1perf_event_open(2) System Calls Manual perf_event_open(2)
2
6 perf_event_open - set up performance monitoring
7
9 Standard C library (libc, -lc)
10
12 #include <linux/perf_event.h> /* Definition of PERF_* constants */
13 #include <linux/hw_breakpoint.h> /* Definition of HW_* constants */
14 #include <sys/syscall.h> /* Definition of SYS_* constants */
15 #include <unistd.h>
16 int syscall(SYS_perf_event_open, struct perf_event_attr *attr,
18 pid_t pid, int cpu, int group_fd, unsigned long flags);
19 Note: glibc provides no wrapper for perf_event_open(), necessitating
21 the use of syscall(2).
22
24 Given a list of parameters, perf_event_open() returns a file descrip‐
25 tor, for use in subsequent system calls (read(2), mmap(2), prctl(2),
26 fcntl(2), etc.).
27 A call to perf_event_open() creates a file descriptor that allows mea‐
29 suring performance information. Each file descriptor corresponds to
30 one event that is measured; these can be grouped together to measure
31 multiple events simultaneously.
32 Events can be enabled and disabled in two ways: via ioctl(2) and via
34 prctl(2). When an event is disabled it does not count or generate
35 overflows but does continue to exist and maintain its count value.
36 Events come in two flavors: counting and sampled. A counting event is
38 one that is used for counting the aggregate number of events that oc‐
39 cur. In general, counting event results are gathered with a read(2)
40 call. A sampling event periodically writes measurements to a buffer
41 that can then be accessed via mmap(2).
42 Arguments
44 The pid and cpu arguments allow specifying which process and CPU to
45 monitor:
46 pid == 0 and cpu == -1
48 This measures the calling process/thread on any CPU.
49 pid == 0 and cpu >= 0
51 This measures the calling process/thread only when running on
52 the specified CPU.
53 pid > 0 and cpu == -1
55 This measures the specified process/thread on any CPU.
56 pid > 0 and cpu >= 0
58 This measures the specified process/thread only when running on
59 the specified CPU.
60 pid == -1 and cpu >= 0
62 This measures all processes/threads on the specified CPU. This
63 requires CAP_PERFMON (since Linux 5.8) or CAP_SYS_ADMIN capabil‐
64 ity or a /proc/sys/kernel/perf_event_paranoid value of less than
65 1.
66 pid == -1 and cpu == -1
68 This setting is invalid and will return an error.
69 When pid is greater than zero, permission to perform this system call
71 is governed by CAP_PERFMON (since Linux 5.9) and a ptrace access mode
72 PTRACE_MODE_READ_REALCREDS check on older Linux versions; see
73 ptrace(2).
74 The group_fd argument allows event groups to be created. An event
76 group has one event which is the group leader. The leader is created
77 first, with group_fd = -1. The rest of the group members are created
78 with subsequent perf_event_open() calls with group_fd being set to the
79 file descriptor of the group leader. (A single event on its own is
80 created with group_fd = -1 and is considered to be a group with only 1
81 member.) An event group is scheduled onto the CPU as a unit: it will
82 be put onto the CPU only if all of the events in the group can be put
83 onto the CPU. This means that the values of the member events can be
84 meaningfully compared —added, divided (to get ratios), and so on— with
85 each other, since they have counted events for the same set of executed
86 instructions.
87 The flags argument is formed by ORing together zero or more of the fol‐
89 lowing values:
90 PERF_FLAG_FD_CLOEXEC (since Linux 3.14)
92 This flag enables the close-on-exec flag for the created event
93 file descriptor, so that the file descriptor is automatically
94 closed on execve(2). Setting the close-on-exec flags at cre‐
95 ation time, rather than later with fcntl(2), avoids potential
96 race conditions where the calling thread invokes
97 perf_event_open() and fcntl(2) at the same time as another
98 thread calls fork(2) then execve(2).
99 PERF_FLAG_FD_NO_GROUP
101 This flag tells the event to ignore the group_fd parameter ex‐
102 cept for the purpose of setting up output redirection using the
103 PERF_FLAG_FD_OUTPUT flag.
104 PERF_FLAG_FD_OUTPUT (broken since Linux 2.6.35)
106 This flag re-routes the event's sampled output to instead be in‐
107 cluded in the mmap buffer of the event specified by group_fd.
108 PERF_FLAG_PID_CGROUP (since Linux 2.6.39)
110 This flag activates per-container system-wide monitoring. A
111 container is an abstraction that isolates a set of resources for
112 finer-grained control (CPUs, memory, etc.). In this mode, the
113 event is measured only if the thread running on the monitored
114 CPU belongs to the designated container (cgroup). The cgroup is
115 identified by passing a file descriptor opened on its directory
116 in the cgroupfs filesystem. For instance, if the cgroup to mon‐
117 itor is called test, then a file descriptor opened on
118 /dev/cgroup/test (assuming cgroupfs is mounted on /dev/cgroup)
119 must be passed as the pid parameter. cgroup monitoring is
120 available only for system-wide events and may therefore require
121 extra permissions.
122 The perf_event_attr structure provides detailed configuration informa‐
124 tion for the event being created.
125 struct perf_event_attr {
127 __u32 type; /* Type of event */
128 __u32 size; /* Size of attribute structure */
129 __u64 config; /* Type-specific configuration */
130 union {
132 __u64 sample_period; /* Period of sampling */
133 __u64 sample_freq; /* Frequency of sampling */
134 };
135 __u64 sample_type; /* Specifies values included in sample */
137 __u64 read_format; /* Specifies values returned in read */
138 __u64 disabled : 1, /* off by default */
140 inherit : 1, /* children inherit it */
141 pinned : 1, /* must always be on PMU */
142 exclusive : 1, /* only group on PMU */
143 exclude_user : 1, /* don't count user */
144 exclude_kernel : 1, /* don't count kernel */
145 exclude_hv : 1, /* don't count hypervisor */
146 exclude_idle : 1, /* don't count when idle */
147 mmap : 1, /* include mmap data */
148 comm : 1, /* include comm data */
149 freq : 1, /* use freq, not period */
150 inherit_stat : 1, /* per task counts */
151 enable_on_exec : 1, /* next exec enables */
152 task : 1, /* trace fork/exit */
153 watermark : 1, /* wakeup_watermark */
154 precise_ip : 2, /* skid constraint */
155 mmap_data : 1, /* non-exec mmap data */
156 sample_id_all : 1, /* sample_type all events */
157 exclude_host : 1, /* don't count in host */
158 exclude_guest : 1, /* don't count in guest */
159 exclude_callchain_kernel : 1,
160 /* exclude kernel callchains */
161 exclude_callchain_user : 1,
162 /* exclude user callchains */
163 mmap2 : 1, /* include mmap with inode data */
164 comm_exec : 1, /* flag comm events that are
165 due to exec */
166 use_clockid : 1, /* use clockid for time fields */
167 context_switch : 1, /* context switch data */
168 write_backward : 1, /* Write ring buffer from end
169 to beginning */
170 namespaces : 1, /* include namespaces data */
171 ksymbol : 1, /* include ksymbol events */
172 bpf_event : 1, /* include bpf events */
173 aux_output : 1, /* generate AUX records
174 instead of events */
175 cgroup : 1, /* include cgroup events */
176 text_poke : 1, /* include text poke events */
177 build_id : 1, /* use build id in mmap2 events */
178 inherit_thread : 1, /* children only inherit */
179 /* if cloned with CLONE_THREAD */
180 remove_on_exec : 1, /* event is removed from task
181 on exec */
182 sigtrap : 1, /* send synchronous SIGTRAP
183 on event */
184 __reserved_1 : 26;
186 union {
188 __u32 wakeup_events; /* wakeup every n events */
189 __u32 wakeup_watermark; /* bytes before wakeup */
190 };
191 __u32 bp_type; /* breakpoint type */
193 union {
195 __u64 bp_addr; /* breakpoint address */
196 __u64 kprobe_func; /* for perf_kprobe */
197 __u64 uprobe_path; /* for perf_uprobe */
198 __u64 config1; /* extension of config */
199 };
200 union {
202 __u64 bp_len; /* breakpoint length */
203 __u64 kprobe_addr; /* with kprobe_func == NULL */
204 __u64 probe_offset; /* for perf_[k,u]probe */
205 __u64 config2; /* extension of config1 */
206 };
207 __u64 branch_sample_type; /* enum perf_branch_sample_type */
208 __u64 sample_regs_user; /* user regs to dump on samples */
209 __u32 sample_stack_user; /* size of stack to dump on
210 samples */
211 __s32 clockid; /* clock to use for time fields */
212 __u64 sample_regs_intr; /* regs to dump on samples */
213 __u32 aux_watermark; /* aux bytes before wakeup */
214 __u16 sample_max_stack; /* max frames in callchain */
215 __u16 __reserved_2; /* align to u64 */
216 __u32 aux_sample_size; /* max aux sample size */
217 __u32 __reserved_3; /* align to u64 */
218 __u64 sig_data; /* user data for sigtrap */
219 };
221 The fields of the perf_event_attr structure are described in more de‐
223 tail below:
224 type This field specifies the overall event type. It has one of the
226 following values:
227 PERF_TYPE_HARDWARE
229 This indicates one of the "generalized" hardware events
230 provided by the kernel. See the config field definition
231 for more details.
232 PERF_TYPE_SOFTWARE
234 This indicates one of the software-defined events pro‐
235 vided by the kernel (even if no hardware support is
236 available).
237 PERF_TYPE_TRACEPOINT
239 This indicates a tracepoint provided by the kernel trace‐
240 point infrastructure.
241 PERF_TYPE_HW_CACHE
243 This indicates a hardware cache event. This has a spe‐
244 cial encoding, described in the config field definition.
245 PERF_TYPE_RAW
247 This indicates a "raw" implementation-specific event in
248 the config field.
249 PERF_TYPE_BREAKPOINT (since Linux 2.6.33)
251 This indicates a hardware breakpoint as provided by the
252 CPU. Breakpoints can be read/write accesses to an ad‐
253 dress as well as execution of an instruction address.
254 dynamic PMU
256 Since Linux 2.6.38, perf_event_open() can support multi‐
257 ple PMUs. To enable this, a value exported by the kernel
258 can be used in the type field to indicate which PMU to
259 use. The value to use can be found in the sysfs filesys‐
260 tem: there is a subdirectory per PMU instance under
261 /sys/bus/event_source/devices. In each subdirectory
262 there is a type file whose content is an integer that can
263 be used in the type field. For instance,
264 /sys/bus/event_source/devices/cpu/type contains the value
265 for the core CPU PMU, which is usually 4.
266 kprobe and uprobe (since Linux 4.17)
268 These two dynamic PMUs create a kprobe/uprobe and attach
269 it to the file descriptor generated by perf_event_open.
270 The kprobe/uprobe will be destroyed on the destruction of
271 the file descriptor. See fields kprobe_func, up‐
272 robe_path, kprobe_addr, and probe_offset for more de‐
273 tails.
274 size The size of the perf_event_attr structure for forward/backward
276 compatibility. Set this using sizeof(struct perf_event_attr) to
277 allow the kernel to see the struct size at the time of compila‐
278 tion.
279 The related define PERF_ATTR_SIZE_VER0 is set to 64; this was
281 the size of the first published struct. PERF_ATTR_SIZE_VER1 is
282 72, corresponding to the addition of breakpoints in Linux
283 2.6.33. PERF_ATTR_SIZE_VER2 is 80 corresponding to the addition
284 of branch sampling in Linux 3.4. PERF_ATTR_SIZE_VER3 is 96 cor‐
285 responding to the addition of sample_regs_user and sam‐
286 ple_stack_user in Linux 3.7. PERF_ATTR_SIZE_VER4 is 104 corre‐
287 sponding to the addition of sample_regs_intr in Linux 3.19.
288 PERF_ATTR_SIZE_VER5 is 112 corresponding to the addition of
289 aux_watermark in Linux 4.1.
290 config This specifies which event you want, in conjunction with the
292 type field. The config1 and config2 fields are also taken into
293 account in cases where 64 bits is not enough to fully specify
294 the event. The encoding of these fields are event dependent.
295 There are various ways to set the config field that are depen‐
297 dent on the value of the previously described type field. What
298 follows are various possible settings for config separated out
299 by type.
300 If type is PERF_TYPE_HARDWARE, we are measuring one of the gen‐
302 eralized hardware CPU events. Not all of these are available on
303 all platforms. Set config to one of the following:
304 PERF_COUNT_HW_CPU_CYCLES
306 Total cycles. Be wary of what happens during CPU
307 frequency scaling.
