2 * Performance counters:
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_COUNTER_H
15 #define _LINUX_PERF_COUNTER_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 counter event types, used by the
40 * attr.event_id parameter of the sys_perf_counter_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 counters:
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" counters provided by the kernel, even if the hardware
93 * does not support performance counters. These counters 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_counter_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_GROUP
= 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,
125 PERF_SAMPLE_MAX
= 1U << 10, /* non-ABI */
129 * Bits that can be set in attr.read_format to request that
130 * reads on the counter should return the indicated quantities,
131 * in increasing order of bit value, after the counter value.
133 enum perf_counter_read_format
{
134 PERF_FORMAT_TOTAL_TIME_ENABLED
= 1U << 0,
135 PERF_FORMAT_TOTAL_TIME_RUNNING
= 1U << 1,
136 PERF_FORMAT_ID
= 1U << 2,
138 PERF_FORMAT_MAX
= 1U << 3, /* non-ABI */
141 #define PERF_ATTR_SIZE_VER0 64 /* sizeof first published struct */
144 * Hardware event to monitor via a performance monitoring counter:
146 struct perf_counter_attr
{
149 * Major type: hardware/software/tracepoint/etc.
154 * Size of the attr structure, for fwd/bwd compat.
159 * Type specific configuration information.
171 __u64 disabled
: 1, /* off by default */
172 inherit
: 1, /* children inherit it */
173 pinned
: 1, /* must always be on PMU */
174 exclusive
: 1, /* only group on PMU */
175 exclude_user
: 1, /* don't count user */
176 exclude_kernel
: 1, /* ditto kernel */
177 exclude_hv
: 1, /* ditto hypervisor */
178 exclude_idle
: 1, /* don't count when idle */
179 mmap
: 1, /* include mmap data */
180 comm
: 1, /* include comm data */
181 freq
: 1, /* use freq, not period */
182 inherit_stat
: 1, /* per task counts */
183 enable_on_exec
: 1, /* next exec enables */
187 __u32 wakeup_events
; /* wakeup every n events */
194 * Ioctls that can be done on a perf counter fd:
196 #define PERF_COUNTER_IOC_ENABLE _IO ('$', 0)
197 #define PERF_COUNTER_IOC_DISABLE _IO ('$', 1)
198 #define PERF_COUNTER_IOC_REFRESH _IO ('$', 2)
199 #define PERF_COUNTER_IOC_RESET _IO ('$', 3)
200 #define PERF_COUNTER_IOC_PERIOD _IOW('$', 4, u64)
202 enum perf_counter_ioc_flags
{
203 PERF_IOC_FLAG_GROUP
= 1U << 0,
207 * Structure of the page that can be mapped via mmap
209 struct perf_counter_mmap_page
{
210 __u32 version
; /* version number of this structure */
211 __u32 compat_version
; /* lowest version this is compat with */
214 * Bits needed to read the hw counters in user-space.
224 * count = pmc_read(pc->index - 1);
225 * count += pc->offset;
230 * } while (pc->lock != seq);
232 * NOTE: for obvious reason this only works on self-monitoring
235 __u32 lock
; /* seqlock for synchronization */
236 __u32 index
; /* hardware counter identifier */
237 __s64 offset
; /* add to hardware counter value */
238 __u64 time_enabled
; /* time counter active */
239 __u64 time_running
; /* time counter on cpu */
242 * Hole for extension of the self monitor capabilities
245 __u64 __reserved
[123]; /* align to 1k */
248 * Control data for the mmap() data buffer.
250 * User-space reading the @data_head value should issue an rmb(), on
251 * SMP capable platforms, after reading this value -- see
252 * perf_counter_wakeup().
254 * When the mapping is PROT_WRITE the @data_tail value should be
255 * written by userspace to reflect the last read data. In this case
256 * the kernel will not over-write unread data.
