PERF_EVENT_OPEN(2) Linux Programmer's Manual PERF_EVENT_OPEN(2)


perf_event_open - set up performance monitoring



int perf_event_open(struct perf_event_attr * attr ,
pid_t pid , int cpu , int group_fd ,
unsigned long flags );
Note: There is no glibc wrapper for this system call; see NOTES.


Given a list of parameters, perf_event_open() returns a file descriptor, for use in subsequent system calls ( read(2), mmap(2), prctl(2), fcntl(2), etc.).

A call to perf_event_open() creates a file descriptor that allows measuring performance information. Each file descriptor corresponds to one event that is measured; these can be grouped together to measure multiple events simultaneously.

Events can be enabled and disabled in two ways: via ioctl(2) and via prctl(2) . When an event is disabled it does not count or generate overflows but does continue to exist and maintain its count value.

Events come in two flavors: counting and sampled. A counting event is one that is used for counting the aggregate number of events that occur. In general, counting event results are gathered with a read(2) call. A sampling event periodically writes measurements to a buffer that can then be accessed via mmap(2) .


The argument pid allows events to be attached to processes in various ways. If pid is 0, measurements happen on the current thread, if pid is greater than 0, the process indicated by pid is measured, and if pid is -1, all processes are counted.
The cpu argument allows measurements to be specific to a CPU. If cpu is greater than or equal to 0, measurements are restricted to the specified CPU; if cpu is -1, the events are measured on all CPUs.

Note that the combination of pid == -1 and cpu == -1 is not valid.

A pid > 0 and cpu == -1 setting measures per-process and follows that process to whatever CPU the process gets scheduled to. Per-process events can be created by any user.

A pid == -1 and cpu >= 0 setting is per-CPU and measures all processes on the specified CPU. Per-CPU events need the CAP_SYS_ADMIN capability or a /proc/sys/kernel/perf_event_paranoid value of less than 1.

The group_fd argument allows event groups to be created. An event group has one event which is the group leader. The leader is created first, with group_fd = -1. The rest of the group members are created with subsequent perf_event_open() calls with group_fd being set to the fd of the group leader. (A single event on its own is created with group_fd = -1 and is considered to be a group with only 1 member.) An event group is scheduled onto the CPU as a unit: it will be put onto the CPU only if all of the events in the group can be put onto the CPU. This means that the values of the member events can be meaningfully compared, added, divided (to get ratios), etc., with each other, since they have counted events for the same set of executed instructions.

The flags argument is formed by ORing together zero or more of the following values:

This flag allows creating an event as part of an event group but having no group leader. It is unclear why this is useful.
This flag re-routes the output from an event to the group leader.
PERF_FLAG_PID_CGROUP (Since Linux 2.6.39).
This flag activates per-container system-wide monitoring. A container is an abstraction that isolates a set of resources for finer grain control (CPUs, memory, etc...). In this mode, the event is measured only if the thread running on the monitored CPU belongs to the designated container (cgroup). The cgroup is identified by passing a file descriptor opened on its directory in the cgroupfs filesystem. For instance, if the cgroup to monitor is called test, then a file descriptor opened on /dev/cgroup/test (assuming cgroupfs is mounted on /dev/cgroup) must be passed as the pid parameter. cgroup monitoring is available only for system-wide events and may therefore require extra permissions.

The perf_event_attr structure provides detailed configuration information for the event being created.


struct perf_event_attr {
__u32 type; /* Type of event */
__u32 size; /* Size of attribute structure */
__u64 config; /* Type-specific configuration */

union {
__u64 sample_period; /* Period of sampling */
__u64 sample_freq; /* Frequency of sampling */

__u64 sample_type; /* Specifies values included in sample */
__u64 read_format; /* Specifies values returned in read */

__u64 disabled : 1, /* off by default */
inherit : 1, /* children inherit it */
pinned : 1, /* must always be on PMU */
exclusive : 1, /* only group on PMU */
exclude_user : 1, /* don't count user */
exclude_kernel : 1, /* don't count kernel */
exclude_hv : 1, /* don't count hypervisor */
exclude_idle : 1, /* don't count when idle */
mmap : 1, /* include mmap data */
comm : 1, /* include comm data */
freq : 1, /* use freq, not period */
inherit_stat : 1, /* per task counts */
enable_on_exec : 1, /* next exec enables */
task : 1, /* trace fork/exit */
watermark : 1, /* wakeup_watermark */
precise_ip : 2, /* skid constraint */
mmap_data : 1, /* non-exec mmap data */
sample_id_all : 1, /* sample_type all events */
exclude_host : 1, /* don't count in host */
exclude_guest : 1, /* don't count in guest */
exclude_callchain_kernel : 1,
/* exclude kernel callchains */
exclude_callchain_user : 1,
/* exclude user callchains */
__reserved_1 : 41;

union {
__u32 wakeup_events; /* wakeup every n events */
__u32 wakeup_watermark; /* bytes before wakeup */

__u32 bp_type; /* breakpoint type */

union {
__u64 bp_addr; /* breakpoint address */
__u64 config1; /* extension of config */

union {
__u64 bp_len; /* breakpoint length */
__u64 config2; /* extension of config1 */
__u64 branch_sample_type; /* enum perf_branch_sample_type */
__u64 sample_regs_user; /* user regs to dump on samples */
__u32 sample_stack_user; /* size of stack to dump on
samples */
__u32 __reserved_2; /* Align to u64 */