308 PERF_COUNT_HW_INSTRUCTIONS
310 Retired instructions. Be careful, these can be af‐
311 fected by various issues, most notably hardware in‐
312 terrupt counts.
313 PERF_COUNT_HW_CACHE_REFERENCES
315 Cache accesses. Usually this indicates Last Level
316 Cache accesses but this may vary depending on your
317 CPU. This may include prefetches and coherency mes‐
318 sages; again this depends on the design of your CPU.
319 PERF_COUNT_HW_CACHE_MISSES
321 Cache misses. Usually this indicates Last Level
322 Cache misses; this is intended to be used in con‐
323 junction with the PERF_COUNT_HW_CACHE_REFERENCES
324 event to calculate cache miss rates.
325 PERF_COUNT_HW_BRANCH_INSTRUCTIONS
327 Retired branch instructions. Prior to Linux 2.6.35,
328 this used the wrong event on AMD processors.
329 PERF_COUNT_HW_BRANCH_MISSES
331 Mispredicted branch instructions.
332 PERF_COUNT_HW_BUS_CYCLES
334 Bus cycles, which can be different from total cy‐
335 cles.
336 PERF_COUNT_HW_STALLED_CYCLES_FRONTEND (since Linux 3.0)
338 Stalled cycles during issue.
339 PERF_COUNT_HW_STALLED_CYCLES_BACKEND (since Linux 3.0)
341 Stalled cycles during retirement.
342 PERF_COUNT_HW_REF_CPU_CYCLES (since Linux 3.3)
344 Total cycles; not affected by CPU frequency scaling.
345 If type is PERF_TYPE_SOFTWARE, we are measuring software events
347 provided by the kernel. Set config to one of the following:
348 PERF_COUNT_SW_CPU_CLOCK
350 This reports the CPU clock, a high-resolution per-
351 CPU timer.
352 PERF_COUNT_SW_TASK_CLOCK
354 This reports a clock count specific to the task that
355 is running.
356 PERF_COUNT_SW_PAGE_FAULTS
358 This reports the number of page faults.
359 PERF_COUNT_SW_CONTEXT_SWITCHES
361 This counts context switches. Until Linux 2.6.34,
362 these were all reported as user-space events, after
363 that they are reported as happening in the kernel.
364 PERF_COUNT_SW_CPU_MIGRATIONS
366 This reports the number of times the process has mi‐
367 grated to a new CPU.
368 PERF_COUNT_SW_PAGE_FAULTS_MIN
370 This counts the number of minor page faults. These
371 did not require disk I/O to handle.
372 PERF_COUNT_SW_PAGE_FAULTS_MAJ
374 This counts the number of major page faults. These
375 required disk I/O to handle.
376 PERF_COUNT_SW_ALIGNMENT_FAULTS (since Linux 2.6.33)
378 This counts the number of alignment faults. These
379 happen when unaligned memory accesses happen; the
380 kernel can handle these but it reduces performance.
381 This happens only on some architectures (never on
382 x86).
383 PERF_COUNT_SW_EMULATION_FAULTS (since Linux 2.6.33)
385 This counts the number of emulation faults. The
386 kernel sometimes traps on unimplemented instructions
387 and emulates them for user space. This can nega‐
388 tively impact performance.
389 PERF_COUNT_SW_DUMMY (since Linux 3.12)
391 This is a placeholder event that counts nothing.
392 Informational sample record types such as mmap or
393 comm must be associated with an active event. This
394 dummy event allows gathering such records without
395 requiring a counting event.
396 PERF_COUNT_SW_BPF_OUTPUT (since Linux 4.4)
398 This is used to generate raw sample data from BPF.
399 BPF programs can write to this event using
400 bpf_perf_event_output helper.
401 PERF_COUNT_SW_CGROUP_SWITCHES (since Linux 5.13)
403 This counts context switches to a task in a differ‐
404 ent cgroup. In other words, if the next task is in
405 the same cgroup, it won't count the switch.
406 If type is PERF_TYPE_TRACEPOINT, then we are measuring kernel
408 tracepoints. The value to use in config can be obtained from
409 under debugfs tracing/events/*/*/id if ftrace is enabled in the
410 kernel.
411 If type is PERF_TYPE_HW_CACHE, then we are measuring a hardware
413 CPU cache event. To calculate the appropriate config value, use
414 the following equation:
415 config = (perf_hw_cache_id) |
417 (perf_hw_cache_op_id << 8) |
418 (perf_hw_cache_op_result_id << 16);
419 where perf_hw_cache_id is one of:
421 PERF_COUNT_HW_CACHE_L1D
423 for measuring Level 1 Data Cache
424 PERF_COUNT_HW_CACHE_L1I
426 for measuring Level 1 Instruction Cache
427 PERF_COUNT_HW_CACHE_LL
429 for measuring Last-Level Cache
430 PERF_COUNT_HW_CACHE_DTLB
432 for measuring the Data TLB
433 PERF_COUNT_HW_CACHE_ITLB
435 for measuring the Instruction TLB
436 PERF_COUNT_HW_CACHE_BPU
438 for measuring the branch prediction unit
439 PERF_COUNT_HW_CACHE_NODE (since Linux 3.1)
441 for measuring local memory accesses
442 and perf_hw_cache_op_id is one of:
444 PERF_COUNT_HW_CACHE_OP_READ
446 for read accesses
447 PERF_COUNT_HW_CACHE_OP_WRITE
449 for write accesses
450 PERF_COUNT_HW_CACHE_OP_PREFETCH
452 for prefetch accesses
453 and perf_hw_cache_op_result_id is one of:
455 PERF_COUNT_HW_CACHE_RESULT_ACCESS
457 to measure accesses
458 PERF_COUNT_HW_CACHE_RESULT_MISS
460 to measure misses
461 If type is PERF_TYPE_RAW, then a custom "raw" config value is
463 needed. Most CPUs support events that are not covered by the
464 "generalized" events. These are implementation defined; see
465 your CPU manual (for example the Intel Volume 3B documentation
466 or the AMD BIOS and Kernel Developer Guide). The libpfm4 li‐
467 brary can be used to translate from the name in the architec‐
468 tural manuals to the raw hex value perf_event_open() expects in
469 this field.
470 If type is PERF_TYPE_BREAKPOINT, then leave config set to zero.
472 Its parameters are set in other places.
473 If type is kprobe or uprobe, set retprobe (bit 0 of config, see
475 /sys/bus/event_source/devices/[k,u]probe/format/retprobe) for
476 kretprobe/uretprobe. See fields kprobe_func, uprobe_path,
477 kprobe_addr, and probe_offset for more details.
478 kprobe_func, uprobe_path, kprobe_addr, and probe_offset
480 These fields describe the kprobe/uprobe for dynamic PMUs kprobe
481 and uprobe. For kprobe: use kprobe_func and probe_offset, or
482 use kprobe_addr and leave kprobe_func as NULL. For uprobe: use
483 uprobe_path and probe_offset.
484 sample_period, sample_freq
486 A "sampling" event is one that generates an overflow notifica‐
487 tion every N events, where N is given by sample_period. A sam‐
488 pling event has sample_period > 0. When an overflow occurs, re‐
489 quested data is recorded in the mmap buffer. The sample_type
490 field controls what data is recorded on each overflow.
491 sample_freq can be used if you wish to use frequency rather than
493 period. In this case, you set the freq flag. The kernel will
494 adjust the sampling period to try and achieve the desired rate.
495 The rate of adjustment is a timer tick.
496 sample_type
498 The various bits in this field specify which values to include
499 in the sample. They will be recorded in a ring-buffer, which is
500 available to user space using mmap(2). The order in which the
501 values are saved in the sample are documented in the MMAP Layout
502 subsection below; it is not the enum perf_event_sample_format
503 order.
504 PERF_SAMPLE_IP
506 Records instruction pointer.
507 PERF_SAMPLE_TID
509 Records the process and thread IDs.
510 PERF_SAMPLE_TIME
512 Records a timestamp.
513 PERF_SAMPLE_ADDR
515 Records an address, if applicable.
516 PERF_SAMPLE_READ
518 Record counter values for all events in a group, not just
519 the group leader.
520 PERF_SAMPLE_CALLCHAIN
522 Records the callchain (stack backtrace).
523 PERF_SAMPLE_ID
525 Records a unique ID for the opened event's group leader.
526 PERF_SAMPLE_CPU
528 Records CPU number.
529 PERF_SAMPLE_PERIOD
531 Records the current sampling period.
532 PERF_SAMPLE_STREAM_ID
534 Records a unique ID for the opened event. Unlike
535 PERF_SAMPLE_ID the actual ID is returned, not the group
536 leader. This ID is the same as the one returned by
537 PERF_FORMAT_ID.
538 PERF_SAMPLE_RAW
540 Records additional data, if applicable. Usually returned
541 by tracepoint events.
542 PERF_SAMPLE_BRANCH_STACK (since Linux 3.4)
544 This provides a record of recent branches, as provided by
545 CPU branch sampling hardware (such as Intel Last Branch
546 Record). Not all hardware supports this feature.
547 See the branch_sample_type field for how to filter which
549 branches are reported.
550 PERF_SAMPLE_REGS_USER (since Linux 3.7)
552 Records the current user-level CPU register state (the
553 values in the process before the kernel was called).
554 PERF_SAMPLE_STACK_USER (since Linux 3.7)
556 Records the user level stack, allowing stack unwinding.
557 PERF_SAMPLE_WEIGHT (since Linux 3.10)
559 Records a hardware provided weight value that expresses
560 how costly the sampled event was. This allows the hard‐
561 ware to highlight expensive events in a profile.
562 PERF_SAMPLE_DATA_SRC (since Linux 3.10)
564 Records the data source: where in the memory hierarchy
565 the data associated with the sampled instruction came
566 from. This is available only if the underlying hardware
567 supports this feature.
568 PERF_SAMPLE_IDENTIFIER (since Linux 3.12)
570 Places the SAMPLE_ID value in a fixed position in the
571 record, either at the beginning (for sample events) or at
572 the end (if a non-sample event).
573 This was necessary because a sample stream may have
575 records from various different event sources with differ‐
576 ent sample_type settings. Parsing the event stream prop‐
577 erly was not possible because the format of the record
578 was needed to find SAMPLE_ID, but the format could not be
579 found without knowing what event the sample belonged to
580 (causing a circular dependency).
581 The PERF_SAMPLE_IDENTIFIER setting makes the event stream
583 always parsable by putting SAMPLE_ID in a fixed location,
584 even though it means having duplicate SAMPLE_ID values in
585 records.
586 PERF_SAMPLE_TRANSACTION (since Linux 3.13)
588 Records reasons for transactional memory abort events
589 (for example, from Intel TSX transactional memory sup‐
590 port).
591 The precise_ip setting must be greater than 0 and a
593 transactional memory abort event must be measured or no
594 values will be recorded. Also note that some perf_event
595 measurements, such as sampled cycle counting, may cause
596 extraneous aborts (by causing an interrupt during a
597 transaction).
598 PERF_SAMPLE_REGS_INTR (since Linux 3.19)
600 Records a subset of the current CPU register state as
601 specified by sample_regs_intr. Unlike PERF_SAM‐
602 PLE_REGS_USER the register values will return kernel reg‐
603 ister state if the overflow happened while kernel code is
604 running. If the CPU supports hardware sampling of regis‐
605 ter state (i.e., PEBS on Intel x86) and precise_ip is set
606 higher than zero then the register values returned are
607 those captured by hardware at the time of the sampled in‐
608 struction's retirement.
609 PERF_SAMPLE_PHYS_ADDR (since Linux 4.13)
611 Records physical address of data like in PERF_SAM‐
612 PLE_ADDR.
613 PERF_SAMPLE_CGROUP (since Linux 5.7)
615 Records (perf_event) cgroup ID of the process. This cor‐
616 responds to the id field in the PERF_RECORD_CGROUP event.
617 PERF_SAMPLE_DATA_PAGE_SIZE (since Linux 5.11)
619 Records page size of data like in PERF_SAMPLE_ADDR.
620 PERF_SAMPLE_CODE_PAGE_SIZE (since Linux 5.11)
622 Records page size of ip like in PERF_SAMPLE_IP.
623 PERF_SAMPLE_WEIGHT_STRUCT (since Linux 5.12)
625 Records hardware provided weight values like in PERF_SAM‐
626 PLE_WEIGHT, but it can represent multiple values in a
627 struct. This shares the same space as PERF_SAM‐
628 PLE_WEIGHT, so users can apply either of those, not both.
629 It has the following format and the meaning of each field
630 is dependent on the hardware implementation.
