2 Performance Counters for Linux
3 ------------------------------
5 Performance counters are special hardware registers available on most modern
6 CPUs. These registers count the number of certain types of hw events: such
7 as instructions executed, cachemisses suffered, or branches mis-predicted -
8 without slowing down the kernel or applications. These registers can also
9 trigger interrupts when a threshold number of events have passed - and can
10 thus be used to profile the code that runs on that CPU.
12 The Linux Performance Counter subsystem provides an abstraction of these
13 hardware capabilities. It provides per task and per CPU counters, counter
14 groups, and it provides event capabilities on top of those. It
15 provides "virtual" 64-bit counters, regardless of the width of the
16 underlying hardware counters.
18 Performance counters are accessed via special file descriptors.
19 There's one file descriptor per virtual counter used.
21 The special file descriptor is opened via the perf_event_open()
24 int sys_perf_event_open(struct perf_event_attr *hw_event_uptr,
25 pid_t pid, int cpu, int group_fd,
28 The syscall returns the new fd. The fd can be used via the normal
29 VFS system calls: read() can be used to read the counter, fcntl()
30 can be used to set the blocking mode, etc.
32 Multiple counters can be kept open at a time, and the counters
35 When creating a new counter fd, 'perf_event_attr' is:
37 struct perf_event_attr {
39 * The MSB of the config word signifies if the rest contains cpu
40 * specific (raw) counter configuration data, if unset, the next
41 * 7 bits are an event type and the rest of the bits are the event
50 __u64 disabled : 1, /* off by default */
51 inherit : 1, /* children inherit it */
52 pinned : 1, /* must always be on PMU */
53 exclusive : 1, /* only group on PMU */
54 exclude_user : 1, /* don't count user */
55 exclude_kernel : 1, /* ditto kernel */
56 exclude_hv : 1, /* ditto hypervisor */
57 exclude_idle : 1, /* don't count when idle */
58 mmap : 1, /* include mmap data */
59 munmap : 1, /* include munmap data */
60 comm : 1, /* include comm data */
64 __u32 extra_config_len;
65 __u32 wakeup_events; /* wakeup every n events */
71 The 'config' field specifies what the counter should count. It
72 is divided into 3 bit-fields:
74 raw_type: 1 bit (most significant bit) 0x8000_0000_0000_0000
75 type: 7 bits (next most significant) 0x7f00_0000_0000_0000
76 event_id: 56 bits (least significant) 0x00ff_ffff_ffff_ffff
78 If 'raw_type' is 1, then the counter will count a hardware event
79 specified by the remaining 63 bits of event_config. The encoding is
82 If 'raw_type' is 0, then the 'type' field says what kind of counter
83 this is, with the following encoding:
85 enum perf_event_types {
86 PERF_TYPE_HARDWARE = 0,
87 PERF_TYPE_SOFTWARE = 1,
88 PERF_TYPE_TRACEPOINT = 2,
91 A counter of PERF_TYPE_HARDWARE will count the hardware event
92 specified by 'event_id':
95 * Generalized performance counter event types, used by the hw_event.event_id
96 * parameter of the sys_perf_event_open() syscall:
100 * Common hardware events, generalized by the kernel:
102 PERF_COUNT_HW_CPU_CYCLES = 0,
103 PERF_COUNT_HW_INSTRUCTIONS = 1,
104 PERF_COUNT_HW_CACHE_REFERENCES = 2,
105 PERF_COUNT_HW_CACHE_MISSES = 3,
106 PERF_COUNT_HW_BRANCH_INSTRUCTIONS = 4,
107 PERF_COUNT_HW_BRANCH_MISSES = 5,
108 PERF_COUNT_HW_BUS_CYCLES = 6,
111 These are standardized types of events that work relatively uniformly
112 on all CPUs that implement Performance Counters support under Linux,
113 although there may be variations (e.g., different CPUs might count
114 cache references and misses at different levels of the cache hierarchy).
