2 * Performance counters:
4 * Copyright(C) 2008, Thomas Gleixner <tglx@linutronix.de>
5 * Copyright(C) 2008, Red Hat, Inc., Ingo Molnar
7 * Data type definitions, declarations, prototypes.
9 * Started by: Thomas Gleixner and Ingo Molnar
11 * For licencing details see kernel-base/COPYING
13 #ifndef _LINUX_PERF_COUNTER_H
14 #define _LINUX_PERF_COUNTER_H
16 #include <linux/types.h>
17 #include <linux/ioctl.h>
18 #include <asm/byteorder.h>
21 * User-space ABI bits:
27 enum perf_event_types
{
28 PERF_TYPE_HARDWARE
= 0,
29 PERF_TYPE_SOFTWARE
= 1,
30 PERF_TYPE_TRACEPOINT
= 2,
33 * available TYPE space, raw is the max value.
40 * Generalized performance counter event types, used by the attr.event_id
41 * parameter of the sys_perf_counter_open() syscall:
45 * Common hardware events, generalized by the kernel:
47 PERF_COUNT_CPU_CYCLES
= 0,
48 PERF_COUNT_INSTRUCTIONS
= 1,
49 PERF_COUNT_CACHE_REFERENCES
= 2,
50 PERF_COUNT_CACHE_MISSES
= 3,
51 PERF_COUNT_BRANCH_INSTRUCTIONS
= 4,
52 PERF_COUNT_BRANCH_MISSES
= 5,
53 PERF_COUNT_BUS_CYCLES
= 6,
55 PERF_HW_EVENTS_MAX
= 7,
59 * Special "software" counters provided by the kernel, even if the hardware
60 * does not support performance counters. These counters measure various
61 * physical and sw events of the kernel (and allow the profiling of them as
65 PERF_COUNT_CPU_CLOCK
= 0,
66 PERF_COUNT_TASK_CLOCK
= 1,
67 PERF_COUNT_PAGE_FAULTS
= 2,
68 PERF_COUNT_CONTEXT_SWITCHES
= 3,
69 PERF_COUNT_CPU_MIGRATIONS
= 4,
70 PERF_COUNT_PAGE_FAULTS_MIN
= 5,
71 PERF_COUNT_PAGE_FAULTS_MAJ
= 6,
73 PERF_SW_EVENTS_MAX
= 7,
76 #define __PERF_COUNTER_MASK(name) \
77 (((1ULL << PERF_COUNTER_##name##_BITS) - 1) << \
78 PERF_COUNTER_##name##_SHIFT)
80 #define PERF_COUNTER_RAW_BITS 1
81 #define PERF_COUNTER_RAW_SHIFT 63
82 #define PERF_COUNTER_RAW_MASK __PERF_COUNTER_MASK(RAW)
84 #define PERF_COUNTER_CONFIG_BITS 63
85 #define PERF_COUNTER_CONFIG_SHIFT 0
86 #define PERF_COUNTER_CONFIG_MASK __PERF_COUNTER_MASK(CONFIG)
88 #define PERF_COUNTER_TYPE_BITS 7
89 #define PERF_COUNTER_TYPE_SHIFT 56
90 #define PERF_COUNTER_TYPE_MASK __PERF_COUNTER_MASK(TYPE)
92 #define PERF_COUNTER_EVENT_BITS 56
93 #define PERF_COUNTER_EVENT_SHIFT 0
94 #define PERF_COUNTER_EVENT_MASK __PERF_COUNTER_MASK(EVENT)
97 * Bits that can be set in attr.sample_type to request information
98 * in the overflow packets.
