4 * Copyright (C) 2008-2009, Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2011, Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2011, 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 <uapi/linux/perf_event.h>
20 * Kernel-internal data types and definitions:
23 #ifdef CONFIG_PERF_EVENTS
24 # include <linux/cgroup.h>
25 # include <asm/perf_event.h>
26 # include <asm/local64.h>
29 struct perf_guest_info_callbacks
{
30 int (*is_in_guest
)(void);
31 int (*is_user_mode
)(void);
32 unsigned long (*get_guest_ip
)(void);
35 #ifdef CONFIG_HAVE_HW_BREAKPOINT
36 #include <asm/hw_breakpoint.h>
39 #include <linux/list.h>
40 #include <linux/mutex.h>
41 #include <linux/rculist.h>
42 #include <linux/rcupdate.h>
43 #include <linux/spinlock.h>
44 #include <linux/hrtimer.h>
46 #include <linux/pid_namespace.h>
47 #include <linux/workqueue.h>
48 #include <linux/ftrace.h>
49 #include <linux/cpu.h>
50 #include <linux/irq_work.h>
51 #include <linux/static_key.h>
52 #include <linux/atomic.h>
53 #include <linux/sysfs.h>
54 #include <linux/perf_regs.h>
55 #include <asm/local.h>
57 struct perf_callchain_entry
{
59 __u64 ip
[PERF_MAX_STACK_DEPTH
];
62 struct perf_raw_record
{
68 * single taken branch record layout:
70 * from: source instruction (may not always be a branch insn)
72 * mispred: branch target was mispredicted
73 * predicted: branch target was predicted
75 * support for mispred, predicted is optional. In case it
76 * is not supported mispred = predicted = 0.
78 struct perf_branch_entry
{
81 __u64 mispred
:1, /* target mispredicted */
82 predicted
:1,/* target predicted */
87 * branch stack layout:
88 * nr: number of taken branches stored in entries[]
90 * Note that nr can vary from sample to sample
91 * branches (to, from) are stored from most recent
92 * to least recent, i.e., entries[0] contains the most
95 struct perf_branch_stack
{
97 struct perf_branch_entry entries
[0];
100 struct perf_regs_user
{
102 struct pt_regs
*regs
;
108 * extra PMU register associated with an event
110 struct hw_perf_event_extra
{
111 u64 config
; /* register value */
112 unsigned int reg
; /* register address or index */
113 int alloc
; /* extra register already allocated */
114 int idx
; /* index in shared_regs->regs[] */
118 * struct hw_perf_event - performance event hardware details:
120 struct hw_perf_event
{
121 #ifdef CONFIG_PERF_EVENTS
123 struct { /* hardware */
126 unsigned long config_base
;
127 unsigned long event_base
;
128 int event_base_rdpmc
;
132 struct hw_perf_event_extra extra_reg
;
133 struct hw_perf_event_extra branch_reg
;
135 struct { /* software */
136 struct hrtimer hrtimer
;
138 #ifdef CONFIG_HAVE_HW_BREAKPOINT
139 struct { /* breakpoint */
140 struct arch_hw_breakpoint info
;
141 struct list_head bp_list
;
143 * Crufty hack to avoid the chicken and egg
144 * problem hw_breakpoint has with context
145 * creation and event initalization.
147 struct task_struct
*bp_target
;
152 local64_t prev_count
;
155 local64_t period_left
;
160 u64 freq_count_stamp
;
165 * hw_perf_event::state flags
167 #define PERF_HES_STOPPED 0x01 /* the counter is stopped */
168 #define PERF_HES_UPTODATE 0x02 /* event->count up-to-date */
169 #define PERF_HES_ARCH 0x04
174 * Common implementation detail of pmu::{start,commit,cancel}_txn
176 #define PERF_EVENT_TXN 0x1
179 * struct pmu - generic performance monitoring unit
182 struct list_head entry
;
185 const struct attribute_group
**attr_groups
;
189 int * __percpu pmu_disable_count
;
190 struct perf_cpu_context
* __percpu pmu_cpu_context
;
194 * Fully disable/enable this PMU, can be used to protect from the PMI
195 * as well as for lazy/batch writing of the MSRs.
