnr]
1963 .B PERF_SAMPLE_BRANCH_STACK
1964 is enabled, then a 64-bit value indicating
1965 the number of records is included, followed by
1967 .I perf_branch_entry
1968 structures which each include the fields:
1972 This indicates the source instruction (may not be a branch).
1978 The branch target was mispredicted.
1981 The branch target was predicted.
1983 .IR in_tx " (since Linux 3.11)"
1984 .\" commit 135c5612c460f89657c4698fe2ea753f6f667963
1985 The branch was in a transactional memory transaction.
1987 .IR abort " (since Linux 3.11)"
1988 .\" commit 135c5612c460f89657c4698fe2ea753f6f667963
1989 The branch was in an aborted transactional memory transaction.
1992 The entries are from most to least recent, so the first entry
1993 has the most recent branch.
1999 is optional; if not supported, both
2002 The type of branches recorded is specified by the
2003 .I branch_sample_type
2008 .IR abi ", " regs[weight(mask)]
2010 .B PERF_SAMPLE_REGS_USER
2011 is enabled, then the user CPU registers are recorded.
2016 .BR PERF_SAMPLE_REGS_ABI_NONE ", " PERF_SAMPLE_REGS_ABI_32 " or "
2017 .BR PERF_SAMPLE_REGS_ABI_64 .
2021 field is an array of the CPU registers that were specified by
2025 The number of values is the number of bits set in the
2029 .IR size ", " data[size] ", " dyn_size
2031 .B PERF_SAMPLE_STACK_USER
2032 is enabled, then the user stack is recorded.
2033 This can be used to generate stack backtraces.
2035 is the size requested by the user in
2036 .I sample_stack_user
2037 or else the maximum record size.
2039 is the stack data (a raw dump of the memory pointed to by the
2040 stack pointer at the time of sampling).
2042 is the amount of data actually dumped (can be less than
2047 .B PERF_SAMPLE_WEIGHT
2048 is enabled, then a 64-bit value provided by the hardware
2049 is recorded that indicates how costly the event was.
2050 This allows expensive events to stand out more clearly
2055 .B PERF_SAMPLE_DATA_SRC
2056 is enabled, then a 64-bit value is recorded that is made up of
2057 the following fields:
2061 Type of opcode, a bitwise combination of:
2072 .B PERF_MEM_OP_STORE
2075 .B PERF_MEM_OP_PFETCH
2084 Memory hierarchy level hit or miss, a bitwise combination of
2085 the following, shifted left by
2086 .BR PERF_MEM_LVL_SHIFT :
2097 .B PERF_MEM_LVL_MISS
2112 .B PERF_MEM_LVL_LOC_RAM
2115 .B PERF_MEM_LVL_REM_RAM1
2118 .B PERF_MEM_LVL_REM_RAM2
2121 .B PERF_MEM_LVL_REM_CCE1
2124 .B PERF_MEM_LVL_REM_CCE2
2136 Snoop mode, a bitwise combination of the following, shifted left by
2137 .BR PERF_MEM_SNOOP_SHIFT :
2142 .B PERF_MEM_SNOOP_NA
2145 .B PERF_MEM_SNOOP_NONE
2148 .B PERF_MEM_SNOOP_HIT
2151 .B PERF_MEM_SNOOP_MISS
2154 .B PERF_MEM_SNOOP_HITM
2160 Lock instruction, a bitwise combination of the following, shifted left by
2161 .BR PERF_MEM_LOCK_SHIFT :
2169 .B PERF_MEM_LOCK_LOCKED
2175 TLB access hit or miss, a bitwise combination of the following, shifted
2177 .BR PERF_MEM_TLB_SHIFT :
2188 .B PERF_MEM_TLB_MISS
2208 .B PERF_SAMPLE_TRANSACTION
2209 flag is set, then a 64-bit field is recorded describing
2210 the sources of any transactional memory aborts.
2212 The field is a bitwise combination of the following values:
2216 Abort from an elision type transaction (Intel-CPU-specific).
2218 .B PERF_TXN_TRANSACTION
2219 Abort from a generic transaction.
2222 Synchronous abort (related to the reported instruction).
2225 Asynchronous abort (not related to the reported instruction).
2228 Retryable abort (retrying the transaction may have succeeded).
2230 .B PERF_TXN_CONFLICT
2231 Abort due to memory conflicts with other threads.
2233 .B PERF_TXN_CAPACITY_WRITE
2234 Abort due to write capacity overflow.
2236 .B PERF_TXN_CAPACITY_READ
2237 Abort due to read capacity overflow.
2240 In addition, a user-specified abort code can be obtained from
2241 the high 32 bits of the field by shifting right by
2242 .B PERF_TXN_ABORT_SHIFT
2244 .BR PERF_TXN_ABORT_MASK .
