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,
126 * Bits that can be set in attr.read_format to request that
127 * reads on the counter should return the indicated quantities,
128 * in increasing order of bit value, after the counter value.
130 enum perf_counter_read_format
{
131 PERF_FORMAT_TOTAL_TIME_ENABLED
= 1U << 0,
132 PERF_FORMAT_TOTAL_TIME_RUNNING
= 1U << 1,
133 PERF_FORMAT_ID
= 1U << 2,
137 * Hardware event to monitor via a performance monitoring counter:
139 struct perf_counter_attr
{
141 * Major type: hardware/software/tracepoint/etc.
147 * Type specific configuration information.
159 __u64 disabled
: 1, /* off by default */
160 inherit
: 1, /* children inherit it */
161 pinned
: 1, /* must always be on PMU */
162 exclusive
: 1, /* only group on PMU */
163 exclude_user
: 1, /* don't count user */
164 exclude_kernel
: 1, /* ditto kernel */
165 exclude_hv
: 1, /* ditto hypervisor */
166 exclude_idle
: 1, /* don't count when idle */
167 mmap
: 1, /* include mmap data */
168 comm
: 1, /* include comm data */
169 freq
: 1, /* use freq, not period */
173 __u32 wakeup_events
; /* wakeup every n events */
180 * Ioctls that can be done on a perf counter fd:
182 #define PERF_COUNTER_IOC_ENABLE _IO ('$', 0)
183 #define PERF_COUNTER_IOC_DISABLE _IO ('$', 1)
184 #define PERF_COUNTER_IOC_REFRESH _IO ('$', 2)
185 #define PERF_COUNTER_IOC_RESET _IO ('$', 3)
186 #define PERF_COUNTER_IOC_PERIOD _IOW('$', 4, u64)
188 enum perf_counter_ioc_flags
{
189 PERF_IOC_FLAG_GROUP
= 1U << 0,
193 * Structure of the page that can be mapped via mmap
195 struct perf_counter_mmap_page
{
196 __u32 version
; /* version number of this structure */
197 __u32 compat_version
; /* lowest version this is compat with */
200 * Bits needed to read the hw counters in user-space.
210 * count = pmc_read(pc->index - 1);
211 * count += pc->offset;
216 * } while (pc->lock != seq);
218 * NOTE: for obvious reason this only works on self-monitoring
221 __u32 lock
; /* seqlock for synchronization */
222 __u32 index
; /* hardware counter identifier */
223 __s64 offset
; /* add to hardware counter value */
226 * Control data for the mmap() data buffer.
228 * User-space reading this value should issue an rmb(), on SMP capable
229 * platforms, after reading this value -- see perf_counter_wakeup().
231 __u64 data_head
; /* head in the data section */
234 #define PERF_EVENT_MISC_CPUMODE_MASK (3 << 0)
235 #define PERF_EVENT_MISC_CPUMODE_UNKNOWN (0 << 0)
236 #define PERF_EVENT_MISC_KERNEL (1 << 0)
237 #define PERF_EVENT_MISC_USER (2 << 0)
238 #define PERF_EVENT_MISC_HYPERVISOR (3 << 0)
239 #define PERF_EVENT_MISC_OVERFLOW (1 << 2)
241 struct perf_event_header
{
247 enum perf_event_type
{
250 * The MMAP events record the PROT_EXEC mappings so that we can
251 * correlate userspace IPs to code. They have the following structure:
254 * struct perf_event_header header;
267 * struct perf_event_header header;
277 * struct perf_event_header header;
283 PERF_EVENT_PERIOD
= 4,
287 * struct perf_event_header header;
291 PERF_EVENT_THROTTLE
= 5,
292 PERF_EVENT_UNTHROTTLE
= 6,
296 * struct perf_event_header header;
303 * When header.misc & PERF_EVENT_MISC_OVERFLOW the event_type field
304 * will be PERF_RECORD_*
307 * struct perf_event_header header;
309 * { u64 ip; } && PERF_RECORD_IP
310 * { u32 pid, tid; } && PERF_RECORD_TID
311 * { u64 time; } && PERF_RECORD_TIME
312 * { u64 addr; } && PERF_RECORD_ADDR
313 * { u64 config; } && PERF_RECORD_CONFIG
314 * { u32 cpu, res; } && PERF_RECORD_CPU
317 * { u64 id, val; } cnt[nr]; } && PERF_RECORD_GROUP
323 * u64 ips[nr]; } && PERF_RECORD_CALLCHAIN
330 * Kernel-internal data types and definitions:
333 #ifdef CONFIG_PERF_COUNTERS
334 # include <asm/perf_counter.h>
337 #include <linux/list.h>
338 #include <linux/mutex.h>
339 #include <linux/rculist.h>
340 #include <linux/rcupdate.h>
341 #include <linux/spinlock.h>
342 #include <linux/hrtimer.h>
343 #include <linux/fs.h>
344 #include <linux/pid_namespace.h>
345 #include <asm/atomic.h>
350 * struct hw_perf_counter - performance counter hardware details:
352 struct hw_perf_counter
{