197 void (*pmu_enable
) (struct pmu
*pmu
); /* optional */
198 void (*pmu_disable
) (struct pmu
*pmu
); /* optional */
201 * Try and initialize the event for this PMU.
202 * Should return -ENOENT when the @event doesn't match this PMU.
204 int (*event_init
) (struct perf_event
*event
);
206 #define PERF_EF_START 0x01 /* start the counter when adding */
207 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
208 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
211 * Adds/Removes a counter to/from the PMU, can be done inside
212 * a transaction, see the ->*_txn() methods.
214 int (*add
) (struct perf_event
*event
, int flags
);
215 void (*del
) (struct perf_event
*event
, int flags
);
218 * Starts/Stops a counter present on the PMU. The PMI handler
219 * should stop the counter when perf_event_overflow() returns
220 * !0. ->start() will be used to continue.
222 void (*start
) (struct perf_event
*event
, int flags
);
223 void (*stop
) (struct perf_event
*event
, int flags
);
226 * Updates the counter value of the event.
228 void (*read
) (struct perf_event
*event
);
231 * Group events scheduling is treated as a transaction, add
232 * group events as a whole and perform one schedulability test.
233 * If the test fails, roll back the whole group
235 * Start the transaction, after this ->add() doesn't need to
236 * do schedulability tests.
238 void (*start_txn
) (struct pmu
*pmu
); /* optional */
240 * If ->start_txn() disabled the ->add() schedulability test
241 * then ->commit_txn() is required to perform one. On success
242 * the transaction is closed. On error the transaction is kept
243 * open until ->cancel_txn() is called.
245 int (*commit_txn
) (struct pmu
*pmu
); /* optional */
247 * Will cancel the transaction, assumes ->del() is called
248 * for each successful ->add() during the transaction.
250 void (*cancel_txn
) (struct pmu
*pmu
); /* optional */
253 * Will return the value for perf_event_mmap_page::index for this event,
254 * if no implementation is provided it will default to: event->hw.idx + 1.
256 int (*event_idx
) (struct perf_event
*event
); /*optional */
259 * flush branch stack on context-switches (needed in cpu-wide mode)
261 void (*flush_branch_stack
) (void);
265 * enum perf_event_active_state - the states of a event
267 enum perf_event_active_state
{
268 PERF_EVENT_STATE_ERROR
= -2,
269 PERF_EVENT_STATE_OFF
= -1,
270 PERF_EVENT_STATE_INACTIVE
= 0,
271 PERF_EVENT_STATE_ACTIVE
= 1,
275 struct perf_sample_data
;
277 typedef void (*perf_overflow_handler_t
)(struct perf_event
*,
278 struct perf_sample_data
*,
279 struct pt_regs
*regs
);
281 enum perf_group_flag
{
282 PERF_GROUP_SOFTWARE
= 0x1,
285 #define SWEVENT_HLIST_BITS 8
286 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
288 struct swevent_hlist
{
289 struct hlist_head heads
[SWEVENT_HLIST_SIZE
];
290 struct rcu_head rcu_head
;
293 #define PERF_ATTACH_CONTEXT 0x01
294 #define PERF_ATTACH_GROUP 0x02
295 #define PERF_ATTACH_TASK 0x04
297 #ifdef CONFIG_CGROUP_PERF
299 * perf_cgroup_info keeps track of time_enabled for a cgroup.
300 * This is a per-cpu dynamically allocated data structure.
