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_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 counter 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_counter_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 to monitor via a performance monitoring counter:
163 struct perf_counter_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 */
205 __u32 wakeup_events
; /* wakeup every n events */
212 * Ioctls that can be done on a perf counter fd:
214 #define PERF_COUNTER_IOC_ENABLE _IO ('$', 0)
215 #define PERF_COUNTER_IOC_DISABLE _IO ('$', 1)
216 #define PERF_COUNTER_IOC_REFRESH _IO ('$', 2)
217 #define PERF_COUNTER_IOC_RESET _IO ('$', 3)
218 #define PERF_COUNTER_IOC_PERIOD _IOW('$', 4, u64)
220 enum perf_counter_ioc_flags
{
221 PERF_IOC_FLAG_GROUP
= 1U << 0,
225 * Structure of the page that can be mapped via mmap
227 struct perf_counter_mmap_page
{
228 __u32 version
; /* version number of this structure */
229 __u32 compat_version
; /* lowest version this is compat with */
232 * Bits needed to read the hw counters in user-space.
242 * count = pmc_read(pc->index - 1);
243 * count += pc->offset;
248 * } while (pc->lock != seq);
250 * NOTE: for obvious reason this only works on self-monitoring
253 __u32 lock
; /* seqlock for synchronization */
254 __u32 index
; /* hardware counter identifier */
255 __s64 offset
; /* add to hardware counter value */
256 __u64 time_enabled
; /* time counter active */
257 __u64 time_running
; /* time counter on cpu */
260 * Hole for extension of the self monitor capabilities
263 __u64 __reserved
[123]; /* align to 1k */
266 * Control data for the mmap() data buffer.
268 * User-space reading the @data_head value should issue an rmb(), on
269 * SMP capable platforms, after reading this value -- see
270 * perf_counter_wakeup().
272 * When the mapping is PROT_WRITE the @data_tail value should be
273 * written by userspace to reflect the last read data. In this case
274 * the kernel will not over-write unread data.
276 __u64 data_head
; /* head in the data section */
277 __u64 data_tail
; /* user-space written tail */
280 #define PERF_EVENT_MISC_CPUMODE_MASK (3 << 0)
281 #define PERF_EVENT_MISC_CPUMODE_UNKNOWN (0 << 0)
282 #define PERF_EVENT_MISC_KERNEL (1 << 0)
283 #define PERF_EVENT_MISC_USER (2 << 0)
284 #define PERF_EVENT_MISC_HYPERVISOR (3 << 0)
286 struct perf_event_header
{
292 enum perf_event_type
{
295 * The MMAP events record the PROT_EXEC mappings so that we can
296 * correlate userspace IPs to code. They have the following structure:
299 * struct perf_event_header header;
312 * struct perf_event_header header;
321 * struct perf_event_header header;
331 * struct perf_event_header header;
340 * struct perf_event_header header;
346 PERF_EVENT_THROTTLE
= 5,
347 PERF_EVENT_UNTHROTTLE
= 6,
351 * struct perf_event_header header;
360 * struct perf_event_header header;
363 * struct read_format values;
370 * struct perf_event_header header;
372 * { u64 ip; } && PERF_SAMPLE_IP
373 * { u32 pid, tid; } && PERF_SAMPLE_TID
374 * { u64 time; } && PERF_SAMPLE_TIME
375 * { u64 addr; } && PERF_SAMPLE_ADDR
376 * { u64 id; } && PERF_SAMPLE_ID
377 * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID
378 * { u32 cpu, res; } && PERF_SAMPLE_CPU
379 * { u64 period; } && PERF_SAMPLE_PERIOD
381 * { struct read_format values; } && PERF_SAMPLE_READ
384 * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN
387 * # The RAW record below is opaque data wrt the ABI
389 * # That is, the ABI doesn't make any promises wrt to
390 * # the stability of its content, it may vary depending
391 * # on event, hardware, kernel version and phase of
394 * # In other words, PERF_SAMPLE_RAW contents are not an ABI.