The fields of the perf_event_attr structure are described in more detail below:

This field specifies the overall event type. It has one of the following values:
This indicates one of the "generalized" hardware events provided by the kernel. See the config field definition for more details.
This indicates one of the software-defined events provided by the kernel (even if no hardware support is available).
This indicates a tracepoint provided by the kernel tracepoint infrastructure.
This indicates a hardware cache event. This has a special encoding, described in the config field definition.
This indicates a "raw" implementation-specific event in the config field.
PERF_TYPE_BREAKPOINT (Since Linux 2.6.33)
This indicates a hardware breakpoint as provided by the CPU. Breakpoints can be read/write accesses to an address as well as execution of an instruction address.
dynamic PMU
Since Linux 2.6.39, perf_event_open() can support multiple PMUs. To enable this, a value exported by the kernel can be used in the type field to indicate which PMU to use. The value to use can be found in the sysfs filesystem: there is a subdirectory per PMU instance under /sys/bus/event_source/devices. In each sub-directory there is a type file whose content is an integer that can be used in the type field. For instance, /sys/bus/event_source/devices/cpu/type contains the value for the core CPU PMU, which is usually 4.
The size of the perf_event_attr structure for forward/backward compatibility. Set this using sizeof(struct perf_event_attr) to allow the kernel to see the struct size at the time of compilation.
The related define PERF_ATTR_SIZE_VER0 is set to 64; this was the size of the first published struct. PERF_ATTR_SIZE_VER1 is 72, corresponding to the addition of breakpoints in Linux 2.6.33. PERF_ATTR_SIZE_VER2 is 80 corresponding to the addition of branch sampling in Linux 3.4. PERF_ATR_SIZE_VER3 is 96 corresponding to the addition of sample_regs_user and sample_stack_user in Linux 3.7.
This specifies which event you want, in conjunction with the type field. The config1 and config2 fields are also taken into account in cases where 64 bits is not enough to fully specify the event. The encoding of these fields are event dependent.
The most significant bit (bit 63) of config signifies CPU-specific (raw) counter configuration data; if the most significant bit is unset, the next 7 bits are an event type and the rest of the bits are the event identifier.
There are various ways to set the config field that are dependent on the value of the previously described type field. What follows are various possible settings for config separated out by type.
If type is PERF_TYPE_HARDWARE, we are measuring one of the generalized hardware CPU events. Not all of these are available on all platforms. Set config to one of the following:
Total cycles. Be wary of what happens during CPU frequency scaling
Retired instructions. Be careful, these can be affected by various issues, most notably hardware interrupt counts
Cache accesses. Usually this indicates Last Level Cache accesses but this may vary depending on your CPU. This may include prefetches and coherency messages; again this depends on the design of your CPU.
Cache misses. Usually this indicates Last Level Cache misses; this is intended to be used in conjunction with the PERF_COUNT_HW_CACHE_REFERENCES event to calculate cache miss rates.
Retired branch instructions. Prior to Linux 2.6.34, this used the wrong event on AMD processors.
Mispredicted branch instructions.
Bus cycles, which can be different from total cycles.
Stalled cycles during issue.
Stalled cycles during retirement.
Total cycles; not affected by CPU frequency scaling.
If type is PERF_TYPE_SOFTWARE, we are measuring software events provided by the kernel. Set config to one of the following:
This reports the CPU clock, a high-resolution per-CPU timer.
This reports a clock count specific to the task that is running.
This reports the number of page faults.
This counts context switches. Until Linux 2.6.34, these were all reported as user-space events, after that they are reported as happening in the kernel.
This reports the number of times the process has migrated to a new CPU.
This counts the number of minor page faults. These did not require disk I/O to handle.
This counts the number of major page faults. These required disk I/O to handle.
This counts the number of alignment faults. These happen when unaligned memory accesses happen; the kernel can handle these but it reduces performance. This happens only on some architectures (never on x86).
This counts the number of emulation faults. The kernel sometimes traps on unimplemented instructions and emulates them for user space. This can negatively impact performance.
If type is PERF_TYPE_TRACEPOINT, then we are measuring kernel tracepoints. The value to use in config can be obtained from under debugfs tracing/events/*/*/id if ftrace is enabled in the kernel.
If type is PERF_TYPE_HW_CACHE, then we are measuring a hardware CPU cache event. To calculate the appropriate config value use the following equation:

(perf_hw_cache_id) | (perf_hw_cache_op_id << 8) |
(perf_hw_cache_op_result_id << 16)

where perf_hw_cache_id is one of:

for measuring Level 1 Data Cache
for measuring Level 1 Instruction Cache
for measuring Last-Level Cache
for measuring the Data TLB
for measuring the Instruction TLB
for measuring the branch prediction unit
for measuring local memory accesses