302 struct perf_cgroup_info
{
308 struct cgroup_subsys_state css
;
309 struct perf_cgroup_info
*info
; /* timing info, one per cpu */
316 * struct perf_event - performance event kernel representation:
319 #ifdef CONFIG_PERF_EVENTS
320 struct list_head group_entry
;
321 struct list_head event_entry
;
322 struct list_head sibling_list
;
323 struct hlist_node hlist_entry
;
326 struct perf_event
*group_leader
;
329 enum perf_event_active_state state
;
330 unsigned int attach_state
;
332 atomic64_t child_count
;
335 * These are the total time in nanoseconds that the event
336 * has been enabled (i.e. eligible to run, and the task has
337 * been scheduled in, if this is a per-task event)
338 * and running (scheduled onto the CPU), respectively.
340 * They are computed from tstamp_enabled, tstamp_running and
341 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
343 u64 total_time_enabled
;
344 u64 total_time_running
;
347 * These are timestamps used for computing total_time_enabled
348 * and total_time_running when the event is in INACTIVE or
349 * ACTIVE state, measured in nanoseconds from an arbitrary point
351 * tstamp_enabled: the notional time when the event was enabled
352 * tstamp_running: the notional time when the event was scheduled on
353 * tstamp_stopped: in INACTIVE state, the notional time when the
354 * event was scheduled off.
361 * timestamp shadows the actual context timing but it can
362 * be safely used in NMI interrupt context. It reflects the
363 * context time as it was when the event was last scheduled in.
365 * ctx_time already accounts for ctx->timestamp. Therefore to
366 * compute ctx_time for a sample, simply add perf_clock().
370 struct perf_event_attr attr
;
374 struct hw_perf_event hw
;
376 struct perf_event_context
*ctx
;
377 atomic_long_t refcount
;
380 * These accumulate total time (in nanoseconds) that children
381 * events have been enabled and running, respectively.
383 atomic64_t child_total_time_enabled
;
384 atomic64_t child_total_time_running
;
387 * Protect attach/detach and child_list:
389 struct mutex child_mutex
;
390 struct list_head child_list
;
391 struct perf_event
*parent
;
396 struct list_head owner_entry
;
397 struct task_struct
*owner
;
400 struct mutex mmap_mutex
;
403 struct user_struct
*mmap_user
;
404 struct ring_buffer
*rb
;
405 struct list_head rb_entry
;
408 wait_queue_head_t waitq
;
409 struct fasync_struct
*fasync
;
411 /* delayed work for NMIs and such */
415 struct irq_work pending
;
417 atomic_t event_limit
;
419 void (*destroy
)(struct perf_event
*);
420 struct rcu_head rcu_head
;
422 struct pid_namespace
*ns
;
425 perf_overflow_handler_t overflow_handler
;
426 void *overflow_handler_context
;
428 #ifdef CONFIG_EVENT_TRACING
429 struct ftrace_event_call
*tp_event
;
430 struct event_filter
*filter
;
431 #ifdef CONFIG_FUNCTION_TRACER
432 struct ftrace_ops ftrace_ops
;
436 #ifdef CONFIG_CGROUP_PERF
437 struct perf_cgroup
*cgrp
; /* cgroup event is attach to */
438 int cgrp_defer_enabled
;
441 #endif /* CONFIG_PERF_EVENTS */
444 enum perf_event_context_type
{
450 * struct perf_event_context - event context structure
452 * Used as a container for task events and CPU events as well:
454 struct perf_event_context
{
456 enum perf_event_context_type type
;
458 * Protect the states of the events in the list,
459 * nr_active, and the list:
463 * Protect the list of events. Locking either mutex or lock
464 * is sufficient to ensure the list doesn't change; to change
465 * the list you need to lock both the mutex and the spinlock.
469 struct list_head pinned_groups
;
470 struct list_head flexible_groups
;
471 struct list_head event_list
;
479 struct task_struct
*task
;
482 * Context clock, runs when context enabled.
488 * These fields let us detect when two contexts have both
489 * been cloned (inherited) from a common ancestor.