100 enum perf_counter_sample_format
{
101 PERF_SAMPLE_IP
= 1U << 0,
102 PERF_SAMPLE_TID
= 1U << 1,
103 PERF_SAMPLE_TIME
= 1U << 2,
104 PERF_SAMPLE_ADDR
= 1U << 3,
105 PERF_SAMPLE_GROUP
= 1U << 4,
106 PERF_SAMPLE_CALLCHAIN
= 1U << 5,
107 PERF_SAMPLE_ID
= 1U << 6,
108 PERF_SAMPLE_CPU
= 1U << 7,
109 PERF_SAMPLE_PERIOD
= 1U << 8,
113 * Bits that can be set in attr.read_format to request that
114 * reads on the counter should return the indicated quantities,
115 * in increasing order of bit value, after the counter value.
117 enum perf_counter_read_format
{
118 PERF_FORMAT_TOTAL_TIME_ENABLED
= 1U << 0,
119 PERF_FORMAT_TOTAL_TIME_RUNNING
= 1U << 1,
120 PERF_FORMAT_ID
= 1U << 2,
124 * Hardware event to monitor via a performance monitoring counter:
126 struct perf_counter_attr
{
128 * The MSB of the config word signifies if the rest contains cpu
129 * specific (raw) counter configuration data, if unset, the next
130 * 7 bits are an event type and the rest of the bits are the event
143 __u64 disabled
: 1, /* off by default */
144 inherit
: 1, /* children inherit it */
145 pinned
: 1, /* must always be on PMU */
146 exclusive
: 1, /* only group on PMU */
147 exclude_user
: 1, /* don't count user */
148 exclude_kernel
: 1, /* ditto kernel */
149 exclude_hv
: 1, /* ditto hypervisor */
150 exclude_idle
: 1, /* don't count when idle */
151 mmap
: 1, /* include mmap data */
152 comm
: 1, /* include comm data */
153 freq
: 1, /* use freq, not period */
157 __u32 wakeup_events
; /* wakeup every n events */
165 * Ioctls that can be done on a perf counter fd:
167 #define PERF_COUNTER_IOC_ENABLE _IO ('$', 0)
168 #define PERF_COUNTER_IOC_DISABLE _IO ('$', 1)
169 #define PERF_COUNTER_IOC_REFRESH _IO ('$', 2)
170 #define PERF_COUNTER_IOC_RESET _IO ('$', 3)
171 #define PERF_COUNTER_IOC_PERIOD _IOW('$', 4, u64)
173 enum perf_counter_ioc_flags
{
174 PERF_IOC_FLAG_GROUP
= 1U << 0,
178 * Structure of the page that can be mapped via mmap
180 struct perf_counter_mmap_page
{
181 __u32 version
; /* version number of this structure */
182 __u32 compat_version
; /* lowest version this is compat with */
185 * Bits needed to read the hw counters in user-space.
195 * count = pmc_read(pc->index - 1);
196 * count += pc->offset;
201 * } while (pc->lock != seq);
203 * NOTE: for obvious reason this only works on self-monitoring
206 __u32 lock
; /* seqlock for synchronization */
207 __u32 index
; /* hardware counter identifier */
208 __s64 offset
; /* add to hardware counter value */
211 * Control data for the mmap() data buffer.
213 * User-space reading this value should issue an rmb(), on SMP capable
214 * platforms, after reading this value -- see perf_counter_wakeup().