2247 .B PERF_RECORD_MMAP2
2248 This record includes extended information on
2250 calls returning executable mappings.
2251 The format is similar to that of the
2253 record, but includes extra values that allow uniquely identifying
2259 struct perf_event_header header;
2272 struct sample_id sample_id;
2284 is the address of the allocated memory.
2287 is the length of the allocated memory.
2290 is the page offset of the allocated memory.
2293 is the major ID of the underlying device.
2296 is the minor ID of the underlying device.
2299 is the inode number.
2302 is the inode generation.
2305 is the protection information.
2308 is the flags information.
2311 is a string describing the backing of the allocated memory.
2314 .SS Overflow handling
2315 Events can be set to notify when a threshold is crossed,
2316 indicating an overflow.
2317 Overflow conditions can be captured by monitoring the
2318 event file descriptor with
2323 Alternately, a SIGIO signal handler can be created and
2324 the event configured with
2326 to generate SIGIO signals.
2328 Overflows are generated only by sampling events
2330 must have a nonzero value).
2332 There are two ways to generate overflow notifications.
2334 The first is to set a
2338 value that will trigger if a certain number of samples
2339 or bytes have been written to the mmap ring buffer.
2344 The other way is by use of the
2345 .B PERF_EVENT_IOC_REFRESH
2347 This ioctl adds to a counter that decrements each time the event overflows.
2351 once the counter reaches 0
2354 the underlying event is disabled.
2356 Starting with Linux 3.18,
2357 .\" commit 179033b3e064d2cd3f5f9945e76b0a0f0fbf4883
2359 is indicated if the event being monitored is attached to a different
2360 process and that process exits.
2361 .SS rdpmc instruction
2362 Starting with Linux 3.4 on x86, you can use the
2363 .\" commit c7206205d00ab375839bd6c7ddb247d600693c09
2365 instruction to get low-latency reads without having to enter the kernel.
2368 is not necessarily faster than other methods for reading event values.
2370 Support for this can be detected with the
2372 field in the mmap page; documentation on how
2373 to calculate event values can be found in that section.
2374 .SS perf_event ioctl calls
2376 Various ioctls act on
2377 .BR perf_event_open ()
2380 .B PERF_EVENT_IOC_ENABLE
2381 This enables the individual event or event group specified by the
2382 file descriptor argument.
2385 .B PERF_IOC_FLAG_GROUP
2386 bit is set in the ioctl argument, then all events in a group are
2387 enabled, even if the event specified is not the group leader
2390 .B PERF_EVENT_IOC_DISABLE
2391 This disables the individual counter or event group specified by the
2392 file descriptor argument.
2394 Enabling or disabling the leader of a group enables or disables the
2395 entire group; that is, while the group leader is disabled, none of the
2396 counters in the group will count.
2397 Enabling or disabling a member of a group other than the leader
2398 affects only that counter; disabling a non-leader
2399 stops that counter from counting but doesn't affect any other counter.
2402 .B PERF_IOC_FLAG_GROUP
2403 bit is set in the ioctl argument, then all events in a group are
2404 disabled, even if the event specified is not the group leader
2407 .B PERF_EVENT_IOC_REFRESH
2408 Non-inherited overflow counters can use this
2409 to enable a counter for a number of overflows specified by the argument,
2410 after which it is disabled.
2411 Subsequent calls of this ioctl add the argument value to the current
2413 An overflow notification with
2415 set will happen on each overflow until the
2416 count reaches 0; when that happens a notification with
2418 set is sent and the event is disabled.
2419 Using an argument of 0 is considered undefined behavior.
2421 .B PERF_EVENT_IOC_RESET
2422 Reset the event count specified by the
2423 file descriptor argument to zero.
2424 This resets only the counts; there is no way to reset the
2432 .B PERF_IOC_FLAG_GROUP
2433 bit is set in the ioctl argument, then all events in a group are
2434 reset, even if the event specified is not the group leader
2437 .B PERF_EVENT_IOC_PERIOD
2438 This updates the overflow period for the event.
2440 Since Linux 3.7 (on ARM)
2441 .\" commit 3581fe0ef37ce12ac7a4f74831168352ae848edc
2442 and Linux 3.14 (all other architectures),
2443 .\" commit bad7192b842c83e580747ca57104dd51fe08c223
2444 the new period takes effect immediately.