258 __u64 data_head
; /* head in the data section */
259 __u64 data_tail
; /* user-space written tail */
262 #define PERF_EVENT_MISC_CPUMODE_MASK (3 << 0)
263 #define PERF_EVENT_MISC_CPUMODE_UNKNOWN (0 << 0)
264 #define PERF_EVENT_MISC_KERNEL (1 << 0)
265 #define PERF_EVENT_MISC_USER (2 << 0)
266 #define PERF_EVENT_MISC_HYPERVISOR (3 << 0)
268 struct perf_event_header
{
274 enum perf_event_type
{
277 * The MMAP events record the PROT_EXEC mappings so that we can
278 * correlate userspace IPs to code. They have the following structure:
281 * struct perf_event_header header;
294 * struct perf_event_header header;
303 * struct perf_event_header header;
313 * struct perf_event_header header;
319 PERF_EVENT_THROTTLE
= 5,
320 PERF_EVENT_UNTHROTTLE
= 6,
324 * struct perf_event_header header;
332 * struct perf_event_header header;
335 * { u64 time_enabled; } && PERF_FORMAT_ENABLED
336 * { u64 time_running; } && PERF_FORMAT_RUNNING
337 * { u64 parent_id; } && PERF_FORMAT_ID
344 * struct perf_event_header header;
346 * { u64 ip; } && PERF_SAMPLE_IP
347 * { u32 pid, tid; } && PERF_SAMPLE_TID
348 * { u64 time; } && PERF_SAMPLE_TIME
349 * { u64 addr; } && PERF_SAMPLE_ADDR
350 * { u64 id; } && PERF_SAMPLE_ID
351 * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID
352 * { u32 cpu, res; } && PERF_SAMPLE_CPU
353 * { u64 period; } && PERF_SAMPLE_PERIOD
356 * { u64 id, val; } cnt[nr]; } && PERF_SAMPLE_GROUP
359 * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN
362 PERF_EVENT_SAMPLE
= 9,
364 PERF_EVENT_MAX
, /* non-ABI */
367 enum perf_callchain_context
{
368 PERF_CONTEXT_HV
= (__u64
)-32,
369 PERF_CONTEXT_KERNEL
= (__u64
)-128,
370 PERF_CONTEXT_USER
= (__u64
)-512,
372 PERF_CONTEXT_GUEST
= (__u64
)-2048,
373 PERF_CONTEXT_GUEST_KERNEL
= (__u64
)-2176,
374 PERF_CONTEXT_GUEST_USER
= (__u64
)-2560,
376 PERF_CONTEXT_MAX
= (__u64
)-4095,
381 * Kernel-internal data types and definitions:
384 #ifdef CONFIG_PERF_COUNTERS
385 # include <asm/perf_counter.h>
388 #include <linux/list.h>
389 #include <linux/mutex.h>
390 #include <linux/rculist.h>
391 #include <linux/rcupdate.h>
392 #include <linux/spinlock.h>
393 #include <linux/hrtimer.h>
394 #include <linux/fs.h>
395 #include <linux/pid_namespace.h>
396 #include <asm/atomic.h>
398 #define PERF_MAX_STACK_DEPTH 255
400 struct perf_callchain_entry
{
402 __u64 ip
[PERF_MAX_STACK_DEPTH
];
408 * struct hw_perf_counter - performance counter hardware details:
410 struct hw_perf_counter
{
411 #ifdef CONFIG_PERF_COUNTERS
413 struct { /* hardware */
415 unsigned long config_base
;
416 unsigned long counter_base
;
419 union { /* software */
421 struct hrtimer hrtimer
;
424 atomic64_t prev_count
;
427 atomic64_t period_left
;
439 * struct pmu - generic performance monitoring unit
442 int (*enable
) (struct perf_counter
*counter
);
443 void (*disable
) (struct perf_counter
*counter
);
444 void (*read
) (struct perf_counter
*counter
);
445 void (*unthrottle
) (struct perf_counter
*counter
);
449 * enum perf_counter_active_state - the states of a counter
451 enum perf_counter_active_state
{
452 PERF_COUNTER_STATE_ERROR
= -2,
453 PERF_COUNTER_STATE_OFF
= -1,
454 PERF_COUNTER_STATE_INACTIVE
= 0,
455 PERF_COUNTER_STATE_ACTIVE
= 1,
460 struct perf_mmap_data
{
461 struct rcu_head rcu_head
;
462 int nr_pages
; /* nr of data pages */
463 int writable
; /* are we writable */
464 int nr_locked
; /* nr pages mlocked */
466 atomic_t poll
; /* POLL_ for wakeups */
467 atomic_t events
; /* event limit */
469 atomic_long_t head
; /* write position */
470 atomic_long_t done_head
; /* completed head */
472 atomic_t lock
; /* concurrent writes */
473 atomic_t wakeup
; /* needs a wakeup */
474 atomic_t lost
; /* nr records lost */
476 struct perf_counter_mmap_page
*user_page
;
480 struct perf_pending_entry
{
481 struct perf_pending_entry
*next
;
482 void (*func
)(struct perf_pending_entry
*);
486 * struct perf_counter - performance counter kernel representation:
488 struct perf_counter
{
489 #ifdef CONFIG_PERF_COUNTERS
490 struct list_head list_entry
;
491 struct list_head event_entry
;
492 struct list_head sibling_list
;
494 struct perf_counter
*group_leader
;
495 const struct pmu
*pmu
;
497 enum perf_counter_active_state state
;
501 * These are the total time in nanoseconds that the counter
502 * has been enabled (i.e. eligible to run, and the task has
503 * been scheduled in, if this is a per-task counter)
504 * and running (scheduled onto the CPU), respectively.