491 struct perf_event_context
*parent_ctx
;
495 int nr_cgroups
; /* cgroup evts */
496 int nr_branch_stack
; /* branch_stack evt */
497 struct rcu_head rcu_head
;
501 * Number of contexts where an event can trigger:
502 * task, softirq, hardirq, nmi.
504 #define PERF_NR_CONTEXTS 4
507 * struct perf_event_cpu_context - per cpu event context structure
509 struct perf_cpu_context
{
510 struct perf_event_context ctx
;
511 struct perf_event_context
*task_ctx
;
514 struct list_head rotation_list
;
515 int jiffies_interval
;
516 struct pmu
*unique_pmu
;
517 struct perf_cgroup
*cgrp
;
520 struct perf_output_handle
{
521 struct perf_event
*event
;
522 struct ring_buffer
*rb
;
523 unsigned long wakeup
;
529 #ifdef CONFIG_PERF_EVENTS
531 extern int perf_pmu_register(struct pmu
*pmu
, char *name
, int type
);
532 extern void perf_pmu_unregister(struct pmu
*pmu
);
534 extern int perf_num_counters(void);
535 extern const char *perf_pmu_name(void);
536 extern void __perf_event_task_sched_in(struct task_struct
*prev
,
537 struct task_struct
*task
);
538 extern void __perf_event_task_sched_out(struct task_struct
*prev
,
539 struct task_struct
*next
);
540 extern int perf_event_init_task(struct task_struct
*child
);
541 extern void perf_event_exit_task(struct task_struct
*child
);
542 extern void perf_event_free_task(struct task_struct
*task
);
543 extern void perf_event_delayed_put(struct task_struct
*task
);
544 extern void perf_event_print_debug(void);
545 extern void perf_pmu_disable(struct pmu
*pmu
);
546 extern void perf_pmu_enable(struct pmu
*pmu
);
547 extern int perf_event_task_disable(void);
548 extern int perf_event_task_enable(void);
549 extern int perf_event_refresh(struct perf_event
*event
, int refresh
);
550 extern void perf_event_update_userpage(struct perf_event
*event
);
551 extern int perf_event_release_kernel(struct perf_event
*event
);
552 extern struct perf_event
*
553 perf_event_create_kernel_counter(struct perf_event_attr
*attr
,
555 struct task_struct
*task
,
556 perf_overflow_handler_t callback
,
558 extern void perf_pmu_migrate_context(struct pmu
*pmu
,
559 int src_cpu
, int dst_cpu
);
560 extern u64
perf_event_read_value(struct perf_event
*event
,
561 u64
*enabled
, u64
*running
);
564 struct perf_sample_data
{
581 struct perf_callchain_entry
*callchain
;
582 struct perf_raw_record
*raw
;
583 struct perf_branch_stack
*br_stack
;
584 struct perf_regs_user regs_user
;
588 static inline void perf_sample_data_init(struct perf_sample_data
*data
,
589 u64 addr
, u64 period
)
591 /* remaining struct members initialized in perf_prepare_sample() */
594 data
->br_stack
= NULL
;
595 data
->period
= period
;
596 data
->regs_user
.abi
= PERF_SAMPLE_REGS_ABI_NONE
;
597 data
->regs_user
.regs
= NULL
;
598 data
->stack_user_size
= 0;
601 extern void perf_output_sample(struct perf_output_handle
*handle
,
602 struct perf_event_header
*header
,
603 struct perf_sample_data
*data
,
604 struct perf_event
*event
);
605 extern void perf_prepare_sample(struct perf_event_header
*header
,
606 struct perf_sample_data
*data
,
607 struct perf_event
*event
,
608 struct pt_regs
*regs
);
610 extern int perf_event_overflow(struct perf_event
*event
,
611 struct perf_sample_data
*data
,
612 struct pt_regs
*regs
);
614 static inline bool is_sampling_event(struct perf_event
*event
)
616 return event
->attr
.sample_period
!= 0;
620 * Return 1 for a software event, 0 for a hardware event
622 static inline int is_software_event(struct perf_event
*event
)
624 return event
->pmu
->task_ctx_nr
