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 This flag enables the close-on-exec flag for the created
137 event file descriptor,
138 so that the file descriptor is automatically closed on
140 Setting the close-on-exec flags at creation time, rather than later with
142 avoids potential race conditions where the calling thread invokes
143 .BR perf_event_open ()
146 at the same time as another thread calls
151 .BR PERF_FLAG_FD_NO_GROUP
152 This flag tells the event to ignore the
154 parameter except for the purpose of setting up output redirection
156 .B PERF_FLAG_FD_OUTPUT
159 .BR PERF_FLAG_FD_OUTPUT " (broken since Linux 2.6.35)."
160 This flag re-routes the event's sampled output to instead
161 be included in the mmap buffer of the event specified by
164 .BR PERF_FLAG_PID_CGROUP " (since Linux 2.6.39)."
165 This flag activates per-container system-wide monitoring.
167 is an abstraction that isolates a set of resources for finer-grained
168 control (CPUs, memory, etc.).
169 In this mode, the event is measured
170 only if the thread running on the monitored CPU belongs to the designated
172 The cgroup is identified by passing a file descriptor
173 opened on its directory in the cgroupfs filesystem.
175 cgroup to monitor is called
177 then a file descriptor opened on
179 (assuming cgroupfs is mounted on
181 must be passed as the
184 cgroup monitoring is available only
185 for system-wide events and may therefore require extra permissions.
189 structure provides detailed configuration information
190 for the event being created.
194 struct perf_event_attr {
195 __u32 type; /* Type of event */
196 __u32 size; /* Size of attribute structure */
197 __u64 config; /* Type-specific configuration */
200 __u64 sample_period; /* Period of sampling */
201 __u64 sample_freq; /* Frequency of sampling */
204 __u64 sample_type; /* Specifies values included in sample */
205 __u64 read_format; /* Specifies values returned in read */
207 __u64 disabled : 1, /* off by default */
208 inherit : 1, /* children inherit it */
209 pinned : 1, /* must always be on PMU */
210 exclusive : 1, /* only group on PMU */
211 exclude_user : 1, /* don't count user */
212 exclude_kernel : 1, /* don't count kernel */
213 exclude_hv : 1, /* don't count hypervisor */
214 exclude_idle : 1, /* don't count when idle */
215 mmap : 1, /* include mmap data */
216 comm : 1, /* include comm data */
217 freq : 1, /* use freq, not period */
218 inherit_stat : 1, /* per task counts */
219 enable_on_exec : 1, /* next exec enables */
220 task : 1, /* trace fork/exit */
221 watermark : 1, /* wakeup_watermark */
222 precise_ip : 2, /* skid constraint */
223 mmap_data : 1, /* non-exec mmap data */
224 sample_id_all : 1, /* sample_type all events */
225 exclude_host : 1, /* don't count in host */
226 exclude_guest : 1, /* don't count in guest */
227 exclude_callchain_kernel : 1,
228 /* exclude kernel callchains */
229 exclude_callchain_user : 1,
230 /* exclude user callchains */
231 mmap2 : 1, /* include mmap with inode data */
232 comm_exec : 1, /* flag comm events that are due to exec */
236 __u32 wakeup_events; /* wakeup every n events */
237 __u32 wakeup_watermark; /* bytes before wakeup */
240 __u32 bp_type; /* breakpoint type */
243 __u64 bp_addr; /* breakpoint address */
244 __u64 config1; /* extension of config */
248 __u64 bp_len; /* breakpoint length */
249 __u64 config2; /* extension of config1 */
251 __u64 branch_sample_type; /* enum perf_branch_sample_type */
252 __u64 sample_regs_user; /* user regs to dump on samples */
253 __u32 sample_stack_user; /* size of stack to dump on
255 __u32 __reserved_2; /* Align to u64 */
263 structure are described in more detail below:
266 This field specifies the overall event type.
267 It has one of the following values:
270 .B PERF_TYPE_HARDWARE
271 This indicates one of the "generalized" hardware events provided
275 field definition for more details.
277 .B PERF_TYPE_SOFTWARE
278 This indicates one of the software-defined events provided by the kernel
279 (even if no hardware support is available).
281 .B PERF_TYPE_TRACEPOINT
282 This indicates a tracepoint
283 provided by the kernel tracepoint infrastructure.
285 .B PERF_TYPE_HW_CACHE
286 This indicates a hardware cache event.
287 This has a special encoding, described in the
292 This indicates a "raw" implementation-specific event in the
295 .BR PERF_TYPE_BREAKPOINT " (since Linux 2.6.33)"
296 This indicates a hardware breakpoint as provided by the CPU.
297 Breakpoints can be read/write accesses to an address as well as
298 execution of an instruction address.
302 .BR perf_event_open ()
303 can support multiple PMUs.
304 To enable this, a value exported by the kernel can be used in the
306 field to indicate which PMU to use.
307 The value to use can be found in the sysfs filesystem:
308 there is a subdirectory per PMU instance under
309 .IR /sys/bus/event_source/devices .
310 In each subdirectory there is a
312 file whose content is an integer that can be used in the
316 .I /sys/bus/event_source/devices/cpu/type
317 contains the value for the core CPU PMU, which is usually 4.
323 structure for forward/backward compatibility.
325 .I sizeof(struct perf_event_attr)
326 to allow the kernel to see
327 the struct size at the time of compilation.
330 .B PERF_ATTR_SIZE_VER0
331 is set to 64; this was the size of the first published struct.
332 .B PERF_ATTR_SIZE_VER1
333 is 72, corresponding to the addition of breakpoints in Linux 2.6.33.
334 .B PERF_ATTR_SIZE_VER2
335 is 80 corresponding to the addition of branch sampling in Linux 3.4.
336 .B PERF_ATTR_SIZE_VER3
337 is 96 corresponding to the addition
345 This specifies which event you want, in conjunction with
350 .IR config1 " and " config2
351 fields are also taken into account in cases where 64 bits is not
352 enough to fully specify the event.
353 The encoding of these fields are event dependent.
355 The most significant bit (bit 63) of
357 signifies CPU-specific (raw) counter configuration data;
358 if the most significant bit is unset, the next 7 bits are an event
359 type and the rest of the bits are the event identifier.
361 There are various ways to set the
363 field that are dependent on the value of the previously
367 What follows are various possible settings for
375 .BR PERF_TYPE_HARDWARE ,
376 we are measuring one of the generalized hardware CPU events.
377 Not all of these are available on all platforms.
380 to one of the following:
383 .B PERF_COUNT_HW_CPU_CYCLES
385 Be wary of what happens during CPU frequency scaling.
387 .B PERF_COUNT_HW_INSTRUCTIONS
388 Retired instructions.
389 Be careful, these can be affected by various
390 issues, most notably hardware interrupt counts.
392 .B PERF_COUNT_HW_CACHE_REFERENCES
394 Usually this indicates Last Level Cache accesses but this may
395 vary depending on your CPU.
396 This may include prefetches and coherency messages; again this
397 depends on the design of your CPU.
399 .B PERF_COUNT_HW_CACHE_MISSES
401 Usually this indicates Last Level Cache misses; this is intended to be
402 used in conjunction with the
403 .B PERF_COUNT_HW_CACHE_REFERENCES
404 event to calculate cache miss rates.
406 .B PERF_COUNT_HW_BRANCH_INSTRUCTIONS
407 Retired branch instructions.
408 Prior to Linux 2.6.34, this used
409 the wrong event on AMD processors.