and perf_hw_cache_op_id is one of

for read accesses
for write accesses
for prefetch accesses

and perf_hw_cache_op_result_id is one of

to measure accesses
to measure misses
If type is PERF_TYPE_RAW, then a custom "raw" config value is needed. Most CPUs support events that are not covered by the "generalized" events. These are implementation defined; see your CPU manual (for example the Intel Volume 3B documentation or the AMD BIOS and Kernel Developer Guide). The libpfm4 library can be used to translate from the name in the architectural manuals to the raw hex value perf_event_open() expects in this field.
If type is PERF_TYPE_BREAKPOINT, then leave config set to zero. Its parameters are set in other places.
sample_period, sample_freq
A "sampling" counter is one that generates an interrupt every N events, where N is given by sample_period. A sampling counter has sample_period > 0. When an overflow interrupt occurs, requested data is recorded in the mmap buffer. The sample_type field controls what data is recorded on each interrupt.
sample_freq can be used if you wish to use frequency rather than period. In this case you set the freq flag. The kernel will adjust the sampling period to try and achieve the desired rate. The rate of adjustment is a timer tick.
The various bits in this field specify which values to include in the sample. They will be recorded in a ring-buffer, which is available to user space using mmap(2). The order in which the values are saved in the sample are documented in the MMAP Layout subsection below; it is not the enum perf_event_sample_format order.
Records instruction pointer.
Records the process and thread IDs.
Records a timestamp.
Records an address, if applicable.
Record counter values for all events in a group, not just the group leader.
Records the callchain (stack backtrace).
Records a unique ID for the opened event's group leader.
Records CPU number.
Records the current sampling period.
Records a unique ID for the opened event. Unlike PERF_SAMPLE_ID the actual ID is returned, not the group leader. This ID is the same as the one returned by PERF_FORMAT_ID.
Records additional data, if applicable. Usually returned by tracepoint events.
This provides a record of recent branches, as provided by CPU branch sampling hardware (such as Intel Last Branch Record). Not all hardware supports this feature.
See the branch_sample_type field for how to filter which branches are reported.
Records the current user-level CPU register state (the values in the process before the kernel was called).
Records the user level stack, allowing stack unwinding.
PERF_SAMPLE_WEIGHT (Since Linux 3.10)
Records a hardware provided weight value that expresses how costly the sampled event was. This allows the hardware to highlight expensive events in a profile.
PERF_SAMPLE_DATA_SRC (Since Linux 3.10)
Records the data source: where in the memory hierarchy the data associated with the sampled instruction came from. This is only available if the underlying hardware supports this feature.
This field specifies the format of the data returned by read(2) on a perf_event_open() file descriptor.
Adds the 64-bit time_enabled field. This can be used to calculate estimated totals if the PMU is overcommitted and multiplexing is happening.
Adds the 64-bit time_running field. This can be used to calculate estimated totals if the PMU is overcommitted and multiplexing is happening.
Adds a 64-bit unique value that corresponds to the event group.
Allows all counter values in an event group to be read with one read.
The disabled bit specifies whether the counter starts out disabled or enabled. If disabled, the event can later be enabled by ioctl(2), prctl(2), or enable_on_exec.
The inherit bit specifies that this counter should count events of child tasks as well as the task specified. This applies only to new children, not to any existing children at the time the counter is created (nor to any new children of existing children).
Inherit does not work for some combinations of read_formats, such as PERF_FORMAT_GROUP.
The pinned bit specifies that the counter should always be on the CPU if at all possible. It applies only to hardware counters and only to group leaders. If a pinned counter cannot be put onto the CPU (e.g., because there are not enough hardware counters or because of a conflict with some other event), then the counter goes into an 'error' state, where reads return end-of-file (i.e., read(2) returns 0) until the counter is subsequently enabled or disabled.
The exclusive bit specifies that when this counter's group is on the CPU, it should be the only group using the CPU's counters. In the future this may allow monitoring programs to support PMU features that need to run alone so that they do not disrupt other hardware counters.
If this bit is set, the count excludes events that happen in user space.
If this bit is set, the count excludes events that happen in kernel-space.
If this bit is set, the count excludes events that happen in the hypervisor. This is mainly for PMUs that have built-in support for handling this (such as POWER). Extra support is needed for handling hypervisor measurements on most machines.
If set, don't count when the CPU is idle.
The mmap bit enables recording of exec mmap events.
The comm bit enables tracking of process command name as modified by the exec(2) and prctl(PR_SET_NAME) system calls. Unfortunately for tools, there is no way to distinguish one system call versus the other.
If this bit is set, then sample_frequency not sample_period is used when setting up the sampling interval.
This bit enables saving of event counts on context switch for inherited tasks. This is meaningful only if the inherit field is set.
If this bit is set, a counter is automatically enabled after a call to exec(2).
If this bit is set, then fork/exit notifications are included in the ring buffer.
If set, have a sampling interrupt happen when we cross the wakeup_watermark boundary. Otherwise interrupts happen after wakeup_events samples.
precise_ip (Since Linux 2.6.35)
This controls the amount of skid. Skid is how many instructions execute between an event of interest happening and the kernel being able to stop and record the event. Smaller skid is better and allows more accurate reporting of which events correspond to which instructions, but hardware is often limited with how small this can be.
The values of this are the following:
0 -
SAMPLE_IP can have arbitrary skid
1 -
SAMPLE_IP must have constant skid
2 -
SAMPLE_IP requested to have 0 skid
3 -
SAMPLE_IP must have 0 skid. See also PERF_RECORD_MISC_EXACT_IP.
mmap_data (Since Linux 2.6.36)
The counterpart of the mmap field, but enables including data mmap events in the ring-buffer.
sample_id_all (Since Linux 2.6.38)
If set, then TID, TIME, ID, CPU, and STREAM_ID can additionally be included in non- PERF_RECORD_SAMPLEs if the corresponding sample_type is selected.
exclude_host (Since Linux 3.2)
Do not measure time spent in VM host
exclude_guest (Since Linux 3.2)
Do not measure time spent in VM guest
exclude_callchain_kernel (Since Linux 3.7)
Do not include kernel callchains.
exclude_callchain_user (Since Linux 3.7)
Do not include user callchains.
wakeup_events, wakeup_watermark
This union sets how many samples ( wakeup_events) or bytes ( wakeup_watermark) happen before an overflow signal happens. Which one is used is selected by the watermark bitflag.
wakeup_events only counts PERF_RECORD_SAMPLE record types. To receive a signal for every incoming PERF_RECORD type set wakeup_watermark to 1.
bp_type (Since Linux 2.6.33)
This chooses the breakpoint type. It is one of:
no breakpoint
count when we read the memory location
count when we write the memory location
count when we read or write the memory location
count when we execute code at the memory location