== perf_sw_context
;
627 extern struct static_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
629 extern void __perf_sw_event(u32
, u64
, struct pt_regs
*, u64
);
631 #ifndef perf_arch_fetch_caller_regs
632 static inline void perf_arch_fetch_caller_regs(struct pt_regs
*regs
, unsigned long ip
) { }
636 * Take a snapshot of the regs. Skip ip and frame pointer to
637 * the nth caller. We only need a few of the regs:
638 * - ip for PERF_SAMPLE_IP
639 * - cs for user_mode() tests
640 * - bp for callchains
641 * - eflags, for future purposes, just in case
643 static inline void perf_fetch_caller_regs(struct pt_regs
*regs
)
645 memset(regs
, 0, sizeof(*regs
));
647 perf_arch_fetch_caller_regs(regs
, CALLER_ADDR0
);
650 static __always_inline
void
651 perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
653 struct pt_regs hot_regs
;
655 if (static_key_false(&perf_swevent_enabled
[event_id
])) {
657 perf_fetch_caller_regs(&hot_regs
);
660 __perf_sw_event(event_id
, nr
, regs
, addr
);
664 extern struct static_key_deferred perf_sched_events
;
666 static inline void perf_event_task_sched_in(struct task_struct
*prev
,
667 struct task_struct
*task
)
669 if (static_key_false(&perf_sched_events
.key
))
670 __perf_event_task_sched_in(prev
, task
);
673 static inline void perf_event_task_sched_out(struct task_struct
*prev
,
674 struct task_struct
*next
)
676 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES
, 1, NULL
, 0);
678 if (static_key_false(&perf_sched_events
.key
))
679 __perf_event_task_sched_out(prev
, next
);
682 extern void perf_event_mmap(struct vm_area_struct
*vma
);
683 extern struct perf_guest_info_callbacks
*perf_guest_cbs
;
684 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*callbacks
);
685 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*callbacks
);
687 extern void perf_event_comm(struct task_struct
*tsk
);
688 extern void perf_event_fork(struct task_struct
*tsk
);
691 DECLARE_PER_CPU(struct perf_callchain_entry
, perf_callchain_entry
);
693 extern void perf_callchain_user(struct perf_callchain_entry
*entry
, struct pt_regs
*regs
);
694 extern void perf_callchain_kernel(struct perf_callchain_entry
*entry
, struct pt_regs
*regs
);
696 static inline void perf_callchain_store(struct perf_callchain_entry
*entry
, u64 ip
)
698 if (entry
->nr
< PERF_MAX_STACK_DEPTH
)
699 entry
->ip
[entry
->nr
++] = ip
;
702 extern int sysctl_perf_event_paranoid
;
703 extern int sysctl_perf_event_mlock
;
704 extern int sysctl_perf_event_sample_rate
;
706 extern int perf_proc_update_handler(struct ctl_table
*table
, int write
,
707 void __user
*buffer
, size_t *lenp
,
710 static inline bool perf_paranoid_tracepoint_raw(void)
712 return sysctl_perf_event_paranoid
> -1;
715 static inline bool perf_paranoid_cpu(void)
717 return sysctl_perf_event_paranoid
> 0;
720 static inline bool perf_paranoid_kernel(void)
722 return sysctl_perf_event_paranoid
> 1;
725 extern void perf_event_init(void);
726 extern void perf_tp_event(u64 addr
, u64 count
, void *record
,
727 int entry_size
, struct pt_regs
*regs
,
728 struct hlist_head
*head
, int rctx
,
729 struct task_struct
*task
);
730 extern void perf_bp_event(struct perf_event
*event
, void *data
);
732 #ifndef perf_misc_flags
733 # define perf_misc_flags(regs) \
734 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
735 # define perf_instruction_pointer(regs) instruction_pointer(regs)
738 static inline bool has_branch_stack(struct perf_event
*event
)
740 return event
->attr
.sample_type
& PERF_SAMPLE_BRANCH_STACK