411 .B PERF_COUNT_HW_BRANCH_MISSES
412 Mispredicted branch instructions.
414 .B PERF_COUNT_HW_BUS_CYCLES
415 Bus cycles, which can be different from total cycles.
417 .BR PERF_COUNT_HW_STALLED_CYCLES_FRONTEND " (since Linux 3.0)"
418 Stalled cycles during issue.
420 .BR PERF_COUNT_HW_STALLED_CYCLES_BACKEND " (since Linux 3.0)"
421 Stalled cycles during retirement.
423 .BR PERF_COUNT_HW_REF_CPU_CYCLES " (since Linux 3.3)"
424 Total cycles; not affected by CPU frequency scaling.
430 .BR PERF_TYPE_SOFTWARE ,
431 we are measuring software events provided by the kernel.
434 to one of the following:
437 .B PERF_COUNT_SW_CPU_CLOCK
438 This reports the CPU clock, a high-resolution per-CPU timer.
440 .B PERF_COUNT_SW_TASK_CLOCK
441 This reports a clock count specific to the task that is running.
443 .B PERF_COUNT_SW_PAGE_FAULTS
444 This reports the number of page faults.
446 .B PERF_COUNT_SW_CONTEXT_SWITCHES
447 This counts context switches.
448 Until Linux 2.6.34, these were all reported as user-space
449 events, after that they are reported as happening in the kernel.
451 .B PERF_COUNT_SW_CPU_MIGRATIONS
452 This reports the number of times the process
453 has migrated to a new CPU.
455 .B PERF_COUNT_SW_PAGE_FAULTS_MIN
456 This counts the number of minor page faults.
457 These did not require disk I/O to handle.
459 .B PERF_COUNT_SW_PAGE_FAULTS_MAJ
460 This counts the number of major page faults.
461 These required disk I/O to handle.
463 .BR PERF_COUNT_SW_ALIGNMENT_FAULTS " (since Linux 2.6.33)"
464 This counts the number of alignment faults.
465 These happen when unaligned memory accesses happen; the kernel
466 can handle these but it reduces performance.
467 This happens only on some architectures (never on x86).
469 .BR PERF_COUNT_SW_EMULATION_FAULTS " (since Linux 2.6.33)"
470 This counts the number of emulation faults.
471 The kernel sometimes traps on unimplemented instructions
472 and emulates them for user space.
473 This can negatively impact performance.
475 .BR PERF_COUNT_SW_DUMMY " (since Linux 3.12)"
476 This is a placeholder event that counts nothing.
477 Informational sample record types such as mmap or comm
478 must be associated with an active event.
479 This dummy event allows gathering such records without requiring
487 .BR PERF_TYPE_TRACEPOINT ,
488 then we are measuring kernel tracepoints.
491 can be obtained from under debugfs
492 .I tracing/events/*/*/id
493 if ftrace is enabled in the kernel.
500 .BR PERF_TYPE_HW_CACHE ,
501 then we are measuring a hardware CPU cache event.
502 To calculate the appropriate
504 value use the following equation:
508 (perf_hw_cache_id) | (perf_hw_cache_op_id << 8) |
509 (perf_hw_cache_op_result_id << 16)
517 .B PERF_COUNT_HW_CACHE_L1D
518 for measuring Level 1 Data Cache
520 .B PERF_COUNT_HW_CACHE_L1I
521 for measuring Level 1 Instruction Cache
523 .B PERF_COUNT_HW_CACHE_LL
524 for measuring Last-Level Cache
526 .B PERF_COUNT_HW_CACHE_DTLB
527 for measuring the Data TLB
529 .B PERF_COUNT_HW_CACHE_ITLB
530 for measuring the Instruction TLB
532 .B PERF_COUNT_HW_CACHE_BPU
533 for measuring the branch prediction unit
535 .BR PERF_COUNT_HW_CACHE_NODE " (since Linux 3.0)"
536 for measuring local memory accesses
540 .I perf_hw_cache_op_id
544 .B PERF_COUNT_HW_CACHE_OP_READ
547 .B PERF_COUNT_HW_CACHE_OP_WRITE
550 .B PERF_COUNT_HW_CACHE_OP_PREFETCH
551 for prefetch accesses
555 .I perf_hw_cache_op_result_id
559 .B PERF_COUNT_HW_CACHE_RESULT_ACCESS
562 .B PERF_COUNT_HW_CACHE_RESULT_MISS
574 Most CPUs support events that are not covered by the "generalized" events.
575 These are implementation defined; see your CPU manual (for example
576 the Intel Volume 3B documentation or the AMD BIOS and Kernel Developer
578 The libpfm4 library can be used to translate from the name in the
579 architectural manuals to the raw hex value
580 .BR perf_event_open ()
581 expects in this field.
586 .BR PERF_TYPE_BREAKPOINT ,
590 Its parameters are set in other places.
593 .IR sample_period ", " sample_freq
594 A "sampling" counter is one that generates an interrupt
595 every N events, where N is given by
597 A sampling counter has
598 .IR sample_period " > 0."
599 When an overflow interrupt occurs, requested data is recorded
603 field controls what data is recorded on each interrupt.
606 can be used if you wish to use frequency rather than period.
607 In this case, you set the
610 The kernel will adjust the sampling period
611 to try and achieve the desired rate.
612 The rate of adjustment is a
616 The various bits in this field specify which values to include
618 They will be recorded in a ring-buffer,
619 which is available to user space using
621 The order in which the values are saved in the
622 sample are documented in the MMAP Layout subsection below;
624 .I "enum perf_event_sample_format"
629 Records instruction pointer.
632 Records the process and thread IDs.
638 Records an address, if applicable.
641 Record counter values for all events in a group, not just the group leader.
643 .B PERF_SAMPLE_CALLCHAIN
644 Records the callchain (stack backtrace).
647 Records a unique ID for the opened event's group leader.
652 .B PERF_SAMPLE_PERIOD
653 Records the current sampling period.
655 .B PERF_SAMPLE_STREAM_ID
656 Records a unique ID for the opened event.
659 the actual ID is returned, not the group leader.
660 This ID is the same as the one returned by
664 Records additional data, if applicable.
665 Usually returned by tracepoint events.
667 .BR PERF_SAMPLE_BRANCH_STACK " (since Linux 3.4)"
668 This provides a record of recent branches, as provided
669 by CPU branch sampling hardware (such as Intel Last Branch Record).
670 Not all hardware supports this feature.
673 .I branch_sample_type
674 field for how to filter which branches are reported.
676 .BR PERF_SAMPLE_REGS_USER " (since Linux 3.7)"
677 Records the current user-level CPU register state
678 (the values in the process before the kernel was called).
680 .BR PERF_SAMPLE_STACK_USER " (since Linux 3.7)"
681 Records the user level stack, allowing stack unwinding.
683 .BR PERF_SAMPLE_WEIGHT " (since Linux 3.10)"
684 Records a hardware provided weight value that expresses how
685 costly the sampled event was.
686 This allows the hardware to highlight expensive events in
689 .BR PERF_SAMPLE_DATA_SRC " (since Linux 3.10)"
690 Records the data source: where in the memory hierarchy
691 the data associated with the sampled instruction came from.
692 This is only available if the underlying hardware
693 supports this feature.
695 .BR PERF_SAMPLE_IDENTIFIER " (since Linux 3.12)"
698 value in a fixed position in the record,
699 either at the beginning (for sample events) or at the end
700 (if a non-sample event).
