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 For each of the optional features a perf_file_section it placed after the data
47 section if the feature bit is set in the perf_header flags bitset. The
48 respective perf_file_section points to the data of the additional header and
51 Some headers consist of strings, which are defined like this:
53 struct perf_header_string {
55 char string[len]; /* zero terminated */
58 Some headers consist of a sequence of strings, which start with a
60 struct perf_header_string_list {
62 struct perf_header_string strings[nr]; /* variable length records */
65 The bits are the flags bits in a 256 bit bitmap starting with
66 flags. These define the valid bits:
68 HEADER_RESERVED = 0, /* always cleared */
69 HEADER_FIRST_FEATURE = 1,
70 HEADER_TRACING_DATA = 1,
76 The header consists of an sequence of build_id_event. The size of each record
77 is defined by header.size (see perf_event.h). Each event defines a ELF build id
78 for a executable file name for a pid. An ELF build id is a unique identifier
79 assigned by the linker to an executable.
81 struct build_id_event {
82 struct perf_event_header header;
85 char filename[header.size - offsetof(struct build_id_event, filename)];
90 A perf_header_string with the hostname where the data was collected
95 A perf_header_string with the os release where the data was collected
100 A perf_header_string with the perf user tool version where the
101 data was collected. This is the same as the version of the source tree
102 the perf tool was built from.
106 A perf_header_string with the CPU architecture (uname -m)
110 A structure defining the number of CPUs.
113 uint32_t nr_cpus_available; /* CPUs not yet onlined */
114 uint32_t nr_cpus_online;
119 A perf_header_string with description of the CPU. On x86 this is the model name
124 A perf_header_string with the exact CPU type. On x86 this is
125 vendor,family,model,stepping. For example: GenuineIntel,6,69,1
127 HEADER_TOTAL_MEM = 10,
129 An uint64_t with the total memory in bytes.
133 A perf_header_string_list with the perf arg-vector used to collect the data.
135 HEADER_EVENT_DESC = 12,
137 Another description of the perf_event_attrs, more detailed than header.attrs
138 including IDs and names. See perf_event.h or the man page for a description
139 of a struct perf_event_attr.
142 uint32_t nr; /* number of events */
143 uint32_t attr_size; /* size of each perf_event_attr */
145 struct perf_event_attr attr; /* size of attr_size */
147 struct perf_header_string event_string;
148 uint64_t ids[nr_ids];
149 } events[nr]; /* Variable length records */
152 HEADER_CPU_TOPOLOGY = 13,
154 String lists defining the core and CPU threads topology.
155 The string lists are followed by a variable length array
156 which contains core_id and socket_id of each cpu.
157 The number of entries can be determined by the size of the
158 section minus the sizes of both string lists.
161 struct perf_header_string_list cores; /* Variable length */
162 struct perf_header_string_list threads; /* Variable length */
166 } cpus[nr]; /* Variable length records */
171 sibling threads : 0-1
172 sibling threads : 2-3
174 HEADER_NUMA_TOPOLOGY = 14,
176 A list of NUMA node descriptions
184 struct perf_header_string cpus;
185 } nodes[nr]; /* Variable length records */
188 HEADER_BRANCH_STACK = 15,
190 Not implemented in perf.
192 HEADER_PMU_MAPPINGS = 16,
194 A list of PMU structures, defining the different PMUs supported by perf.
200 struct perf_header_string pmu_name;
201 } [nr]; /* Variable length records */
204 HEADER_GROUP_DESC = 17,
206 Description of counter groups ({...} in perf syntax)
211 struct perf_header_string string;
214 } [nr]; /* Variable length records */
217 HEADER_AUXTRACE = 18,
219 Define additional auxtrace areas in the perf.data. auxtrace is used to store
220 undecoded hardware tracing information, such as Intel Processor Trace data.
223 * struct auxtrace_index_entry - indexes a AUX area tracing event within a
225 * @file_offset: offset within the perf.data file
226 * @sz: size of the event
228 struct auxtrace_index_entry {
233 #define PERF_AUXTRACE_INDEX_ENTRY_COUNT 256
236 * struct auxtrace_index - index of AUX area tracing events within a perf.data
238 * @list: linking a number of arrays of entries
239 * @nr: number of entries
240 * @entries: array of entries
242 struct auxtrace_index {
243 struct list_head list;
245 struct auxtrace_index_entry entries[PERF_AUXTRACE_INDEX_ENTRY_COUNT];
250 This is merely a flag signifying that the data section contains data
251 recorded from perf stat record.
255 Description of the cache hierarchy. Based on the Linux sysfs format
256 in /sys/devices/system/cpu/cpu*/cache/
258 u32 version Currently always 1
259 u32 number_of_cache_levels
266 struct perf_header_string type;
267 struct perf_header_string size;
268 struct perf_header_string map;
269 }[number_of_cache_levels];
271 HEADER_SAMPLE_TIME = 21,
273 Two uint64_t for the time of first sample and the time of last sample.
275 HEADER_COMPRESSED = 27,
285 Indicates that trace contains records of PERF_RECORD_COMPRESSED type
286 that have perf_events records in compressed form.
288 other bits are reserved and should ignored for now
289 HEADER_FEAT_BITS = 256,
293 This is an array of perf_event_attrs, each attr_size bytes long, which defines
294 each event collected. See perf_event.h or the man page for a detailed
299 This section is the bulk of the file. It consist of a stream of perf_events
300 describing events. This matches the format generated by the kernel.
