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 <asm/perf_event.h>
25 # include <asm/local64.h>
28 struct perf_guest_info_callbacks
{
29 int (*is_in_guest
)(void);
30 int (*is_user_mode
)(void);
31 unsigned long (*get_guest_ip
)(void);
34 #ifdef CONFIG_HAVE_HW_BREAKPOINT
35 #include <asm/hw_breakpoint.h>
38 #include <linux/list.h>
39 #include <linux/mutex.h>
40 #include <linux/rculist.h>
41 #include <linux/rcupdate.h>
42 #include <linux/spinlock.h>
43 #include <linux/hrtimer.h>
45 #include <linux/pid_namespace.h>
46 #include <linux/workqueue.h>
47 #include <linux/ftrace.h>
48 #include <linux/cpu.h>
49 #include <linux/irq_work.h>
50 #include <linux/static_key.h>
51 #include <linux/jump_label_ratelimit.h>
52 #include <linux/atomic.h>
53 #include <linux/sysfs.h>
54 #include <linux/perf_regs.h>
55 #include <linux/workqueue.h>
56 #include <linux/cgroup.h>
57 #include <asm/local.h>
59 struct perf_callchain_entry
{
61 __u64 ip
[PERF_MAX_STACK_DEPTH
];
64 struct perf_raw_record
{
70 * branch stack layout:
71 * nr: number of taken branches stored in entries[]
73 * Note that nr can vary from sample to sample
74 * branches (to, from) are stored from most recent
75 * to least recent, i.e., entries[0] contains the most
78 struct perf_branch_stack
{
80 struct perf_branch_entry entries
[0];
86 * extra PMU register associated with an event
88 struct hw_perf_event_extra
{
89 u64 config
; /* register value */
90 unsigned int reg
; /* register address or index */
91 int alloc
; /* extra register already allocated */
92 int idx
; /* index in shared_regs->regs[] */
96 * struct hw_perf_event - performance event hardware details:
98 struct hw_perf_event
{
99 #ifdef CONFIG_PERF_EVENTS
101 struct { /* hardware */
104 unsigned long config_base
;
105 unsigned long event_base
;
106 int event_base_rdpmc
;
111 struct hw_perf_event_extra extra_reg
;
112 struct hw_perf_event_extra branch_reg
;
114 struct { /* software */
115 struct hrtimer hrtimer
;
117 struct { /* tracepoint */
118 /* for tp_event->class */
119 struct list_head tp_list
;
121 struct { /* intel_cqm */
124 struct list_head cqm_events_entry
;
125 struct list_head cqm_groups_entry
;
126 struct list_head cqm_group_entry
;
128 struct { /* itrace */
131 #ifdef CONFIG_HAVE_HW_BREAKPOINT
132 struct { /* breakpoint */
134 * Crufty hack to avoid the chicken and egg
135 * problem hw_breakpoint has with context
136 * creation and event initalization.
138 struct arch_hw_breakpoint info
;
139 struct list_head bp_list
;
144 * If the event is a per task event, this will point to the task in
145 * question. See the comment in perf_event_alloc().
147 struct task_struct
*target
;
150 * hw_perf_event::state flags; used to track the PERF_EF_* state.
152 #define PERF_HES_STOPPED 0x01 /* the counter is stopped */
153 #define PERF_HES_UPTODATE 0x02 /* event->count up-to-date */
154 #define PERF_HES_ARCH 0x04
159 * The last observed hardware counter value, updated with a
160 * local64_cmpxchg() such that pmu::read() can be called nested.
162 local64_t prev_count
;
165 * The period to start the next sample with.
170 * The period we started this sample with.
175 * However much is left of the current period; note that this is
176 * a full 64bit value and allows for generation of periods longer
177 * than hardware might allow.
179 local64_t period_left
;
182 * State for throttling the event, see __perf_event_overflow() and
183 * perf_adjust_freq_unthr_context().
189 * State for freq target events, see __perf_event_overflow() and
190 * perf_adjust_freq_unthr_context().
