5 This document describes the on-disk perf.data format, generated by perf record
6 or perf inject and consumed by the other perf tools.
8 On a high level perf.data contains the events generated by the PMUs, plus metadata.
10 All fields are in native-endian of the machine that generated the perf.data.
12 When perf is writing to a pipe it uses a special version of the file
13 format that does not rely on seeking to adjust data offsets. This
14 format is described in "Pipe-mode data" section. The pipe data version can be
15 augmented with additional events using perf inject.
17 The file starts with a perf_header:
20 char magic[8]; /* PERFILE2 */
21 uint64_t size; /* size of the header */
22 uint64_t attr_size; /* size of an attribute in attrs */
23 struct perf_file_section attrs;
24 struct perf_file_section data;
25 struct perf_file_section event_types;
30 The magic number identifies the perf file and the version. Current perf versions
31 use PERFILE2. Old perf versions generated a version 1 format (PERFFILE). Version 1
32 is not described here. The magic number also identifies the endian. When the
33 magic value is 64bit byte swapped compared the file is in non-native
36 A perf_file_section contains a pointer to another section of the perf file.
37 The header contains three such pointers: for attributes, data and event types.
39 struct perf_file_section {
40 uint64_t offset; /* offset from start of file */
41 uint64_t size; /* size of the section */
46 The header is followed by different optional headers, described by the bits set
47 in flags. Only headers for which the bit is set are included. Each header
48 consists of a perf_file_section located after the initial header.
49 The respective perf_file_section points to the data of the additional
50 header and defines its size.
52 Some headers consist of strings, which are defined like this:
54 struct perf_header_string {
56 char string[len]; /* zero terminated */
59 Some headers consist of a sequence of strings, which start with a
61 struct perf_header_string_list {
63 struct perf_header_string strings[nr]; /* variable length records */
66 The bits are the flags bits in a 256 bit bitmap starting with
67 flags. These define the valid bits:
69 HEADER_RESERVED = 0, /* always cleared */
70 HEADER_FIRST_FEATURE = 1,
71 HEADER_TRACING_DATA = 1,
77 The header consists of an sequence of build_id_event. The size of each record
78 is defined by header.size (see perf_event.h). Each event defines a ELF build id
79 for a executable file name for a pid. An ELF build id is a unique identifier
80 assigned by the linker to an executable.
82 struct build_id_event {
83 struct perf_event_header header;
86 char filename[header.size - offsetof(struct build_id_event, filename)];
91 A perf_header_string with the hostname where the data was collected
96 A perf_header_string with the os release where the data was collected
101 A perf_header_string with the perf user tool version where the
102 data was collected. This is the same as the version of the source tree
103 the perf tool was built from.
107 A perf_header_string with the CPU architecture (uname -m)
111 A structure defining the number of CPUs.
114 uint32_t nr_cpus_available; /* CPUs not yet onlined */
115 uint32_t nr_cpus_online;
120 A perf_header_string with description of the CPU. On x86 this is the model name
125 A perf_header_string with the exact CPU type. On x86 this is
126 vendor,family,model,stepping. For example: GenuineIntel,6,69,1
128 HEADER_TOTAL_MEM = 10,
130 An uint64_t with the total memory in bytes.
134 A perf_header_string with the perf command line used to collect the data.
136 HEADER_EVENT_DESC = 12,
138 Another description of the perf_event_attrs, more detailed than header.attrs
139 including IDs and names. See perf_event.h or the man page for a description
140 of a struct perf_event_attr.
143 uint32_t nr; /* number of events */
144 uint32_t attr_size; /* size of each perf_event_attr */
146 struct perf_event_attr attr; /* size of attr_size */
148 struct perf_header_string event_string;
149 uint64_t ids[nr_ids];
150 } events[nr]; /* Variable length records */
153 HEADER_CPU_TOPOLOGY = 13,
155 String lists defining the core and CPU threads topology.
156 The string lists are followed by a variable length array
157 which contains core_id and socket_id of each cpu.