;
743 extern int perf_output_begin(struct perf_output_handle
*handle
,
744 struct perf_event
*event
, unsigned int size
);
745 extern void perf_output_end(struct perf_output_handle
*handle
);
746 extern unsigned int perf_output_copy(struct perf_output_handle
*handle
,
747 const void *buf
, unsigned int len
);
748 extern unsigned int perf_output_skip(struct perf_output_handle
*handle
,
750 extern int perf_swevent_get_recursion_context(void);
751 extern void perf_swevent_put_recursion_context(int rctx
);
752 extern void perf_event_enable(struct perf_event
*event
);
753 extern void perf_event_disable(struct perf_event
*event
);
754 extern int __perf_event_disable(void *info
);
755 extern void perf_event_task_tick(void);
758 perf_event_task_sched_in(struct task_struct
*prev
,
759 struct task_struct
*task
) { }
761 perf_event_task_sched_out(struct task_struct
*prev
,
762 struct task_struct
*next
) { }
763 static inline int perf_event_init_task(struct task_struct
*child
) { return 0; }
764 static inline void perf_event_exit_task(struct task_struct
*child
) { }
765 static inline void perf_event_free_task(struct task_struct
*task
) { }
766 static inline void perf_event_delayed_put(struct task_struct
*task
) { }
767 static inline void perf_event_print_debug(void) { }
768 static inline int perf_event_task_disable(void) { return -EINVAL
; }
769 static inline int perf_event_task_enable(void) { return -EINVAL
; }
770 static inline int perf_event_refresh(struct perf_event
*event
, int refresh
)
776 perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
) { }
778 perf_bp_event(struct perf_event
*event
, void *data
) { }
780 static inline int perf_register_guest_info_callbacks
781 (struct perf_guest_info_callbacks
*callbacks
) { return 0; }
782 static inline int perf_unregister_guest_info_callbacks
783 (struct perf_guest_info_callbacks
*callbacks
) { return 0; }
785 static inline void perf_event_mmap(struct vm_area_struct
*vma
) { }
786 static inline void perf_event_comm(struct task_struct
*tsk
) { }
787 static inline void perf_event_fork(struct task_struct
*tsk
) { }
788 static inline void perf_event_init(void) { }
789 static inline int perf_swevent_get_recursion_context(void) { return -1; }
790 static inline void perf_swevent_put_recursion_context(int rctx
) { }
791 static inline void perf_event_enable(struct perf_event
*event
) { }
792 static inline void perf_event_disable(struct perf_event
*event
) { }
793 static inline int __perf_event_disable(void *info
) { return -1; }
794 static inline void perf_event_task_tick(void) { }
797 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
800 * This has to have a higher priority than migration_notifier in sched.c.
802 #define perf_cpu_notifier(fn) \
804 static struct notifier_block fn##_nb __cpuinitdata = \
805 { .notifier_call = fn, .priority = CPU_PRI_PERF }; \
806 fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE, \
807 (void *)(unsigned long)smp_processor_id()); \
808 fn(&fn##_nb, (unsigned long)CPU_STARTING, \
809 (void *)(unsigned long)smp_processor_id()); \
810 fn(&fn##_nb, (unsigned long)CPU_ONLINE, \
811 (void *)(unsigned long)smp_processor_id()); \
812 register_cpu_notifier(&fn##_nb); \
816 #define PMU_FORMAT_ATTR(_name, _format) \
818 _name##_show(struct device *dev, \
819 struct device_attribute *attr, \
822 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
823 return sprintf(page, _format "\n"); \
826 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
828 #endif /* _LINUX_PERF_EVENT_H */