702 This was necessary because a sample stream may have
703 records from various different event sources with different
706 Parsing the event stream properly was not possible because the
707 format of the record was needed to find
710 the format could not be found without knowing what
711 event the sample belonged to (causing a circular
715 .B PERF_SAMPLE_IDENTIFIER
716 setting makes the event stream always parsable
719 in a fixed location, even though
720 it means having duplicate
724 .BR PERF_SAMPLE_TRANSACTION " (Since Linux 3.13)"
725 Records reasons for transactional memory abort events
726 (for example, from Intel TSX transactional memory support).
730 setting must be greater than 0 and a transactional memory abort
731 event must be measured or no values will be recorded.
732 Also note that some perf_event measurements, such as sampled
733 cycle counting, may cause extraneous aborts (by causing an
734 interrupt during a transaction).
738 This field specifies the format of the data returned by
741 .BR perf_event_open ()
745 .B PERF_FORMAT_TOTAL_TIME_ENABLED
749 This can be used to calculate estimated totals if
750 the PMU is overcommitted and multiplexing is happening.
752 .B PERF_FORMAT_TOTAL_TIME_RUNNING
756 This can be used to calculate estimated totals if
757 the PMU is overcommitted and multiplexing is happening.
760 Adds a 64-bit unique value that corresponds to the event group.
763 Allows all counter values in an event group to be read with one read.
769 bit specifies whether the counter starts out disabled or enabled.
770 If disabled, the event can later be enabled by
776 When creating an event group, typically the group leader is initialized
779 set to 1 and any child events are initialized with
784 being 0, the child events will not start until the group leader
790 bit specifies that this counter should count events of child
791 tasks as well as the task specified.
792 This applies only to new children, not to any existing children at
793 the time the counter is created (nor to any new children of
796 Inherit does not work for some combinations of
799 .BR PERF_FORMAT_GROUP .
804 bit specifies that the counter should always be on the CPU if at all
806 It applies only to hardware counters and only to group leaders.
807 If a pinned counter cannot be put onto the CPU (e.g., because there are
808 not enough hardware counters or because of a conflict with some other
809 event), then the counter goes into an 'error' state, where reads
810 return end-of-file (i.e.,
812 returns 0) until the counter is subsequently enabled or disabled.
817 bit specifies that when this counter's group is on the CPU,
818 it should be the only group using the CPU's counters.
819 In the future this may allow monitoring programs to
820 support PMU features that need to run alone so that they do not
821 disrupt other hardware counters.
823 Note that many unexpected situations may prevent events with the
825 bit set from ever running.
826 This includes any users running a system-wide
827 measurement as well as any kernel use of the performance counters
828 (including the commonly enabled NMI Watchdog Timer interface).
831 If this bit is set, the count excludes events that happen in user space.
834 If this bit is set, the count excludes events that happen in kernel-space.
837 If this bit is set, the count excludes events that happen in the
839 This is mainly for PMUs that have built-in support for handling this
841 Extra support is needed for handling hypervisor measurements on most
845 If set, don't count when the CPU is idle.
850 bit enables generation of
857 This allows tools to notice new executable code being mapped into
858 a program (dynamic shared libraries for example)
859 so that addresses can be mapped back to the original code.
864 bit enables tracking of process command name as modified by the
867 .BR prctl (PR_SET_NAME)
868 system calls as well as writing to
869 .IR /proc/self/comm .
872 flag is also successfully set (possible since Linux 3.16),
874 .B PERF_RECORD_MISC_COMM_EXEC
875 can be used to differentiate the
877 case from the others.
880 If this bit is set, then
884 is used when setting up the sampling interval.
887 This bit enables saving of event counts on context switch for
889 This is meaningful only if the
894 If this bit is set, a counter is automatically
895 enabled after a call to
899 If this bit is set, then
900 fork/exit notifications are included in the ring buffer.
903 If set, have a sampling interrupt happen when we cross the
906 Otherwise, interrupts happen after
910 .IR "precise_ip" " (since Linux 2.6.35)"
911 This controls the amount of skid.
912 Skid is how many instructions
913 execute between an event of interest happening and the kernel
914 being able to stop and record the event.
916 better and allows more accurate reporting of which events
917 correspond to which instructions, but hardware is often limited
918 with how small this can be.
920 The values of this are the following:
925 can have arbitrary skid.
929 must have constant skid.
933 requested to have 0 skid.
939 .BR PERF_RECORD_MISC_EXACT_IP .
942 .IR "mmap_data" " (since Linux 2.6.36)"
943 The counterpart of the
946 This enables generation of
950 calls that do not have
952 set (for example data and SysV shared memory).
954 .IR "sample_id_all" " (since Linux 2.6.38)"
955 If set, then TID, TIME, ID, STREAM_ID, and CPU can
956 additionally be included in
957 .RB non- PERF_RECORD_SAMPLE s
963 .B PERF_SAMPLE_IDENTIFIER
964 is specified, then an additional ID value is included
965 as the last value to ease parsing the record stream.
968 value appearing twice.
970 The layout is described by this pseudo-structure:
974 { u32 pid, tid; } /* if PERF_SAMPLE_TID set */
975 { u64 time; } /* if PERF_SAMPLE_TIME set */
976 { u64 id; } /* if PERF_SAMPLE_ID set */
977 { u64 stream_id;} /* if PERF_SAMPLE_STREAM_ID set */
978 { u32 cpu, res; } /* if PERF_SAMPLE_CPU set */
979 { u64 id; } /* if PERF_SAMPLE_IDENTIFIER set */
983 .IR "exclude_host" " (since Linux 3.2)"
984 Do not measure time spent in VM host.
986 .IR "exclude_guest" " (since Linux 3.2)"
987 Do not measure time spent in VM guest.
989 .IR "exclude_callchain_kernel" " (since Linux 3.7)"
990 Do not include kernel callchains.
992 .IR "exclude_callchain_user" " (since Linux 3.7)"
993 Do not include user callchains.
995 .IR "mmap2" " (since Linux 3.16)"
996 Generate an extended executable mmap record that contains enough
997 additional information to uniquely identify shared mappings.
1000 flag must also be set for this to work.
1002 .IR "comm_exec" " (since Linux 3.16)"
1003 This is purely a feature-detection flag, it does not change
1005 If this flag can successfully be set, then, when
1008 .B PERF_RECORD_MISC_COMM_EXEC
1009 flag will be set in the
1011 field of a comm record header if the rename event being
1012 reported was caused by a call to
1014 This allows tools to distinguish between the various
1015 types of process renaming.
1017 .IR "wakeup_events" ", " "wakeup_watermark"
1018 This union sets how many samples
1019 .RI ( wakeup_events )
1021 .RI ( wakeup_watermark )
1022 happen before an overflow signal happens.
1023 Which one is used is selected by the
1029 .B PERF_RECORD_SAMPLE
1031 To receive a signal for every incoming
1037 .IR "bp_type" " (since Linux 2.6.33)"
1038 This chooses the breakpoint type.
1042 .BR HW_BREAKPOINT_EMPTY
1046 Count when we read the memory location.
1049 Count when we write the memory location.
1051 .BR HW_BREAKPOINT_RW
1052 Count when we read or write the memory location.
1055 Count when we execute code at the memory location.
1057 The values can be combined via a bitwise or, but the
1067 .IR "bp_addr" " (since Linux 2.6.33)"
1069 address of the breakpoint.
1070 For execution breakpoints this is the memory address of the instruction
1071 of interest; for read and write breakpoints it is the memory address
1072 of the memory location of interest.
1074 .IR "config1" " (since Linux 2.6.39)"
1076 is used for setting events that need an extra register or otherwise
1077 do not fit in the regular config field.