The values can be combined via a bitwise or, but the combination of HW_BREAKPOINT_R or HW_BREAKPOINT_W with HW_BREAKPOINT_X is not allowed.

bp_addr (Since Linux 2.6.33)
bp_addr address of the breakpoint. For execution breakpoints this is the memory address of the instruction of interest; for read and write breakpoints it is the memory address of the memory location of interest.
config1 (Since Linux 2.6.39)
config1 is used for setting events that need an extra register or otherwise do not fit in the regular config field. Raw OFFCORE_EVENTS on Nehalem/Westmere/SandyBridge use this field on 3.3 and later kernels.
bp_len (Since Linux 2.6.33)
bp_len is the length of the breakpoint being measured if type is PERF_TYPE_BREAKPOINT. Options are HW_BREAKPOINT_LEN_1, HW_BREAKPOINT_LEN_2, HW_BREAKPOINT_LEN_4, HW_BREAKPOINT_LEN_8. For an execution breakpoint, set this to sizeof(long).
config2 (Since Linux 2.6.39)
config2 is a further extension of the config1 field.
branch_sample_type (Since Linux 3.4)
If PERF_SAMPLE_BRANCH_STACK is enabled, then this specifies what branches to include in the branch record. If the user does not set privilege level explicitly, the kernel will use the event's privilege level. Event and branch privilege levels do not have to match. The value is formed by ORing together zero or more of the following values, although PERF_SAMPLE_BRANCH_ANY covers all branch types.
Branch target is in user space
Branch target is in kernel space
Branch target is in hypervisor
Any branch type.
Any call branch
Any return branch
Indirect calls
User, kernel, and hv
sample_regs_user (Since Linux 3.7)
This bitmask defines the set of user CPU registers to dump on samples. The layout of the register mask is architecture specific and described in the kernel header arch/ARCH/include/uapi/asm/perf_regs.h.
sample_stack_user (Since Linux 3.7)
This defines the size of the user stack to dump if PERF_SAMPLE_STACK_USER is specified.

Reading results

Once a perf_event_open() file descriptor has been opened, the values of the events can be read from the file descriptor. The values that are there are specified by the read_format field in the attr structure at open time.
If you attempt to read into a buffer that is not big enough to hold the data ENOSPC is returned
Here is the layout of the data returned by a read:
If PERF_FORMAT_GROUP was specified to allow reading all events in a group at once:

struct read_format {
u64 nr; /* The number of events */
u64 time_enabled; /* if PERF_FORMAT_TOTAL_TIME_ENABLED */
u64 time_running; /* if PERF_FORMAT_TOTAL_TIME_RUNNING */
u64 value; /* The value of the event */
u64 id; /* if PERF_FORMAT_ID */
} values[nr];

If PERF_FORMAT_GROUP was not specified:

struct read_format {
u64 value; /* The value of the event */
u64 time_enabled; /* if PERF_FORMAT_TOTAL_TIME_ENABLED */
u64 time_running; /* if PERF_FORMAT_TOTAL_TIME_RUNNING */
u64 id; /* if PERF_FORMAT_ID */

The values read are as follows:

The number of events in this file descriptor. Only available if PERF_FORMAT_GROUP was specified.
time_enabled, time_running
Total time the event was enabled and running. Normally these are the same. If more events are started than available counter slots on the PMU, then multiplexing happens and events run only part of the time. In that case the time_enabled and time running values can be used to scale an estimated value for the count.
An unsigned 64-bit value containing the counter result.
A globally unique value for this particular event, only there if PERF_FORMAT_ID was specified in read_format.

MMAP layout

When using perf_event_open() in sampled mode, asynchronous events (like counter overflow or PROT_EXEC mmap tracking) are logged into a ring-buffer. This ring-buffer is created and accessed through mmap(2).
The mmap size should be 1+2^n pages, where the first page is a metadata page ( struct perf_event_mmap_page) that contains various bits of information such as where the ring-buffer head is.
Before kernel 2.6.39, there is a bug that means you must allocate a mmap ring buffer when sampling even if you do not plan to access it.
The structure of the first metadata mmap page is as follows:

struct perf_event_mmap_page {
__u32 version; /* version number of this structure */
__u32 compat_version; /* lowest version this is compat with */
__u32 lock; /* seqlock for synchronization */
__u32 index; /* hardware counter identifier */
__s64 offset; /* add to hardware counter value */
__u64 time_enabled; /* time event active */
__u64 time_running; /* time event on CPU */
union {
__u64 capabilities;
__u64 cap_usr_time : 1,
cap_usr_rdpmc : 1,
__u16 pmc_width;
__u16 time_shift;
__u32 time_mult;
__u64 time_offset;
__u64 __reserved[120]; /* Pad to 1k */
__u64 data_head; /* head in the data section */
__u64 data_tail; /* user-space written tail */