193 u64 freq_count_stamp
;
200 * Common implementation detail of pmu::{start,commit,cancel}_txn
202 #define PERF_EVENT_TXN 0x1
205 * pmu::capabilities flags
207 #define PERF_PMU_CAP_NO_INTERRUPT 0x01
208 #define PERF_PMU_CAP_NO_NMI 0x02
209 #define PERF_PMU_CAP_AUX_NO_SG 0x04
210 #define PERF_PMU_CAP_AUX_SW_DOUBLEBUF 0x08
211 #define PERF_PMU_CAP_EXCLUSIVE 0x10
212 #define PERF_PMU_CAP_ITRACE 0x20
215 * struct pmu - generic performance monitoring unit
218 struct list_head entry
;
220 struct module
*module
;
222 const struct attribute_group
**attr_groups
;
227 * various common per-pmu feature flags
231 int * __percpu pmu_disable_count
;
232 struct perf_cpu_context
* __percpu pmu_cpu_context
;
233 atomic_t exclusive_cnt
; /* < 0: cpu; > 0: tsk */
235 int hrtimer_interval_ms
;
238 * Fully disable/enable this PMU, can be used to protect from the PMI
239 * as well as for lazy/batch writing of the MSRs.
241 void (*pmu_enable
) (struct pmu
*pmu
); /* optional */
242 void (*pmu_disable
) (struct pmu
*pmu
); /* optional */
245 * Try and initialize the event for this PMU.
248 * -ENOENT -- @event is not for this PMU
250 * -ENODEV -- @event is for this PMU but PMU not present
251 * -EBUSY -- @event is for this PMU but PMU temporarily unavailable
252 * -EINVAL -- @event is for this PMU but @event is not valid
253 * -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
254 * -EACCESS -- @event is for this PMU, @event is valid, but no privilidges
256 * 0 -- @event is for this PMU and valid
258 * Other error return values are allowed.
260 int (*event_init
) (struct perf_event
*event
);
263 * Notification that the event was mapped or unmapped. Called
264 * in the context of the mapping task.
266 void (*event_mapped
) (struct perf_event
*event
); /*optional*/
267 void (*event_unmapped
) (struct perf_event
*event
); /*optional*/
270 * Flags for ->add()/->del()/ ->start()/->stop(). There are
271 * matching hw_perf_event::state flags.
273 #define PERF_EF_START 0x01 /* start the counter when adding */
274 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
275 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
278 * Adds/Removes a counter to/from the PMU, can be done inside a
279 * transaction, see the ->*_txn() methods.
281 * The add/del callbacks will reserve all hardware resources required
282 * to service the event, this includes any counter constraint
285 * Called with IRQs disabled and the PMU disabled on the CPU the event
288 * ->add() called without PERF_EF_START should result in the same state
289 * as ->add() followed by ->stop().
291 * ->del() must always PERF_EF_UPDATE stop an event. If it calls
292 * ->stop() that must deal with already being stopped without
295 int (*add
) (struct perf_event
*event
, int flags
);
296 void (*del
) (struct perf_event
*event
, int flags
);
299 * Starts/Stops a counter present on the PMU.
301 * The PMI handler should stop the counter when perf_event_overflow()
302 * returns !0. ->start() will be used to continue.
304 * Also used to change the sample period.
306 * Called with IRQs disabled and the PMU disabled on the CPU the event
307 * is on -- will be called from NMI context with the PMU generates
310 * ->stop() with PERF_EF_UPDATE will read the counter and update
311 * period/count values like ->read() would.
313 * ->start() with PERF_EF_RELOAD will reprogram the the counter
314 * value, must be preceded by a ->stop() with PERF_EF_UPDATE.
316 void (*start
) (struct perf_event
*event
, int flags
);
317 void (*stop
) (struct perf_event
*event
, int flags
);
320 * Updates the counter value of the event.
322 * For sampling capable PMUs this will also update the software period
323 * hw_perf_event::period_left field.
325 void (*read
) (struct perf_event
*event
);
328 * Group events scheduling is treated as a transaction, add
329 * group events as a whole and perform one schedulability test.
330 * If the test fails, roll back the whole group
332 * Start the transaction, after this ->add() doesn't need to
333 * do schedulability tests.
335 void (*start_txn
) (struct pmu
*pmu
); /* optional */
337 * If ->start_txn() disabled the ->add() schedulability test
338 * then ->commit_txn() is required to perform one. On success
339 * the transaction is closed. On error the transaction is kept
340 * open until ->cancel_txn() is called.
342 int (*commit_txn
) (struct pmu
*pmu
); /* optional */
344 * Will cancel the transaction, assumes ->del() is called
345 * for each successful ->add() during the transaction.
347 void (*cancel_txn
) (struct pmu
*pmu
); /* optional */
350 * Will return the value for perf_event_mmap_page::index for this event,
351 * if no implementation is provided it will default to: event->hw.idx + 1.
