4 * Copyright (C) 2008-2009, Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009, Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009, Red Hat, Inc., Peter Zijlstra
8 * Data type definitions, declarations, prototypes.
10 * Started by: Thomas Gleixner and Ingo Molnar
12 * For licencing details see kernel-base/COPYING
14 #ifndef _LINUX_PERF_EVENT_H
15 #define _LINUX_PERF_EVENT_H
17 #include <linux/types.h>
18 #include <linux/ioctl.h>
19 #include <asm/byteorder.h>
22 * User-space ABI bits:
29 PERF_TYPE_HARDWARE
= 0,
30 PERF_TYPE_SOFTWARE
= 1,
31 PERF_TYPE_TRACEPOINT
= 2,
32 PERF_TYPE_HW_CACHE
= 3,
35 PERF_TYPE_MAX
, /* non-ABI */
39 * Generalized performance event event_id types, used by the
40 * attr.event_id parameter of the sys_perf_event_open()
45 * Common hardware events, generalized by the kernel:
47 PERF_COUNT_HW_CPU_CYCLES
= 0,
48 PERF_COUNT_HW_INSTRUCTIONS
= 1,
49 PERF_COUNT_HW_CACHE_REFERENCES
= 2,
50 PERF_COUNT_HW_CACHE_MISSES
= 3,
51 PERF_COUNT_HW_BRANCH_INSTRUCTIONS
= 4,
52 PERF_COUNT_HW_BRANCH_MISSES
= 5,
53 PERF_COUNT_HW_BUS_CYCLES
= 6,
55 PERF_COUNT_HW_MAX
, /* non-ABI */
59 * Generalized hardware cache events:
61 * { L1-D, L1-I, LLC, ITLB, DTLB, BPU } x
62 * { read, write, prefetch } x
63 * { accesses, misses }
65 enum perf_hw_cache_id
{
66 PERF_COUNT_HW_CACHE_L1D
= 0,
67 PERF_COUNT_HW_CACHE_L1I
= 1,
68 PERF_COUNT_HW_CACHE_LL
= 2,
69 PERF_COUNT_HW_CACHE_DTLB
= 3,
70 PERF_COUNT_HW_CACHE_ITLB
= 4,
71 PERF_COUNT_HW_CACHE_BPU
= 5,
73 PERF_COUNT_HW_CACHE_MAX
, /* non-ABI */
76 enum perf_hw_cache_op_id
{
77 PERF_COUNT_HW_CACHE_OP_READ
= 0,
78 PERF_COUNT_HW_CACHE_OP_WRITE
= 1,
79 PERF_COUNT_HW_CACHE_OP_PREFETCH
= 2,
81 PERF_COUNT_HW_CACHE_OP_MAX
, /* non-ABI */
84 enum perf_hw_cache_op_result_id
{
85 PERF_COUNT_HW_CACHE_RESULT_ACCESS
= 0,
86 PERF_COUNT_HW_CACHE_RESULT_MISS
= 1,
88 PERF_COUNT_HW_CACHE_RESULT_MAX
, /* non-ABI */
92 * Special "software" events provided by the kernel, even if the hardware
93 * does not support performance events. These events measure various
94 * physical and sw events of the kernel (and allow the profiling of them as
98 PERF_COUNT_SW_CPU_CLOCK
= 0,
99 PERF_COUNT_SW_TASK_CLOCK
= 1,
100 PERF_COUNT_SW_PAGE_FAULTS
= 2,
101 PERF_COUNT_SW_CONTEXT_SWITCHES
= 3,
102 PERF_COUNT_SW_CPU_MIGRATIONS
= 4,
103 PERF_COUNT_SW_PAGE_FAULTS_MIN
= 5,
104 PERF_COUNT_SW_PAGE_FAULTS_MAJ
= 6,
106 PERF_COUNT_SW_MAX
, /* non-ABI */
110 * Bits that can be set in attr.sample_type to request information
111 * in the overflow packets.