398 * char data[size];}&& PERF_SAMPLE_RAW
401 PERF_EVENT_SAMPLE
= 9,
403 PERF_EVENT_MAX
, /* non-ABI */
406 enum perf_callchain_context
{
407 PERF_CONTEXT_HV
= (__u64
)-32,
408 PERF_CONTEXT_KERNEL
= (__u64
)-128,
409 PERF_CONTEXT_USER
= (__u64
)-512,
411 PERF_CONTEXT_GUEST
= (__u64
)-2048,
412 PERF_CONTEXT_GUEST_KERNEL
= (__u64
)-2176,
413 PERF_CONTEXT_GUEST_USER
= (__u64
)-2560,
415 PERF_CONTEXT_MAX
= (__u64
)-4095,
420 * Kernel-internal data types and definitions:
423 #ifdef CONFIG_PERF_COUNTERS
424 # include <asm/perf_counter.h>
427 #include <linux/list.h>
428 #include <linux/mutex.h>
429 #include <linux/rculist.h>
430 #include <linux/rcupdate.h>
431 #include <linux/spinlock.h>
432 #include <linux/hrtimer.h>
433 #include <linux/fs.h>
434 #include <linux/pid_namespace.h>
435 #include <asm/atomic.h>
437 #define PERF_MAX_STACK_DEPTH 255
439 struct perf_callchain_entry
{
441 __u64 ip
[PERF_MAX_STACK_DEPTH
];
444 struct perf_raw_record
{
452 * struct hw_perf_counter - performance counter hardware details:
454 struct hw_perf_counter
{
455 #ifdef CONFIG_PERF_COUNTERS
457 struct { /* hardware */
459 unsigned long config_base
;
460 unsigned long counter_base
;
463 union { /* software */
465 struct hrtimer hrtimer
;
468 atomic64_t prev_count
;
471 atomic64_t period_left
;
483 * struct pmu - generic performance monitoring unit
486 int (*enable
) (struct perf_counter
*counter
);
487 void (*disable
) (struct perf_counter
*counter
);
488 void (*read
) (struct perf_counter
*counter
);
489 void (*unthrottle
) (struct perf_counter
*counter
);
493 * enum perf_counter_active_state - the states of a counter
495 enum perf_counter_active_state
{
496 PERF_COUNTER_STATE_ERROR
= -2,
497 PERF_COUNTER_STATE_OFF
= -1,
498 PERF_COUNTER_STATE_INACTIVE
= 0,
499 PERF_COUNTER_STATE_ACTIVE
= 1,
504 struct perf_mmap_data
{
505 struct rcu_head rcu_head
;
506 int nr_pages
; /* nr of data pages */
507 int writable
; /* are we writable */
508 int nr_locked
; /* nr pages mlocked */
510 atomic_t poll
; /* POLL_ for wakeups */
511 atomic_t events
; /* event limit */
513 atomic_long_t head
; /* write position */
514 atomic_long_t done_head
; /* completed head */
516 atomic_t lock
; /* concurrent writes */
517 atomic_t wakeup
; /* needs a wakeup */
518 atomic_t lost
; /* nr records lost */
520 struct perf_counter_mmap_page
*user_page
;
524 struct perf_pending_entry
{
525 struct perf_pending_entry
*next
;
526 void (*func
)(struct perf_pending_entry
*);
530 * struct perf_counter - performance counter kernel representation:
532 struct perf_counter
{
533 #ifdef CONFIG_PERF_COUNTERS
534 struct list_head list_entry
;
535 struct list_head event_entry
;
536 struct list_head sibling_list
;
538 struct perf_counter
*group_leader
;
539 const struct pmu
*pmu
;
541 enum perf_counter_active_state state
;
545 * These are the total time in nanoseconds that the counter
546 * has been enabled (i.e. eligible to run, and the task has
547 * been scheduled in, if this is a per-task counter)
548 * and running (scheduled onto the CPU), respectively.