115 If a CPU is not able to count the selected event, then the system call
118 More hw_event_types are supported as well, but they are CPU-specific
119 and accessed as raw events. For example, to count "External bus
120 cycles while bus lock signal asserted" events on Intel Core CPUs, pass
121 in a 0x4064 event_id value and set hw_event.raw_type to 1.
123 A counter of type PERF_TYPE_SOFTWARE will count one of the available
124 software events, selected by 'event_id':
127 * Special "software" counters provided by the kernel, even if the hardware
128 * does not support performance counters. These counters measure various
129 * physical and sw events of the kernel (and allow the profiling of them as
133 PERF_COUNT_SW_CPU_CLOCK = 0,
134 PERF_COUNT_SW_TASK_CLOCK = 1,
135 PERF_COUNT_SW_PAGE_FAULTS = 2,
136 PERF_COUNT_SW_CONTEXT_SWITCHES = 3,
137 PERF_COUNT_SW_CPU_MIGRATIONS = 4,
138 PERF_COUNT_SW_PAGE_FAULTS_MIN = 5,
139 PERF_COUNT_SW_PAGE_FAULTS_MAJ = 6,
140 PERF_COUNT_SW_ALIGNMENT_FAULTS = 7,
141 PERF_COUNT_SW_EMULATION_FAULTS = 8,
144 Counters of the type PERF_TYPE_TRACEPOINT are available when the ftrace event
145 tracer is available, and event_id values can be obtained from
146 /debug/tracing/events/*/*/id
149 Counters come in two flavours: counting counters and sampling
150 counters. A "counting" counter is one that is used for counting the
151 number of events that occur, and is characterised by having
155 A read() on a counter returns the current value of the counter and possible
156 additional values as specified by 'read_format', each value is a u64 (8 bytes)
160 * Bits that can be set in hw_event.read_format to request that
161 * reads on the counter should return the indicated quantities,
162 * in increasing order of bit value, after the counter value.
164 enum perf_event_read_format {
165 PERF_FORMAT_TOTAL_TIME_ENABLED = 1,
166 PERF_FORMAT_TOTAL_TIME_RUNNING = 2,
169 Using these additional values one can establish the overcommit ratio for a
170 particular counter allowing one to take the round-robin scheduling effect
174 A "sampling" counter is one that is set up to generate an interrupt
175 every N events, where N is given by 'irq_period'. A sampling counter
176 has irq_period > 0. The record_type controls what data is recorded on each
180 * Bits that can be set in hw_event.record_type to request information
181 * in the overflow packets.
183 enum perf_event_record_format {
184 PERF_RECORD_IP = 1U << 0,
185 PERF_RECORD_TID = 1U << 1,
186 PERF_RECORD_TIME = 1U << 2,
187 PERF_RECORD_ADDR = 1U << 3,
188 PERF_RECORD_GROUP = 1U << 4,
189 PERF_RECORD_CALLCHAIN = 1U << 5,
192 Such (and other) events will be recorded in a ring-buffer, which is
193 available to user-space using mmap() (see below).
195 The 'disabled' bit specifies whether the counter starts out disabled
196 or enabled. If it is initially disabled, it can be enabled by ioctl
197 or prctl (see below).
199 The 'inherit' bit, if set, specifies that this counter should count
200 events on descendant tasks as well as the task specified. This only
201 applies to new descendents, not to any existing descendents at the
202 time the counter is created (nor to any new descendents of existing
205 The 'pinned' bit, if set, specifies that the counter should always be
206 on the CPU if at all possible. It only applies to hardware counters
207 and only to group leaders. If a pinned counter cannot be put onto the
208 CPU (e.g. because there are not enough hardware counters or because of
209 a conflict with some other event), then the counter goes into an
210 'error' state, where reads return end-of-file (i.e. read() returns 0)
211 until the counter is subsequently enabled or disabled.