113 enum perf_event_sample_format
{
114 PERF_SAMPLE_IP
= 1U << 0,
115 PERF_SAMPLE_TID
= 1U << 1,
116 PERF_SAMPLE_TIME
= 1U << 2,
117 PERF_SAMPLE_ADDR
= 1U << 3,
118 PERF_SAMPLE_READ
= 1U << 4,
119 PERF_SAMPLE_CALLCHAIN
= 1U << 5,
120 PERF_SAMPLE_ID
= 1U << 6,
121 PERF_SAMPLE_CPU
= 1U << 7,
122 PERF_SAMPLE_PERIOD
= 1U << 8,
123 PERF_SAMPLE_STREAM_ID
= 1U << 9,
124 PERF_SAMPLE_RAW
= 1U << 10,
126 PERF_SAMPLE_MAX
= 1U << 11, /* non-ABI */
130 * The format of the data returned by read() on a perf event fd,
131 * as specified by attr.read_format:
133 * struct read_format {
135 * { u64 time_enabled; } && PERF_FORMAT_ENABLED
136 * { u64 time_running; } && PERF_FORMAT_RUNNING
137 * { u64 id; } && PERF_FORMAT_ID
138 * } && !PERF_FORMAT_GROUP
141 * { u64 time_enabled; } && PERF_FORMAT_ENABLED
142 * { u64 time_running; } && PERF_FORMAT_RUNNING
144 * { u64 id; } && PERF_FORMAT_ID
146 * } && PERF_FORMAT_GROUP
149 enum perf_event_read_format
{
150 PERF_FORMAT_TOTAL_TIME_ENABLED
= 1U << 0,
151 PERF_FORMAT_TOTAL_TIME_RUNNING
= 1U << 1,
152 PERF_FORMAT_ID
= 1U << 2,
153 PERF_FORMAT_GROUP
= 1U << 3,
155 PERF_FORMAT_MAX
= 1U << 4, /* non-ABI */
158 #define PERF_ATTR_SIZE_VER0 64 /* sizeof first published struct */
161 * Hardware event_id to monitor via a performance monitoring event:
163 struct perf_event_attr
{
166 * Major type: hardware/software/tracepoint/etc.
171 * Size of the attr structure, for fwd/bwd compat.
176 * Type specific configuration information.
188 __u64 disabled
: 1, /* off by default */
189 inherit
: 1, /* children inherit it */
190 pinned
: 1, /* must always be on PMU */
191 exclusive
: 1, /* only group on PMU */
192 exclude_user
: 1, /* don't count user */
193 exclude_kernel
: 1, /* ditto kernel */
194 exclude_hv
: 1, /* ditto hypervisor */
195 exclude_idle
: 1, /* don't count when idle */
196 mmap
: 1, /* include mmap data */
197 comm
: 1, /* include comm data */
198 freq
: 1, /* use freq, not period */
199 inherit_stat
: 1, /* per task counts */
200 enable_on_exec
: 1, /* next exec enables */
201 task
: 1, /* trace fork/exit */
202 watermark
: 1, /* wakeup_watermark */
207 __u32 wakeup_events
; /* wakeup every n events */
208 __u32 wakeup_watermark
; /* bytes before wakeup */
216 * Ioctls that can be done on a perf event fd:
218 #define PERF_EVENT_IOC_ENABLE _IO ('$', 0)
219 #define PERF_EVENT_IOC_DISABLE _IO ('$', 1)
220 #define PERF_EVENT_IOC_REFRESH _IO ('$', 2)
221 #define PERF_EVENT_IOC_RESET _IO ('$', 3)
222 #define PERF_EVENT_IOC_PERIOD _IOW('$', 4, u64)
223 #define PERF_EVENT_IOC_SET_OUTPUT _IO ('$', 5)
224 #define PERF_EVENT_IOC_SET_FILTER _IOW('$', 6, char *)
226 enum perf_event_ioc_flags
{
227 PERF_IOC_FLAG_GROUP
= 1U << 0,
231 * Structure of the page that can be mapped via mmap
233 struct perf_event_mmap_page
{
234 __u32 version
; /* version number of this structure */
235 __u32 compat_version
; /* lowest version this is compat with */
238 * Bits needed to read the hw events in user-space.