158 The number of entries can be determined by the size of the
159 section minus the sizes of both string lists.
162 struct perf_header_string_list cores; /* Variable length */
163 struct perf_header_string_list threads; /* Variable length */
167 } cpus[nr]; /* Variable length records */
172 sibling threads : 0-1
173 sibling threads : 2-3
175 HEADER_NUMA_TOPOLOGY = 14,
177 A list of NUMA node descriptions
185 struct perf_header_string cpus;
186 } nodes[nr]; /* Variable length records */
189 HEADER_BRANCH_STACK = 15,
191 Not implemented in perf.
193 HEADER_PMU_MAPPINGS = 16,
195 A list of PMU structures, defining the different PMUs supported by perf.
201 struct perf_header_string pmu_name;
202 } [nr]; /* Variable length records */
205 HEADER_GROUP_DESC = 17,
207 Description of counter groups ({...} in perf syntax)
212 struct perf_header_string string;
215 } [nr]; /* Variable length records */
218 HEADER_AUXTRACE = 18,
220 Define additional auxtrace areas in the perf.data. auxtrace is used to store
221 undecoded hardware tracing information, such as Intel Processor Trace data.
224 * struct auxtrace_index_entry - indexes a AUX area tracing event within a
226 * @file_offset: offset within the perf.data file
227 * @sz: size of the event
229 struct auxtrace_index_entry {
234 #define PERF_AUXTRACE_INDEX_ENTRY_COUNT 256
237 * struct auxtrace_index - index of AUX area tracing events within a perf.data
239 * @list: linking a number of arrays of entries
240 * @nr: number of entries
241 * @entries: array of entries
243 struct auxtrace_index {
244 struct list_head list;
246 struct auxtrace_index_entry entries[PERF_AUXTRACE_INDEX_ENTRY_COUNT];
251 This is merely a flag signifying that the data section contains data
252 recorded from perf stat record.
256 Description of the cache hierarchy. Based on the Linux sysfs format
257 in /sys/devices/system/cpu/cpu*/cache/
259 u32 version Currently always 1
260 u32 number_of_cache_levels
267 struct perf_header_string type;
268 struct perf_header_string size;
269 struct perf_header_string map;
270 }[number_of_cache_levels];
272 HEADER_SAMPLE_TIME = 21,
274 Two uint64_t for the time of first sample and the time of last sample.
276 other bits are reserved and should ignored for now
277 HEADER_FEAT_BITS = 256,
281 This is an array of perf_event_attrs, each attr_size bytes long, which defines
282 each event collected. See perf_event.h or the man page for a detailed
287 This section is the bulk of the file. It consist of a stream of perf_events
288 describing events. This matches the format generated by the kernel.
289 See perf_event.h or the manpage for a detailed description.
291 Some notes on parsing:
295 The events are not necessarily in time stamp order, as they can be
296 collected in parallel on different CPUs. If the events should be
297 processed in time order they need to be sorted first. It is possible
298 to only do a partial sort using the FINISHED_ROUND event header (see
299 below). perf record guarantees that there is no reordering over a
304 When the event stream contains multiple events each event is identified
305 by an ID. This can be either through the PERF_SAMPLE_ID or the
306 PERF_SAMPLE_IDENTIFIER header. The PERF_SAMPLE_IDENTIFIER header is
307 at a fixed offset from the event header, which allows reliable
308 parsing of the header. Relying on ID may be ambiguous.