216 __u64 data_head
; /* head in the data section */
219 #define PERF_EVENT_MISC_CPUMODE_MASK (3 << 0)
220 #define PERF_EVENT_MISC_CPUMODE_UNKNOWN (0 << 0)
221 #define PERF_EVENT_MISC_KERNEL (1 << 0)
222 #define PERF_EVENT_MISC_USER (2 << 0)
223 #define PERF_EVENT_MISC_HYPERVISOR (3 << 0)
224 #define PERF_EVENT_MISC_OVERFLOW (1 << 2)
226 struct perf_event_header
{
232 enum perf_event_type
{
235 * The MMAP events record the PROT_EXEC mappings so that we can
236 * correlate userspace IPs to code. They have the following structure:
239 * struct perf_event_header header;
252 * struct perf_event_header header;
262 * struct perf_event_header header;
268 PERF_EVENT_PERIOD
= 4,
272 * struct perf_event_header header;
276 PERF_EVENT_THROTTLE
= 5,
277 PERF_EVENT_UNTHROTTLE
= 6,
281 * struct perf_event_header header;
288 * When header.misc & PERF_EVENT_MISC_OVERFLOW the event_type field
289 * will be PERF_RECORD_*
292 * struct perf_event_header header;
294 * { u64 ip; } && PERF_RECORD_IP
295 * { u32 pid, tid; } && PERF_RECORD_TID
296 * { u64 time; } && PERF_RECORD_TIME
297 * { u64 addr; } && PERF_RECORD_ADDR
298 * { u64 config; } && PERF_RECORD_CONFIG
299 * { u32 cpu, res; } && PERF_RECORD_CPU
302 * { u64 id, val; } cnt[nr]; } && PERF_RECORD_GROUP
308 * u64 ips[nr]; } && PERF_RECORD_CALLCHAIN
315 * Kernel-internal data types and definitions:
318 #ifdef CONFIG_PERF_COUNTERS
319 # include <asm/perf_counter.h>
322 #include <linux/list.h>
323 #include <linux/mutex.h>
324 #include <linux/rculist.h>
325 #include <linux/rcupdate.h>
326 #include <linux/spinlock.h>
327 #include <linux/hrtimer.h>
328 #include <linux/fs.h>
329 #include <linux/pid_namespace.h>
330 #include <asm/atomic.h>
334 static inline u64
perf_event_raw(struct perf_counter_attr
*attr
)
336 return attr
->config
& PERF_COUNTER_RAW_MASK
;
339 static inline u64
perf_event_config(struct perf_counter_attr
*attr
)
341 return attr
->config
& PERF_COUNTER_CONFIG_MASK
;
344 static inline u64
perf_event_type(struct perf_counter_attr
*attr
)
346 return (attr
->config
& PERF_COUNTER_TYPE_MASK
) >>
347 PERF_COUNTER_TYPE_SHIFT
;
350 static inline u64
perf_event_id(struct perf_counter_attr
*attr
)
352 return attr
->config
& PERF_COUNTER_EVENT_MASK
;
356 * struct hw_perf_counter - performance counter hardware details:
358 struct hw_perf_counter
{
359 #ifdef CONFIG_PERF_COUNTERS
361 struct { /* hardware */
363 unsigned long config_base
;
364 unsigned long counter_base
;
367 union { /* software */
369 struct hrtimer hrtimer
;
372 atomic64_t prev_count
;
374 atomic64_t period_left
;
385 * struct pmu - generic performance monitoring unit
388 int (*enable
) (struct perf_counter
*counter
);
389 void (*disable
) (struct perf_counter
*counter
);
390 void (*read
) (struct perf_counter
*counter
);
391 void (*unthrottle
) (struct perf_counter
*counter
);
395 * enum perf_counter_active_state - the states of a counter
397 enum perf_counter_active_state
{
398 PERF_COUNTER_STATE_ERROR
= -2,
399 PERF_COUNTER_STATE_OFF
= -1,
400 PERF_COUNTER_STATE_INACTIVE
= 0,
401 PERF_COUNTER_STATE_ACTIVE
= 1,
406 struct perf_mmap_data
{
407 struct rcu_head rcu_head
;
408 int nr_pages
; /* nr of data pages */
409 int nr_locked
; /* nr pages mlocked */
411 atomic_t poll
; /* POLL_ for wakeups */
412 atomic_t events
; /* event limit */
414 atomic_long_t head
; /* write position */
415 atomic_long_t done_head
; /* completed head */
417 atomic_t lock
; /* concurrent writes */
419 atomic_t wakeup
; /* needs a wakeup */
421 struct perf_counter_mmap_page
*user_page
;
425 struct perf_pending_entry
{
426 struct perf_pending_entry
*next
;
427 void (*func
)(struct perf_pending_entry
*);
431 * struct perf_counter - performance counter kernel representation:
433 struct perf_counter
{
434 #ifdef CONFIG_PERF_COUNTERS
435 struct list_head list_entry
;
436 struct list_head event_entry
;
437 struct list_head sibling_list
;
439 struct perf_counter
*group_leader
;
440 const struct pmu
*pmu
;
442 enum perf_counter_active_state state
;
446 * These are the total time in nanoseconds that the counter
447 * has been enabled (i.e. eligible to run, and the task has
448 * been scheduled in, if this is a per-task counter)
449 * and running (scheduled onto the CPU), respectively.