The following looks at the fields in the perf_event_mmap_page structure in more detail:
Version number of this structure.
The lowest version this is compatible with.
A seqlock for synchronization.
A unique hardware counter identifier.
Add this to hardware counter value??
Time the event was active.
Time the event was running.
User time capability
If the hardware supports user-space read of performance counters without syscall (this is the "rdpmc" instruction on x86), then the following code can be used to do a read:

u32 seq, time_mult, time_shift, idx, width;
u64 count, enabled, running;
u64 cyc, time_offset;
s64 pmc = 0;

do {
seq = pc->lock;
enabled = pc->time_enabled;
running = pc->time_running;

if (pc->cap_usr_time && enabled != running) {
cyc = rdtsc();
time_offset = pc->time_offset;
time_mult = pc->time_mult;
time_shift = pc->time_shift;

idx = pc->index;
count = pc->offset;

if (pc->cap_usr_rdpmc && idx) {
width = pc->pmc_width;
pmc = rdpmc(idx - 1);

} while (pc->lock != seq);

If cap_usr_rdpmc, this field provides the bit-width of the value read using the rdpmc or equivalent instruction. This can be used to sign extend the result like:

pmc <<= 64 - pmc_width;
pmc >>= 64 - pmc_width; // signed shift right
count += pmc;

time_shift, time_mult, time_offset
If cap_usr_time, these fields can be used to compute the time delta since time_enabled (in nanoseconds) using rdtsc or similar.

u64 quot, rem;
u64 delta;
quot = (cyc >> time_shift);
rem = cyc & ((1 << time_shift) - 1);
delta = time_offset + quot * time_mult +
((rem * time_mult) >> time_shift);


Where time_offset, time_mult, time_shift, and cyc are read in the seqcount loop described above. This delta can then be added to enabled and possible running (if idx), improving the scaling:

enabled += delta;
if (idx)
running += delta;
quot = count / running;
rem = count % running;
count = quot * enabled + (rem * enabled) / running;

This points to the head of the data section. The value continuously increases, it does not wrap. The value needs to be manually wrapped by the size of the mmap buffer before accessing the samples.
On SMP-capable platforms, after reading the data_head value, user space should issue an rmb().
When the mapping is PROT_WRITE, the data_tail value should be written by user space to reflect the last read data. In this case the kernel will not over-write unread data.

The following 2^n ring-buffer pages have the layout described below.


If perf_event_attr.sample_id_all is set, then all event types will have the sample_type selected fields related to where/when (identity) an event took place (TID, TIME, ID, CPU, STREAM_ID) described in PERF_RECORD_SAMPLE below, it will be stashed just after the perf_event_header and the fields already present for the existing fields, i.e., at the end of the payload. That way a newer perf.data file will be supported by older perf tools, with these new optional fields being ignored.


The mmap values start with a header:


struct perf_event_header {
__u32 type;
__u16 misc;
__u16 size;


Below, we describe the perf_event_header fields in more detail.

The type value is one of the below. The values in the corresponding record (that follows the header) depend on the type selected as shown.
The MMAP events record the PROT_EXEC mappings so that we can correlate user-space IPs to code. They have the following structure:

struct {
struct perf_event_header header;
u32 pid, tid;
u64 addr;
u64 len;
u64 pgoff;
char filename[];

This record indicates when events are lost.

struct {
struct perf_event_header header;
u64 id;
u64 lost;

is the unique event ID for the samples that were lost.
is the number of events that were lost.
This record indicates a change in the process name.

struct {
struct perf_event_header header;
u32 pid, tid;
char comm[];

This record indicates a process exit event.

struct {
struct perf_event_header header;
u32 pid, ppid;
u32 tid, ptid;
u64 time;

This record indicates a throttle/unthrottle event.

struct {
struct perf_event_header header;
u64 time;
u64 id;
u64 stream_id;

This record indicates a fork event.

struct {
struct perf_event_header header;
u32 pid, ppid;
u32 tid, ptid;
u64 time;

This record indicates a read event.

struct {
struct perf_event_header header;
u32 pid, tid;
struct read_format values;

This record indicates a sample.