1078 Raw OFFCORE_EVENTS on Nehalem/Westmere/SandyBridge use this field
1079 on 3.3 and later kernels.
1081 .IR "bp_len" " (since Linux 2.6.33)"
1083 is the length of the breakpoint being measured if
1086 .BR PERF_TYPE_BREAKPOINT .
1088 .BR HW_BREAKPOINT_LEN_1 ,
1089 .BR HW_BREAKPOINT_LEN_2 ,
1090 .BR HW_BREAKPOINT_LEN_4 ,
1091 .BR HW_BREAKPOINT_LEN_8 .
1092 For an execution breakpoint, set this to
1095 .IR "config2" " (since Linux 2.6.39)"
1098 is a further extension of the
1102 .IR "branch_sample_type" " (since Linux 3.4)"
1104 .B PERF_SAMPLE_BRANCH_STACK
1105 is enabled, then this specifies what branches to include
1106 in the branch record.
1108 The first part of the value is the privilege level, which
1109 is a combination of one of the following values.
1110 If the user does not set privilege level explicitly, the kernel
1111 will use the event's privilege level.
1112 Event and branch privilege levels do not have to match.
1115 .B PERF_SAMPLE_BRANCH_USER
1116 Branch target is in user space.
1118 .B PERF_SAMPLE_BRANCH_KERNEL
1119 Branch target is in kernel space.
1121 .B PERF_SAMPLE_BRANCH_HV
1122 Branch target is in hypervisor.
1124 .B PERF_SAMPLE_BRANCH_PLM_ALL
1125 A convenience value that is the three preceding values ORed together.
1128 In addition to the privilege value, at least one or more of the
1129 following bits must be set.
1132 .B PERF_SAMPLE_BRANCH_ANY
1135 .B PERF_SAMPLE_BRANCH_ANY_CALL
1138 .B PERF_SAMPLE_BRANCH_ANY_RETURN
1141 .B PERF_SAMPLE_BRANCH_IND_CALL
1144 .BR PERF_SAMPLE_BRANCH_COND " (since Linux 3.16)"
1145 Conditional branches.
1147 .BR PERF_SAMPLE_BRANCH_ABORT_TX " (since Linux 3.11)"
1148 Transactional memory aborts.
1150 .BR PERF_SAMPLE_BRANCH_IN_TX " (since Linux 3.11)"
1151 Branch in transactional memory transaction.
1153 .BR PERF_SAMPLE_BRANCH_NO_TX " (since Linux 3.11)"
1154 Branch not in transactional memory transaction.
1158 .IR "sample_regs_user" " (since Linux 3.7)"
1159 This bit mask defines the set of user CPU registers to dump on samples.
1160 The layout of the register mask is architecture-specific and
1161 described in the kernel header
1162 .IR arch/ARCH/include/uapi/asm/perf_regs.h .
1164 .IR "sample_stack_user" " (since Linux 3.7)"
1165 This defines the size of the user stack to dump if
1166 .B PERF_SAMPLE_STACK_USER
1170 .BR perf_event_open ()
1171 file descriptor has been opened, the values
1172 of the events can be read from the file descriptor.
1173 The values that are there are specified by the
1177 structure at open time.
1179 If you attempt to read into a buffer that is not big enough to hold the
1184 Here is the layout of the data returned by a read:
1187 .B PERF_FORMAT_GROUP
1188 was specified to allow reading all events in a group at once:
1192 struct read_format {
1193 u64 nr; /* The number of events */
1194 u64 time_enabled; /* if PERF_FORMAT_TOTAL_TIME_ENABLED */
1195 u64 time_running; /* if PERF_FORMAT_TOTAL_TIME_RUNNING */
1197 u64 value; /* The value of the event */
1198 u64 id; /* if PERF_FORMAT_ID */
1205 .B PERF_FORMAT_GROUP
1212 struct read_format {
1213 u64 value; /* The value of the event */
1214 u64 time_enabled; /* if PERF_FORMAT_TOTAL_TIME_ENABLED */
1215 u64 time_running; /* if PERF_FORMAT_TOTAL_TIME_RUNNING */
1216 u64 id; /* if PERF_FORMAT_ID */
1221 The values read are as follows:
1224 The number of events in this file descriptor.
1226 .B PERF_FORMAT_GROUP
1229 .IR time_enabled ", " time_running
1230 Total time the event was enabled and running.
1231 Normally these are the same.
1232 If more events are started,
1233 then available counter slots on the PMU, then multiplexing
1234 happens and events run only part of the time.
1239 values can be used to scale an estimated value for the count.
1242 An unsigned 64-bit value containing the counter result.
1245 A globally unique value for this particular event, only there if
1251 .BR perf_event_open ()
1252 in sampled mode, asynchronous events
1253 (like counter overflow or
1256 are logged into a ring-buffer.
1257 This ring-buffer is created and accessed through
1260 The mmap size should be 1+2^n pages, where the first page is a
1262 .RI ( "struct perf_event_mmap_page" )
1263 that contains various
1264 bits of information such as where the ring-buffer head is.
1266 Before kernel 2.6.39, there is a bug that means you must allocate a mmap
1267 ring buffer when sampling even if you do not plan to access it.
1269 The structure of the first metadata mmap page is as follows:
1273 struct perf_event_mmap_page {
1274 __u32 version; /* version number of this structure */
1275 __u32 compat_version; /* lowest version this is compat with */
1276 __u32 lock; /* seqlock for synchronization */
1277 __u32 index; /* hardware counter identifier */
1278 __s64 offset; /* add to hardware counter value */
1279 __u64 time_enabled; /* time event active */
1280 __u64 time_running; /* time event on CPU */
1284 __u64 cap_usr_time / cap_usr_rdpmc / cap_bit0 : 1,
1285 cap_bit0_is_deprecated : 1,
1288 cap_user_time_zero : 1,
1295 __u64 __reserved[120]; /* Pad to 1k */
1296 __u64 data_head; /* head in the data section */
1297 __u64 data_tail; /* user-space written tail */
1302 The following list describes the fields in the
1303 .I perf_event_mmap_page
1304 structure in more detail:
1307 Version number of this structure.
1310 The lowest version this is compatible with.
1313 A seqlock for synchronization.
1316 A unique hardware counter identifier.
1319 When using rdpmc for reads this offset value
1320 must be added to the one returned by rdpmc to get
1321 the current total event count.
1324 Time the event was active.
1327 Time the event was running.
1329 .IR cap_usr_time " / " cap_usr_rdpmc " / " cap_bit0 " (since Linux 3.4)"
1330 There was a bug in the definition of
1334 from Linux 3.4 until Linux 3.11.
1335 Both bits were defined to point to the same location, so it was
1336 impossible to know if
1342 Starting with 3.12 these are renamed to
1344 and you should use the new
1351 .IR cap_bit0_is_deprecated " (since Linux 3.12)"
1352 If set, this bit indicates that the kernel supports
1353 the properly separated
1359 If not-set, it indicates an older kernel where
1363 map to the same bit and thus both features should
1364 be used with caution.
1367 .IR cap_user_rdpmc " (since Linux 3.12)"
1368 If the hardware supports user-space read of performance counters
1369 without syscall (this is the "rdpmc" instruction on x86), then
1370 the following code can be used to do a read:
1374 u32 seq, time_mult, time_shift, idx, width;
1375 u64 count, enabled, running;
1376 u64 cyc, time_offset;
1381 enabled = pc\->time_enabled;
1382 running = pc\->time_running;
1384 if (pc\->cap_usr_time && enabled != running) {
1386 time_offset = pc\->time_offset;
1387 time_mult = pc\->time_mult;
1388 time_shift = pc\->time_shift;
1392 count = pc\->offset;
1394 if (pc\->cap_usr_rdpmc && idx) {
1395 width = pc\->pmc_width;
1396 count += rdpmc(idx \- 1);
1400 } while (pc\->lock != seq);
1404 .IR cap_user_time " (since Linux 3.12)"
1405 This bit indicates the hardware has a constant, nonstop
1406 timestamp counter (TSC on x86).