213 The 'exclusive' bit, if set, specifies that when this counter's group
214 is on the CPU, it should be the only group using the CPU's counters.
215 In future, this will allow sophisticated monitoring programs to supply
216 extra configuration information via 'extra_config_len' to exploit
217 advanced features of the CPU's Performance Monitor Unit (PMU) that are
218 not otherwise accessible and that might disrupt other hardware
221 The 'exclude_user', 'exclude_kernel' and 'exclude_hv' bits provide a
222 way to request that counting of events be restricted to times when the
223 CPU is in user, kernel and/or hypervisor mode.
225 The 'mmap' and 'munmap' bits allow recording of PROT_EXEC mmap/munmap
226 operations, these can be used to relate userspace IP addresses to actual
227 code, even after the mapping (or even the whole process) is gone,
228 these events are recorded in the ring-buffer (see below).
230 The 'comm' bit allows tracking of process comm data on process creation.
231 This too is recorded in the ring-buffer (see below).
233 The 'pid' parameter to the perf_event_open() system call allows the
234 counter to be specific to a task:
236 pid == 0: if the pid parameter is zero, the counter is attached to the
239 pid > 0: the counter is attached to a specific task (if the current task
240 has sufficient privilege to do so)
242 pid < 0: all tasks are counted (per cpu counters)
244 The 'cpu' parameter allows a counter to be made specific to a CPU:
246 cpu >= 0: the counter is restricted to a specific CPU
247 cpu == -1: the counter counts on all CPUs
249 (Note: the combination of 'pid == -1' and 'cpu == -1' is not valid.)
251 A 'pid > 0' and 'cpu == -1' counter is a per task counter that counts
252 events of that task and 'follows' that task to whatever CPU the task
253 gets schedule to. Per task counters can be created by any user, for
256 A 'pid == -1' and 'cpu == x' counter is a per CPU counter that counts
257 all events on CPU-x. Per CPU counters need CAP_SYS_ADMIN privilege.
259 The 'flags' parameter is currently unused and must be zero.
261 The 'group_fd' parameter allows counter "groups" to be set up. A
262 counter group has one counter which is the group "leader". The leader
263 is created first, with group_fd = -1 in the perf_event_open call
264 that creates it. The rest of the group members are created
265 subsequently, with group_fd giving the fd of the group leader.
266 (A single counter on its own is created with group_fd = -1 and is
267 considered to be a group with only 1 member.)
269 A counter group is scheduled onto the CPU as a unit, that is, it will
270 only be put onto the CPU if all of the counters in the group can be
271 put onto the CPU. This means that the values of the member counters
272 can be meaningfully compared, added, divided (to get ratios), etc.,
273 with each other, since they have counted events for the same set of
274 executed instructions.
277 Like stated, asynchronous events, like counter overflow or PROT_EXEC mmap
278 tracking are logged into a ring-buffer. This ring-buffer is created and
279 accessed through mmap().
281 The mmap size should be 1+2^n pages, where the first page is a meta-data page
282 (struct perf_event_mmap_page) that contains various bits of information such
283 as where the ring-buffer head is.
286 * Structure of the page that can be mapped via mmap
288 struct perf_event_mmap_page {
289 __u32 version; /* version number of this structure */
290 __u32 compat_version; /* lowest version this is compat with */
293 * Bits needed to read the hw counters in user-space.
303 * count = pmc_read(pc->index - 1);
304 * count += pc->offset;
309 * } while (pc->lock != seq);
311 * NOTE: for obvious reason this only works on self-monitoring
314 __u32 lock; /* seqlock for synchronization */
315 __u32 index; /* hardware counter identifier */
316 __s64 offset; /* add to hardware counter value */
319 * Control data for the mmap() data buffer.
321 * User-space reading this value should issue an rmb(), on SMP capable
322 * platforms, after reading this value -- see perf_event_wakeup().