353 #ifdef CONFIG_PERF_COUNTERS
355 struct { /* hardware */
357 unsigned long config_base
;
358 unsigned long counter_base
;
361 union { /* software */
363 struct hrtimer hrtimer
;
366 atomic64_t prev_count
;
369 atomic64_t period_left
;
381 * struct pmu - generic performance monitoring unit
384 int (*enable
) (struct perf_counter
*counter
);
385 void (*disable
) (struct perf_counter
*counter
);
386 void (*read
) (struct perf_counter
*counter
);
387 void (*unthrottle
) (struct perf_counter
*counter
);
391 * enum perf_counter_active_state - the states of a counter
393 enum perf_counter_active_state
{
394 PERF_COUNTER_STATE_ERROR
= -2,
395 PERF_COUNTER_STATE_OFF
= -1,
396 PERF_COUNTER_STATE_INACTIVE
= 0,
397 PERF_COUNTER_STATE_ACTIVE
= 1,
402 struct perf_mmap_data
{
403 struct rcu_head rcu_head
;
404 int nr_pages
; /* nr of data pages */
405 int nr_locked
; /* nr pages mlocked */
407 atomic_t poll
; /* POLL_ for wakeups */
408 atomic_t events
; /* event limit */
410 atomic_long_t head
; /* write position */
411 atomic_long_t done_head
; /* completed head */
413 atomic_t lock
; /* concurrent writes */
415 atomic_t wakeup
; /* needs a wakeup */
417 struct perf_counter_mmap_page
*user_page
;
421 struct perf_pending_entry
{
422 struct perf_pending_entry
*next
;
423 void (*func
)(struct perf_pending_entry
*);
427 * struct perf_counter - performance counter kernel representation:
429 struct perf_counter
{
430 #ifdef CONFIG_PERF_COUNTERS
431 struct list_head list_entry
;
432 struct list_head event_entry
;
433 struct list_head sibling_list
;
435 struct perf_counter
*group_leader
;
436 const struct pmu
*pmu
;
438 enum perf_counter_active_state state
;
442 * These are the total time in nanoseconds that the counter
443 * has been enabled (i.e. eligible to run, and the task has
444 * been scheduled in, if this is a per-task counter)
445 * and running (scheduled onto the CPU), respectively.
447 * They are computed from tstamp_enabled, tstamp_running and
448 * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
450 u64 total_time_enabled
;
451 u64 total_time_running
;
454 * These are timestamps used for computing total_time_enabled
455 * and total_time_running when the counter is in INACTIVE or
456 * ACTIVE state, measured in nanoseconds from an arbitrary point
458 * tstamp_enabled: the notional time when the counter was enabled
459 * tstamp_running: the notional time when the counter was scheduled on
460 * tstamp_stopped: in INACTIVE state, the notional time when the
461 * counter was scheduled off.
467 struct perf_counter_attr attr
;
468 struct hw_perf_counter hw
;
470 struct perf_counter_context
*ctx
;
474 * These accumulate total time (in nanoseconds) that children
475 * counters have been enabled and running, respectively.
477 atomic64_t child_total_time_enabled
;
478 atomic64_t child_total_time_running
;
481 * Protect attach/detach and child_list:
483 struct mutex child_mutex
;
484 struct list_head child_list
;
485 struct perf_counter
*parent
;
490 struct list_head owner_entry
;
491 struct task_struct
*owner
;
494 struct mutex mmap_mutex
;
496 struct perf_mmap_data
*data
;
499 wait_queue_head_t waitq
;
500 struct fasync_struct
*fasync
;
502 /* delayed work for NMIs and such */
506 struct perf_pending_entry pending
;
508 atomic_t event_limit
;
510 void (*destroy
)(struct perf_counter
*);
511 struct rcu_head rcu_head
;
513 struct pid_namespace
*ns
;
519 * struct perf_counter_context - counter context structure
521 * Used as a container for task counters and CPU counters as well:
523 struct perf_counter_context
{
525 * Protect the states of the counters in the list,
526 * nr_active, and the list:
530 * Protect the list of counters. Locking either mutex or lock
531 * is sufficient to ensure the list doesn't change; to change
532 * the list you need to lock both the mutex and the spinlock.
536 struct list_head counter_list
;
537 struct list_head event_list
;
542 struct task_struct
*task
;
545 * Context clock, runs when context enabled.
551 * These fields let us detect when two contexts have both
552 * been cloned (inherited) from a common ancestor.