1408 .IR cap_user_time_zero " (since Linux 3.12)"
1409 Indicates the presence of
1411 which allows mapping timestamp values to
1417 this field provides the bit-width of the value
1418 read using the rdpmc or equivalent instruction.
1419 This can be used to sign extend the result like:
1423 pmc <<= 64 \- pmc_width;
1424 pmc >>= 64 \- pmc_width; // signed shift right
1429 .IR time_shift ", " time_mult ", " time_offset
1433 these fields can be used to compute the time
1434 delta since time_enabled (in nanoseconds) using rdtsc or similar.
1439 quot = (cyc >> time_shift);
1440 rem = cyc & ((1 << time_shift) \- 1);
1441 delta = time_offset + quot * time_mult +
1442 ((rem * time_mult) >> time_shift);
1452 seqcount loop described above.
1453 This delta can then be added to
1454 enabled and possible running (if idx), improving the scaling:
1460 quot = count / running;
1461 rem = count % running;
1462 count = quot * enabled + (rem * enabled) / running;
1465 .IR time_zero " (since Linux 3.12)"
1468 .I cap_usr_time_zero
1469 is set, then the hardware clock (the TSC timestamp counter on x86)
1470 can be calculated from the
1471 .IR time_zero ", " time_mult ", and " time_shift " values:"
1474 time = timestamp - time_zero;
1475 quot = time / time_mult;
1476 rem = time % time_mult;
1477 cyc = (quot << time_shift) + (rem << time_shift) / time_mult;
1483 quot = cyc >> time_shift;
1484 rem = cyc & ((1 << time_shift) - 1);
1485 timestamp = time_zero + quot * time_mult +
1486 ((rem * time_mult) >> time_shift);
1490 This points to the head of the data section.
1491 The value continuously increases, it does not wrap.
1492 The value needs to be manually wrapped by the size of the mmap buffer
1493 before accessing the samples.
1495 On SMP-capable platforms, after reading the
1498 user space should issue an rmb().
1505 value should be written by user space to reflect the last read data.
1506 In this case, the kernel will not overwrite unread data.
1508 The following 2^n ring-buffer pages have the layout described below.
1511 .I perf_event_attr.sample_id_all
1512 is set, then all event types will
1513 have the sample_type selected fields related to where/when (identity)
1514 an event took place (TID, TIME, ID, CPU, STREAM_ID) described in
1515 .B PERF_RECORD_SAMPLE
1516 below, it will be stashed just after the
1517 .I perf_event_header
1518 and the fields already present for the existing
1519 fields, that is, at the end of the payload.
1520 That way a newer perf.data
1521 file will be supported by older perf tools, with these new optional
1522 fields being ignored.
1524 The mmap values start with a header:
1528 struct perf_event_header {
1536 Below, we describe the
1537 .I perf_event_header
1538 fields in more detail.
1539 For ease of reading,
1540 the fields with shorter descriptions are presented first.
1543 This indicates the size of the record.
1548 field contains additional information about the sample.
1550 The CPU mode can be determined from this value by masking with
1551 .B PERF_RECORD_MISC_CPUMODE_MASK
1552 and looking for one of the following (note these are not
1553 bit masks, only one can be set at a time):
1556 .B PERF_RECORD_MISC_CPUMODE_UNKNOWN
1559 .B PERF_RECORD_MISC_KERNEL
1560 Sample happened in the kernel.
1562 .B PERF_RECORD_MISC_USER
1563 Sample happened in user code.
1565 .B PERF_RECORD_MISC_HYPERVISOR
1566 Sample happened in the hypervisor.
1568 .B PERF_RECORD_MISC_GUEST_KERNEL
1569 Sample happened in the guest kernel.
1571 .B PERF_RECORD_MISC_GUEST_USER
1572 Sample happened in guest user code.
1576 In addition, one of the following bits can be set:
1578 .B PERF_RECORD_MISC_MMAP_DATA
1579 This is set when the mapping is not executable;
1580 otherwise the mapping is executable.
1582 .B PERF_RECORD_MISC_COMM_EXEC
1585 record on kernels more recent than Linux 3.16
1586 if a process name change was caused by an
1590 .B PERF_RECORD_MISC_MMAP_DATA
1591 since the two values would not be set in the same record.
1593 .B PERF_RECORD_MISC_EXACT_IP
1594 This indicates that the content of
1597 to the actual instruction that triggered the event.
1599 .IR perf_event_attr.precise_ip .
1601 .B PERF_RECORD_MISC_EXT_RESERVED
1602 This indicates there is extended data available (currently not used).
1608 value is one of the below.
1609 The values in the corresponding record (that follows the header)
1617 The MMAP events record the
1619 mappings so that we can correlate
1620 user-space IPs to code.
1621 They have the following structure:
1626 struct perf_event_header header;
1644 is the address of the allocated memory.
1646 is the length of the allocated memory.
1648 is the page offset of the allocated memory.
1650 is a string describing the backing of the allocated memory.
1654 This record indicates when events are lost.
1659 struct perf_event_header header;
1662 struct sample_id sample_id;
1669 is the unique event ID for the samples that were lost.
1672 is the number of events that were lost.
1676 This record indicates a change in the process name.
1681 struct perf_event_header header;
1685 struct sample_id sample_id;
1698 is a string containing the new name of the process.
1702 This record indicates a process exit event.
1707 struct perf_event_header header;
1711 struct sample_id sample_id;
1716 .BR PERF_RECORD_THROTTLE ", " PERF_RECORD_UNTHROTTLE
1717 This record indicates a throttle/unthrottle event.
1722 struct perf_event_header header;
1726 struct sample_id sample_id;
1732 This record indicates a fork event.
1737 struct perf_event_header header;
1741 struct sample_id sample_id;
1747 This record indicates a read event.
1752 struct perf_event_header header;
1754 struct read_format values;
1755 struct sample_id sample_id;
1760 .B PERF_RECORD_SAMPLE
1761 This record indicates a sample.
1766 struct perf_event_header header;
1767 u64 sample_id; /* if PERF_SAMPLE_IDENTIFIER */
1768 u64 ip; /* if PERF_SAMPLE_IP */
1769 u32 pid, tid; /* if PERF_SAMPLE_TID */
1770 u64 time; /* if PERF_SAMPLE_TIME */
1771 u64 addr; /* if PERF_SAMPLE_ADDR */
1772 u64 id; /* if PERF_SAMPLE_ID */
1773 u64 stream_id; /* if PERF_SAMPLE_STREAM_ID */
1774 u32 cpu, res; /* if PERF_SAMPLE_CPU */
1775 u64 period; /* if PERF_SAMPLE_PERIOD */
1776 struct read_format v; /* if PERF_SAMPLE_READ */
1777 u64 nr; /* if PERF_SAMPLE_CALLCHAIN */
1778 u64 ips[nr]; /* if PERF_SAMPLE_CALLCHAIN */
1779 u32 size; /* if PERF_SAMPLE_RAW */
1780 char data[size]; /* if PERF_SAMPLE_RAW */
1781 u64 bnr; /* if PERF_SAMPLE_BRANCH_STACK */
1782 struct perf_branch_entry lbr[bnr];
1783 /* if PERF_SAMPLE_BRANCH_STACK */
1784 u64 abi; /* if PERF_SAMPLE_REGS_USER */
1785 u64 regs[weight(mask)];
1786 /* if PERF_SAMPLE_REGS_USER */
1787 u64 size; /* if PERF_SAMPLE_STACK_USER */
1788 char data[size]; /* if PERF_SAMPLE_STACK_USER */
1789 u64 dyn_size; /* if PERF_SAMPLE_STACK_USER */
1790 u64 weight; /* if PERF_SAMPLE_WEIGHT */
1791 u64 data_src; /* if PERF_SAMPLE_DATA_SRC */
1792 u64 transaction;/* if PERF_SAMPLE_TRANSACTION */
1799 .B PERF_SAMPLE_IDENTIFIER
1800 is enabled, a 64-bit unique ID is included.