324 __u32 data_head; /* head in the data section */
327 NOTE: the hw-counter userspace bits are arch specific and are currently only
328 implemented on powerpc.
330 The following 2^n pages are the ring-buffer which contains events of the form:
332 #define PERF_RECORD_MISC_KERNEL (1 << 0)
333 #define PERF_RECORD_MISC_USER (1 << 1)
334 #define PERF_RECORD_MISC_OVERFLOW (1 << 2)
336 struct perf_event_header {
342 enum perf_event_type {
345 * The MMAP events record the PROT_EXEC mappings so that we can
346 * correlate userspace IPs to code. They have the following structure:
349 * struct perf_event_header header;
358 PERF_RECORD_MMAP = 1,
359 PERF_RECORD_MUNMAP = 2,
363 * struct perf_event_header header;
369 PERF_RECORD_COMM = 3,
372 * When header.misc & PERF_RECORD_MISC_OVERFLOW the event_type field
373 * will be PERF_RECORD_*
376 * struct perf_event_header header;
378 * { u64 ip; } && PERF_RECORD_IP
379 * { u32 pid, tid; } && PERF_RECORD_TID
380 * { u64 time; } && PERF_RECORD_TIME
381 * { u64 addr; } && PERF_RECORD_ADDR
384 * { u64 event, val; } cnt[nr]; } && PERF_RECORD_GROUP
390 * u64 ips[nr]; } && PERF_RECORD_CALLCHAIN
395 NOTE: PERF_RECORD_CALLCHAIN is arch specific and currently only implemented
398 Notification of new events is possible through poll()/select()/epoll() and
399 fcntl() managing signals.
401 Normally a notification is generated for every page filled, however one can
402 additionally set perf_event_attr.wakeup_events to generate one every
403 so many counter overflow events.
405 Future work will include a splice() interface to the ring-buffer.
408 Counters can be enabled and disabled in two ways: via ioctl and via
409 prctl. When a counter is disabled, it doesn't count or generate
410 events but does continue to exist and maintain its count value.
412 An individual counter can be enabled with
414 ioctl(fd, PERF_EVENT_IOC_ENABLE, 0);
418 ioctl(fd, PERF_EVENT_IOC_DISABLE, 0);
420 For a counter group, pass PERF_IOC_FLAG_GROUP as the third argument.
421 Enabling or disabling the leader of a group enables or disables the
422 whole group; that is, while the group leader is disabled, none of the
423 counters in the group will count. Enabling or disabling a member of a
424 group other than the leader only affects that counter - disabling an
425 non-leader stops that counter from counting but doesn't affect any
428 Additionally, non-inherited overflow counters can use
430 ioctl(fd, PERF_EVENT_IOC_REFRESH, nr);
432 to enable a counter for 'nr' events, after which it gets disabled again.
434 A process can enable or disable all the counter groups that are
435 attached to it, using prctl:
437 prctl(PR_TASK_PERF_EVENTS_ENABLE);
439 prctl(PR_TASK_PERF_EVENTS_DISABLE);
441 This applies to all counters on the current process, whether created
442 by this process or by another, and doesn't affect any counters that
443 this process has created on other processes. It only enables or
444 disables the group leaders, not any other members in the groups.
450 If your architecture does not have hardware performance metrics, you can
451 still use the generic software counters based on hrtimers for sampling.
453 So to start with, in order to add HAVE_PERF_EVENTS to your Kconfig, you
454 will need at least this:
455 - asm/perf_event.h - a basic stub will suffice at first
456 - support for atomic64 types (and associated helper functions)
457 - set_perf_event_pending() implemented
459 If your architecture does have hardware capabilities, you can override the
460 weak stub hw_perf_event_init() to register hardware counters.
462 Architectures that have d-cache aliassing issues, such as Sparc and ARM,
463 should select PERF_USE_VMALLOC in order to avoid these for perf mmap().