631 union perf_sample_weight {
633 u64 full; /* PERF_SAMPLE_WEIGHT */
634 struct { /* PERF_SAMPLE_WEIGHT_STRUCT */
635 u32 var1_dw;
636 u16 var2_w;
637 u16 var3_w;
638 };
639 };
640 read_format
642 This field specifies the format of the data returned by read(2)
643 on a perf_event_open() file descriptor.
644 PERF_FORMAT_TOTAL_TIME_ENABLED
646 Adds the 64-bit time_enabled field. This can be used to
647 calculate estimated totals if the PMU is overcommitted
648 and multiplexing is happening.
649 PERF_FORMAT_TOTAL_TIME_RUNNING
651 Adds the 64-bit time_running field. This can be used to
652 calculate estimated totals if the PMU is overcommitted
653 and multiplexing is happening.
654 PERF_FORMAT_ID
656 Adds a 64-bit unique value that corresponds to the event
657 group.
658 PERF_FORMAT_GROUP
660 Allows all counter values in an event group to be read
661 with one read.
662 PERF_FORMAT_LOST (since Linux 6.0)
664 Adds a 64-bit value that is the number of lost samples
665 for this event. This would be only meaningful when sam‐
666 ple_period or sample_freq is set.
667 disabled
669 The disabled bit specifies whether the counter starts out dis‐
670 abled or enabled. If disabled, the event can later be enabled
671 by ioctl(2), prctl(2), or enable_on_exec.
672 When creating an event group, typically the group leader is ini‐
674 tialized with disabled set to 1 and any child events are ini‐
675 tialized with disabled set to 0. Despite disabled being 0, the
676 child events will not start until the group leader is enabled.
677 inherit
679 The inherit bit specifies that this counter should count events
680 of child tasks as well as the task specified. This applies only
681 to new children, not to any existing children at the time the
682 counter is created (nor to any new children of existing chil‐
683 dren).
684 Inherit does not work for some combinations of read_format val‐
686 ues, such as PERF_FORMAT_GROUP.
687 pinned The pinned bit specifies that the counter should always be on
689 the CPU if at all possible. It applies only to hardware coun‐
690 ters and only to group leaders. If a pinned counter cannot be
691 put onto the CPU (e.g., because there are not enough hardware
692 counters or because of a conflict with some other event), then
693 the counter goes into an 'error' state, where reads return end-
694 of-file (i.e., read(2) returns 0) until the counter is subse‐
695 quently enabled or disabled.
696 exclusive
698 The exclusive bit specifies that when this counter's group is on
699 the CPU, it should be the only group using the CPU's counters.
700 In the future this may allow monitoring programs to support PMU
701 features that need to run alone so that they do not disrupt
702 other hardware counters.
703 Note that many unexpected situations may prevent events with the
705 exclusive bit set from ever running. This includes any users
706 running a system-wide measurement as well as any kernel use of
707 the performance counters (including the commonly enabled NMI
708 Watchdog Timer interface).
709 exclude_user
711 If this bit is set, the count excludes events that happen in
712 user space.
713 exclude_kernel
715 If this bit is set, the count excludes events that happen in
716 kernel space.
717 exclude_hv
719 If this bit is set, the count excludes events that happen in the
720 hypervisor. This is mainly for PMUs that have built-in support
721 for handling this (such as POWER). Extra support is needed for
722 handling hypervisor measurements on most machines.
723 exclude_idle
725 If set, don't count when the CPU is running the idle task.
726 While you can currently enable this for any event type, it is
727 ignored for all but software events.
728 mmap The mmap bit enables generation of PERF_RECORD_MMAP samples for
730 every mmap(2) call that has PROT_EXEC set. This allows tools to
731 notice new executable code being mapped into a program (dynamic
732 shared libraries for example) so that addresses can be mapped
733 back to the original code.
734 comm The comm bit enables tracking of process command name as modi‐
736 fied by the execve(2) and prctl(PR_SET_NAME) system calls as
737 well as writing to /proc/self/comm. If the comm_exec flag is
738 also successfully set (possible since Linux 3.16), then the misc
739 flag PERF_RECORD_MISC_COMM_EXEC can be used to differentiate the
740 execve(2) case from the others.
741 freq If this bit is set, then sample_frequency not sample_period is
743 used when setting up the sampling interval.
744 inherit_stat
746 This bit enables saving of event counts on context switch for
747 inherited tasks. This is meaningful only if the inherit field
748 is set.
749 enable_on_exec
751 If this bit is set, a counter is automatically enabled after a
752 call to execve(2).
753 task If this bit is set, then fork/exit notifications are included in
755 the ring buffer.
756 watermark
758 If set, have an overflow notification happen when we cross the
759 wakeup_watermark boundary. Otherwise, overflow notifications
760 happen after wakeup_events samples.
761 precise_ip (since Linux 2.6.35)
763 This controls the amount of skid. Skid is how many instructions
764 execute between an event of interest happening and the kernel
765 being able to stop and record the event. Smaller skid is better
766 and allows more accurate reporting of which events correspond to
767 which instructions, but hardware is often limited with how small
768 this can be.
769 The possible values of this field are the following:
771 0 SAMPLE_IP can have arbitrary skid.
773 1 SAMPLE_IP must have constant skid.
775 2 SAMPLE_IP requested to have 0 skid.
777 3 SAMPLE_IP must have 0 skid. See also the description of
779 PERF_RECORD_MISC_EXACT_IP.
780 mmap_data (since Linux 2.6.36)
782 This is the counterpart of the mmap field. This enables genera‐
783 tion of PERF_RECORD_MMAP samples for mmap(2) calls that do not
784 have PROT_EXEC set (for example data and SysV shared memory).
785 sample_id_all (since Linux 2.6.38)
787 If set, then TID, TIME, ID, STREAM_ID, and CPU can additionally
788 be included in non-PERF_RECORD_SAMPLEs if the corresponding sam‐
789 ple_type is selected.
790 If PERF_SAMPLE_IDENTIFIER is specified, then an additional ID
792 value is included as the last value to ease parsing the record
793 stream. This may lead to the id value appearing twice.
794 The layout is described by this pseudo-structure:
796 struct sample_id {
798 { u32 pid, tid; } /* if PERF_SAMPLE_TID set */
799 { u64 time; } /* if PERF_SAMPLE_TIME set */
800 { u64 id; } /* if PERF_SAMPLE_ID set */
801 { u64 stream_id;} /* if PERF_SAMPLE_STREAM_ID set */
802 { u32 cpu, res; } /* if PERF_SAMPLE_CPU set */
803 { u64 id; } /* if PERF_SAMPLE_IDENTIFIER set */
804 };
805 exclude_host (since Linux 3.2)
807 When conducting measurements that include processes running VM
808 instances (i.e., have executed a KVM_RUN ioctl(2)), only measure
809 events happening inside a guest instance. This is only meaning‐
810 ful outside the guests; this setting does not change counts
811 gathered inside of a guest. Currently, this functionality is
812 x86 only.
813 exclude_guest (since Linux 3.2)
815 When conducting measurements that include processes running VM
816 instances (i.e., have executed a KVM_RUN ioctl(2)), do not mea‐
817 sure events happening inside guest instances. This is only
818 meaningful outside the guests; this setting does not change
819 counts gathered inside of a guest. Currently, this functional‐
820 ity is x86 only.
821 exclude_callchain_kernel (since Linux 3.7)
823 Do not include kernel callchains.
824 exclude_callchain_user (since Linux 3.7)
826 Do not include user callchains.
827 mmap2 (since Linux 3.16)
829 Generate an extended executable mmap record that contains enough
830 additional information to uniquely identify shared mappings.
831 The mmap flag must also be set for this to work.
832 comm_exec (since Linux 3.16)
834 This is purely a feature-detection flag, it does not change ker‐
835 nel behavior. If this flag can successfully be set, then, when
836 comm is enabled, the PERF_RECORD_MISC_COMM_EXEC flag will be set
837 in the misc field of a comm record header if the rename event
838 being reported was caused by a call to execve(2). This allows
839 tools to distinguish between the various types of process renam‐
840 ing.
841 use_clockid (since Linux 4.1)
843 This allows selecting which internal Linux clock to use when
844 generating timestamps via the clockid field. This can make it
845 easier to correlate perf sample times with timestamps generated
846 by other tools.
847 context_switch (since Linux 4.3)
849 This enables the generation of PERF_RECORD_SWITCH records when a
850 context switch occurs. It also enables the generation of
851 PERF_RECORD_SWITCH_CPU_WIDE records when sampling in CPU-wide
852 mode. This functionality is in addition to existing tracepoint
853 and software events for measuring context switches. The advan‐
854 tage of this method is that it will give full information even
855 with strict perf_event_paranoid settings.
856 write_backward (since Linux 4.6)
858 This causes the ring buffer to be written from the end to the
859 beginning. This is to support reading from overwritable ring
860 buffer.
861 namespaces (since Linux 4.11)
863 This enables the generation of PERF_RECORD_NAMESPACES records
864 when a task enters a new namespace. Each namespace has a combi‐
865 nation of device and inode numbers.
866 ksymbol (since Linux 5.0)
868 This enables the generation of PERF_RECORD_KSYMBOL records when
869 new kernel symbols are registered or unregistered. This is ana‐
870 lyzing dynamic kernel functions like eBPF.
871 bpf_event (since Linux 5.0)
873 This enables the generation of PERF_RECORD_BPF_EVENT records
874 when an eBPF program is loaded or unloaded.
875 aux_output (since Linux 5.4)
877 This allows normal (non-AUX) events to generate data for AUX
878 events if the hardware supports it.
879 cgroup (since Linux 5.7)
881 This enables the generation of PERF_RECORD_CGROUP records when a
882 new cgroup is created (and activated).
883 text_poke (since Linux 5.8)
885 This enables the generation of PERF_RECORD_TEXT_POKE records
886 when there's a change to the kernel text (i.e., self-modifying
887 code).
888 build_id (since Linux 5.12)
890 This changes the contents in the PERF_RECORD_MMAP2 to have a
891 build-id instead of device and inode numbers.
892 inherit_thread (since Linux 5.13)
894 This disables the inheritance of the event to a child process.
895 Only new threads in the same process (which is cloned with
896 CLONE_THREAD) will inherit the event.
897 remove_on_exec (since Linux 5.13)
899 This closes the event when it starts a new process image by ex‐
900 ecve(2).
901 sigtrap (since Linux 5.13)
903 This enables synchronous signal delivery of SIGTRAP on event
904 overflow.
905 wakeup_events, wakeup_watermark
907 This union sets how many samples (wakeup_events) or bytes
908 (wakeup_watermark) happen before an overflow notification hap‐
909 pens. Which one is used is selected by the watermark bit flag.
910 wakeup_events counts only PERF_RECORD_SAMPLE record types. To
912 receive overflow notification for all PERF_RECORD types choose
913 watermark and set wakeup_watermark to 1.
914 Prior to Linux 3.0, setting wakeup_events to 0 resulted in no
916 overflow notifications; more recent kernels treat 0 the same as
917 1.
918 bp_type (since Linux 2.6.33)
920 This chooses the breakpoint type. It is one of:
921 HW_BREAKPOINT_EMPTY
923 No breakpoint.
924 HW_BREAKPOINT_R
926 Count when we read the memory location.
927 HW_BREAKPOINT_W
929 Count when we write the memory location.
930 HW_BREAKPOINT_RW
932 Count when we read or write the memory location.
933 HW_BREAKPOINT_X
935 Count when we execute code at the memory location.
936 The values can be combined via a bitwise or, but the combination
938 of HW_BREAKPOINT_R or HW_BREAKPOINT_W with HW_BREAKPOINT_X is
939 not allowed.
940 bp_addr (since Linux 2.6.33)
942 This is the address of the breakpoint. For execution break‐
943 points, this is the memory address of the instruction of inter‐
944 est; for read and write breakpoints, it is the memory address of
945 the memory location of interest.
946 config1 (since Linux 2.6.39)
948 config1 is used for setting events that need an extra register
949 or otherwise do not fit in the regular config field. Raw OFF‐
950 CORE_EVENTS on Nehalem/Westmere/SandyBridge use this field on
951 Linux 3.3 and later kernels.
952 bp_len (since Linux 2.6.33)
954 bp_len is the length of the breakpoint being measured if type is
955 PERF_TYPE_BREAKPOINT. Options are HW_BREAKPOINT_LEN_1,
956 HW_BREAKPOINT_LEN_2, HW_BREAKPOINT_LEN_4, and HW_BREAK‐
957 POINT_LEN_8. For an execution breakpoint, set this to
958 sizeof(long).
959 config2 (since Linux 2.6.39)
961 config2 is a further extension of the config1 field.