1801 This is a duplication of the
1804 value, but included at the beginning of the sample
1805 so parsers can easily obtain the value.
1810 is enabled, then a 64-bit instruction
1811 pointer value is included.
1816 is enabled, then a 32-bit process ID
1817 and 32-bit thread ID are included.
1822 is enabled, then a 64-bit timestamp
1824 This is obtained via local_clock() which is a hardware timestamp
1825 if available and the jiffies value if not.
1830 is enabled, then a 64-bit address is included.
1831 This is usually the address of a tracepoint,
1832 breakpoint, or software event; otherwise the value is 0.
1837 is enabled, a 64-bit unique ID is included.
1838 If the event is a member of an event group, the group leader ID is returned.
1839 This ID is the same as the one returned by
1840 .BR PERF_FORMAT_ID .
1844 .B PERF_SAMPLE_STREAM_ID
1845 is enabled, a 64-bit unique ID is included.
1848 the actual ID is returned, not the group leader.
1849 This ID is the same as the one returned by
1850 .BR PERF_FORMAT_ID .
1855 is enabled, this is a 32-bit value indicating
1856 which CPU was being used, in addition to a reserved (unused)
1861 .B PERF_SAMPLE_PERIOD
1862 is enabled, a 64-bit value indicating
1863 the current sampling period is written.
1868 is enabled, a structure of type read_format
1869 is included which has values for all events in the event group.
1870 The values included depend on the
1873 .BR perf_event_open ()
1878 .B PERF_SAMPLE_CALLCHAIN
1879 is enabled, then a 64-bit number is included
1880 which indicates how many following 64-bit instruction pointers will
1882 This is the current callchain.
1884 .IR size ", " data[size]
1887 is enabled, then a 32-bit value indicating size
1888 is included followed by an array of 8-bit values of length size.
1889 The values are padded with 0 to have 64-bit alignment.
1891 This RAW record data is opaque with respect to the ABI.
1892 The ABI doesn't make any promises with respect to the stability
1893 of its content, it may vary depending
1894 on event, hardware, and kernel version.
1896 .IR bnr ", " lbr[bnr]
1898 .B PERF_SAMPLE_BRANCH_STACK
1899 is enabled, then a 64-bit value indicating
1900 the number of records is included, followed by
1902 .I perf_branch_entry
1903 structures which each include the fields:
1907 This indicates the source instruction (may not be a branch).
1913 The branch target was mispredicted.
1916 The branch target was predicted.
1918 .IR in_tx " (since Linux 3.11)"
1919 The branch was in a transactional memory transaction.
1921 .IR abort " (since Linux 3.11)"
1922 The branch was in an aborted transactional memory transaction.
1925 The entries are from most to least recent, so the first entry
1926 has the most recent branch.
1932 is optional; if not supported, both
1935 The type of branches recorded is specified by the
1936 .I branch_sample_type
1941 .IR abi ", " regs[weight(mask)]
1943 .B PERF_SAMPLE_REGS_USER
1944 is enabled, then the user CPU registers are recorded.
1949 .BR PERF_SAMPLE_REGS_ABI_NONE ", " PERF_SAMPLE_REGS_ABI_32 " or "
1950 .BR PERF_SAMPLE_REGS_ABI_64 .
1954 field is an array of the CPU registers that were specified by
1958 The number of values is the number of bits set in the
1962 .IR size ", " data[size] ", " dyn_size
1964 .B PERF_SAMPLE_STACK_USER
1965 is enabled, then the user stack is recorded.
1966 This can be used to generate stack backtraces.
1968 is the size requested by the user in
1969 .I sample_stack_user
1970 or else the maximum record size.
1972 is the stack data (a raw dump of the memory pointed to by the
1973 stack pointer at the time of sampling).
1975 is the amount of data actually dumped (can be less than
1980 .B PERF_SAMPLE_WEIGHT
1981 is enabled, then a 64-bit value provided by the hardware
1982 is recorded that indicates how costly the event was.
1983 This allows expensive events to stand out more clearly
1988 .B PERF_SAMPLE_DATA_SRC
1989 is enabled, then a 64-bit value is recorded that is made up of
1990 the following fields:
1994 Type of opcode, a bitwise combination of:
2005 .B PERF_MEM_OP_STORE
2008 .B PERF_MEM_OP_PFETCH
2017 Memory hierarchy level hit or miss, a bitwise combination of
2018 the following, shifted left by
2019 .BR PERF_MEM_LVL_SHIFT :
2030 .B PERF_MEM_LVL_MISS
2045 .B PERF_MEM_LVL_LOC_RAM
2048 .B PERF_MEM_LVL_REM_RAM1
2051 .B PERF_MEM_LVL_REM_RAM2
2054 .B PERF_MEM_LVL_REM_CCE1
2057 .B PERF_MEM_LVL_REM_CCE2
2069 Snoop mode, a bitwise combination of the following, shifted left by
2070 .BR PERF_MEM_SNOOP_SHIFT :
2075 .B PERF_MEM_SNOOP_NA
2078 .B PERF_MEM_SNOOP_NONE
2081 .B PERF_MEM_SNOOP_HIT
2084 .B PERF_MEM_SNOOP_MISS
2087 .B PERF_MEM_SNOOP_HITM
2093 Lock instruction, a bitwise combination of the following, shifted left by
2094 .BR PERF_MEM_LOCK_SHIFT :
2102 .B PERF_MEM_LOCK_LOCKED
2108 TLB access hit or miss, a bitwise combination of the following, shifted
2110 .BR PERF_MEM_TLB_SHIFT :
2121 .B PERF_MEM_TLB_MISS
2141 .B PERF_SAMPLE_TRANSACTION
2142 flag is set, then a 64-bit field is recorded describing
2143 the sources of any transactional memory aborts.
2145 The field is a bitwise combination of the following values:
2149 Abort from an elision type transaction (Intel-CPU-specific).
2151 .B PERF_TXN_TRANSACTION
2152 Abort from a generic transaction.
2155 Synchronous abort (related to the reported instruction).
2158 Asynchronous abort (not related to the reported instruction).
2161 Retryable abort (retrying the transaction may have succeeded).
2163 .B PERF_TXN_CONFLICT
2164 Abort due to memory conflicts with other threads.
2166 .B PERF_TXN_CAPACITY_WRITE
2167 Abort due to write capacity overflow.
2169 .B PERF_TXN_CAPACITY_READ
2170 Abort due to read capacity overflow.
2173 In addition, a user-specified abort code can be obtained from
2174 the high 32 bits of the field by shifting right by
2175 .B PERF_TXN_ABORT_SHIFT
2177 .BR PERF_TXN_ABORT_MASK .