struct {
struct perf_event_header header;
u64 ip; /* if PERF_SAMPLE_IP */
u32 pid, tid; /* if PERF_SAMPLE_TID */
u64 time; /* if PERF_SAMPLE_TIME */
u64 addr; /* if PERF_SAMPLE_ADDR */
u64 id; /* if PERF_SAMPLE_ID */
u64 stream_id; /* if PERF_SAMPLE_STREAM_ID */
u32 cpu, res; /* if PERF_SAMPLE_CPU */
u64 period; /* if PERF_SAMPLE_PERIOD */
struct read_format v; /* if PERF_SAMPLE_READ */
u64 nr; /* if PERF_SAMPLE_CALLCHAIN */
u64 ips[nr]; /* if PERF_SAMPLE_CALLCHAIN */
u32 size; /* if PERF_SAMPLE_RAW */
char data[size]; /* if PERF_SAMPLE_RAW */
u64 bnr; /* if PERF_SAMPLE_BRANCH_STACK */
struct perf_branch_entry lbr[bnr];
u64 abi; /* if PERF_SAMPLE_REGS_USER */
u64 regs[weight(mask)];
u64 size; /* if PERF_SAMPLE_STACK_USER */
char data[size]; /* if PERF_SAMPLE_STACK_USER */
u64 dyn_size; /* if PERF_SAMPLE_STACK_USER */
u64 weight; /* if PERF_SAMPLE_WEIGHT */
u64 data_src; /* if PERF_SAMPLE_DATA_SRC */
If PERF_SAMPLE_IP is enabled, then a 64-bit instruction pointer value is included.
pid, tid
If PERF_SAMPLE_TID is enabled, then a 32-bit process ID and 32-bit thread ID are included.
If PERF_SAMPLE_TIME is enabled, then a 64-bit timestamp is included. This is obtained via local_clock() which is a hardware timestamp if available and the jiffies value if not.
If PERF_SAMPLE_ADDR is enabled, then a 64-bit address is included. This is usually the address of a tracepoint, breakpoint, or software event; otherwise the value is 0.
If PERF_SAMPLE_ID is enabled, a 64-bit unique ID is included. If the event is a member of an event group, the group leader ID is returned. This ID is the same as the one returned by PERF_FORMAT_ID.
If PERF_SAMPLE_STREAM_ID is enabled, a 64-bit unique ID is included. Unlike PERF_SAMPLE_ID the actual ID is returned, not the group leader. This ID is the same as the one returned by PERF_FORMAT_ID.
cpu, res
If PERF_SAMPLE_CPU is enabled, this is a 32-bit value indicating which CPU was being used, in addition to a reserved (unused) 32-bit value.
If PERF_SAMPLE_PERIOD is enabled, a 64-bit value indicating the current sampling period is written.
If PERF_SAMPLE_READ is enabled, a structure of type read_format is included which has values for all events in the event group. The values included depend on the read_format value used at perf_event_open() time.
nr, ips[nr]
If PERF_SAMPLE_CALLCHAIN is enabled, then a 64-bit number is included which indicates how many following 64-bit instruction pointers will follow. This is the current callchain.
size, data[size]
If PERF_SAMPLE_RAW is enabled, then a 32-bit value indicating size is included followed by an array of 8-bit values of length size. The values are padded with 0 to have 64-bit alignment.
This RAW record data is opaque with respect to the ABI. The ABI doesn't make any promises with respect to the stability of its content, it may vary depending on event, hardware, and kernel version.
bnr, lbr[bnr]
If PERF_SAMPLE_BRANCH_STACK is enabled, then a 64-bit value indicating the number of records is included, followed by bnr perf_branch_entry structures which each include the fields:
indicating the source instruction (may not be a branch)
the branch target
the branch target was mispredicted
the branch target was predicted.
The entries are from most to least recent, so the first entry has the most recent branch.
Support for mispred and predicted is optional; if not supported, both values will be 0.
abi, regs[weight(mask)]
If PERF_SAMPLE_REGS_USER is enabled, then the user CPU registers are recorded.
The regs field is an array of the CPU registers that were specified by the sample_regs_user attr field. The number of values is the number of bits set in the sample_regs_user bitmask.
size, data[size], dyn_size
If PERF_SAMPLE_STACK_USER is enabled, then record the user stack to enable backtracing. size is the size requested by the user in stack_user_size or else the maximum record size. data is the stack data. dyn_size is the amount of data actually dumped (can be less than size).
If PERF_SAMPLE_WEIGHT is enabled, then a 64 bit value provided by the hardware is recorded that indicates how costly the event was. This allows expensive events to stand out more clearly in profiles.
If PERF_SAMPLE_DATA_SRC is enabled, then a 64 bit value is recorded that is made up of the following fields:
type of opcode, a bitwise combination of PERF_MEM_OP_NA (not available), PERF_MEM_OP_LOAD (load instruction), PERF_MEM_OP_STORE (store instruction), PERF_MEM_OP_PFETCH (prefetch), and PERF_MEM_OP_EXEC (executable code).
memory hierarchy level hit or miss, a bitwise combination of PERF_MEM_LVL_NA (not available), PERF_MEM_LVL_HIT (hit), PERF_MEM_LVL_MISS (miss), PERF_MEM_LVL_L1 (level 1 cache), PERF_MEM_LVL_LFB (line fill buffer), PERF_MEM_LVL_L2 (level 2 cache), PERF_MEM_LVL_L3 (level 3 cache), PERF_MEM_LVL_LOC_RAM (local DRAM), PERF_MEM_LVL_REM_RAM1 (remote DRAM 1 hop), PERF_MEM_LVL_REM_RAM2 (remote DRAM 2 hops), PERF_MEM_LVL_REM_CCE1 (remote cache 1 hop), PERF_MEM_LVL_REM_CCE2 (remote cache 2 hops), PERF_MEM_LVL_IO (I/O memory), and PERF_MEM_LVL_UNC (uncached memory).
snoop mode, a bitwise combination of PERF_MEM_SNOOP_NA (not available), PERF_MEM_SNOOP_NONE (no snoop), PERF_MEM_SNOOP_HIT (snoop hit), PERF_MEM_SNOOP_MISS (snoop miss), and PERF_MEM_SNOOP_HITM (snoop hit modified).
lock instruction, a bitwise combination of PERF_MEM_LOCK_NA (not available) and PERF_MEM_LOCK_LOCKED (locked transaction).
tlb access hit or miss, a bitwise combination of PERF_MEM_TLB_NA (not available), PERF_MEM_TLB_HIT (hit), PERF_MEM_TLB_MISS (miss), PERF_MEM_TLB_L1 (level 1 TLB), PERF_MEM_TLB_L2 (level 2 TLB), PERF_MEM_TLB_WK (hardware walker), and PERF_MEM_TLB_OS (OS fault handler).
The misc field contains additional information about the sample.
The CPU mode can be determined from this value by masking with PERF_RECORD_MISC_CPUMODE_MASK and looking for one of the following (note these are not bit masks, only one can be set at a time):
Unknown CPU mode.
Sample happened in the kernel.
Sample happened in user code.
Sample happened in the hypervisor.
Sample happened in the guest kernel.
Sample happened in guest user code.
In addition, one of the following bits can be set:
This is set when the mapping is not executable; otherwise the mapping is executable.
This indicates that the content of PERF_SAMPLE_IP points to the actual instruction that triggered the event. See also perf_event_attr.precise_ip.
This indicates there is extended data available (currently not used).
This indicates the size of the record.