353 int (*event_idx
) (struct perf_event
*event
); /*optional */
356 * context-switches callback
358 void (*sched_task
) (struct perf_event_context
*ctx
,
361 * PMU specific data size
363 size_t task_ctx_size
;
367 * Return the count value for a counter.
369 u64 (*count
) (struct perf_event
*event
); /*optional*/
372 * Set up pmu-private data structures for an AUX area
374 void *(*setup_aux
) (int cpu
, void **pages
,
375 int nr_pages
, bool overwrite
);
379 * Free pmu-private AUX data structures
381 void (*free_aux
) (void *aux
); /* optional */
384 * Filter events for PMU-specific reasons.
386 int (*filter_match
) (struct perf_event
*event
); /* optional */
390 * enum perf_event_active_state - the states of a event
392 enum perf_event_active_state
{
393 PERF_EVENT_STATE_EXIT
= -3,
394 PERF_EVENT_STATE_ERROR
= -2,
395 PERF_EVENT_STATE_OFF
= -1,
396 PERF_EVENT_STATE_INACTIVE
= 0,
397 PERF_EVENT_STATE_ACTIVE
= 1,
401 struct perf_sample_data
;
403 typedef void (*perf_overflow_handler_t
)(struct perf_event
*,
404 struct perf_sample_data
*,
405 struct pt_regs
*regs
);
407 enum perf_group_flag
{
408 PERF_GROUP_SOFTWARE
= 0x1,
411 #define SWEVENT_HLIST_BITS 8
412 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
414 struct swevent_hlist
{
415 struct hlist_head heads
[SWEVENT_HLIST_SIZE
];
416 struct rcu_head rcu_head
;
419 #define PERF_ATTACH_CONTEXT 0x01
420 #define PERF_ATTACH_GROUP 0x02
421 #define PERF_ATTACH_TASK 0x04
422 #define PERF_ATTACH_TASK_DATA 0x08
428 * struct perf_event - performance event kernel representation:
431 #ifdef CONFIG_PERF_EVENTS
433 * entry onto perf_event_context::event_list;
434 * modifications require ctx->lock
435 * RCU safe iterations.
437 struct list_head event_entry
;
440 * XXX: group_entry and sibling_list should be mutually exclusive;
441 * either you're a sibling on a group, or you're the group leader.
442 * Rework the code to always use the same list element.
444 * Locked for modification by both ctx->mutex and ctx->lock; holding
445 * either sufficies for read.
447 struct list_head group_entry
;
448 struct list_head sibling_list
;
451 * We need storage to track the entries in perf_pmu_migrate_context; we
452 * cannot use the event_entry because of RCU and we want to keep the
453 * group in tact which avoids us using the other two entries.
455 struct list_head migrate_entry
;
457 struct hlist_node hlist_entry
;
458 struct list_head active_entry
;
461 struct perf_event
*group_leader
;
464 enum perf_event_active_state state
;
465 unsigned int attach_state
;
467 atomic64_t child_count
;
470 * These are the total time in nanoseconds that the event
471 * has been enabled (i.e. eligible to run, and the task has
472 * been scheduled in, if this is a per-task event)
473 * and running (scheduled onto the CPU), respectively.
475 * They are computed from tstamp_enabled, tstamp_running and
476 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
478 u64 total_time_enabled
;
479 u64 total_time_running
;
482 * These are timestamps used for computing total_time_enabled
483 * and total_time_running when the event is in INACTIVE or
484 * ACTIVE state, measured in nanoseconds from an arbitrary point
486 * tstamp_enabled: the notional time when the event was enabled
487 * tstamp_running: the notional time when the event was scheduled on
488 * tstamp_stopped: in INACTIVE state, the notional time when the
489 * event was scheduled off.
496 * timestamp shadows the actual context timing but it can
497 * be safely used in NMI interrupt context. It reflects the
498 * context time as it was when the event was last scheduled in.
500 * ctx_time already accounts for ctx->timestamp. Therefore to
501 * compute ctx_time for a sample, simply add perf_clock().
505 struct perf_event_attr attr
;
509 struct hw_perf_event hw
;
511 struct perf_event_context
*ctx
;
512 atomic_long_t refcount
;
515 * These accumulate total time (in nanoseconds) that children
516 * events have been enabled and running, respectively.