2180 .B PERF_RECORD_MMAP2
2181 This record includes extended information on
2183 calls returning executable mappings.
2184 The format is similar to that of the
2186 record, but includes extra values that allow uniquely identifying
2192 struct perf_event_header header;
2205 struct sample_id sample_id;
2217 is the address of the allocated memory.
2220 is the length of the allocated memory.
2223 is the page offset of the allocated memory.
2226 is the major ID of the underlying device.
2229 is the minor ID of the underlying device.
2232 is the inode number.
2235 is the inode generation.
2238 is the protection information.
2241 is the flags information.
2244 is a string describing the backing of the allocated memory.
2248 Events can be set to deliver a signal when a threshold is crossed.
2250 .\" The following sentence doesn't seem to make sense.
2251 .\" These system calls do not set up signal handlers.
2252 The signal handler is set up using the
2260 To generate signals, sampling must be enabled
2262 must have a nonzero value).
2264 There are two ways to generate signals.
2266 The first is to set a
2270 value that will generate a signal if a certain number of samples
2271 or bytes have been written to the mmap ring buffer.
2272 In this case, a signal of type
2276 The other way is by use of the
2277 .B PERF_EVENT_IOC_REFRESH
2279 This ioctl adds to a counter that decrements each time the event overflows.
2282 signal is sent on overflow, but
2283 once the value reaches 0, a signal is sent of type
2286 the underlying event is disabled.
2288 Note: on newer kernels (since at least as early as Linux 3.2),
2289 .\" FIXME . Find out when this was introduced
2290 a signal is provided for every overflow, even if
2293 .SS rdpmc instruction
2294 Starting with Linux 3.4 on x86, you can use the
2296 instruction to get low-latency reads without having to enter the kernel.
2299 is not necessarily faster than other methods for reading event values.
2301 Support for this can be detected with the
2303 field in the mmap page; documentation on how
2304 to calculate event values can be found in that section.
2305 .SS perf_event ioctl calls
2307 Various ioctls act on
2308 .BR perf_event_open ()
2311 .B PERF_EVENT_IOC_ENABLE
2312 This enables the individual event or event group specified by the
2313 file descriptor argument.
2316 .B PERF_IOC_FLAG_GROUP
2317 bit is set in the ioctl argument, then all events in a group are
2318 enabled, even if the event specified is not the group leader
2321 .B PERF_EVENT_IOC_DISABLE
2322 This disables the individual counter or event group specified by the
2323 file descriptor argument.
2325 Enabling or disabling the leader of a group enables or disables the
2326 entire group; that is, while the group leader is disabled, none of the
2327 counters in the group will count.
2328 Enabling or disabling a member of a group other than the leader
2329 affects only that counter; disabling a non-leader
2330 stops that counter from counting but doesn't affect any other counter.
2333 .B PERF_IOC_FLAG_GROUP
2334 bit is set in the ioctl argument, then all events in a group are
2335 disabled, even if the event specified is not the group leader
2338 .B PERF_EVENT_IOC_REFRESH
2339 Non-inherited overflow counters can use this
2340 to enable a counter for a number of overflows specified by the argument,
2341 after which it is disabled.
2342 Subsequent calls of this ioctl add the argument value to the current
2346 set will happen on each overflow until the
2347 count reaches 0; when that happens a signal with
2349 set is sent and the event is disabled.
2350 Using an argument of 0 is considered undefined behavior.
2352 .B PERF_EVENT_IOC_RESET
2353 Reset the event count specified by the
2354 file descriptor argument to zero.
2355 This resets only the counts; there is no way to reset the
2363 .B PERF_IOC_FLAG_GROUP
2364 bit is set in the ioctl argument, then all events in a group are
2365 reset, even if the event specified is not the group leader
2368 .B PERF_EVENT_IOC_PERIOD
2369 This updates the overflow period for the event.
2371 Since Linux 3.7 (on ARM) and Linux 3.14 (all other architectures),
2372 the new period takes effect immediately.
2373 On older kernels, the new period did not take effect until
2374 after the next overflow.
2376 The argument is a pointer to a 64-bit value containing the
2379 Prior to Linux 2.6.36 this ioctl always failed due to a bug
2383 .B PERF_EVENT_IOC_SET_OUTPUT
2384 This tells the kernel to report event notifications to the specified
2385 file descriptor rather than the default one.
2386 The file descriptors must all be on the same CPU.
2388 The argument specifies the desired file descriptor, or \-1 if
2389 output should be ignored.
2391 .BR PERF_EVENT_IOC_SET_FILTER " (since Linux 2.6.33)"
2392 This adds an ftrace filter to this event.
2394 The argument is a pointer to the desired ftrace filter.
2396 .BR PERF_EVENT_IOC_ID " (since Linux 3.12)"
2397 This returns the event ID value for the given event file descriptor.
2399 The argument is a pointer to a 64-bit unsigned integer
2402 A process can enable or disable all the event groups that are
2403 attached to it using the
2405 .B PR_TASK_PERF_EVENTS_ENABLE
2407 .B PR_TASK_PERF_EVENTS_DISABLE
2409 This applies to all counters on the calling process, whether created by
2410 this process or by another, and does not affect any counters that this
2411 process has created on other processes.
2412 It enables or disables only
2413 the group leaders, not any other members in the groups.
2414 .SS perf_event related configuration files
2416 .I /proc/sys/kernel/
2419 .I /proc/sys/kernel/perf_event_paranoid
2422 .I perf_event_paranoid
2423 file can be set to restrict access to the performance counters.
2426 only allow user-space measurements.
2428 allow both kernel and user measurements (default).
2430 allow access to CPU-specific data but not raw tracepoint samples.
2435 The existence of the
2436 .I perf_event_paranoid
2437 file is the official method for determining if a kernel supports
2438 .BR perf_event_open ().
2440 .I /proc/sys/kernel/perf_event_max_sample_rate
2442 This sets the maximum sample rate.
2443 Setting this too high can allow
2444 users to sample at a rate that impacts overall machine performance
2445 and potentially lock up the machine.
2446 The default value is
2447 100000 (samples per second).
2449 .I /proc/sys/kernel/perf_event_mlock_kb
2451 Maximum number of pages an unprivileged user can
2453 The default is 516 (kB).
2457 .I /sys/bus/event_source/devices/
2459 Since Linux 2.6.34, the kernel supports having multiple PMUs
2460 available for monitoring.
2461 Information on how to program these PMUs can be found under
2462 .IR /sys/bus/event_source/devices/ .
2463 Each subdirectory corresponds to a different PMU.
2465 .IR /sys/bus/event_source/devices/*/type " (since Linux 2.6.38)"
2466 This contains an integer that can be used in the
2470 to indicate that you wish to use this PMU.
2472 .IR /sys/bus/event_source/devices/*/rdpmc " (since Linux 3.4)"
2473 If this file is 1, then direct user-space access to the
2474 performance counter registers is allowed via the rdpmc instruction.
2475 This can be disabled by echoing 0 to the file.
2477 .IR /sys/bus/event_source/devices/*/format/ " (since Linux 3.4)"
2478 This subdirectory contains information on the architecture-specific
2479 subfields available for programming the various
2485 The content of each file is the name of the config field, followed
2486 by a colon, followed by a series of integer bit ranges separated by
2488 For example, the file
2490 may contain the value
2491 .I config1:1,6-10,44
2492 which indicates that event is an attribute that occupies bits 1,6-10, and 44
2494 .IR perf_event_attr::config1 .