248 * count = pmc_read(pc->index - 1);
249 * count += pc->offset;
254 * } while (pc->lock != seq);
256 * NOTE: for obvious reason this only works on self-monitoring
259 __u32 lock
; /* seqlock for synchronization */
260 __u32 index
; /* hardware event identifier */
261 __s64 offset
; /* add to hardware event value */
262 __u64 time_enabled
; /* time event active */
263 __u64 time_running
; /* time event on cpu */
266 * Hole for extension of the self monitor capabilities
269 __u64 __reserved
[123]; /* align to 1k */
272 * Control data for the mmap() data buffer.
274 * User-space reading the @data_head value should issue an rmb(), on
275 * SMP capable platforms, after reading this value -- see
276 * perf_event_wakeup().
278 * When the mapping is PROT_WRITE the @data_tail value should be
279 * written by userspace to reflect the last read data. In this case
280 * the kernel will not over-write unread data.
282 __u64 data_head
; /* head in the data section */
283 __u64 data_tail
; /* user-space written tail */
286 #define PERF_RECORD_MISC_CPUMODE_MASK (3 << 0)
287 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN (0 << 0)
288 #define PERF_RECORD_MISC_KERNEL (1 << 0)
289 #define PERF_RECORD_MISC_USER (2 << 0)
290 #define PERF_RECORD_MISC_HYPERVISOR (3 << 0)
292 struct perf_event_header
{
298 enum perf_event_type
{
301 * The MMAP events record the PROT_EXEC mappings so that we can
302 * correlate userspace IPs to code. They have the following structure:
305 * struct perf_event_header header;
314 PERF_RECORD_MMAP
= 1,
318 * struct perf_event_header header;
323 PERF_RECORD_LOST
= 2,
327 * struct perf_event_header header;
333 PERF_RECORD_COMM
= 3,
337 * struct perf_event_header header;
343 PERF_RECORD_EXIT
= 4,
347 * struct perf_event_header header;
353 PERF_RECORD_THROTTLE
= 5,
354 PERF_RECORD_UNTHROTTLE
= 6,
358 * struct perf_event_header header;
364 PERF_RECORD_FORK
= 7,
368 * struct perf_event_header header;
371 * struct read_format values;
374 PERF_RECORD_READ
= 8,
378 * struct perf_event_header header;
380 * { u64 ip; } && PERF_SAMPLE_IP
381 * { u32 pid, tid; } && PERF_SAMPLE_TID
382 * { u64 time; } && PERF_SAMPLE_TIME
383 * { u64 addr; } && PERF_SAMPLE_ADDR
384 * { u64 id; } && PERF_SAMPLE_ID
385 * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID
386 * { u32 cpu, res; } && PERF_SAMPLE_CPU
387 * { u64 period; } && PERF_SAMPLE_PERIOD
389 * { struct read_format values; } && PERF_SAMPLE_READ
392 * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN
395 * # The RAW record below is opaque data wrt the ABI
397 * # That is, the ABI doesn't make any promises wrt to
398 * # the stability of its content, it may vary depending
399 * # on event, hardware, kernel version and phase of
402 * # In other words, PERF_SAMPLE_RAW contents are not an ABI.