962 branch_sample_type (since Linux 3.4)
964 If PERF_SAMPLE_BRANCH_STACK is enabled, then this specifies what
965 branches to include in the branch record.
966 The first part of the value is the privilege level, which is a
968 combination of one of the values listed below. If the user does
969 not set privilege level explicitly, the kernel will use the
970 event's privilege level. Event and branch privilege levels do
971 not have to match.
972 PERF_SAMPLE_BRANCH_USER
974 Branch target is in user space.
975 PERF_SAMPLE_BRANCH_KERNEL
977 Branch target is in kernel space.
978 PERF_SAMPLE_BRANCH_HV
980 Branch target is in hypervisor.
981 PERF_SAMPLE_BRANCH_PLM_ALL
983 A convenience value that is the three preceding values
984 ORed together.
985 In addition to the privilege value, at least one or more of the
987 following bits must be set.
988 PERF_SAMPLE_BRANCH_ANY
990 Any branch type.
991 PERF_SAMPLE_BRANCH_ANY_CALL
993 Any call branch (includes direct calls, indirect calls,
994 and far jumps).
995 PERF_SAMPLE_BRANCH_IND_CALL
997 Indirect calls.
998 PERF_SAMPLE_BRANCH_CALL (since Linux 4.4)
1000 Direct calls.
1001 PERF_SAMPLE_BRANCH_ANY_RETURN
1003 Any return branch.
1004 PERF_SAMPLE_BRANCH_IND_JUMP (since Linux 4.2)
1006 Indirect jumps.
1007 PERF_SAMPLE_BRANCH_COND (since Linux 3.16)
1009 Conditional branches.
1010 PERF_SAMPLE_BRANCH_ABORT_TX (since Linux 3.11)
1012 Transactional memory aborts.
1013 PERF_SAMPLE_BRANCH_IN_TX (since Linux 3.11)
1015 Branch in transactional memory transaction.
1016 PERF_SAMPLE_BRANCH_NO_TX (since Linux 3.11)
1018 Branch not in transactional memory transaction.
1019 PERF_SAMPLE_BRANCH_CALL_STACK (since Linux 4.1) Branch is
1020 part of a hardware-generated call stack. This requires
1021 hardware support, currently only found on Intel x86
1022 Haswell or newer.
1023 sample_regs_user (since Linux 3.7)
1025 This bit mask defines the set of user CPU registers to dump on
1026 samples. The layout of the register mask is architecture-spe‐
1027 cific and is described in the kernel header file arch/ARCH/in‐
1028 clude/uapi/asm/perf_regs.h.
1029 sample_stack_user (since Linux 3.7)
1031 This defines the size of the user stack to dump if PERF_SAM‐
1032 PLE_STACK_USER is specified.
1033 clockid (since Linux 4.1)
1035 If use_clockid is set, then this field selects which internal
1036 Linux timer to use for timestamps. The available timers are de‐
1037 fined in linux/time.h, with CLOCK_MONOTONIC, CLOCK_MONO‐
1038 TONIC_RAW, CLOCK_REALTIME, CLOCK_BOOTTIME, and CLOCK_TAI cur‐
1039 rently supported.
1040 aux_watermark (since Linux 4.1)
1042 This specifies how much data is required to trigger a
1043 PERF_RECORD_AUX sample.
1044 sample_max_stack (since Linux 4.8)
1046 When sample_type includes PERF_SAMPLE_CALLCHAIN, this field
1047 specifies how many stack frames to report when generating the
1048 callchain.
1049 aux_sample_size (since Linux 5.5)
1051 When PERF_SAMPLE_AUX flag is set, specify the desired size of
1052 AUX data. Note that it can get smaller data than the specified
1053 size.
1054 sig_data (since Linux 5.13)
1056 This data will be copied to user's signal handler (through
1057 si_perf in the siginfo_t) to disambiguate which event triggered
1058 the signal.
1059 Reading results
1061 Once a perf_event_open() file descriptor has been opened, the values of
1062 the events can be read from the file descriptor. The values that are
1063 there are specified by the read_format field in the attr structure at
1064 open time.
1065 If you attempt to read into a buffer that is not big enough to hold the
1067 data, the error ENOSPC results.
1068 Here is the layout of the data returned by a read:
1070 • If PERF_FORMAT_GROUP was specified to allow reading all events in a
1072 group at once:
1073 struct read_format {
1075 u64 nr; /* The number of events */
1076 u64 time_enabled; /* if PERF_FORMAT_TOTAL_TIME_ENABLED */
1077 u64 time_running; /* if PERF_FORMAT_TOTAL_TIME_RUNNING */
1078 struct {
1079 u64 value; /* The value of the event */
1080 u64 id; /* if PERF_FORMAT_ID */
1081 u64 lost; /* if PERF_FORMAT_LOST */
1082 } values[nr];
1083 };
1084 • If PERF_FORMAT_GROUP was not specified:
1086 struct read_format {
1088 u64 value; /* The value of the event */
1089 u64 time_enabled; /* if PERF_FORMAT_TOTAL_TIME_ENABLED */
1090 u64 time_running; /* if PERF_FORMAT_TOTAL_TIME_RUNNING */
1091 u64 id; /* if PERF_FORMAT_ID */
1092 u64 lost; /* if PERF_FORMAT_LOST */
1093 };
1094 The values read are as follows:
1096 nr The number of events in this file descriptor. Available only if
1098 PERF_FORMAT_GROUP was specified.
1099 time_enabled, time_running
1101 Total time the event was enabled and running. Normally these
1102 values are the same. Multiplexing happens if the number of
1103 events is more than the number of available PMU counter slots.
1104 In that case the events run only part of the time and the
1105 time_enabled and time running values can be used to scale an es‐
1106 timated value for the count.
1107 value An unsigned 64-bit value containing the counter result.
1109 id A globally unique value for this particular event; only present
1111 if PERF_FORMAT_ID was specified in read_format.
1112 lost The number of lost samples of this event; only present if
1114 PERF_FORMAT_LOST was specified in read_format.
1115 MMAP layout
1117 When using perf_event_open() in sampled mode, asynchronous events (like
1118 counter overflow or PROT_EXEC mmap tracking) are logged into a ring-
1119 buffer. This ring-buffer is created and accessed through mmap(2).
1120 The mmap size should be 1+2^n pages, where the first page is a metadata
1122 page (struct perf_event_mmap_page) that contains various bits of infor‐
1123 mation such as where the ring-buffer head is.
1124 Before Linux 2.6.39, there is a bug that means you must allocate an
1126 mmap ring buffer when sampling even if you do not plan to access it.
1127 The structure of the first metadata mmap page is as follows:
1129 struct perf_event_mmap_page {
1131 __u32 version; /* version number of this structure */
1132 __u32 compat_version; /* lowest version this is compat with */
1133 __u32 lock; /* seqlock for synchronization */
1134 __u32 index; /* hardware counter identifier */
1135 __s64 offset; /* add to hardware counter value */
1136 __u64 time_enabled; /* time event active */
1137 __u64 time_running; /* time event on CPU */
1138 union {
1139 __u64 capabilities;
1140 struct {
1141 __u64 cap_usr_time / cap_usr_rdpmc / cap_bit0 : 1,
1142 cap_bit0_is_deprecated : 1,
1143 cap_user_rdpmc : 1,
1144 cap_user_time : 1,
1145 cap_user_time_zero : 1,
1146 };
1147 };
1148 __u16 pmc_width;
1149 __u16 time_shift;
1150 __u32 time_mult;
1151 __u64 time_offset;
1152 __u64 __reserved[120]; /* Pad to 1 k */
1153 __u64 data_head; /* head in the data section */
1154 __u64 data_tail; /* user-space written tail */
1155 __u64 data_offset; /* where the buffer starts */
1156 __u64 data_size; /* data buffer size */
1157 __u64 aux_head;
1158 __u64 aux_tail;
1159 __u64 aux_offset;
1160 __u64 aux_size;
1161 }
1163 The following list describes the fields in the perf_event_mmap_page
1165 structure in more detail:
1166 version
1168 Version number of this structure.
1169 compat_version
1171 The lowest version this is compatible with.
1172 lock A seqlock for synchronization.
1174 index A unique hardware counter identifier.
1176 offset When using rdpmc for reads this offset value must be added to
1178 the one returned by rdpmc to get the current total event count.
1179 time_enabled
1181 Time the event was active.
1182 time_running
1184 Time the event was running.
1185 cap_usr_time / cap_usr_rdpmc / cap_bit0 (since Linux 3.4)
1187 There was a bug in the definition of cap_usr_time and
1188 cap_usr_rdpmc from Linux 3.4 until Linux 3.11. Both bits were
1189 defined to point to the same location, so it was impossible to
1190 know if cap_usr_time or cap_usr_rdpmc were actually set.
1191 Starting with Linux 3.12, these are renamed to cap_bit0 and you
1193 should use the cap_user_time and cap_user_rdpmc fields instead.
1194 cap_bit0_is_deprecated (since Linux 3.12)
1196 If set, this bit indicates that the kernel supports the properly
1197 separated cap_user_time and cap_user_rdpmc bits.
1198 If not-set, it indicates an older kernel where cap_usr_time and
1200 cap_usr_rdpmc map to the same bit and thus both features should
1201 be used with caution.
1202 cap_user_rdpmc (since Linux 3.12)
1204 If the hardware supports user-space read of performance counters
1205 without syscall (this is the "rdpmc" instruction on x86), then
1206 the following code can be used to do a read:
1207 u32 seq, time_mult, time_shift, idx, width;
1209 u64 count, enabled, running;
1210 u64 cyc, time_offset;
1211 do {
1213 seq = pc->lock;
1214 barrier();
1215 enabled = pc->time_enabled;
1216 running = pc->time_running;
1217 if (pc->cap_usr_time && enabled != running) {
1219 cyc = rdtsc();
1220 time_offset = pc->time_offset;
1221 time_mult = pc->time_mult;
1222 time_shift = pc->time_shift;
1223 }
1224 idx = pc->index;
1226 count = pc->offset;
1227 if (pc->cap_usr_rdpmc && idx) {
1229 width = pc->pmc_width;
1230 count += rdpmc(idx - 1);
1231 }
1232 barrier();
1234 } while (pc->lock != seq);
1235 cap_user_time (since Linux 3.12)
1237 This bit indicates the hardware has a constant, nonstop time‐
1238 stamp counter (TSC on x86).
1239 cap_user_time_zero (since Linux 3.12)
1241 Indicates the presence of time_zero which allows mapping time‐
1242 stamp values to the hardware clock.
1243 pmc_width
1245 If cap_usr_rdpmc, this field provides the bit-width of the value
1246 read using the rdpmc or equivalent instruction. This can be
1247 used to sign extend the result like:
1248 pmc <<= 64 - pmc_width;
1250 pmc >>= 64 - pmc_width; // signed shift right
1251 count += pmc;
1252 time_shift, time_mult, time_offset
1254 If cap_usr_time, these fields can be used to compute the time
1256 delta since time_enabled (in nanoseconds) using rdtsc or simi‐
1257 lar.
1258 u64 quot, rem;
1260 u64 delta;
1261 quot = cyc >> time_shift;
1263 rem = cyc & (((u64)1 << time_shift) - 1);
1264 delta = time_offset + quot * time_mult +
1265 ((rem * time_mult) >> time_shift);
1266 Where time_offset, time_mult, time_shift, and cyc are read in
1268 the seqcount loop described above. This delta can then be added
1269 to enabled and possible running (if idx), improving the scaling:
1270 enabled += delta;
1272 if (idx)
1273 running += delta;
1274 quot = count / running;
1275 rem = count % running;
1276 count = quot * enabled + (rem * enabled) / running;
1277 time_zero (since Linux 3.12)
1279 If cap_usr_time_zero is set, then the hardware clock (the TSC
1281 timestamp counter on x86) can be calculated from the time_zero,
1282 time_mult, and time_shift values:
1283 time = timestamp - time_zero;
1285 quot = time / time_mult;
1286 rem = time % time_mult;
1287 cyc = (quot << time_shift) + (rem << time_shift) / time_mult;
1288 And vice versa:
1290 quot = cyc >> time_shift;
1292 rem = cyc & (((u64)1 << time_shift) - 1);
1293 timestamp = time_zero + quot * time_mult +
1294 ((rem * time_mult) >> time_shift);
1295 data_head
1297 This points to the head of the data section. The value continu‐
1298 ously increases, it does not wrap. The value needs to be manu‐
1299 ally wrapped by the size of the mmap buffer before accessing the
1300 samples.
1301 On SMP-capable platforms, after reading the data_head value,
1303 user space should issue an rmb().