2445 On older kernels, the new period did not take effect until
2446 after the next overflow.
2448 The argument is a pointer to a 64-bit value containing the
2451 Prior to Linux 2.6.36
2452 .\" commit ad0cf3478de8677f720ee06393b3147819568d6a
2453 this ioctl always failed due to a bug
2457 .B PERF_EVENT_IOC_SET_OUTPUT
2458 This tells the kernel to report event notifications to the specified
2459 file descriptor rather than the default one.
2460 The file descriptors must all be on the same CPU.
2462 The argument specifies the desired file descriptor, or \-1 if
2463 output should be ignored.
2465 .BR PERF_EVENT_IOC_SET_FILTER " (since Linux 2.6.33)"
2466 .\" commit 6fb2915df7f0747d9044da9dbff5b46dc2e20830
2467 This adds an ftrace filter to this event.
2469 The argument is a pointer to the desired ftrace filter.
2471 .BR PERF_EVENT_IOC_ID " (since Linux 3.12)"
2472 .\" commit cf4957f17f2a89984915ea808876d9c82225b862
2473 This returns the event ID value for the given event file descriptor.
2475 The argument is a pointer to a 64-bit unsigned integer
2478 A process can enable or disable all the event groups that are
2479 attached to it using the
2481 .B PR_TASK_PERF_EVENTS_ENABLE
2483 .B PR_TASK_PERF_EVENTS_DISABLE
2485 This applies to all counters on the calling process, whether created by
2486 this process or by another, and does not affect any counters that this
2487 process has created on other processes.
2488 It enables or disables only
2489 the group leaders, not any other members in the groups.
2490 .SS perf_event related configuration files
2492 .I /proc/sys/kernel/
2495 .I /proc/sys/kernel/perf_event_paranoid
2498 .I perf_event_paranoid
2499 file can be set to restrict access to the performance counters.
2502 allow only user-space measurements.
2504 allow both kernel and user measurements (default).
2506 allow access to CPU-specific data but not raw tracepoint samples.
2511 The existence of the
2512 .I perf_event_paranoid
2513 file is the official method for determining if a kernel supports
2514 .BR perf_event_open ().
2516 .I /proc/sys/kernel/perf_event_max_sample_rate
2518 This sets the maximum sample rate.
2519 Setting this too high can allow
2520 users to sample at a rate that impacts overall machine performance
2521 and potentially lock up the machine.
2522 The default value is
2523 100000 (samples per second).
2525 .I /proc/sys/kernel/perf_event_mlock_kb
2527 Maximum number of pages an unprivileged user can
2529 The default is 516 (kB).
2533 .I /sys/bus/event_source/devices/
2535 Since Linux 2.6.34, the kernel supports having multiple PMUs
2536 available for monitoring.
2537 Information on how to program these PMUs can be found under
2538 .IR /sys/bus/event_source/devices/ .
2539 Each subdirectory corresponds to a different PMU.
2541 .IR /sys/bus/event_source/devices/*/type " (since Linux 2.6.38)"
2542 .\" commit abe43400579d5de0078c2d3a760e6598e183f871
2543 This contains an integer that can be used in the
2547 to indicate that you wish to use this PMU.
2549 .IR /sys/bus/event_source/devices/*/rdpmc " (since Linux 3.4)"
2550 .\" commit 0c9d42ed4cee2aa1dfc3a260b741baae8615744f
2551 If this file is 1, then direct user-space access to the
2552 performance counter registers is allowed via the rdpmc instruction.
2553 This can be disabled by echoing 0 to the file.
2555 .IR /sys/bus/event_source/devices/*/format/ " (since Linux 3.4)"
2556 .\" commit 641cc938815dfd09f8fa1ec72deb814f0938ac33
2557 This subdirectory contains information on the architecture-specific
2558 subfields available for programming the various
2564 The content of each file is the name of the config field, followed
2565 by a colon, followed by a series of integer bit ranges separated by
2567 For example, the file
2569 may contain the value
2570 .I config1:1,6-10,44
2571 which indicates that event is an attribute that occupies bits 1,6-10, and 44
2573 .IR perf_event_attr::config1 .
2575 .IR /sys/bus/event_source/devices/*/events/ " (since Linux 3.4)"
2576 .\" commit 641cc938815dfd09f8fa1ec72deb814f0938ac33
2577 This subdirectory contains files with predefined events.
2578 The contents are strings describing the event settings
2579 expressed in terms of the fields found in the previously mentioned
2582 These are not necessarily complete lists of all events supported by
2583 a PMU, but usually a subset of events deemed useful or interesting.