406 * char data[size];}&& PERF_SAMPLE_RAW
409 PERF_RECORD_SAMPLE
= 9,
411 PERF_RECORD_MAX
, /* non-ABI */
414 enum perf_callchain_context
{
415 PERF_CONTEXT_HV
= (__u64
)-32,
416 PERF_CONTEXT_KERNEL
= (__u64
)-128,
417 PERF_CONTEXT_USER
= (__u64
)-512,
419 PERF_CONTEXT_GUEST
= (__u64
)-2048,
420 PERF_CONTEXT_GUEST_KERNEL
= (__u64
)-2176,
421 PERF_CONTEXT_GUEST_USER
= (__u64
)-2560,
423 PERF_CONTEXT_MAX
= (__u64
)-4095,
426 #define PERF_FLAG_FD_NO_GROUP (1U << 0)
427 #define PERF_FLAG_FD_OUTPUT (1U << 1)
431 * Kernel-internal data types and definitions:
434 #ifdef CONFIG_PERF_EVENTS
435 # include <asm/perf_event.h>
438 #include <linux/list.h>
439 #include <linux/mutex.h>
440 #include <linux/rculist.h>
441 #include <linux/rcupdate.h>
442 #include <linux/spinlock.h>
443 #include <linux/hrtimer.h>
444 #include <linux/fs.h>
445 #include <linux/pid_namespace.h>
446 #include <linux/workqueue.h>
447 #include <asm/atomic.h>
449 #define PERF_MAX_STACK_DEPTH 255
451 struct perf_callchain_entry
{
453 __u64 ip
[PERF_MAX_STACK_DEPTH
];
456 struct perf_raw_record
{
464 * struct hw_perf_event - performance event hardware details:
466 struct hw_perf_event
{
467 #ifdef CONFIG_PERF_EVENTS
469 struct { /* hardware */
471 unsigned long config_base
;
472 unsigned long event_base
;
475 union { /* software */
477 struct hrtimer hrtimer
;
480 atomic64_t prev_count
;
483 atomic64_t period_left
;
495 * struct pmu - generic performance monitoring unit
498 int (*enable
) (struct perf_event
*event
);
499 void (*disable
) (struct perf_event
*event
);
500 void (*read
) (struct perf_event
*event
);
501 void (*unthrottle
) (struct perf_event
*event
);
505 * enum perf_event_active_state - the states of a event
507 enum perf_event_active_state
{
508 PERF_EVENT_STATE_ERROR
= -2,
509 PERF_EVENT_STATE_OFF
= -1,
510 PERF_EVENT_STATE_INACTIVE
= 0,
511 PERF_EVENT_STATE_ACTIVE
= 1,
516 struct perf_mmap_data
{
517 struct rcu_head rcu_head
;
518 #ifdef CONFIG_PERF_USE_VMALLOC
519 struct work_struct work
;
522 int nr_pages
; /* nr of data pages */
523 int writable
; /* are we writable */
524 int nr_locked
; /* nr pages mlocked */
526 atomic_t poll
; /* POLL_ for wakeups */
527 atomic_t events
; /* event_id limit */
529 atomic_long_t head
; /* write position */
530 atomic_long_t done_head
; /* completed head */
532 atomic_t lock
; /* concurrent writes */
533 atomic_t wakeup
; /* needs a wakeup */
534 atomic_t lost
; /* nr records lost */
536 long watermark
; /* wakeup watermark */
538 struct perf_event_mmap_page
*user_page
;
542 struct perf_pending_entry
{
543 struct perf_pending_entry
*next
;
544 void (*func
)(struct perf_pending_entry
*);
548 * struct perf_event - performance event kernel representation:
551 #ifdef CONFIG_PERF_EVENTS
552 struct list_head group_entry
;
553 struct list_head event_entry
;
554 struct list_head sibling_list
;
556 struct perf_event
*group_leader
;
557 struct perf_event
*output
;
558 const struct pmu
*pmu
;
560 enum perf_event_active_state state
;
564 * These are the total time in nanoseconds that the event
565 * has been enabled (i.e. eligible to run, and the task has
566 * been scheduled in, if this is a per-task event)
567 * and running (scheduled onto the CPU), respectively.
569 * They are computed from tstamp_enabled, tstamp_running and
570 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
572 u64 total_time_enabled
;
573 u64 total_time_running
;
576 * These are timestamps used for computing total_time_enabled
577 * and total_time_running when the event is in INACTIVE or
578 * ACTIVE state, measured in nanoseconds from an arbitrary point
580 * tstamp_enabled: the notional time when the event was enabled
581 * tstamp_running: the notional time when the event was scheduled on
582 * tstamp_stopped: in INACTIVE state, the notional time when the
583 * event was scheduled off.