518 atomic64_t child_total_time_enabled
;
519 atomic64_t child_total_time_running
;
522 * Protect attach/detach and child_list:
524 struct mutex child_mutex
;
525 struct list_head child_list
;
526 struct perf_event
*parent
;
531 struct list_head owner_entry
;
532 struct task_struct
*owner
;
535 struct mutex mmap_mutex
;
538 struct ring_buffer
*rb
;
539 struct list_head rb_entry
;
540 unsigned long rcu_batches
;
544 wait_queue_head_t waitq
;
545 struct fasync_struct
*fasync
;
547 /* delayed work for NMIs and such */
551 struct irq_work pending
;
553 atomic_t event_limit
;
555 void (*destroy
)(struct perf_event
*);
556 struct rcu_head rcu_head
;
558 struct pid_namespace
*ns
;
562 perf_overflow_handler_t overflow_handler
;
563 void *overflow_handler_context
;
565 #ifdef CONFIG_EVENT_TRACING
566 struct trace_event_call
*tp_event
;
567 struct event_filter
*filter
;
568 #ifdef CONFIG_FUNCTION_TRACER
569 struct ftrace_ops ftrace_ops
;
573 #ifdef CONFIG_CGROUP_PERF
574 struct perf_cgroup
*cgrp
; /* cgroup event is attach to */
575 int cgrp_defer_enabled
;
578 #endif /* CONFIG_PERF_EVENTS */
582 * struct perf_event_context - event context structure
584 * Used as a container for task events and CPU events as well:
586 struct perf_event_context
{
589 * Protect the states of the events in the list,
590 * nr_active, and the list:
594 * Protect the list of events. Locking either mutex or lock
595 * is sufficient to ensure the list doesn't change; to change
596 * the list you need to lock both the mutex and the spinlock.
600 struct list_head active_ctx_list
;
601 struct list_head pinned_groups
;
602 struct list_head flexible_groups
;
603 struct list_head event_list
;
611 struct task_struct
*task
;
614 * Context clock, runs when context enabled.
620 * These fields let us detect when two contexts have both
621 * been cloned (inherited) from a common ancestor.
623 struct perf_event_context
*parent_ctx
;
627 int nr_cgroups
; /* cgroup evts */
628 void *task_ctx_data
; /* pmu specific data */
629 struct rcu_head rcu_head
;
631 struct delayed_work orphans_remove
;
632 bool orphans_remove_sched
;
636 * Number of contexts where an event can trigger:
637 * task, softirq, hardirq, nmi.
639 #define PERF_NR_CONTEXTS 4
642 * struct perf_event_cpu_context - per cpu event context structure
644 struct perf_cpu_context
{
645 struct perf_event_context ctx
;
646 struct perf_event_context
*task_ctx
;
650 raw_spinlock_t hrtimer_lock
;
651 struct hrtimer hrtimer
;
652 ktime_t hrtimer_interval
;
653 unsigned int hrtimer_active
;
655 struct pmu
*unique_pmu
;
656 struct perf_cgroup
*cgrp
;
659 struct perf_output_handle
{
660 struct perf_event
*event
;
661 struct ring_buffer
*rb
;
662 unsigned long wakeup
;
671 #ifdef CONFIG_CGROUP_PERF
674 * perf_cgroup_info keeps track of time_enabled for a cgroup.
675 * This is a per-cpu dynamically allocated data structure.
677 struct perf_cgroup_info
{
683 struct cgroup_subsys_state css
;
684 struct perf_cgroup_info __percpu
*info
;
688 * Must ensure cgroup is pinned (css_get) before calling
689 * this function. In other words, we cannot call this function
690 * if there is no cgroup event for the current CPU context.