554 struct perf_counter_context
*parent_ctx
;
558 struct rcu_head rcu_head
;
562 * struct perf_counter_cpu_context - per cpu counter context structure
564 struct perf_cpu_context
{
565 struct perf_counter_context ctx
;
566 struct perf_counter_context
*task_ctx
;
572 * Recursion avoidance:
574 * task, softirq, irq, nmi context
579 #ifdef CONFIG_PERF_COUNTERS
582 * Set by architecture code:
584 extern int perf_max_counters
;
586 extern const struct pmu
*hw_perf_counter_init(struct perf_counter
*counter
);
588 extern void perf_counter_task_sched_in(struct task_struct
*task
, int cpu
);
589 extern void perf_counter_task_sched_out(struct task_struct
*task
,
590 struct task_struct
*next
, int cpu
);
591 extern void perf_counter_task_tick(struct task_struct
*task
, int cpu
);
592 extern int perf_counter_init_task(struct task_struct
*child
);
593 extern void perf_counter_exit_task(struct task_struct
*child
);
594 extern void perf_counter_free_task(struct task_struct
*task
);
595 extern void perf_counter_do_pending(void);
596 extern void perf_counter_print_debug(void);
597 extern void __perf_disable(void);
598 extern bool __perf_enable(void);
599 extern void perf_disable(void);
600 extern void perf_enable(void);
601 extern int perf_counter_task_disable(void);
602 extern int perf_counter_task_enable(void);
603 extern int hw_perf_group_sched_in(struct perf_counter
*group_leader
,
604 struct perf_cpu_context
*cpuctx
,
605 struct perf_counter_context
*ctx
, int cpu
);
606 extern void perf_counter_update_userpage(struct perf_counter
*counter
);
608 struct perf_sample_data
{
609 struct pt_regs
*regs
;
614 extern int perf_counter_overflow(struct perf_counter
*counter
, int nmi
,
615 struct perf_sample_data
*data
);
618 * Return 1 for a software counter, 0 for a hardware counter
620 static inline int is_software_counter(struct perf_counter
*counter
)
622 return (counter
->attr
.type
!= PERF_TYPE_RAW
) &&
623 (counter
->attr
.type
!= PERF_TYPE_HARDWARE
);
626 extern void perf_swcounter_event(u32
, u64
, int, struct pt_regs
*, u64
);
628 extern void __perf_counter_mmap(struct vm_area_struct
*vma
);
630 static inline void perf_counter_mmap(struct vm_area_struct
*vma
)
632 if (vma
->vm_flags
& VM_EXEC
)
633 __perf_counter_mmap(vma
);
636 extern void perf_counter_comm(struct task_struct
*tsk
);
637 extern void perf_counter_fork(struct task_struct
*tsk
);
639 extern void perf_counter_task_migration(struct task_struct
*task
, int cpu
);
641 #define MAX_STACK_DEPTH 255
643 struct perf_callchain_entry
{
648 u64 ip
[MAX_STACK_DEPTH
];
651 extern struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
);
653 extern int sysctl_perf_counter_paranoid
;
654 extern int sysctl_perf_counter_mlock
;
655 extern int sysctl_perf_counter_sample_rate
;
657 extern void perf_counter_init(void);
659 #ifndef perf_misc_flags
660 #define perf_misc_flags(regs) (user_mode(regs) ? PERF_EVENT_MISC_USER : \
661 PERF_EVENT_MISC_KERNEL)
662 #define perf_instruction_pointer(regs) instruction_pointer(regs)
667 perf_counter_task_sched_in(struct task_struct
*task
, int cpu
) { }
669 perf_counter_task_sched_out(struct task_struct
*task
,
670 struct task_struct
*next
, int cpu
) { }
672 perf_counter_task_tick(struct task_struct
*task
, int cpu
) { }
673 static inline int perf_counter_init_task(struct task_struct
*child
) { return 0; }
674 static inline void perf_counter_exit_task(struct task_struct
*child
) { }
675 static inline void perf_counter_free_task(struct task_struct
*task
) { }
676 static inline void perf_counter_do_pending(void) { }
677 static inline void perf_counter_print_debug(void) { }
678 static inline void perf_disable(void) { }
679 static inline void perf_enable(void) { }
680 static inline int perf_counter_task_disable(void) { return -EINVAL
; }
681 static inline int perf_counter_task_enable(void) { return -EINVAL
; }
684 perf_swcounter_event(u32 event
, u64 nr
, int nmi
,
685 struct pt_regs
*regs
, u64 addr
) { }
687 static inline void perf_counter_mmap(struct vm_area_struct
*vma
) { }
688 static inline void perf_counter_comm(struct task_struct
*tsk
) { }
689 static inline void perf_counter_fork(struct task_struct
*tsk
) { }
690 static inline void perf_counter_init(void) { }
691 static inline void perf_counter_task_migration(struct task_struct
*task
,
695 #endif /* __KERNEL__ */
696 #endif /* _LINUX_PERF_COUNTER_H */