2585 The content of each file is a list of attribute names
2586 separated by commas.
2587 Each entry has an optional value (either hex or decimal).
2588 If no value is specified, then it is assumed to be a single-bit
2589 field with a value of 1.
2590 An example entry may look like this:
2591 .IR event=0x2,inv,ldlat=3 .
2593 .I /sys/bus/event_source/devices/*/uevent
2594 This file is the standard kernel device interface
2595 for injecting hotplug events.
2597 .IR /sys/bus/event_source/devices/*/cpumask " (since Linux 3.7)"
2598 .\" commit 314d9f63f385096580e9e2a06eaa0745d92fe4ac
2601 file contains a comma-separated list of integers that
2602 indicate a representative CPU number for each socket (package)
2604 This is needed when setting up uncore or northbridge events, as
2605 those PMUs present socket-wide events.
2608 .BR perf_event_open ()
2609 returns the new file descriptor, or \-1 if an error occurred
2612 is set appropriately).
2614 The errors returned by
2615 .BR perf_event_open ()
2616 can be inconsistent, and may
2617 vary across processor architectures and performance monitoring units.
2625 .BR PERF_ATTR_SIZE_VER0 ),
2626 too big (larger than the page size),
2627 or larger than the kernel supports and the extra bytes are not zero.
2633 field is overwritten by the kernel to be the size of the structure
2637 Returned when the requested event requires
2639 permissions (or a more permissive perf_event paranoid setting).
2640 Some common cases where an unprivileged process
2641 may encounter this error:
2642 attaching to a process owned by a different user;
2643 monitoring all processes on a given CPU (i.e., specifying the
2648 when the paranoid setting requires it.
2653 file descriptor is not valid, or, if
2654 .B PERF_FLAG_PID_CGROUP
2656 the cgroup file descriptor in
2663 pointer points at an invalid memory address.
2666 Returned if the specified event is invalid.
2667 There are many possible reasons for this.
2668 A not-exhaustive list:
2670 is higher than the maximum setting;
2673 to monitor does not exist;
2680 value is out of range;
2684 set and the event is not a group leader;
2687 values are out of range or set reserved bits;
2688 the generic event selected is not supported; or
2689 there is not enough room to add the selected event.
2692 Each opened event uses one file descriptor.
2693 If a large number of events are opened the per-user file
2694 descriptor limit (often 1024) will be hit and no more
2695 events can be created.
2698 Returned when the event involves a feature not supported
2704 setting is not valid.
2705 This error is also returned for
2706 some unsupported generic events.
2709 Prior to Linux 3.3, if there was not enough room for the event,
2710 .\" commit aa2bc1ade59003a379ffc485d6da2d92ea3370a6
2713 In Linux 3.3, this was changed to
2716 is still returned if you try to add more breakpoint events
2717 than supported by the hardware.
2721 .B PERF_SAMPLE_STACK_USER
2724 and it is not supported by hardware.
2727 Returned if an event requiring a specific hardware feature is
2728 requested but there is no hardware support.
2729 This includes requesting low-skid events if not supported,
2730 branch tracing if it is not available, sampling if no PMU
2731 interrupt is available, and branch stacks for software events.
2734 Returned on many (but not all) architectures when an unsupported
2735 .IR exclude_hv ", " exclude_idle ", " exclude_user ", or " exclude_kernel
2736 setting is specified.
2738 It can also happen, as with
2740 when the requested event requires
2742 permissions (or a more permissive perf_event paranoid setting).
2743 This includes setting a breakpoint on a kernel address,
2744 and (since Linux 3.13) setting a kernel function-trace tracepoint.
2745 .\" commit a4e95fc2cbb31d70a65beffeaf8773f881328c34
2748 Returned if attempting to attach to a process that does not exist.
2750 .BR perf_event_open ()
2751 was introduced in Linux 2.6.31 but was called
2752 .\" commit 0793a61d4df8daeac6492dbf8d2f3e5713caae5e
2753 .BR perf_counter_open ().
2754 It was renamed in Linux 2.6.32.
2755 .\" commit cdd6c482c9ff9c55475ee7392ec8f672eddb7be6
2758 .BR perf_event_open ()
2759 system call Linux- specific
2760 and should not be used in programs intended to be portable.
2762 Glibc does not provide a wrapper for this system call; call it using
2764 See the example below.
2766 The official way of knowing if
2767 .BR perf_event_open ()
2768 support is enabled is checking
2769 for the existence of the file
2770 .IR /proc/sys/kernel/perf_event_paranoid .