589 struct perf_event_attr attr
;
590 struct hw_perf_event hw
;
592 struct perf_event_context
*ctx
;
596 * These accumulate total time (in nanoseconds) that children
597 * events have been enabled and running, respectively.
599 atomic64_t child_total_time_enabled
;
600 atomic64_t child_total_time_running
;
603 * Protect attach/detach and child_list:
605 struct mutex child_mutex
;
606 struct list_head child_list
;
607 struct perf_event
*parent
;
612 struct list_head owner_entry
;
613 struct task_struct
*owner
;
616 struct mutex mmap_mutex
;
618 struct perf_mmap_data
*data
;
621 wait_queue_head_t waitq
;
622 struct fasync_struct
*fasync
;
624 /* delayed work for NMIs and such */
628 struct perf_pending_entry pending
;
630 atomic_t event_limit
;
632 void (*destroy
)(struct perf_event
*);
633 struct rcu_head rcu_head
;
635 struct pid_namespace
*ns
;
638 #ifdef CONFIG_EVENT_PROFILE
639 struct event_filter
*filter
;
642 #endif /* CONFIG_PERF_EVENTS */
646 * struct perf_event_context - event context structure
648 * Used as a container for task events and CPU events as well:
650 struct perf_event_context
{
652 * Protect the states of the events in the list,
653 * nr_active, and the list:
657 * Protect the list of events. Locking either mutex or lock
658 * is sufficient to ensure the list doesn't change; to change
659 * the list you need to lock both the mutex and the spinlock.
663 struct list_head group_list
;
664 struct list_head event_list
;
670 struct task_struct
*task
;
673 * Context clock, runs when context enabled.
679 * These fields let us detect when two contexts have both
680 * been cloned (inherited) from a common ancestor.
682 struct perf_event_context
*parent_ctx
;
686 struct rcu_head rcu_head
;
690 * struct perf_event_cpu_context - per cpu event context structure
692 struct perf_cpu_context
{
693 struct perf_event_context ctx
;
694 struct perf_event_context
*task_ctx
;
700 * Recursion avoidance:
702 * task, softirq, irq, nmi context
707 struct perf_output_handle
{
708 struct perf_event
*event
;
709 struct perf_mmap_data
*data
;
711 unsigned long offset
;
718 #ifdef CONFIG_PERF_EVENTS
721 * Set by architecture code:
723 extern int perf_max_events
;
725 extern const struct pmu
*hw_perf_event_init(struct perf_event
*event
);
727 extern void perf_event_task_sched_in(struct task_struct
*task
, int cpu
);
728 extern void perf_event_task_sched_out(struct task_struct
*task
,
729 struct task_struct
*next
, int cpu
);
730 extern void perf_event_task_tick(struct task_struct
*task
, int cpu
);
731 extern int perf_event_init_task(struct task_struct
*child
);
732 extern void perf_event_exit_task(struct task_struct
*child
);
733 extern void perf_event_free_task(struct task_struct
*task
);
734 extern void set_perf_event_pending(void);
735 extern void perf_event_do_pending(void);
736 extern void perf_event_print_debug(void);
737 extern void __perf_disable(void);
738 extern bool __perf_enable(void);
739 extern void perf_disable(void);
740 extern void perf_enable(void);
741 extern int perf_event_task_disable(void);
742 extern int perf_event_task_enable(void);
743 extern int hw_perf_group_sched_in(struct perf_event
*group_leader
,
744 struct perf_cpu_context
*cpuctx
,
745 struct perf_event_context
*ctx
, int cpu
);
746 extern void perf_event_update_userpage(struct perf_event
*event
);
747 extern int perf_event_release_kernel(struct perf_event
*event
);
748 extern struct perf_event
*
749 perf_event_create_kernel_counter(struct perf_event_attr
*attr
,
752 extern u64
perf_event_read_value(struct perf_event
*event
);
754 struct perf_sample_data
{
771 struct perf_callchain_entry
*callchain
;