451 * They are computed from tstamp_enabled, tstamp_running and
452 * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
454 u64 total_time_enabled
;
455 u64 total_time_running
;
458 * These are timestamps used for computing total_time_enabled
459 * and total_time_running when the counter is in INACTIVE or
460 * ACTIVE state, measured in nanoseconds from an arbitrary point
462 * tstamp_enabled: the notional time when the counter was enabled
463 * tstamp_running: the notional time when the counter was scheduled on
464 * tstamp_stopped: in INACTIVE state, the notional time when the
465 * counter was scheduled off.
471 struct perf_counter_attr attr
;
472 struct hw_perf_counter hw
;
474 struct perf_counter_context
*ctx
;
478 * These accumulate total time (in nanoseconds) that children
479 * counters have been enabled and running, respectively.
481 atomic64_t child_total_time_enabled
;
482 atomic64_t child_total_time_running
;
485 * Protect attach/detach and child_list:
487 struct mutex child_mutex
;
488 struct list_head child_list
;
489 struct perf_counter
*parent
;
494 struct list_head owner_entry
;
495 struct task_struct
*owner
;
498 struct mutex mmap_mutex
;
500 struct perf_mmap_data
*data
;
503 wait_queue_head_t waitq
;
504 struct fasync_struct
*fasync
;
506 /* delayed work for NMIs and such */
510 struct perf_pending_entry pending
;
512 atomic_t event_limit
;
514 void (*destroy
)(struct perf_counter
*);
515 struct rcu_head rcu_head
;
517 struct pid_namespace
*ns
;
523 * struct perf_counter_context - counter context structure
525 * Used as a container for task counters and CPU counters as well:
527 struct perf_counter_context
{
529 * Protect the states of the counters in the list,
530 * nr_active, and the list:
534 * Protect the list of counters. Locking either mutex or lock
535 * is sufficient to ensure the list doesn't change; to change
536 * the list you need to lock both the mutex and the spinlock.
540 struct list_head counter_list
;
541 struct list_head event_list
;
546 struct task_struct
*task
;
549 * Context clock, runs when context enabled.
555 * These fields let us detect when two contexts have both
556 * been cloned (inherited) from a common ancestor.