2776 is needed to properly get overflow signals in threads.
2777 This was introduced in Linux 2.6.32.
2778 .\" commit ba0a6c9f6fceed11c6a99e8326f0477fe383e6b5
2780 Prior to Linux 2.6.33 (at least for x86),
2781 .\" commit b690081d4d3f6a23541493f1682835c3cd5c54a1
2782 the kernel did not check
2783 if events could be scheduled together until read time.
2784 The same happens on all known kernels if the NMI watchdog is enabled.
2785 This means to see if a given set of events works you have to
2786 .BR perf_event_open (),
2787 start, then read before you know for sure you
2788 can get valid measurements.
2790 Prior to Linux 2.6.34, event constraints were not enforced by the kernel.
2791 In that case, some events would silently return "0" if the kernel
2792 scheduled them in an improper counter slot.
2793 .\" FIXME: cannot find a kernel commit for this one
2795 Prior to Linux 2.6.34, there was a bug when multiplexing where the
2796 wrong results could be returned.
2797 .\" commit 45e16a6834b6af098702e5ea6c9a40de42ff77d8
2799 Kernels from Linux 2.6.35 to Linux 2.6.39 can quickly crash the kernel if
2800 "inherit" is enabled and many threads are started.
2801 .\" commit 38b435b16c36b0d863efcf3f07b34a6fac9873fd
2803 Prior to Linux 2.6.35,
2804 .\" commit 050735b08ca8a016bbace4445fa025b88fee770b
2805 .B PERF_FORMAT_GROUP
2806 did not work with attached processes.
2808 In older Linux 2.6 versions,
2809 refreshing an event group leader refreshed all siblings,
2810 and refreshing with a parameter of 0 enabled infinite refresh.
2811 This behavior is unsupported and should not be relied on.
2813 There is a bug in the kernel code between
2814 Linux 2.6.36 and Linux 3.0 that ignores the
2815 "watermark" field and acts as if a wakeup_event
2816 was chosen if the union has a
2817 nonzero value in it.
2818 .\" commit 4ec8363dfc1451f8c8f86825731fe712798ada02
2820 From Linux 2.6.31 to Linux 3.4, the
2821 .B PERF_IOC_FLAG_GROUP
2822 ioctl argument was broken and would repeatedly operate
2823 on the event specified rather than iterating across
2824 all sibling events in a group.
2825 .\" commit 724b6daa13e100067c30cfc4d1ad06629609dc4e
2827 From Linux 3.4 to Linux 3.11, the mmap
2828 .\" commit fa7315871046b9a4c48627905691dbde57e51033
2832 bits mapped to the same location.
2833 Code should migrate to the new
2839 Always double-check your results!
2840 Various generalized events have had wrong values.
2841 For example, retired branches measured
2842 the wrong thing on AMD machines until Linux 2.6.35.
2843 .\" commit f287d332ce835f77a4f5077d2c0ef1e3f9ea42d2
2845 The following is a short example that measures the total
2846 instruction count of a call to
2854 #include <sys/ioctl.h>
2855 #include <linux/perf_event.h>
2856 #include <asm/unistd.h>
2859 perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
2860 int cpu, int group_fd, unsigned long flags)
2864 ret = syscall(__NR_perf_event_open, hw_event, pid, cpu,
2870 main(int argc, char **argv)
2872 struct perf_event_attr pe;
2876 memset(&pe, 0, sizeof(struct perf_event_attr));
2877 pe.type = PERF_TYPE_HARDWARE;
2878 pe.size = sizeof(struct perf_event_attr);
2879 pe.config = PERF_COUNT_HW_INSTRUCTIONS;
2881 pe.exclude_kernel = 1;
2884 fd = perf_event_open(&pe, 0, \-1, \-1, 0);
2886 fprintf(stderr, "Error opening leader %llx\\n", pe.config);
2890 ioctl(fd, PERF_EVENT_IOC_RESET, 0);
2891 ioctl(fd, PERF_EVENT_IOC_ENABLE, 0);
2893 printf("Measuring instruction count for this printf\\n");
2895 ioctl(fd, PERF_EVENT_IOC_DISABLE, 0);
2896 read(fd, &count, sizeof(long long));
2898 printf("Used %lld instructions\\n", count);
2910 This page is part of release 3.79 of the Linux
2913 A description of the project,
2914 information about reporting bugs,
2915 and the latest version of this page,
2917 \%http://www.kernel.org/doc/man\-pages/.