1304 data_tail
1306 When the mapping is PROT_WRITE, the data_tail value should be
1307 written by user space to reflect the last read data. In this
1308 case, the kernel will not overwrite unread data.
1309 data_offset (since Linux 4.1)
1311 Contains the offset of the location in the mmap buffer where
1312 perf sample data begins.
1313 data_size (since Linux 4.1)
1315 Contains the size of the perf sample region within the mmap buf‐
1316 fer.
1317 aux_head, aux_tail, aux_offset, aux_size (since Linux 4.1)
1319 The AUX region allows mmap(2)-ing a separate sample buffer for
1320 high-bandwidth data streams (separate from the main perf sample
1321 buffer). An example of a high-bandwidth stream is instruction
1322 tracing support, as is found in newer Intel processors.
1323 To set up an AUX area, first aux_offset needs to be set with an
1325 offset greater than data_offset+data_size and aux_size needs to
1326 be set to the desired buffer size. The desired offset and size
1327 must be page aligned, and the size must be a power of two.
1328 These values are then passed to mmap in order to map the AUX
1329 buffer. Pages in the AUX buffer are included as part of the
1330 RLIMIT_MEMLOCK resource limit (see setrlimit(2)), and also as
1331 part of the perf_event_mlock_kb allowance.
1332 By default, the AUX buffer will be truncated if it will not fit
1334 in the available space in the ring buffer. If the AUX buffer is
1335 mapped as a read only buffer, then it will operate in ring buf‐
1336 fer mode where old data will be overwritten by new. In over‐
1337 write mode, it might not be possible to infer where the new data
1338 began, and it is the consumer's job to disable measurement while
1339 reading to avoid possible data races.
1340 The aux_head and aux_tail ring buffer pointers have the same be‐
1342 havior and ordering rules as the previous described data_head
1343 and data_tail.
1344 The following 2^n ring-buffer pages have the layout described below.
1346 If perf_event_attr.sample_id_all is set, then all event types will have
1348 the sample_type selected fields related to where/when (identity) an
1349 event took place (TID, TIME, ID, CPU, STREAM_ID) described in
1350 PERF_RECORD_SAMPLE below, it will be stashed just after the
1351 perf_event_header and the fields already present for the existing
1352 fields, that is, at the end of the payload. This allows a newer
1353 perf.data file to be supported by older perf tools, with the new op‐
1354 tional fields being ignored.
1355 The mmap values start with a header:
1357 struct perf_event_header {
1359 __u32 type;
1360 __u16 misc;
1361 __u16 size;
1362 };
1363 Below, we describe the perf_event_header fields in more detail. For
1365 ease of reading, the fields with shorter descriptions are presented
1366 first.
1367 size This indicates the size of the record.
1369 misc The misc field contains additional information about the sample.
1371 The CPU mode can be determined from this value by masking with
1373 PERF_RECORD_MISC_CPUMODE_MASK and looking for one of the follow‐
1374 ing (note these are not bit masks, only one can be set at a
1375 time):
1376 PERF_RECORD_MISC_CPUMODE_UNKNOWN
1378 Unknown CPU mode.
1379 PERF_RECORD_MISC_KERNEL
1381 Sample happened in the kernel.
1382 PERF_RECORD_MISC_USER
1384 Sample happened in user code.
1385 PERF_RECORD_MISC_HYPERVISOR
1387 Sample happened in the hypervisor.
1388 PERF_RECORD_MISC_GUEST_KERNEL (since Linux 2.6.35)
1390 Sample happened in the guest kernel.
1391 PERF_RECORD_MISC_GUEST_USER (since Linux 2.6.35)
1393 Sample happened in guest user code.
1394 Since the following three statuses are generated by different
1396 record types, they alias to the same bit:
1397 PERF_RECORD_MISC_MMAP_DATA (since Linux 3.10)
1399 This is set when the mapping is not executable; otherwise
1400 the mapping is executable.
1401 PERF_RECORD_MISC_COMM_EXEC (since Linux 3.16)
1403 This is set for a PERF_RECORD_COMM record on kernels more
1404 recent than Linux 3.16 if a process name change was
1405 caused by an execve(2) system call.
1406 PERF_RECORD_MISC_SWITCH_OUT (since Linux 4.3)
1408 When a PERF_RECORD_SWITCH or PERF_RECORD_SWITCH_CPU_WIDE
1409 record is generated, this bit indicates that the context
1410 switch is away from the current process (instead of into
1411 the current process).
1412 In addition, the following bits can be set:
1414 PERF_RECORD_MISC_EXACT_IP
1416 This indicates that the content of PERF_SAMPLE_IP points
1417 to the actual instruction that triggered the event. See
1418 also perf_event_attr.precise_ip.
1419 PERF_RECORD_MISC_SWITCH_OUT_PREEMPT (since Linux 4.17)
1421 When a PERF_RECORD_SWITCH or PERF_RECORD_SWITCH_CPU_WIDE
1422 record is generated, this indicates the context switch
1423 was a preemption.
1424 PERF_RECORD_MISC_MMAP_BUILD_ID (since Linux 5.12)
1426 This indicates that the content of PERF_SAMPLE_MMAP2 con‐
1427 tains build-ID data instead of device major and minor
1428 numbers as well as the inode number.
1429 PERF_RECORD_MISC_EXT_RESERVED (since Linux 2.6.35)
1431 This indicates there is extended data available (cur‐
1432 rently not used).
1433 PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT
1435 This bit is not set by the kernel. It is reserved for
1436 the user-space perf utility to indicate that
1437 /proc/pid/maps parsing was taking too long and was
1438 stopped, and thus the mmap records may be truncated.
1439 type The type value is one of the below. The values in the corre‐
1441 sponding record (that follows the header) depend on the type se‐
1442 lected as shown.
1443 PERF_RECORD_MMAP
1445 The MMAP events record the PROT_EXEC mappings so that we can
1446 correlate user-space IPs to code. They have the following
1447 structure:
1448 struct {
1450 struct perf_event_header header;
1451 u32 pid, tid;
1452 u64 addr;
1453 u64 len;
1454 u64 pgoff;
1455 char filename[];
1456 };
1457 pid is the process ID.
1459 tid is the thread ID.
1461 addr is the address of the allocated memory. len is the
1463 length of the allocated memory. pgoff is the page
1464 offset of the allocated memory. filename is a string
1465 describing the backing of the allocated memory.
1466 PERF_RECORD_LOST
1468 This record indicates when events are lost.
1469 struct {
1471 struct perf_event_header header;
1472 u64 id;
1473 u64 lost;
1474 struct sample_id sample_id;
1475 };
1476 id is the unique event ID for the samples that were
1478 lost.
1479 lost is the number of events that were lost.
1481 PERF_RECORD_COMM
1483 This record indicates a change in the process name.
1484 struct {
1486 struct perf_event_header header;
1487 u32 pid;
1488 u32 tid;
1489 char comm[];
1490 struct sample_id sample_id;
1491 };
1492 pid is the process ID.
1494 tid is the thread ID.
1496 comm is a string containing the new name of the process.
1498 PERF_RECORD_EXIT
1500 This record indicates a process exit event.
1501 struct {
1503 struct perf_event_header header;
1504 u32 pid, ppid;
1505 u32 tid, ptid;
1506 u64 time;
1507 struct sample_id sample_id;
1508 };
1509 PERF_RECORD_THROTTLE, PERF_RECORD_UNTHROTTLE
1511 This record indicates a throttle/unthrottle event.
1512 struct {
1514 struct perf_event_header header;
1515 u64 time;
1516 u64 id;
1517 u64 stream_id;
1518 struct sample_id sample_id;
1519 };
1520 PERF_RECORD_FORK
1522 This record indicates a fork event.
1523 struct {
1525 struct perf_event_header header;
1526 u32 pid, ppid;
1527 u32 tid, ptid;
1528 u64 time;
1529 struct sample_id sample_id;
1530 };
1531 PERF_RECORD_READ
1533 This record indicates a read event.
1534 struct {
1536 struct perf_event_header header;
1537 u32 pid, tid;
1538 struct read_format values;
1539 struct sample_id sample_id;
1540 };
1541 PERF_RECORD_SAMPLE
1543 This record indicates a sample.
1544 struct {
1546 struct perf_event_header header;
1547 u64 sample_id; /* if PERF_SAMPLE_IDENTIFIER */
1548 u64 ip; /* if PERF_SAMPLE_IP */
1549 u32 pid, tid; /* if PERF_SAMPLE_TID */
1550 u64 time; /* if PERF_SAMPLE_TIME */
1551 u64 addr; /* if PERF_SAMPLE_ADDR */
1552 u64 id; /* if PERF_SAMPLE_ID */
1553 u64 stream_id; /* if PERF_SAMPLE_STREAM_ID */
1554 u32 cpu, res; /* if PERF_SAMPLE_CPU */
1555 u64 period; /* if PERF_SAMPLE_PERIOD */
1556 struct read_format v;
1557 /* if PERF_SAMPLE_READ */
1558 u64 nr; /* if PERF_SAMPLE_CALLCHAIN */
1559 u64 ips[nr]; /* if PERF_SAMPLE_CALLCHAIN */
1560 u32 size; /* if PERF_SAMPLE_RAW */
1561 char data[size]; /* if PERF_SAMPLE_RAW */
1562 u64 bnr; /* if PERF_SAMPLE_BRANCH_STACK */
1563 struct perf_branch_entry lbr[bnr];
1564 /* if PERF_SAMPLE_BRANCH_STACK */
1565 u64 abi; /* if PERF_SAMPLE_REGS_USER */
1566 u64 regs[weight(mask)];
1567 /* if PERF_SAMPLE_REGS_USER */
1568 u64 size; /* if PERF_SAMPLE_STACK_USER */
1569 char data[size]; /* if PERF_SAMPLE_STACK_USER */
1570 u64 dyn_size; /* if PERF_SAMPLE_STACK_USER &&
1571 size != 0 */
1572 union perf_sample_weight weight;
1573 /* if PERF_SAMPLE_WEIGHT */
1574 /* || PERF_SAMPLE_WEIGHT_STRUCT */
1575 u64 data_src; /* if PERF_SAMPLE_DATA_SRC */
1576 u64 transaction; /* if PERF_SAMPLE_TRANSACTION */
1577 u64 abi; /* if PERF_SAMPLE_REGS_INTR */
1578 u64 regs[weight(mask)];
1579 /* if PERF_SAMPLE_REGS_INTR */
1580 u64 phys_addr; /* if PERF_SAMPLE_PHYS_ADDR */
1581 u64 cgroup; /* if PERF_SAMPLE_CGROUP */
1582 u64 data_page_size;
1583 /* if PERF_SAMPLE_DATA_PAGE_SIZE */
1584 u64 code_page_size;
1585 /* if PERF_SAMPLE_CODE_PAGE_SIZE */
1586 u64 size; /* if PERF_SAMPLE_AUX */
1587 char data[size]; /* if PERF_SAMPLE_AUX */
1588 };
1589 sample_id
1591 If PERF_SAMPLE_IDENTIFIER is enabled, a 64-bit unique ID
1592 is included. This is a duplication of the PERF_SAM‐
1593 PLE_ID id value, but included at the beginning of the
1594 sample so parsers can easily obtain the value.
1595 ip If PERF_SAMPLE_IP is enabled, then a 64-bit instruction
1597 pointer value is included.
1598 pid, tid
1600 If PERF_SAMPLE_TID is enabled, then a 32-bit process ID
1601 and 32-bit thread ID are included.
1602 time
1604 If PERF_SAMPLE_TIME is enabled, then a 64-bit timestamp
1605 is included. This is obtained via local_clock() which
1606 is a hardware timestamp if available and the jiffies
1607 value if not.
1608 addr
1610 If PERF_SAMPLE_ADDR is enabled, then a 64-bit address is
1611 included. This is usually the address of a tracepoint,
1612 breakpoint, or software event; otherwise the value is 0.
1613 id If PERF_SAMPLE_ID is enabled, a 64-bit unique ID is in‐
1615 cluded. If the event is a member of an event group, the
1616 group leader ID is returned. This ID is the same as the
1617 one returned by PERF_FORMAT_ID.
1618 stream_id
1620 If PERF_SAMPLE_STREAM_ID is enabled, a 64-bit unique ID
1621 is included. Unlike PERF_SAMPLE_ID the actual ID is re‐
1622 turned, not the group leader. This ID is the same as
1623 the one returned by PERF_FORMAT_ID.
1624 cpu, res
1626 If PERF_SAMPLE_CPU is enabled, this is a 32-bit value
1627 indicating which CPU was being used, in addition to a
1628 reserved (unused) 32-bit value.