309 IDENTIFIER is only supported by newer Linux kernels.
311 Perf record specific events:
313 In addition to the kernel generated event types perf record adds its
314 own event types (in addition it also synthesizes some kernel events,
315 for example MMAP events)
317 PERF_RECORD_USER_TYPE_START = 64,
318 PERF_RECORD_HEADER_ATTR = 64,
321 struct perf_event_header header;
322 struct perf_event_attr attr;
326 PERF_RECORD_HEADER_EVENT_TYPE = 65, /* deprecated */
328 #define MAX_EVENT_NAME 64
330 struct perf_trace_event_type {
332 char name[MAX_EVENT_NAME];
335 struct event_type_event {
336 struct perf_event_header header;
337 struct perf_trace_event_type event_type;
341 PERF_RECORD_HEADER_TRACING_DATA = 66,
345 struct tracing_data_event {
346 struct perf_event_header header;
350 PERF_RECORD_HEADER_BUILD_ID = 67,
352 Define a ELF build ID for a referenced executable.
354 struct build_id_event; /* See above */
356 PERF_RECORD_FINISHED_ROUND = 68,
358 No event reordering over this header. No payload.
360 PERF_RECORD_ID_INDEX = 69,
362 Map event ids to CPUs and TIDs.
364 struct id_index_entry {
371 struct id_index_event {
372 struct perf_event_header header;
374 struct id_index_entry entries[nr];
377 PERF_RECORD_AUXTRACE_INFO = 70,
379 Auxtrace type specific information. Describe me
381 struct auxtrace_info_event {
382 struct perf_event_header header;
384 uint32_t reserved__; /* For alignment */
388 PERF_RECORD_AUXTRACE = 71,
390 Defines auxtrace data. Followed by the actual data. The contents of
391 the auxtrace data is dependent on the event and the CPU. For example
392 for Intel Processor Trace it contains Processor Trace data generated
395 struct auxtrace_event {
396 struct perf_event_header header;
403 uint32_t reserved__; /* For alignment */
407 struct perf_event_header header;
413 PERF_RECORD_AUXTRACE_ERROR = 72,
415 Describes an error in hardware tracing
417 enum auxtrace_error_type {
418 PERF_AUXTRACE_ERROR_ITRACE = 1,
419 PERF_AUXTRACE_ERROR_MAX
422 #define MAX_AUXTRACE_ERROR_MSG 64
424 struct auxtrace_error_event {
425 struct perf_event_header header;
431 uint32_t reserved__; /* For alignment */
433 char msg[MAX_AUXTRACE_ERROR_MSG];
436 PERF_RECORD_HEADER_FEATURE = 80,
438 Describes a header feature. These are records used in pipe-mode that
439 contain information that otherwise would be in perf.data file's header.
443 Define the event attributes with their IDs.
445 An array bound by the perf_file_section size.
448 struct perf_event_attr attr; /* Size defined by header.attr_size */
449 struct perf_file_section ids;
452 ids points to a array of uint64_t defining the ids for event attr attr.
456 Pipe-mode avoid seeks in the file by removing the perf_file_section and flags
457 from the struct perf_header. The trimmed header is:
459 struct perf_pipe_file_header {
464 The information about attrs, data, and event_types is instead in the
465 synthesized events PERF_RECORD_ATTR, PERF_RECORD_HEADER_TRACING_DATA,
466 PERF_RECORD_HEADER_EVENT_TYPE, and PERF_RECORD_HEADER_FEATURE
467 that are generated by perf record in pipe-mode.
472 include/uapi/linux/perf_event.h
474 This is the canonical description of the kernel generated perf_events
475 and the perf_event_attrs.
479 A manpage describing perf_event and perf_event_attr is here:
480 http://web.eece.maine.edu/~vweaver/projects/perf_events/programming.html
481 This tends to be slightly behind the kernel include, but has better
482 descriptions. An (typically older) version of the man page may be
483 included with the standard Linux man pages, available with "man
488 https://github.com/andikleen/pmu-tools/tree/master/parser
490 A definition of the perf.data format in python "construct" format is available
491 in pmu-tools parser. This allows to read perf.data from python and dump it.
495 The quipper C++ parser is available at
496 http://github.com/google/perf_data_converter/tree/master/src/quipper