506 * They are computed from tstamp_enabled, tstamp_running and
507 * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
509 u64 total_time_enabled
;
510 u64 total_time_running
;
513 * These are timestamps used for computing total_time_enabled
514 * and total_time_running when the counter is in INACTIVE or
515 * ACTIVE state, measured in nanoseconds from an arbitrary point
517 * tstamp_enabled: the notional time when the counter was enabled
518 * tstamp_running: the notional time when the counter was scheduled on
519 * tstamp_stopped: in INACTIVE state, the notional time when the
520 * counter was scheduled off.
526 struct perf_counter_attr attr
;
527 struct hw_perf_counter hw
;
529 struct perf_counter_context
*ctx
;
533 * These accumulate total time (in nanoseconds) that children
534 * counters have been enabled and running, respectively.
536 atomic64_t child_total_time_enabled
;
537 atomic64_t child_total_time_running
;
540 * Protect attach/detach and child_list:
542 struct mutex child_mutex
;
543 struct list_head child_list
;
544 struct perf_counter
*parent
;
549 struct list_head owner_entry
;
550 struct task_struct
*owner
;
553 struct mutex mmap_mutex
;
555 struct perf_mmap_data
*data
;
558 wait_queue_head_t waitq
;
559 struct fasync_struct
*fasync
;
561 /* delayed work for NMIs and such */
565 struct perf_pending_entry pending
;
567 atomic_t event_limit
;
569 void (*destroy
)(struct perf_counter
*);
570 struct rcu_head rcu_head
;
572 struct pid_namespace
*ns
;
578 * struct perf_counter_context - counter context structure
580 * Used as a container for task counters and CPU counters as well:
582 struct perf_counter_context
{
584 * Protect the states of the counters in the list,
585 * nr_active, and the list:
589 * Protect the list of counters. Locking either mutex or lock
590 * is sufficient to ensure the list doesn't change; to change
591 * the list you need to lock both the mutex and the spinlock.
595 struct list_head counter_list
;
596 struct list_head event_list
;
602 struct task_struct
*task
;
605 * Context clock, runs when context enabled.
611 * These fields let us detect when two contexts have both
612 * been cloned (inherited) from a common ancestor.