550 * They are computed from tstamp_enabled, tstamp_running and
551 * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
553 u64 total_time_enabled
;
554 u64 total_time_running
;
557 * These are timestamps used for computing total_time_enabled
558 * and total_time_running when the counter is in INACTIVE or
559 * ACTIVE state, measured in nanoseconds from an arbitrary point
561 * tstamp_enabled: the notional time when the counter was enabled
562 * tstamp_running: the notional time when the counter was scheduled on
563 * tstamp_stopped: in INACTIVE state, the notional time when the
564 * counter was scheduled off.
570 struct perf_counter_attr attr
;
571 struct hw_perf_counter hw
;
573 struct perf_counter_context
*ctx
;
577 * These accumulate total time (in nanoseconds) that children
578 * counters have been enabled and running, respectively.
580 atomic64_t child_total_time_enabled
;
581 atomic64_t child_total_time_running
;
584 * Protect attach/detach and child_list:
586 struct mutex child_mutex
;
587 struct list_head child_list
;
588 struct perf_counter
*parent
;
593 struct list_head owner_entry
;
594 struct task_struct
*owner
;
597 struct mutex mmap_mutex
;
599 struct perf_mmap_data
*data
;
602 wait_queue_head_t waitq
;
603 struct fasync_struct
*fasync
;
605 /* delayed work for NMIs and such */
609 struct perf_pending_entry pending
;
611 atomic_t event_limit
;
613 void (*destroy
)(struct perf_counter
*);
614 struct rcu_head rcu_head
;
616 struct pid_namespace
*ns
;
622 * struct perf_counter_context - counter context structure
624 * Used as a container for task counters and CPU counters as well:
626 struct perf_counter_context
{
628 * Protect the states of the counters in the list,
629 * nr_active, and the list:
633 * Protect the list of counters. Locking either mutex or lock
634 * is sufficient to ensure the list doesn't change; to change
635 * the list you need to lock both the mutex and the spinlock.
639 struct list_head counter_list
;
640 struct list_head event_list
;
646 struct task_struct
*task
;
649 * Context clock, runs when context enabled.
655 * These fields let us detect when two contexts have both
656 * been cloned (inherited) from a common ancestor.
658 struct perf_counter_context
*parent_ctx
;
662 struct rcu_head rcu_head
;
666 * struct perf_counter_cpu_context - per cpu counter context structure
668 struct perf_cpu_context
{
669 struct perf_counter_context ctx
;
670 struct perf_counter_context
*task_ctx
;
676 * Recursion avoidance:
678 * task, softirq, irq, nmi context
683 #ifdef CONFIG_PERF_COUNTERS
686 * Set by architecture code:
688 extern int perf_max_counters
;
690 extern const struct pmu
*hw_perf_counter_init(struct perf_counter
*counter
);
692 extern void perf_counter_task_sched_in(struct task_struct
*task
, int cpu
);
693 extern void perf_counter_task_sched_out(struct task_struct
*task
,
694 struct task_struct
*next
, int cpu
);
695 extern void perf_counter_task_tick(struct task_struct
*task
, int cpu
);
696 extern int perf_counter_init_task(struct task_struct
*child
);
697 extern void perf_counter_exit_task(struct task_struct
*child
);
698 extern void perf_counter_free_task(struct task_struct
*task
);
699 extern void set_perf_counter_pending(void);
700 extern void perf_counter_do_pending(void);
701 extern void perf_counter_print_debug(void);
702 extern void __perf_disable(void);
703 extern bool __perf_enable(void);
704 extern void perf_disable(void);
705 extern void perf_enable(void);