Signal overflow

Events can be set to deliver a signal when a threshold is crossed. The signal handler is set up using the poll(2), select(2), epoll(2) and fcntl(2), system calls.
To generate signals, sampling must be enabled ( sample_period must have a non-zero value).
There are two ways to generate signals.
The first is to set a wakeup_events or wakeup_watermark value that will generate a signal if a certain number of samples or bytes have been written to the mmap ring buffer. In this case a signal of type POLL_IN is sent.
The other way is by use of the PERF_EVENT_IOC_REFRESH ioctl. This ioctl adds to a counter that decrements each time the event overflows. When non-zero, a POLL_IN signal is sent on overflow, but once the value reaches 0, a signal is sent of type POLL_HUP and the underlying event is disabled.
Note: on newer kernels (definitely noticed with 3.2) a signal is provided for every overflow, even if wakeup_events is not set.

rdpmc instruction

Starting with Linux 3.4 on x86, you can use the rdpmc instruction to get low-latency reads without having to enter the kernel. Note that using rdpmc is not necessarily faster than other methods for reading event values.
Support for this can be detected with the cap_usr_rdpmc field in the mmap page; documentation on how to calculate event values can be found in that section.

perf_event ioctl calls

Various ioctls act on perf_event_open() file descriptors
Enables the individual event or event group specified by the file descriptor argument.
If the PERF_IOC_FLAG_GROUP bit is set in the ioctl argument, then all events in a group are enabled, even if the event specified is not the group leader (but see BUGS).
Disables the individual counter or event group specified by the file descriptor argument.
Enabling or disabling the leader of a group enables or disables the entire group; that is, while the group leader is disabled, none of the counters in the group will count. Enabling or disabling a member of a group other than the leader affects only that counter; disabling a non-leader stops that counter from counting but doesn't affect any other counter.
If the PERF_IOC_FLAG_GROUP bit is set in the ioctl argument, then all events in a group are disabled, even if the event specified is not the group leader (but see BUGS).
Non-inherited overflow counters can use this to enable a counter for a number of overflows specified by the argument, after which it is disabled. Subsequent calls of this ioctl add the argument value to the current count. A signal with POLL_IN set will happen on each overflow until the count reaches 0; when that happens a signal with POLL_HUP set is sent and the event is disabled. Using an argument of 0 is considered undefined behavior.
Reset the event count specified by the file descriptor argument to zero. This resets only the counts; there is no way to reset the multiplexing time_enabled or time_running values.
If the PERF_IOC_FLAG_GROUP bit is set in the ioctl argument, then all events in a group are reset, even if the event specified is not the group leader (but see BUGS).
IOC_PERIOD is the command to update the period; it does not update the current period but instead defers until next.
The argument is a pointer to a 64-bit value containing the desired new period.
This tells the kernel to report event notifications to the specified file descriptor rather than the default one. The file descriptors must all be on the same CPU.
The argument specifies the desired file descriptor, or -1 if output should be ignored.
PERF_EVENT_IOC_SET_FILTER (Since Linux 2.6.33)
This adds an ftrace filter to this event.
The argument is a pointer to the desired ftrace filter.

Using prctl

A process can enable or disable all the event groups that are attached to it using the prctl(2) PR_TASK_PERF_EVENTS_ENABLE and PR_TASK_PERF_EVENTS_DISABLE operations. This applies to all counters on the current process, whether created by this process or by another, and does not affect any counters that this process has created on other processes. It enables or disables only the group leaders, not any other members in the groups.