1629 period
1631 If PERF_SAMPLE_PERIOD is enabled, a 64-bit value indi‐
1632 cating the current sampling period is written.
1633 v If PERF_SAMPLE_READ is enabled, a structure of type
1635 read_format is included which has values for all events
1636 in the event group. The values included depend on the
1637 read_format value used at perf_event_open() time.
1638 nr, ips[nr]
1640 If PERF_SAMPLE_CALLCHAIN is enabled, then a 64-bit num‐
1641 ber is included which indicates how many following
1642 64-bit instruction pointers will follow. This is the
1643 current callchain.
1644 size, data[size]
1646 If PERF_SAMPLE_RAW is enabled, then a 32-bit value indi‐
1647 cating size is included followed by an array of 8-bit
1648 values of length size. The values are padded with 0 to
1649 have 64-bit alignment.
1650 This RAW record data is opaque with respect to the ABI.
1652 The ABI doesn't make any promises with respect to the
1653 stability of its content, it may vary depending on
1654 event, hardware, and kernel version.
1655 bnr, lbr[bnr]
1657 If PERF_SAMPLE_BRANCH_STACK is enabled, then a 64-bit
1658 value indicating the number of records is included, fol‐
1659 lowed by bnr perf_branch_entry structures which each in‐
1660 clude the fields:
1661 from This indicates the source instruction (may not be
1663 a branch).
1664 to The branch target.
1666 mispred
1668 The branch target was mispredicted.
1669 predicted
1671 The branch target was predicted.
1672 in_tx (since Linux 3.11)
1674 The branch was in a transactional memory transac‐
1675 tion.
1676 abort (since Linux 3.11)
1678 The branch was in an aborted transactional memory
1679 transaction.
1680 cycles (since Linux 4.3)
1682 This reports the number of cycles elapsed since
1683 the previous branch stack update.
1684 The entries are from most to least recent, so the first
1686 entry has the most recent branch.
1687 Support for mispred, predicted, and cycles is optional;
1689 if not supported, those values will be 0.
1690 The type of branches recorded is specified by the
1692 branch_sample_type field.
1693 abi, regs[weight(mask)]
1695 If PERF_SAMPLE_REGS_USER is enabled, then the user CPU
1696 registers are recorded.
1697 The abi field is one of PERF_SAMPLE_REGS_ABI_NONE,
1699 PERF_SAMPLE_REGS_ABI_32, or PERF_SAMPLE_REGS_ABI_64.
1700 The regs field is an array of the CPU registers that
1702 were specified by the sample_regs_user attr field. The
1703 number of values is the number of bits set in the sam‐
1704 ple_regs_user bit mask.
1705 size, data[size], dyn_size
1707 If PERF_SAMPLE_STACK_USER is enabled, then the user
1708 stack is recorded. This can be used to generate stack
1709 backtraces. size is the size requested by the user in
1710 sample_stack_user or else the maximum record size. data
1711 is the stack data (a raw dump of the memory pointed to
1712 by the stack pointer at the time of sampling). dyn_size
1713 is the amount of data actually dumped (can be less than
1714 size). Note that dyn_size is omitted if size is 0.
1715 weight
1717 If PERF_SAMPLE_WEIGHT or PERF_SAMPLE_WEIGHT_STRUCT is
1718 enabled, then a 64-bit value provided by the hardware is
1719 recorded that indicates how costly the event was. This
1720 allows expensive events to stand out more clearly in
1721 profiles.
1722 data_src
1724 If PERF_SAMPLE_DATA_SRC is enabled, then a 64-bit value
1725 is recorded that is made up of the following fields:
1726 mem_op
1728 Type of opcode, a bitwise combination of:
1729 PERF_MEM_OP_NA Not available
1731 PERF_MEM_OP_LOAD Load instruction
1732 PERF_MEM_OP_STORE Store instruction
1733 PERF_MEM_OP_PFETCH Prefetch
1734 PERF_MEM_OP_EXEC Executable code
1735 mem_lvl
1737 Memory hierarchy level hit or miss, a bitwise combi‐
1738 nation of the following, shifted left by
1739 PERF_MEM_LVL_SHIFT:
1740 PERF_MEM_LVL_NA Not available
1742 PERF_MEM_LVL_HIT Hit
1743 PERF_MEM_LVL_MISS Miss
1744 PERF_MEM_LVL_L1 Level 1 cache
1745 PERF_MEM_LVL_LFB Line fill buffer
1746 PERF_MEM_LVL_L2 Level 2 cache
1747 PERF_MEM_LVL_L3 Level 3 cache
1748 PERF_MEM_LVL_LOC_RAM Local DRAM
1749 PERF_MEM_LVL_REM_RAM1 Remote DRAM 1 hop
1750 PERF_MEM_LVL_REM_RAM2 Remote DRAM 2 hops
1751 PERF_MEM_LVL_REM_CCE1 Remote cache 1 hop
1752 PERF_MEM_LVL_REM_CCE2 Remote cache 2 hops
1753 PERF_MEM_LVL_IO I/O memory
1754 PERF_MEM_LVL_UNC Uncached memory
1755 mem_snoop
1757 Snoop mode, a bitwise combination of the following,
1758 shifted left by PERF_MEM_SNOOP_SHIFT:
1759 PERF_MEM_SNOOP_NA Not available
1761 PERF_MEM_SNOOP_NONE No snoop
1762 PERF_MEM_SNOOP_HIT Snoop hit
1763 PERF_MEM_SNOOP_MISS Snoop miss
1764 PERF_MEM_SNOOP_HITM Snoop hit modified
1765 mem_lock
1767 Lock instruction, a bitwise combination of the fol‐
1768 lowing, shifted left by PERF_MEM_LOCK_SHIFT:
1769 PERF_MEM_LOCK_NA Not available
1771 PERF_MEM_LOCK_LOCKED Locked transaction
1772 mem_dtlb
1774 TLB access hit or miss, a bitwise combination of the
1775 following, shifted left by PERF_MEM_TLB_SHIFT:
1776 PERF_MEM_TLB_NA Not available
1778 PERF_MEM_TLB_HIT Hit
1779 PERF_MEM_TLB_MISS Miss
1780 PERF_MEM_TLB_L1 Level 1 TLB
1781 PERF_MEM_TLB_L2 Level 2 TLB
1782 PERF_MEM_TLB_WK Hardware walker
1783 PERF_MEM_TLB_OS OS fault handler
1784 transaction
1786 If the PERF_SAMPLE_TRANSACTION flag is set, then a
1787 64-bit field is recorded describing the sources of any
1788 transactional memory aborts.
1789 The field is a bitwise combination of the following val‐
1791 ues:
1792 PERF_TXN_ELISION
1794 Abort from an elision type transaction (Intel-
1795 CPU-specific).
1796 PERF_TXN_TRANSACTION
1798 Abort from a generic transaction.
1799 PERF_TXN_SYNC
1801 Synchronous abort (related to the reported in‐
1802 struction).
1803 PERF_TXN_ASYNC
1805 Asynchronous abort (not related to the reported
1806 instruction).
1807 PERF_TXN_RETRY
1809 Retryable abort (retrying the transaction may
1810 have succeeded).
1811 PERF_TXN_CONFLICT
1813 Abort due to memory conflicts with other threads.
1814 PERF_TXN_CAPACITY_WRITE
1816 Abort due to write capacity overflow.
1817 PERF_TXN_CAPACITY_READ
1819 Abort due to read capacity overflow.
1820 In addition, a user-specified abort code can be obtained
1822 from the high 32 bits of the field by shifting right by
1823 PERF_TXN_ABORT_SHIFT and masking with the value
1824 PERF_TXN_ABORT_MASK.
1825 abi, regs[weight(mask)]
1827 If PERF_SAMPLE_REGS_INTR is enabled, then the user CPU
1828 registers are recorded.
1829 The abi field is one of PERF_SAMPLE_REGS_ABI_NONE,
1831 PERF_SAMPLE_REGS_ABI_32, or PERF_SAMPLE_REGS_ABI_64.
1832 The regs field is an array of the CPU registers that
1834 were specified by the sample_regs_intr attr field. The
1835 number of values is the number of bits set in the sam‐
1836 ple_regs_intr bit mask.
1837 phys_addr
1839 If the PERF_SAMPLE_PHYS_ADDR flag is set, then the
1840 64-bit physical address is recorded.
1841 cgroup
1843 If the PERF_SAMPLE_CGROUP flag is set, then the 64-bit
1844 cgroup ID (for the perf_event subsystem) is recorded.
1845 To get the pathname of the cgroup, the ID should match
1846 to one in a PERF_RECORD_CGROUP.
1847 data_page_size
1849 If the PERF_SAMPLE_DATA_PAGE_SIZE flag is set, then the
1850 64-bit page size value of the data address is recorded.
1851 code_page_size
1853 If the PERF_SAMPLE_CODE_PAGE_SIZE flag is set, then the
1854 64-bit page size value of the ip address is recorded.
1855 size
1857 data[size]
1858 If PERF_SAMPLE_AUX is enabled, a snapshot of the aux
1859 buffer is recorded.
1860 PERF_RECORD_MMAP2
1862 This record includes extended information on mmap(2) calls
1863 returning executable mappings. The format is similar to
1864 that of the PERF_RECORD_MMAP record, but includes extra val‐
1865 ues that allow uniquely identifying shared mappings. De‐
1866 pending on the PERF_RECORD_MISC_MMAP_BUILD_ID bit in the
1867 header, the extra values have different layout and meanings.
1868 struct {
1870 struct perf_event_header header;
1871 u32 pid;
1872 u32 tid;
1873 u64 addr;
1874 u64 len;
1875 u64 pgoff;
1876 union {
1877 struct {
1878 u32 maj;
1879 u32 min;
1880 u64 ino;
1881 u64 ino_generation;
1882 };
1883 struct { /* if PERF_RECORD_MISC_MMAP_BUILD_ID */
1884 u8 build_id_size;
1885 u8 __reserved_1;
1886 u16 __reserved_2;
1887 u8 build_id[20];
1888 };
1889 };
1890 u32 prot;
1891 u32 flags;
1892 char filename[];
1893 struct sample_id sample_id;
1894 };
1895 pid is the process ID.
1897 tid is the thread ID.
1899 addr is the address of the allocated memory.
1901 len is the length of the allocated memory.
1903 pgoff is the page offset of the allocated memory.
1905 maj is the major ID of the underlying device.
1907 min is the minor ID of the underlying device.
1909 ino is the inode number.
1911 ino_generation
1913 is the inode generation.
1914 build_id_size
1916 is the actual size of build_id field (up to 20).
1917 build_id
1919 is a raw data to identify a binary.
1920 prot is the protection information.
1922 flags is the flags information.
1924 filename
1926 is a string describing the backing of the allocated
1927 memory.
1928 PERF_RECORD_AUX (since Linux 4.1)
1930 This record reports that new data is available in the sepa‐
1931 rate AUX buffer region.
1932 struct {
1934 struct perf_event_header header;
1935 u64 aux_offset;
1936 u64 aux_size;
1937 u64 flags;
1938 struct sample_id sample_id;
1939 };
1940 aux_offset
1942 offset in the AUX mmap region where the new data be‐
1943 gins.
1944 aux_size
1946 size of the data made available.
1947 flags describes the AUX update.
1949 PERF_AUX_FLAG_TRUNCATED
1951 if set, then the data returned was truncated
1952 to fit the available buffer size.
1953 PERF_AUX_FLAG_OVERWRITE
1955 if set, then the data returned has overwritten
1956 previous data.
1957 PERF_RECORD_ITRACE_START (since Linux 4.1)
1959 This record indicates which process has initiated an in‐
1960 struction trace event, allowing tools to properly correlate
1961 the instruction addresses in the AUX buffer with the proper
1962 executable.
1963 struct {
1965 struct perf_event_header header;
1966 u32 pid;
1967 u32 tid;
1968 };
1969 pid process ID of the thread starting an instruction
1971 trace.
1972 tid thread ID of the thread starting an instruction
1974 trace.
1975 PERF_RECORD_LOST_SAMPLES (since Linux 4.2)
1977 When using hardware sampling (such as Intel PEBS) this
1978 record indicates some number of samples that may have been
1979 lost.
1980 struct {
1982 struct perf_event_header header;
1983 u64 lost;
1984 struct sample_id sample_id;
1985 };
1986 lost the number of potentially lost samples.
1988 PERF_RECORD_SWITCH (since Linux 4.3)
1990 This record indicates a context switch has happened. The
1991 PERF_RECORD_MISC_SWITCH_OUT bit in the misc field indicates
1992 whether it was a context switch into or away from the cur‐
1993 rent process.