706 extern int perf_counter_task_disable(void);
707 extern int perf_counter_task_enable(void);
708 extern int hw_perf_group_sched_in(struct perf_counter
*group_leader
,
709 struct perf_cpu_context
*cpuctx
,
710 struct perf_counter_context
*ctx
, int cpu
);
711 extern void perf_counter_update_userpage(struct perf_counter
*counter
);
713 struct perf_sample_data
{
714 struct pt_regs
*regs
;
717 struct perf_raw_record
*raw
;
720 extern int perf_counter_overflow(struct perf_counter
*counter
, int nmi
,
721 struct perf_sample_data
*data
);
722 extern void perf_counter_output(struct perf_counter
*counter
, int nmi
,
723 struct perf_sample_data
*data
);
726 * Return 1 for a software counter, 0 for a hardware counter
728 static inline int is_software_counter(struct perf_counter
*counter
)
730 return (counter
->attr
.type
!= PERF_TYPE_RAW
) &&
731 (counter
->attr
.type
!= PERF_TYPE_HARDWARE
) &&
732 (counter
->attr
.type
!= PERF_TYPE_HW_CACHE
);
735 extern atomic_t perf_swcounter_enabled
[PERF_COUNT_SW_MAX
];
737 extern void __perf_swcounter_event(u32
, u64
, int, struct pt_regs
*, u64
);
740 perf_swcounter_event(u32 event
, u64 nr
, int nmi
, struct pt_regs
*regs
, u64 addr
)
742 if (atomic_read(&perf_swcounter_enabled
[event
]))
743 __perf_swcounter_event(event
, nr
, nmi
, regs
, addr
);
746 extern void __perf_counter_mmap(struct vm_area_struct
*vma
);
748 static inline void perf_counter_mmap(struct vm_area_struct
*vma
)
750 if (vma
->vm_flags
& VM_EXEC
)
751 __perf_counter_mmap(vma
);
754 extern void perf_counter_comm(struct task_struct
*tsk
);
755 extern void perf_counter_fork(struct task_struct
*tsk
);
757 extern struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
);
759 extern int sysctl_perf_counter_paranoid
;
760 extern int sysctl_perf_counter_mlock
;
761 extern int sysctl_perf_counter_sample_rate
;
763 extern void perf_counter_init(void);
764 extern void perf_tpcounter_event(int event_id
, u64 addr
, u64 count
,
765 void *record
, int entry_size
);
767 #ifndef perf_misc_flags
768 #define perf_misc_flags(regs) (user_mode(regs) ? PERF_EVENT_MISC_USER : \
769 PERF_EVENT_MISC_KERNEL)
770 #define perf_instruction_pointer(regs) instruction_pointer(regs)
775 perf_counter_task_sched_in(struct task_struct
*task
, int cpu
) { }
777 perf_counter_task_sched_out(struct task_struct
*task
,
778 struct task_struct
*next
, int cpu
) { }
780 perf_counter_task_tick(struct task_struct
*task
, int cpu
) { }
781 static inline int perf_counter_init_task(struct task_struct
*child
) { return 0; }
782 static inline void perf_counter_exit_task(struct task_struct
*child
) { }
783 static inline void perf_counter_free_task(struct task_struct
*task
) { }
784 static inline void perf_counter_do_pending(void) { }
785 static inline void perf_counter_print_debug(void) { }
786 static inline void perf_disable(void) { }
787 static inline void perf_enable(void) { }
788 static inline int perf_counter_task_disable(void) { return -EINVAL
; }
789 static inline int perf_counter_task_enable(void) { return -EINVAL
; }
792 perf_swcounter_event(u32 event
, u64 nr
, int nmi
,
793 struct pt_regs
*regs
, u64 addr
) { }
795 static inline void perf_counter_mmap(struct vm_area_struct
*vma
) { }
796 static inline void perf_counter_comm(struct task_struct
*tsk
) { }
797 static inline void perf_counter_fork(struct task_struct
*tsk
) { }
798 static inline void perf_counter_init(void) { }
801 #endif /* __KERNEL__ */
802 #endif /* _LINUX_PERF_COUNTER_H */