614 struct perf_counter_context
*parent_ctx
;
618 struct rcu_head rcu_head
;
622 * struct perf_counter_cpu_context - per cpu counter context structure
624 struct perf_cpu_context
{
625 struct perf_counter_context ctx
;
626 struct perf_counter_context
*task_ctx
;
632 * Recursion avoidance:
634 * task, softirq, irq, nmi context
639 #ifdef CONFIG_PERF_COUNTERS
642 * Set by architecture code:
644 extern int perf_max_counters
;
646 extern const struct pmu
*hw_perf_counter_init(struct perf_counter
*counter
);
648 extern void perf_counter_task_sched_in(struct task_struct
*task
, int cpu
);
649 extern void perf_counter_task_sched_out(struct task_struct
*task
,
650 struct task_struct
*next
, int cpu
);
651 extern void perf_counter_task_tick(struct task_struct
*task
, int cpu
);
652 extern int perf_counter_init_task(struct task_struct
*child
);
653 extern void perf_counter_exit_task(struct task_struct
*child
);
654 extern void perf_counter_free_task(struct task_struct
*task
);
655 extern void set_perf_counter_pending(void);
656 extern void perf_counter_do_pending(void);
657 extern void perf_counter_print_debug(void);
658 extern void __perf_disable(void);
659 extern bool __perf_enable(void);
660 extern void perf_disable(void);
661 extern void perf_enable(void);
662 extern int perf_counter_task_disable(void);
663 extern int perf_counter_task_enable(void);
664 extern int hw_perf_group_sched_in(struct perf_counter
*group_leader
,
665 struct perf_cpu_context
*cpuctx
,
666 struct perf_counter_context
*ctx
, int cpu
);
667 extern void perf_counter_update_userpage(struct perf_counter
*counter
);
669 struct perf_sample_data
{
670 struct pt_regs
*regs
;
675 extern int perf_counter_overflow(struct perf_counter
*counter
, int nmi
,
676 struct perf_sample_data
*data
);
679 * Return 1 for a software counter, 0 for a hardware counter
681 static inline int is_software_counter(struct perf_counter
*counter
)
683 return (counter
->attr
.type
!= PERF_TYPE_RAW
) &&
684 (counter
->attr
.type
!= PERF_TYPE_HARDWARE
) &&
685 (counter
->attr
.type
!= PERF_TYPE_HW_CACHE
);
688 extern atomic_t perf_swcounter_enabled
[PERF_COUNT_SW_MAX
];
690 extern void __perf_swcounter_event(u32
, u64
, int, struct pt_regs
*, u64
);
693 perf_swcounter_event(u32 event
, u64 nr
, int nmi
, struct pt_regs
*regs
, u64 addr
)
695 if (atomic_read(&perf_swcounter_enabled
[event
]))
696 __perf_swcounter_event(event
, nr
, nmi
, regs
, addr
);
699 extern void __perf_counter_mmap(struct vm_area_struct
*vma
);
701 static inline void perf_counter_mmap(struct vm_area_struct
*vma
)
703 if (vma
->vm_flags
& VM_EXEC
)
704 __perf_counter_mmap(vma
);
707 extern void perf_counter_comm(struct task_struct
*tsk
);
708 extern void perf_counter_fork(struct task_struct
*tsk
);
710 extern struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
);
712 extern int sysctl_perf_counter_paranoid
;
713 extern int sysctl_perf_counter_mlock
;
714 extern int sysctl_perf_counter_sample_rate
;
716 extern void perf_counter_init(void);
718 #ifndef perf_misc_flags
719 #define perf_misc_flags(regs) (user_mode(regs) ? PERF_EVENT_MISC_USER : \
720 PERF_EVENT_MISC_KERNEL)
721 #define perf_instruction_pointer(regs) instruction_pointer(regs)
726 perf_counter_task_sched_in(struct task_struct
*task
, int cpu
) { }
728 perf_counter_task_sched_out(struct task_struct
*task
,
729 struct task_struct
*next
, int cpu
) { }
731 perf_counter_task_tick(struct task_struct
*task
, int cpu
) { }
732 static inline int perf_counter_init_task(struct task_struct
*child
) { return 0; }
733 static inline void perf_counter_exit_task(struct task_struct
*child
) { }
734 static inline void perf_counter_free_task(struct task_struct
*task
) { }
735 static inline void perf_counter_do_pending(void) { }
736 static inline void perf_counter_print_debug(void) { }
737 static inline void perf_disable(void) { }
738 static inline void perf_enable(void) { }
739 static inline int perf_counter_task_disable(void) { return -EINVAL
; }
740 static inline int perf_counter_task_enable(void) { return -EINVAL
; }
743 perf_swcounter_event(u32 event
, u64 nr
, int nmi
,
744 struct pt_regs
*regs
, u64 addr
) { }
746 static inline void perf_counter_mmap(struct vm_area_struct
*vma
) { }
747 static inline void perf_counter_comm(struct task_struct
*tsk
) { }
748 static inline void perf_counter_fork(struct task_struct
*tsk
) { }
749 static inline void perf_counter_init(void) { }
752 #endif /* __KERNEL__ */
753 #endif /* _LINUX_PERF_COUNTER_H */