1994 struct {
1996 struct perf_event_header header;
1997 struct sample_id sample_id;
1998 };
1999 PERF_RECORD_SWITCH_CPU_WIDE (since Linux 4.3)
2001 As with PERF_RECORD_SWITCH this record indicates a context
2002 switch has happened, but it only occurs when sampling in
2003 CPU-wide mode and provides additional information on the
2004 process being switched to/from. The
2005 PERF_RECORD_MISC_SWITCH_OUT bit in the misc field indicates
2006 whether it was a context switch into or away from the cur‐
2007 rent process.
2008 struct {
2010 struct perf_event_header header;
2011 u32 next_prev_pid;
2012 u32 next_prev_tid;
2013 struct sample_id sample_id;
2014 };
2015 next_prev_pid
2017 The process ID of the previous (if switching in) or
2018 next (if switching out) process on the CPU.
2019 next_prev_tid
2021 The thread ID of the previous (if switching in) or
2022 next (if switching out) thread on the CPU.
2023 PERF_RECORD_NAMESPACES (since Linux 4.11)
2025 This record includes various namespace information of a
2026 process.
2027 struct {
2029 struct perf_event_header header;
2030 u32 pid;
2031 u32 tid;
2032 u64 nr_namespaces;
2033 struct { u64 dev, inode } [nr_namespaces];
2034 struct sample_id sample_id;
2035 };
2036 pid is the process ID
2038 tid is the thread ID
2040 nr_namespace
2042 is the number of namespaces in this record
2043 Each namespace has dev and inode fields and is recorded in
2045 the fixed position like below:
2046 NET_NS_INDEX=0
2048 Network namespace
2049 UTS_NS_INDEX=1
2051 UTS namespace
2052 IPC_NS_INDEX=2
2054 IPC namespace
2055 PID_NS_INDEX=3
2057 PID namespace
2058 USER_NS_INDEX=4
2060 User namespace
2061 MNT_NS_INDEX=5
2063 Mount namespace
2064 CGROUP_NS_INDEX=6
2066 Cgroup namespace
2067 PERF_RECORD_KSYMBOL (since Linux 5.0)
2069 This record indicates kernel symbol register/unregister
2070 events.
2071 struct {
2073 struct perf_event_header header;
2074 u64 addr;
2075 u32 len;
2076 u16 ksym_type;
2077 u16 flags;
2078 char name[];
2079 struct sample_id sample_id;
2080 };
2081 addr is the address of the kernel symbol.
2083 len is the length of the kernel symbol.
2085 ksym_type
2087 is the type of the kernel symbol. Currently the fol‐
2088 lowing types are available:
2089 PERF_RECORD_KSYMBOL_TYPE_BPF
2091 The kernel symbol is a BPF function.
2092 flags If the PERF_RECORD_KSYMBOL_FLAGS_UNREGISTER is set,
2094 then this event is for unregistering the kernel sym‐
2095 bol.
2096 PERF_RECORD_BPF_EVENT (since Linux 5.0)
2098 This record indicates BPF program is loaded or unloaded.
2099 struct {
2101 struct perf_event_header header;
2102 u16 type;
2103 u16 flags;
2104 u32 id;
2105 u8 tag[BPF_TAG_SIZE];
2106 struct sample_id sample_id;
2107 };
2108 type is one of the following values:
2110 PERF_BPF_EVENT_PROG_LOAD
2112 A BPF program is loaded
2113 PERF_BPF_EVENT_PROG_UNLOAD
2115 A BPF program is unloaded
2116 id is the ID of the BPF program.
2118 tag is the tag of the BPF program. Currently,
2120 BPF_TAG_SIZE is defined as 8.
2121 PERF_RECORD_CGROUP (since Linux 5.7)
2123 This record indicates a new cgroup is created and activated.
2124 struct {
2126 struct perf_event_header header;
2127 u64 id;
2128 char path[];
2129 struct sample_id sample_id;
2130 };
2131 id is the cgroup identifier. This can be also retrieved
2133 by name_to_handle_at(2) on the cgroup path (as a file
2134 handle).
2135 path is the path of the cgroup from the root.
2137 PERF_RECORD_TEXT_POKE (since Linux 5.8)
2139 This record indicates a change in the kernel text. This in‐
2140 cludes addition and removal of the text and the correspond‐
2141 ing length is zero in this case.
2142 struct {
2144 struct perf_event_header header;
2145 u64 addr;
2146 u16 old_len;
2147 u16 new_len;
2148 u8 bytes[];
2149 struct sample_id sample_id;
2150 };
2151 addr is the address of the change
2153 old_len
2155 is the old length
2156 new_len
2158 is the new length
2159 bytes contains old bytes immediately followed by new bytes.
2161 Overflow handling
2163 Events can be set to notify when a threshold is crossed, indicating an
2164 overflow. Overflow conditions can be captured by monitoring the event
2165 file descriptor with poll(2), select(2), or epoll(7). Alternatively,
2166 the overflow events can be captured via sa signal handler, by enabling
2167 I/O signaling on the file descriptor; see the discussion of the F_SE‐
2168 TOWN and F_SETSIG operations in fcntl(2).
2169 Overflows are generated only by sampling events (sample_period must
2171 have a nonzero value).
2172 There are two ways to generate overflow notifications.
2174 The first is to set a wakeup_events or wakeup_watermark value that will
2176 trigger if a certain number of samples or bytes have been written to
2177 the mmap ring buffer. In this case, POLL_IN is indicated.
2178 The other way is by use of the PERF_EVENT_IOC_REFRESH ioctl. This
2180 ioctl adds to a counter that decrements each time the event overflows.
2181 When nonzero, POLL_IN is indicated, but once the counter reaches 0
2182 POLL_HUP is indicated and the underlying event is disabled.
2183 Refreshing an event group leader refreshes all siblings and refreshing
2185 with a parameter of 0 currently enables infinite refreshes; these be‐
2186 haviors are unsupported and should not be relied on.
2187 Starting with Linux 3.18, POLL_HUP is indicated if the event being mon‐
2189 itored is attached to a different process and that process exits.
2190 rdpmc instruction
2192 Starting with Linux 3.4 on x86, you can use the rdpmc instruction to
2193 get low-latency reads without having to enter the kernel. Note that
2194 using rdpmc is not necessarily faster than other methods for reading
2195 event values.
2196 Support for this can be detected with the cap_usr_rdpmc field in the
2198 mmap page; documentation on how to calculate event values can be found
2199 in that section.
2200 Originally, when rdpmc support was enabled, any process (not just ones
2202 with an active perf event) could use the rdpmc instruction to access
2203 the counters. Starting with Linux 4.0, rdpmc support is only allowed
2204 if an event is currently enabled in a process's context. To restore
2205 the old behavior, write the value 2 to /sys/devices/cpu/rdpmc.
2206 perf_event ioctl calls
2208 Various ioctls act on perf_event_open() file descriptors:
2209 PERF_EVENT_IOC_ENABLE
2211 This enables the individual event or event group specified by
2212 the file descriptor argument.
2213 If the PERF_IOC_FLAG_GROUP bit is set in the ioctl argument,
2215 then all events in a group are enabled, even if the event speci‐
2216 fied is not the group leader (but see BUGS).
2217 PERF_EVENT_IOC_DISABLE
2219 This disables the individual counter or event group specified by
2220 the file descriptor argument.
2221 Enabling or disabling the leader of a group enables or disables
2223 the entire group; that is, while the group leader is disabled,
2224 none of the counters in the group will count. Enabling or dis‐
2225 abling a member of a group other than the leader affects only
2226 that counter; disabling a non-leader stops that counter from
2227 counting but doesn't affect any other counter.
2228 If the PERF_IOC_FLAG_GROUP bit is set in the ioctl argument,
2230 then all events in a group are disabled, even if the event spec‐
2231 ified is not the group leader (but see BUGS).
2232 PERF_EVENT_IOC_REFRESH
2234 Non-inherited overflow counters can use this to enable a counter
2235 for a number of overflows specified by the argument, after which
2236 it is disabled. Subsequent calls of this ioctl add the argument
2237 value to the current count. An overflow notification with
2238 POLL_IN set will happen on each overflow until the count reaches
2239 0; when that happens a notification with POLL_HUP set is sent
2240 and the event is disabled. Using an argument of 0 is considered
2241 undefined behavior.
2242 PERF_EVENT_IOC_RESET
2244 Reset the event count specified by the file descriptor argument
2245 to zero. This resets only the counts; there is no way to reset
2246 the multiplexing time_enabled or time_running values.
2247 If the PERF_IOC_FLAG_GROUP bit is set in the ioctl argument,
2249 then all events in a group are reset, even if the event speci‐
2250 fied is not the group leader (but see BUGS).
2251 PERF_EVENT_IOC_PERIOD
2253 This updates the overflow period for the event.
2254 Since Linux 3.7 (on ARM) and Linux 3.14 (all other architec‐
2256 tures), the new period takes effect immediately. On older ker‐
2257 nels, the new period did not take effect until after the next
2258 overflow.
2259 The argument is a pointer to a 64-bit value containing the de‐
2261 sired new period.
2262 Prior to Linux 2.6.36, this ioctl always failed due to a bug in
2264 the kernel.
2265 PERF_EVENT_IOC_SET_OUTPUT
2267 This tells the kernel to report event notifications to the spec‐
2268 ified file descriptor rather than the default one. The file de‐
2269 scriptors must all be on the same CPU.
2270 The argument specifies the desired file descriptor, or -1 if
2272 output should be ignored.
2273 PERF_EVENT_IOC_SET_FILTER (since Linux 2.6.33)
2275 This adds an ftrace filter to this event.
2276 The argument is a pointer to the desired ftrace filter.
2278 PERF_EVENT_IOC_ID (since Linux 3.12)
2280 This returns the event ID value for the given event file de‐
2281 scriptor.
2282 The argument is a pointer to a 64-bit unsigned integer to hold
2284 the result.
2285 PERF_EVENT_IOC_SET_BPF (since Linux 4.1)
2287 This allows attaching a Berkeley Packet Filter (BPF) program to
2288 an existing kprobe tracepoint event. You need CAP_PERFMON
2289 (since Linux 5.8) or CAP_SYS_ADMIN privileges to use this ioctl.
2290 The argument is a BPF program file descriptor that was created
2292 by a previous bpf(2) system call.
2293 PERF_EVENT_IOC_PAUSE_OUTPUT (since Linux 4.7)
2295 This allows pausing and resuming the event's ring-buffer. A
2296 paused ring-buffer does not prevent generation of samples, but
2297 simply discards them. The discarded samples are considered
2298 lost, and cause a PERF_RECORD_LOST sample to be generated when
2299 possible. An overflow signal may still be triggered by the dis‐
2300 carded sample even though the ring-buffer remains empty.
2301 The argument is an unsigned 32-bit integer. A nonzero value
2303 pauses the ring-buffer, while a zero value resumes the ring-buf‐
2304 fer.
2305 PERF_EVENT_MODIFY_ATTRIBUTES (since Linux 4.17)
2307 This allows modifying an existing event without the overhead of
2308 closing and reopening a new event. Currently this is supported
2309 only for breakpoint events.
2310 The argument is a pointer to a perf_event_attr structure con‐
2312 taining the updated event settings.
2313 PERF_EVENT_IOC_QUERY_BPF (since Linux 4.16)
2315 This allows querying which Berkeley Packet Filter (BPF) programs
2316 are attached to an existing kprobe tracepoint. You can only at‐
2317 tach one BPF program per event, but you can have multiple events
2318 attached to a tracepoint. Querying this value on one tracepoint
2319 event returns the ID of all BPF programs in all events attached
2320 to the tracepoint. You need CAP_PERFMON (since Linux 5.8) or
2321 CAP_SYS_ADMIN privileges to use this ioctl.
2322 The argument is a pointer to a structure
2324 struct perf_event_query_bpf {
2325 __u32 ids_len;
2326 __u32 prog_cnt;
2327 __u32 ids[0];
2328 };
2329 The ids_len field indicates the number of ids that can fit in
2331 the provided ids array. The prog_cnt value is filled in by the
2332 kernel with the number of attached BPF programs. The ids array
2333 is filled with the ID of each attached BPF program. If there
2334 are more programs than will fit in the array, then the kernel
2335 will return ENOSPC and ids_len will indicate the number of pro‐
2336 gram IDs that were successfully copied.
2337 Using prctl(2)
2339 A process can enable or disable all currently open event groups using
2340 the prctl(2) PR_TASK_PERF_EVENTS_ENABLE and PR_TASK_PERF_EVENTS_DISABLE
2341 operations. This applies only to events created locally by the calling
2342 process. This does not apply to events created by other processes at‐
2343 tached to the calling process or inherited events from a parent
2344 process. Only group leaders are enabled and disabled, not any other
2345 members of the groups.