692 static inline struct perf_cgroup
*
693 perf_cgroup_from_task(struct task_struct
*task
)
695 return container_of(task_css(task
, perf_event_cgrp_id
),
696 struct perf_cgroup
, css
);
698 #endif /* CONFIG_CGROUP_PERF */
700 #ifdef CONFIG_PERF_EVENTS
702 extern void *perf_aux_output_begin(struct perf_output_handle
*handle
,
703 struct perf_event
*event
);
704 extern void perf_aux_output_end(struct perf_output_handle
*handle
,
705 unsigned long size
, bool truncated
);
706 extern int perf_aux_output_skip(struct perf_output_handle
*handle
,
708 extern void *perf_get_aux(struct perf_output_handle
*handle
);
710 extern int perf_pmu_register(struct pmu
*pmu
, const char *name
, int type
);
711 extern void perf_pmu_unregister(struct pmu
*pmu
);
713 extern int perf_num_counters(void);
714 extern const char *perf_pmu_name(void);
715 extern void __perf_event_task_sched_in(struct task_struct
*prev
,
716 struct task_struct
*task
);
717 extern void __perf_event_task_sched_out(struct task_struct
*prev
,
718 struct task_struct
*next
);
719 extern int perf_event_init_task(struct task_struct
*child
);
720 extern void perf_event_exit_task(struct task_struct
*child
);
721 extern void perf_event_free_task(struct task_struct
*task
);
722 extern void perf_event_delayed_put(struct task_struct
*task
);
723 extern struct perf_event
*perf_event_get(unsigned int fd
);
724 extern const struct perf_event_attr
*perf_event_attrs(struct perf_event
*event
);
725 extern void perf_event_print_debug(void);
726 extern void perf_pmu_disable(struct pmu
*pmu
);
727 extern void perf_pmu_enable(struct pmu
*pmu
);
728 extern void perf_sched_cb_dec(struct pmu
*pmu
);
729 extern void perf_sched_cb_inc(struct pmu
*pmu
);
730 extern int perf_event_task_disable(void);
731 extern int perf_event_task_enable(void);
732 extern int perf_event_refresh(struct perf_event
*event
, int refresh
);
733 extern void perf_event_update_userpage(struct perf_event
*event
);
734 extern int perf_event_release_kernel(struct perf_event
*event
);
735 extern struct perf_event
*
736 perf_event_create_kernel_counter(struct perf_event_attr
*attr
,
738 struct task_struct
*task
,
739 perf_overflow_handler_t callback
,
741 extern void perf_pmu_migrate_context(struct pmu
*pmu
,
742 int src_cpu
, int dst_cpu
);
743 extern u64
perf_event_read_local(struct perf_event
*event
);
744 extern u64
perf_event_read_value(struct perf_event
*event
,
745 u64
*enabled
, u64
*running
);
748 struct perf_sample_data
{
750 * Fields set by perf_sample_data_init(), group so as to
751 * minimize the cachelines touched.
754 struct perf_raw_record
*raw
;
755 struct perf_branch_stack
*br_stack
;
759 union perf_mem_data_src data_src
;
762 * The other fields, optionally {set,used} by
763 * perf_{prepare,output}_sample().
778 struct perf_callchain_entry
*callchain
;
781 * regs_user may point to task_pt_regs or to regs_user_copy, depending
784 struct perf_regs regs_user
;
785 struct pt_regs regs_user_copy
;
787 struct perf_regs regs_intr
;
789 } ____cacheline_aligned
;
791 /* default value for data source */
792 #define PERF_MEM_NA (PERF_MEM_S(OP, NA) |\
793 PERF_MEM_S(LVL, NA) |\
794 PERF_MEM_S(SNOOP, NA) |\
795 PERF_MEM_S(LOCK, NA) |\
798 static inline void perf_sample_data_init(struct perf_sample_data
*data
,
799 u64 addr
, u64 period
)
801 /* remaining struct members initialized in perf_prepare_sample() */
804 data
->br_stack
= NULL
;
805 data
->period
= period
;
807 data
->data_src
.val
= PERF_MEM_NA
;
811 extern void perf_output_sample(struct perf_output_handle
*handle
,
812 struct perf_event_header
*header
,
813 struct perf_sample_data
*data
,
814 struct perf_event
*event
);
815 extern void perf_prepare_sample(struct perf_event_header
*header
,
816 struct perf_sample_data
*data
,
817 struct perf_event
*event
,
818 struct pt_regs
*regs
);
820 extern int perf_event_overflow(struct perf_event
*event
,
821 struct perf_sample_data
*data
,
822 struct pt_regs
*regs
);
824 extern void perf_event_output(struct perf_event
*event
,
825 struct perf_sample_data
*data
,
826 struct pt_regs
*regs
);
829 perf_event_header__init_id(struct perf_event_header
*header
,
830 struct perf_sample_data
*data
,
831 struct perf_event
*event
);
833 perf_event__output_id_sample(struct perf_event
*event
,