772 struct perf_raw_record
*raw
;
775 extern void perf_output_sample(struct perf_output_handle
*handle
,
776 struct perf_event_header
*header
,
777 struct perf_sample_data
*data
,
778 struct perf_event
*event
);
779 extern void perf_prepare_sample(struct perf_event_header
*header
,
780 struct perf_sample_data
*data
,
781 struct perf_event
*event
,
782 struct pt_regs
*regs
);
784 extern int perf_event_overflow(struct perf_event
*event
, int nmi
,
785 struct perf_sample_data
*data
,
786 struct pt_regs
*regs
);
789 * Return 1 for a software event, 0 for a hardware event
791 static inline int is_software_event(struct perf_event
*event
)
793 return (event
->attr
.type
!= PERF_TYPE_RAW
) &&
794 (event
->attr
.type
!= PERF_TYPE_HARDWARE
) &&
795 (event
->attr
.type
!= PERF_TYPE_HW_CACHE
);
798 extern atomic_t perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
800 extern void __perf_sw_event(u32
, u64
, int, struct pt_regs
*, u64
);
803 perf_sw_event(u32 event_id
, u64 nr
, int nmi
, struct pt_regs
*regs
, u64 addr
)
805 if (atomic_read(&perf_swevent_enabled
[event_id
]))
806 __perf_sw_event(event_id
, nr
, nmi
, regs
, addr
);
809 extern void __perf_event_mmap(struct vm_area_struct
*vma
);
811 static inline void perf_event_mmap(struct vm_area_struct
*vma
)
813 if (vma
->vm_flags
& VM_EXEC
)
814 __perf_event_mmap(vma
);
817 extern void perf_event_comm(struct task_struct
*tsk
);
818 extern void perf_event_fork(struct task_struct
*tsk
);
820 extern struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
);
822 extern int sysctl_perf_event_paranoid
;
823 extern int sysctl_perf_event_mlock
;
824 extern int sysctl_perf_event_sample_rate
;
826 extern void perf_event_init(void);
827 extern void perf_tp_event(int event_id
, u64 addr
, u64 count
,
828 void *record
, int entry_size
);
830 #ifndef perf_misc_flags
831 #define perf_misc_flags(regs) (user_mode(regs) ? PERF_RECORD_MISC_USER : \
832 PERF_RECORD_MISC_KERNEL)
833 #define perf_instruction_pointer(regs) instruction_pointer(regs)
836 extern int perf_output_begin(struct perf_output_handle
*handle
,
837 struct perf_event
*event
, unsigned int size
,
838 int nmi
, int sample
);
839 extern void perf_output_end(struct perf_output_handle
*handle
);
840 extern void perf_output_copy(struct perf_output_handle
*handle
,
841 const void *buf
, unsigned int len
);
844 perf_event_task_sched_in(struct task_struct
*task
, int cpu
) { }
846 perf_event_task_sched_out(struct task_struct
*task
,
847 struct task_struct
*next
, int cpu
) { }
849 perf_event_task_tick(struct task_struct
*task
, int cpu
) { }
850 static inline int perf_event_init_task(struct task_struct
*child
) { return 0; }
851 static inline void perf_event_exit_task(struct task_struct
*child
) { }
852 static inline void perf_event_free_task(struct task_struct
*task
) { }
853 static inline void perf_event_do_pending(void) { }
854 static inline void perf_event_print_debug(void) { }
855 static inline void perf_disable(void) { }
856 static inline void perf_enable(void) { }
857 static inline int perf_event_task_disable(void) { return -EINVAL
; }
858 static inline int perf_event_task_enable(void) { return -EINVAL
; }
861 perf_sw_event(u32 event_id
, u64 nr
, int nmi
,
862 struct pt_regs
*regs
, u64 addr
) { }
864 static inline void perf_event_mmap(struct vm_area_struct
*vma
) { }
865 static inline void perf_event_comm(struct task_struct
*tsk
) { }
866 static inline void perf_event_fork(struct task_struct
*tsk
) { }
867 static inline void perf_event_init(void) { }
871 #define perf_output_put(handle, x) \
872 perf_output_copy((handle), &(x), sizeof(x))
874 #endif /* __KERNEL__ */
875 #endif /* _LINUX_PERF_EVENT_H */