558 struct perf_counter_context
*parent_ctx
;
562 struct rcu_head rcu_head
;
566 * struct perf_counter_cpu_context - per cpu counter context structure
568 struct perf_cpu_context
{
569 struct perf_counter_context ctx
;
570 struct perf_counter_context
*task_ctx
;
576 * Recursion avoidance:
578 * task, softirq, irq, nmi context
583 #ifdef CONFIG_PERF_COUNTERS
586 * Set by architecture code:
588 extern int perf_max_counters
;
590 extern const struct pmu
*hw_perf_counter_init(struct perf_counter
*counter
);
592 extern void perf_counter_task_sched_in(struct task_struct
*task
, int cpu
);
593 extern void perf_counter_task_sched_out(struct task_struct
*task
,
594 struct task_struct
*next
, int cpu
);
595 extern void perf_counter_task_tick(struct task_struct
*task
, int cpu
);
596 extern int perf_counter_init_task(struct task_struct
*child
);
597 extern void perf_counter_exit_task(struct task_struct
*child
);
598 extern void perf_counter_free_task(struct task_struct
*task
);
599 extern void perf_counter_do_pending(void);
600 extern void perf_counter_print_debug(void);
601 extern void __perf_disable(void);
602 extern bool __perf_enable(void);
603 extern void perf_disable(void);
604 extern void perf_enable(void);
605 extern int perf_counter_task_disable(void);
606 extern int perf_counter_task_enable(void);
607 extern int hw_perf_group_sched_in(struct perf_counter
*group_leader
,
608 struct perf_cpu_context
*cpuctx
,
609 struct perf_counter_context
*ctx
, int cpu
);
610 extern void perf_counter_update_userpage(struct perf_counter
*counter
);
612 extern int perf_counter_overflow(struct perf_counter
*counter
,
613 int nmi
, struct pt_regs
*regs
, u64 addr
);
615 * Return 1 for a software counter, 0 for a hardware counter
617 static inline int is_software_counter(struct perf_counter
*counter
)
619 return !perf_event_raw(&counter
->attr
) &&
620 perf_event_type(&counter
->attr
) != PERF_TYPE_HARDWARE
;
623 extern void perf_swcounter_event(u32
, u64
, int, struct pt_regs
*, u64
);
625 extern void __perf_counter_mmap(struct vm_area_struct
*vma
);
627 static inline void perf_counter_mmap(struct vm_area_struct
*vma
)
629 if (vma
->vm_flags
& VM_EXEC
)
630 __perf_counter_mmap(vma
);
633 extern void perf_counter_comm(struct task_struct
*tsk
);
634 extern void perf_counter_fork(struct task_struct
*tsk
);
636 extern void perf_counter_task_migration(struct task_struct
*task
, int cpu
);
638 #define MAX_STACK_DEPTH 255
640 struct perf_callchain_entry
{
641 u16 nr
, hv
, kernel
, user
;
642 u64 ip
[MAX_STACK_DEPTH
];
645 extern struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
);
647 extern int sysctl_perf_counter_priv
;
648 extern int sysctl_perf_counter_mlock
;
649 extern int sysctl_perf_counter_limit
;
651 extern void perf_counter_init(void);
653 #ifndef perf_misc_flags
654 #define perf_misc_flags(regs) (user_mode(regs) ? PERF_EVENT_MISC_USER : \
655 PERF_EVENT_MISC_KERNEL)
656 #define perf_instruction_pointer(regs) instruction_pointer(regs)
661 perf_counter_task_sched_in(struct task_struct
*task
, int cpu
) { }
663 perf_counter_task_sched_out(struct task_struct
*task
,
664 struct task_struct
*next
, int cpu
) { }
666 perf_counter_task_tick(struct task_struct
*task
, int cpu
) { }
667 static inline int perf_counter_init_task(struct task_struct
*child
) { return 0; }
668 static inline void perf_counter_exit_task(struct task_struct
*child
) { }
669 static inline void perf_counter_free_task(struct task_struct
*task
) { }
670 static inline void perf_counter_do_pending(void) { }
671 static inline void perf_counter_print_debug(void) { }
672 static inline void perf_disable(void) { }
673 static inline void perf_enable(void) { }
674 static inline int perf_counter_task_disable(void) { return -EINVAL
; }
675 static inline int perf_counter_task_enable(void) { return -EINVAL
; }
678 perf_swcounter_event(u32 event
, u64 nr
, int nmi
,
679 struct pt_regs
*regs
, u64 addr
) { }
681 static inline void perf_counter_mmap(struct vm_area_struct
*vma
) { }
682 static inline void perf_counter_comm(struct task_struct
*tsk
) { }
683 static inline void perf_counter_fork(struct task_struct
*tsk
) { }
684 static inline void perf_counter_init(void) { }
685 static inline void perf_counter_task_migration(struct task_struct
*task
,
689 #endif /* __KERNEL__ */
690 #endif /* _LINUX_PERF_COUNTER_H */