2346 perf_event related configuration files
2348 Files in /proc/sys/kernel/
2349 /proc/sys/kernel/perf_event_paranoid
2351 The perf_event_paranoid file can be set to restrict access
2352 to the performance counters.
2353 2 allow only user-space measurements (default since
2355 Linux 4.6).
2356 1 allow both kernel and user measurements (default be‐
2357 fore Linux 4.6).
2358 0 allow access to CPU-specific data but not raw trace‐
2359 point samples.
2360 -1 no restrictions.
2361 The existence of the perf_event_paranoid file is the offi‐
2363 cial method for determining if a kernel supports
2364 perf_event_open().
2365 /proc/sys/kernel/perf_event_max_sample_rate
2367 This sets the maximum sample rate. Setting this too high
2368 can allow users to sample at a rate that impacts overall ma‐
2369 chine performance and potentially lock up the machine. The
2370 default value is 100000 (samples per second).
2371 /proc/sys/kernel/perf_event_max_stack
2373 This file sets the maximum depth of stack frame entries re‐
2374 ported when generating a call trace.
2375 /proc/sys/kernel/perf_event_mlock_kb
2377 Maximum number of pages an unprivileged user can mlock(2).
2378 The default is 516 (kB).
2379 Files in /sys/bus/event_source/devices/
2381 Since Linux 2.6.34, the kernel supports having multiple PMUs avail‐
2383 able for monitoring. Information on how to program these PMUs can
2384 be found under /sys/bus/event_source/devices/. Each subdirectory
2385 corresponds to a different PMU.
2386 /sys/bus/event_source/devices/*/type (since Linux 2.6.38)
2388 This contains an integer that can be used in the type field
2389 of perf_event_attr to indicate that you wish to use this
2390 PMU.
2391 /sys/bus/event_source/devices/cpu/rdpmc (since Linux 3.4)
2393 If this file is 1, then direct user-space access to the per‐
2394 formance counter registers is allowed via the rdpmc instruc‐
2395 tion. This can be disabled by echoing 0 to the file.
2396 As of Linux 4.0 the behavior has changed, so that 1 now
2398 means only allow access to processes with active perf
2399 events, with 2 indicating the old allow-anyone-access behav‐
2400 ior.
2401 /sys/bus/event_source/devices/*/format/ (since Linux 3.4)
2403 This subdirectory contains information on the architecture-
2404 specific subfields available for programming the various
2405 config fields in the perf_event_attr struct.
2406 The content of each file is the name of the config field,
2408 followed by a colon, followed by a series of integer bit
2409 ranges separated by commas. For example, the file event may
2410 contain the value config1:1,6-10,44 which indicates that
2411 event is an attribute that occupies bits 1,6–10, and 44 of
2412 perf_event_attr::config1.
2413 /sys/bus/event_source/devices/*/events/ (since Linux 3.4)
2415 This subdirectory contains files with predefined events.
2416 The contents are strings describing the event settings ex‐
2417 pressed in terms of the fields found in the previously men‐
2418 tioned ./format/ directory. These are not necessarily com‐
2419 plete lists of all events supported by a PMU, but usually a
2420 subset of events deemed useful or interesting.
2421 The content of each file is a list of attribute names sepa‐
2423 rated by commas. Each entry has an optional value (either
2424 hex or decimal). If no value is specified, then it is as‐
2425 sumed to be a single-bit field with a value of 1. An exam‐
2426 ple entry may look like this: event=0x2,inv,ldlat=3.
2427 /sys/bus/event_source/devices/*/uevent
2429 This file is the standard kernel device interface for in‐
2430 jecting hotplug events.
2431 /sys/bus/event_source/devices/*/cpumask (since Linux 3.7)
2433 The cpumask file contains a comma-separated list of integers
2434 that indicate a representative CPU number for each socket
2435 (package) on the motherboard. This is needed when setting
2436 up uncore or northbridge events, as those PMUs present
2437 socket-wide events.
2438
2440 On success, perf_event_open() returns the new file descriptor. On er‐
2441 ror, -1 is returned and errno is set to indicate the error.
2442
2444 The errors returned by perf_event_open() can be inconsistent, and may
2445 vary across processor architectures and performance monitoring units.
2446 E2BIG Returned if the perf_event_attr size value is too small (smaller
2448 than PERF_ATTR_SIZE_VER0), too big (larger than the page size),
2449 or larger than the kernel supports and the extra bytes are not
2450 zero. When E2BIG is returned, the perf_event_attr size field is
2451 overwritten by the kernel to be the size of the structure it was
2452 expecting.
2453 EACCES Returned when the requested event requires CAP_PERFMON (since
2455 Linux 5.8) or CAP_SYS_ADMIN permissions (or a more permissive
2456 perf_event paranoid setting). Some common cases where an un‐
2457 privileged process may encounter this error: attaching to a
2458 process owned by a different user; monitoring all processes on a
2459 given CPU (i.e., specifying the pid argument as -1); and not
2460 setting exclude_kernel when the paranoid setting requires it.
2461 EBADF Returned if the group_fd file descriptor is not valid, or, if
2463 PERF_FLAG_PID_CGROUP is set, the cgroup file descriptor in pid
2464 is not valid.
2465 EBUSY (since Linux 4.1)
2467 Returned if another event already has exclusive access to the
2468 PMU.
2469 EFAULT Returned if the attr pointer points at an invalid memory ad‐
2471 dress.
2472 EINTR Returned when trying to mix perf and ftrace handling for a up‐
2474 robe.
2475 EINVAL Returned if the specified event is invalid. There are many pos‐
2477 sible reasons for this. A not-exhaustive list: sample_freq is
2478 higher than the maximum setting; the cpu to monitor does not ex‐
2479 ist; read_format is out of range; sample_type is out of range;
2480 the flags value is out of range; exclusive or pinned set and the
2481 event is not a group leader; the event config values are out of
2482 range or set reserved bits; the generic event selected is not
2483 supported; or there is not enough room to add the selected
2484 event.
2485 EMFILE Each opened event uses one file descriptor. If a large number
2487 of events are opened, the per-process limit on the number of
2488 open file descriptors will be reached, and no more events can be
2489 created.
2490 ENODEV Returned when the event involves a feature not supported by the
2492 current CPU.
2493 ENOENT Returned if the type setting is not valid. This error is also
2495 returned for some unsupported generic events.
2496 ENOSPC Prior to Linux 3.3, if there was not enough room for the event,
2498 ENOSPC was returned. In Linux 3.3, this was changed to EINVAL.
2499 ENOSPC is still returned if you try to add more breakpoint
2500 events than supported by the hardware.
2501 ENOSYS Returned if PERF_SAMPLE_STACK_USER is set in sample_type and it
2503 is not supported by hardware.
2504 EOPNOTSUPP
2506 Returned if an event requiring a specific hardware feature is
2507 requested but there is no hardware support. This includes re‐
2508 questing low-skid events if not supported, branch tracing if it
2509 is not available, sampling if no PMU interrupt is available, and
2510 branch stacks for software events.
2511 EOVERFLOW (since Linux 4.8)
2513 Returned if PERF_SAMPLE_CALLCHAIN is requested and sam‐
2514 ple_max_stack is larger than the maximum specified in
2515 /proc/sys/kernel/perf_event_max_stack.
2516 EPERM Returned on many (but not all) architectures when an unsupported
2518 exclude_hv, exclude_idle, exclude_user, or exclude_kernel set‐
2519 ting is specified.
2520 It can also happen, as with EACCES, when the requested event re‐
2522 quires CAP_PERFMON (since Linux 5.8) or CAP_SYS_ADMIN permis‐
2523 sions (or a more permissive perf_event paranoid setting). This
2524 includes setting a breakpoint on a kernel address, and (since
2525 Linux 3.13) setting a kernel function-trace tracepoint.
2526 ESRCH Returned if attempting to attach to a process that does not ex‐
2528 ist.
2529
2531 Linux.
2532
2534 perf_event_open() was introduced in Linux 2.6.31 but was called
2535 perf_counter_open(). It was renamed in Linux 2.6.32.
2536
2538 The official way of knowing if perf_event_open() support is enabled is
2539 checking for the existence of the file /proc/sys/ker‐
2540 nel/perf_event_paranoid.
2541 CAP_PERFMON capability (since Linux 5.8) provides secure approach to
2543 performance monitoring and observability operations in a system accord‐
2544 ing to the principal of least privilege (POSIX IEEE 1003.1e). Access‐
2545 ing system performance monitoring and observability operations using
2546 CAP_PERFMON rather than the much more powerful CAP_SYS_ADMIN excludes
2547 chances to misuse credentials and makes operations more secure.
2548 CAP_SYS_ADMIN usage for secure system performance monitoring and ob‐
2549 servability is discouraged in favor of the CAP_PERFMON capability.
2550
2552 The F_SETOWN_EX option to fcntl(2) is needed to properly get overflow
2553 signals in threads. This was introduced in Linux 2.6.32.
2554 Prior to Linux 2.6.33 (at least for x86), the kernel did not check if
2556 events could be scheduled together until read time. The same happens
2557 on all known kernels if the NMI watchdog is enabled. This means to see
2558 if a given set of events works you have to perf_event_open(), start,
2559 then read before you know for sure you can get valid measurements.
2560 Prior to Linux 2.6.34, event constraints were not enforced by the ker‐
2562 nel. In that case, some events would silently return "0" if the kernel
2563 scheduled them in an improper counter slot.
2564 Prior to Linux 2.6.34, there was a bug when multiplexing where the
2566 wrong results could be returned.
2567 Kernels from Linux 2.6.35 to Linux 2.6.39 can quickly crash the kernel
2569 if "inherit" is enabled and many threads are started.
2570 Prior to Linux 2.6.35, PERF_FORMAT_GROUP did not work with attached
2572 processes.
2573 There is a bug in the kernel code between Linux 2.6.36 and Linux 3.0
2575 that ignores the "watermark" field and acts as if a wakeup_event was
2576 chosen if the union has a nonzero value in it.
2577 From Linux 2.6.31 to Linux 3.4, the PERF_IOC_FLAG_GROUP ioctl argument
2579 was broken and would repeatedly operate on the event specified rather
2580 than iterating across all sibling events in a group.
2581 From Linux 3.4 to Linux 3.11, the mmap cap_usr_rdpmc and cap_usr_time
2583 bits mapped to the same location. Code should migrate to the new
2584 cap_user_rdpmc and cap_user_time fields instead.
2585 Always double-check your results! Various generalized events have had
2587 wrong values. For example, retired branches measured the wrong thing
2588 on AMD machines until Linux 2.6.35.
2589
2591 The following is a short example that measures the total instruction
2592 count of a call to printf(3).
2593 #include <linux/perf_event.h>
2595 #include <stdio.h>
2596 #include <stdlib.h>
2597 #include <string.h>
2598 #include <sys/ioctl.h>
2599 #include <sys/syscall.h>
2600 #include <unistd.h>
2601 static long
2603 perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
2604 int cpu, int group_fd, unsigned long flags)
2605 {
2606 int ret;
2607 ret = syscall(SYS_perf_event_open, hw_event, pid, cpu,
2609 group_fd, flags);
2610 return ret;
2611 }
2612 int
2614 main(void)
2615 {
2616 int fd;
2617 long long count;
2618 struct perf_event_attr pe;
2619 memset(&pe, 0, sizeof(pe));
2621 pe.type = PERF_TYPE_HARDWARE;
2622 pe.size = sizeof(pe);
2623 pe.config = PERF_COUNT_HW_INSTRUCTIONS;
2624 pe.disabled = 1;
2625 pe.exclude_kernel = 1;
2626 pe.exclude_hv = 1;
2627 fd = perf_event_open(&pe, 0, -1, -1, 0);
2629 if (fd == -1) {
2630 fprintf(stderr, "Error opening leader %llx\n", pe.config);
2631 exit(EXIT_FAILURE);
2632 }
2633 ioctl(fd, PERF_EVENT_IOC_RESET, 0);
2635 ioctl(fd, PERF_EVENT_IOC_ENABLE, 0);
2636 printf("Measuring instruction count for this printf\n");
2638 ioctl(fd, PERF_EVENT_IOC_DISABLE, 0);
2640 read(fd, &count, sizeof(count));
2641 printf("Used %lld instructions\n", count);
2643 close(fd);
2645 }
2646
2648 perf(1), fcntl(2), mmap(2), open(2), prctl(2), read(2)
2649 Documentation/admin-guide/perf-security.rst in the kernel source tree
2651Linux man-pages 6.05 2023-05-03 perf_event_open(2)