301 See perf_event.h or the manpage for a detailed description.
303 Some notes on parsing:
307 The events are not necessarily in time stamp order, as they can be
308 collected in parallel on different CPUs. If the events should be
309 processed in time order they need to be sorted first. It is possible
310 to only do a partial sort using the FINISHED_ROUND event header (see
311 below). perf record guarantees that there is no reordering over a
316 When the event stream contains multiple events each event is identified
317 by an ID. This can be either through the PERF_SAMPLE_ID or the
318 PERF_SAMPLE_IDENTIFIER header. The PERF_SAMPLE_IDENTIFIER header is
319 at a fixed offset from the event header, which allows reliable
320 parsing of the header. Relying on ID may be ambiguous.
321 IDENTIFIER is only supported by newer Linux kernels.
323 Perf record specific events:
325 In addition to the kernel generated event types perf record adds its
326 own event types (in addition it also synthesizes some kernel events,
327 for example MMAP events)
329 PERF_RECORD_USER_TYPE_START = 64,
330 PERF_RECORD_HEADER_ATTR = 64,
333 struct perf_event_header header;
334 struct perf_event_attr attr;
338 PERF_RECORD_HEADER_EVENT_TYPE = 65, /* deprecated */
340 #define MAX_EVENT_NAME 64
342 struct perf_trace_event_type {
344 char name[MAX_EVENT_NAME];
347 struct event_type_event {
348 struct perf_event_header header;
349 struct perf_trace_event_type event_type;
353 PERF_RECORD_HEADER_TRACING_DATA = 66,
357 struct tracing_data_event {
358 struct perf_event_header header;
362 PERF_RECORD_HEADER_BUILD_ID = 67,
364 Define a ELF build ID for a referenced executable.
366 struct build_id_event; /* See above */
368 PERF_RECORD_FINISHED_ROUND = 68,
370 No event reordering over this header. No payload.
372 PERF_RECORD_ID_INDEX = 69,
374 Map event ids to CPUs and TIDs.
376 struct id_index_entry {
383 struct id_index_event {
384 struct perf_event_header header;
386 struct id_index_entry entries[nr];
389 PERF_RECORD_AUXTRACE_INFO = 70,
391 Auxtrace type specific information. Describe me
393 struct auxtrace_info_event {
394 struct perf_event_header header;
396 uint32_t reserved__; /* For alignment */
400 PERF_RECORD_AUXTRACE = 71,
402 Defines auxtrace data. Followed by the actual data. The contents of
403 the auxtrace data is dependent on the event and the CPU. For example
404 for Intel Processor Trace it contains Processor Trace data generated
407 struct auxtrace_event {
408 struct perf_event_header header;
415 uint32_t reserved__; /* For alignment */
419 struct perf_event_header header;
425 PERF_RECORD_AUXTRACE_ERROR = 72,
427 Describes an error in hardware tracing
429 enum auxtrace_error_type {
430 PERF_AUXTRACE_ERROR_ITRACE = 1,
431 PERF_AUXTRACE_ERROR_MAX
434 #define MAX_AUXTRACE_ERROR_MSG 64
436 struct auxtrace_error_event {
437 struct perf_event_header header;
443 uint32_t reserved__; /* For alignment */
445 char msg[MAX_AUXTRACE_ERROR_MSG];
448 PERF_RECORD_HEADER_FEATURE = 80,
450 Describes a header feature. These are records used in pipe-mode that
451 contain information that otherwise would be in perf.data file's header.
453 PERF_RECORD_COMPRESSED = 81,
455 struct compressed_event {
456 struct perf_event_header header;
460 The header is followed by compressed data frame that can be decompressed
461 into array of perf trace records. The size of the entire compressed event
462 record including the header is limited by the max value of header.size.
466 Define the event attributes with their IDs.
468 An array bound by the perf_file_section size.
471 struct perf_event_attr attr; /* Size defined by header.attr_size */
472 struct perf_file_section ids;
475 ids points to a array of uint64_t defining the ids for event attr attr.
479 Pipe-mode avoid seeks in the file by removing the perf_file_section and flags
480 from the struct perf_header. The trimmed header is:
482 struct perf_pipe_file_header {
487 The information about attrs, data, and event_types is instead in the
488 synthesized events PERF_RECORD_ATTR, PERF_RECORD_HEADER_TRACING_DATA,
489 PERF_RECORD_HEADER_EVENT_TYPE, and PERF_RECORD_HEADER_FEATURE
490 that are generated by perf record in pipe-mode.
495 include/uapi/linux/perf_event.h
497 This is the canonical description of the kernel generated perf_events
498 and the perf_event_attrs.
502 A manpage describing perf_event and perf_event_attr is here:
503 http://web.eece.maine.edu/~vweaver/projects/perf_events/programming.html
504 This tends to be slightly behind the kernel include, but has better
505 descriptions. An (typically older) version of the man page may be
506 included with the standard Linux man pages, available with "man
511 https://github.com/andikleen/pmu-tools/tree/master/parser
513 A definition of the perf.data format in python "construct" format is available
514 in pmu-tools parser. This allows to read perf.data from python and dump it.
518 The quipper C++ parser is available at
519 http://github.com/google/perf_data_converter/tree/master/src/quipper