2496 .IR /sys/bus/event_source/devices/*/events/ " (since Linux 3.4)"
2497 This subdirectory contains files with predefined events.
2498 The contents are strings describing the event settings
2499 expressed in terms of the fields found in the previously mentioned
2502 These are not necessarily complete lists of all events supported by
2503 a PMU, but usually a subset of events deemed useful or interesting.
2505 The content of each file is a list of attribute names
2506 separated by commas.
2507 Each entry has an optional value (either hex or decimal).
2508 If no value is specified, then it is assumed to be a single-bit
2509 field with a value of 1.
2510 An example entry may look like this:
2511 .IR event=0x2,inv,ldlat=3 .
2513 .I /sys/bus/event_source/devices/*/uevent
2514 This file is the standard kernel device interface
2515 for injecting hotplug events.
2517 .IR /sys/bus/event_source/devices/*/cpumask " (since Linux 3.7)"
2520 file contains a comma-separated list of integers that
2521 indicate a representative CPU number for each socket (package)
2523 This is needed when setting up uncore or northbridge events, as
2524 those PMUs present socket-wide events.
2527 .BR perf_event_open ()
2528 returns the new file descriptor, or \-1 if an error occurred
2531 is set appropriately).
2533 The errors returned by
2534 .BR perf_event_open ()
2535 can be inconsistent, and may
2536 vary across processor architectures and performance monitoring units.
2544 .BR PERF_ATTR_SIZE_VER0 ),
2545 too big (larger than the page size),
2546 or larger than the kernel supports and the extra bytes are not zero.
2552 field is overwritten by the kernel to be the size of the structure
2556 Returned when the requested event requires
2558 permissions (or a more permissive perf_event paranoid setting).
2559 Some common cases where an unprivileged process
2560 may encounter this error:
2561 attaching to a process owned by a different user;
2562 monitoring all processes on a given CPU (i.e., specifying the
2567 when the paranoid setting requires it.
2572 file descriptor is not valid, or, if
2573 .B PERF_FLAG_PID_CGROUP
2575 the cgroup file descriptor in
2582 pointer points at an invalid memory address.
2585 Returned if the specified event is invalid.
2586 There are many possible reasons for this.
2587 A not-exhaustive list:
2589 is higher than the maximum setting;
2592 to monitor does not exist;
2599 value is out of range;
2603 set and the event is not a group leader;
2606 values are out of range or set reserved bits;
2607 the generic event selected is not supported; or
2608 there is not enough room to add the selected event.
2611 Each opened event uses one file descriptor.
2612 If a large number of events are opened the per-user file
2613 descriptor limit (often 1024) will be hit and no more
2614 events can be created.
2617 Returned when the event involves a feature not supported
2623 setting is not valid.
2624 This error is also returned for
2625 some unsupported generic events.
2628 Prior to Linux 3.3, if there was not enough room for the event,
2631 In Linux 3.3, this was changed to
2634 is still returned if you try to add more breakpoint events
2635 than supported by the hardware.
2639 .B PERF_SAMPLE_STACK_USER
2642 and it is not supported by hardware.
2645 Returned if an event requiring a specific hardware feature is
2646 requested but there is no hardware support.
2647 This includes requesting low-skid events if not supported,
2648 branch tracing if it is not available, sampling if no PMU
2649 interrupt is available, and branch stacks for software events.
2652 Returned on many (but not all) architectures when an unsupported
2653 .IR exclude_hv ", " exclude_idle ", " exclude_user ", or " exclude_kernel
2654 setting is specified.
2656 It can also happen, as with
2658 when the requested event requires
2660 permissions (or a more permissive perf_event paranoid setting).
2661 This includes setting a breakpoint on a kernel address,
2662 and (since Linux 3.13) setting a kernel function-trace tracepoint.
2665 Returned if attempting to attach to a process that does not exist.
2667 .BR perf_event_open ()
2668 was introduced in Linux 2.6.31 but was called
2669 .BR perf_counter_open ().
2670 It was renamed in Linux 2.6.32.
2673 .BR perf_event_open ()
2674 system call Linux- specific
2675 and should not be used in programs intended to be portable.
2677 Glibc does not provide a wrapper for this system call; call it using
2679 See the example below.
2681 The official way of knowing if
2682 .BR perf_event_open ()
2683 support is enabled is checking
2684 for the existence of the file
2685 .IR /proc/sys/kernel/perf_event_paranoid .
2691 is needed to properly get overflow signals in threads.
2692 This was introduced in Linux 2.6.32.
2694 Prior to Linux 2.6.33 (at least for x86), the kernel did not check
2695 if events could be scheduled together until read time.
2696 The same happens on all known kernels if the NMI watchdog is enabled.
2697 This means to see if a given set of events works you have to
2698 .BR perf_event_open (),
2699 start, then read before you know for sure you
2700 can get valid measurements.
2702 Prior to Linux 2.6.34, event constraints were not enforced by the kernel.
2703 In that case, some events would silently return "0" if the kernel
2704 scheduled them in an improper counter slot.
2706 Prior to Linux 2.6.34, there was a bug when multiplexing where the
2707 wrong results could be returned.
2709 Kernels from Linux 2.6.35 to Linux 2.6.39 can quickly crash the kernel if
2710 "inherit" is enabled and many threads are started.
2712 Prior to Linux 2.6.35,
2713 .B PERF_FORMAT_GROUP
2714 did not work with attached processes.
2716 In older Linux 2.6 versions,
2717 refreshing an event group leader refreshed all siblings,
2718 and refreshing with a parameter of 0 enabled infinite refresh.
2719 This behavior is unsupported and should not be relied on.
2721 There is a bug in the kernel code between
2722 Linux 2.6.36 and Linux 3.0 that ignores the
2723 "watermark" field and acts as if a wakeup_event
2724 was chosen if the union has a
2725 nonzero value in it.
2727 From Linux 2.6.31 to Linux 3.4, the
2728 .B PERF_IOC_FLAG_GROUP
2729 ioctl argument was broken and would repeatedly operate
2730 on the event specified rather than iterating across
2731 all sibling events in a group.
2733 From Linux 3.4 to Linux 3.11, the mmap
2737 bits mapped to the same location.
2738 Code should migrate to the new
2744 Always double-check your results!
2745 Various generalized events have had wrong values.
2746 For example, retired branches measured
2747 the wrong thing on AMD machines until Linux 2.6.35.
2749 The following is a short example that measures the total
2750 instruction count of a call to
2758 #include <sys/ioctl.h>
2759 #include <linux/perf_event.h>
2760 #include <asm/unistd.h>
2763 perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
2764 int cpu, int group_fd, unsigned long flags)
2768 ret = syscall(__NR_perf_event_open, hw_event, pid, cpu,
2774 main(int argc, char **argv)
2776 struct perf_event_attr pe;
2780 memset(&pe, 0, sizeof(struct perf_event_attr));
2781 pe.type = PERF_TYPE_HARDWARE;
2782 pe.size = sizeof(struct perf_event_attr);
2783 pe.config = PERF_COUNT_HW_INSTRUCTIONS;
2785 pe.exclude_kernel = 1;
2788 fd = perf_event_open(&pe, 0, \-1, \-1, 0);
2790 fprintf(stderr, "Error opening leader %llx\\n", pe.config);
2794 ioctl(fd, PERF_EVENT_IOC_RESET, 0);
2795 ioctl(fd, PERF_EVENT_IOC_ENABLE, 0);
2797 printf("Measuring instruction count for this printf\\n");
2799 ioctl(fd, PERF_EVENT_IOC_DISABLE, 0);
2800 read(fd, &count, sizeof(long long));
2802 printf("Used %lld instructions\\n", count);