perf_event related configuration files

Files in /proc/sys/kernel/
The perf_event_paranoid file can be set to restrict access to the performance counters.
2 - only allow user-space measurements
1 - (default) allow both kernel and user measurements
0 - allow access to CPU-specific data but not raw tracepoint samples
-1 - no restrictions
The existence of the perf_event_paranoid file is the official method for determining if a kernel supports perf_event_open().
This sets the maximum sample rate. Setting this too high can allow users to sample at a rate that impacts overall machine performance and potentially lock up the machine. The default value is 100000 (samples per second).
Maximum number of pages an unprivileged user can mlock (2) . The default is 516 (kB).
Files in /sys/bus/event_source/devices/
Since Linux 2.6.34 the kernel supports having multiple PMUs available for monitoring. Information on how to program these PMUs can be found under /sys/bus/event_source/devices/. Each subdirectory corresponds to a different PMU.
/sys/bus/event_source/devices/*/type (Since Linux 2.6.38)
This contains an integer that can be used in the type field of perf_event_attr to indicate you wish to use this PMU.
/sys/bus/event_source/devices/*/rdpmc (Since Linux 3.4)
If this file is 1, then direct user-space access to the performance counter registers is allowed via the rdpmc instruction. This can be disabled by echoing 0 to the file.
/sys/bus/event_source/devices/*/format/ (Since Linux 3.4)
This sub-directory contains information on the architecture-specific sub-fields available for programming the various config fields in the perf_event_attr struct.
The content of each file is the name of the config field, followed by a colon, followed by a series of integer bit ranges separated by commas. For example, the file event may contain the value config1:1,6-10,44 which indicates that event is an attribute that occupies bits 1,6-10, and 44 of perf_event_attr::config1.
/sys/bus/event_source/devices/*/events/ (Since Linux 3.4)
This sub-directory contains files with pre-defined events. The contents are strings describing the event settings expressed in terms of the fields found in the previously mentioned ./format/ directory. These are not necessarily complete lists of all events supported by a PMU, but usually a subset of events deemed useful or interesting.
The content of each file is a list of attribute names separated by commas. Each entry has an optional value (either hex or decimal). If no value is specified than it is assumed to be a single-bit field with a value of 1. An example entry may look like this: event=0x2,inv,ldlat=3
This file is the standard kernel device interface for injecting hotplug events.
/sys/bus/event_source/devices/*/cpumask (Since Linux 3.7)
The cpumask file contains a comma-separated list of integers that indicate a representative cpu number for each socket (package) on the motherboard. This is needed when setting up uncore or northbridge events, as those PMUs present socket-wide events.


perf_event_open() returns the new file descriptor, or -1 if an error occurred (in which case, errno is set appropriately).


Returned if the specified event is not available.
Prior to Linux 3.3, if there was not enough room for the event, ENOSPC was returned. Linus did not like this, and this was changed to EINVAL. ENOSPC is still returned if you try to read results into too small of a buffer.


perf_event_open() was introduced in Linux 2.6.31 but was called perf_counter_open() . It was renamed in Linux 2.6.32.


This perf_event_open() system call Linux- specific and should not be used in programs intended to be portable.


Glibc does not provide a wrapper for this system call; call it using syscall(2). See the example below.
The official way of knowing if perf_event_open() support is enabled is checking for the existence of the file /proc/sys/kernel/perf_event_paranoid.


The F_SETOWN_EX option to fcntl(2) is needed to properly get overflow signals in threads. This was introduced in Linux 2.6.32.
Prior to Linux 2.6.33 (at least for x86) the kernel did not check if events could be scheduled together until read time. The same happens on all known kernels if the NMI watchdog is enabled. This means to see if a given set of events works you have to perf_event_open(), start, then read before you know for sure you can get valid measurements.
Prior to Linux 2.6.34 event constraints were not enforced by the kernel. In that case, some events would silently return "0" if the kernel scheduled them in an improper counter slot.
Prior to Linux 2.6.34 there was a bug when multiplexing where the wrong results could be returned.
Kernels from Linux 2.6.35 to Linux 2.6.39 can quickly crash the kernel if "inherit" is enabled and many threads are started.
Prior to Linux 2.6.35, PERF_FORMAT_GROUP did not work with attached processes.
In older Linux 2.6 versions, refreshing an event group leader refreshed all siblings, and refreshing with a parameter of 0 enabled infinite refresh. This behavior is unsupported and should not be relied on.
There is a bug in the kernel code between Linux 2.6.36 and Linux 3.0 that ignores the "watermark" field and acts as if a wakeup_event was chosen if the union has a non-zero value in it.
From Linux 2.6.31 to Linux 3.4, the PERF_IOC_FLAG_GROUP ioctl argument was broken and would repeatedly operate on the event specified rather than iterating across all sibling events in a group.
Always double-check your results! Various generalized events have had wrong values. For example, retired branches measured the wrong thing on AMD machines until Linux 2.6.35.


The following is a short example that measures the total instruction count of a call to printf(3).

#include <stdlib.h>
#include <stdio.h>
#include <unistd.h>
#include <string.h>
#include <sys/ioctl.h>
#include <linux/perf_event.h>
#include <asm/unistd.h>

perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
int cpu, int group_fd, unsigned long flags)
int ret;

ret = syscall(__NR_perf_event_open, hw_event, pid, cpu,
group_fd, flags);
return ret;

main(int argc, char **argv)
struct perf_event_attr pe;
long long count;
int fd;

memset(&pe, 0, sizeof(struct perf_event_attr));
pe.size = sizeof(struct perf_event_attr);
pe.disabled = 1;
pe.exclude_kernel = 1;
pe.exclude_hv = 1;

fd = perf_event_open(&pe, 0, -1, -1, 0);
if (fd == -1) {
fprintf(stderr, "Error opening leader %llx\n", pe.config);

ioctl(fd, PERF_EVENT_IOC_RESET, 0);
ioctl(fd, PERF_EVENT_IOC_ENABLE, 0);

printf("Measuring instruction count for this printf\n");

read(fd, &count, sizeof(long long));

printf("Used %lld instructions\n", count);



fcntl(2), mmap(2), open(2), prctl(2), read(2)


This page is part of release 3.53 of the Linux man-pages project. A description of the project, and information about reporting bugs, can be found at http://www.kernel.org/doc/man-pages/.
2013-07-16 Linux