834 struct perf_output_handle
*handle
,
835 struct perf_sample_data
*sample
);
838 perf_log_lost_samples(struct perf_event
*event
, u64 lost
);
840 static inline bool is_sampling_event(struct perf_event
*event
)
842 return event
->attr
.sample_period
!= 0;
846 * Return 1 for a software event, 0 for a hardware event
848 static inline int is_software_event(struct perf_event
*event
)
850 return event
->pmu
->task_ctx_nr
== perf_sw_context
;
853 extern struct static_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
855 extern void ___perf_sw_event(u32
, u64
, struct pt_regs
*, u64
);
856 extern void __perf_sw_event(u32
, u64
, struct pt_regs
*, u64
);
858 #ifndef perf_arch_fetch_caller_regs
859 static inline void perf_arch_fetch_caller_regs(struct pt_regs
*regs
, unsigned long ip
) { }
863 * Take a snapshot of the regs. Skip ip and frame pointer to
864 * the nth caller. We only need a few of the regs:
865 * - ip for PERF_SAMPLE_IP
866 * - cs for user_mode() tests
867 * - bp for callchains
868 * - eflags, for future purposes, just in case
870 static inline void perf_fetch_caller_regs(struct pt_regs
*regs
)
872 memset(regs
, 0, sizeof(*regs
));
874 perf_arch_fetch_caller_regs(regs
, CALLER_ADDR0
);
877 static __always_inline
void
878 perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
880 if (static_key_false(&perf_swevent_enabled
[event_id
]))
881 __perf_sw_event(event_id
, nr
, regs
, addr
);
884 DECLARE_PER_CPU(struct pt_regs
, __perf_regs
[4]);
887 * 'Special' version for the scheduler, it hard assumes no recursion,
888 * which is guaranteed by us not actually scheduling inside other swevents
889 * because those disable preemption.
891 static __always_inline
void
892 perf_sw_event_sched(u32 event_id
, u64 nr
, u64 addr
)
894 if (static_key_false(&perf_swevent_enabled
[event_id
])) {
895 struct pt_regs
*regs
= this_cpu_ptr(&__perf_regs
[0]);
897 perf_fetch_caller_regs(regs
);
898 ___perf_sw_event(event_id
, nr
, regs
, addr
);
902 extern struct static_key_deferred perf_sched_events
;
904 static __always_inline
bool
905 perf_sw_migrate_enabled(void)
907 if (static_key_false(&perf_swevent_enabled
[PERF_COUNT_SW_CPU_MIGRATIONS
]))
912 static inline void perf_event_task_migrate(struct task_struct
*task
)
914 if (perf_sw_migrate_enabled())
915 task
->sched_migrated
= 1;
918 static inline void perf_event_task_sched_in(struct task_struct
*prev
,
919 struct task_struct
*task
)
921 if (static_key_false(&perf_sched_events
.key
))
922 __perf_event_task_sched_in(prev
, task
);
924 if (perf_sw_migrate_enabled() && task
->sched_migrated
) {
925 struct pt_regs
*regs
= this_cpu_ptr(&__perf_regs
[0]);
927 perf_fetch_caller_regs(regs
);
928 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, regs
, 0);
929 task
->sched_migrated
= 0;
933 static inline void perf_event_task_sched_out(struct task_struct
*prev
,
934 struct task_struct
*next
)
936 perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES
, 1, 0);
938 if (static_key_false(&perf_sched_events
.key
))
939 __perf_event_task_sched_out(prev
, next
);
942 static inline u64
__perf_event_count(struct perf_event
*event
)
944 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
947 extern void perf_event_mmap(struct vm_area_struct
*vma
);
948 extern struct perf_guest_info_callbacks
*perf_guest_cbs
;
949 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*callbacks
);
950 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*callbacks
);
952 extern void perf_event_exec(void);
953 extern void perf_event_comm(struct task_struct
*tsk
, bool exec
);
954 extern void perf_event_fork(struct task_struct
*tsk
);
957 DECLARE_PER_CPU(struct perf_callchain_entry
, perf_callchain_entry
);
959 extern void perf_callchain_user(struct perf_callchain_entry
*entry
, struct pt_regs
*regs
);
960 extern void perf_callchain_kernel(struct perf_callchain_entry
*entry
, struct pt_regs
*regs
);
962 static inline void perf_callchain_store(struct perf_callchain_entry
*entry
, u64 ip
)
964 if (entry
->nr
< PERF_MAX_STACK_DEPTH
)
965 entry
->ip
[entry
->nr
++] = ip
;
968 extern int sysctl_perf_event_paranoid
;
969 extern int sysctl_perf_event_mlock
;
970 extern int sysctl_perf_event_sample_rate
;
971 extern int sysctl_perf_cpu_time_max_percent
;
973 extern void perf_sample_event_took(u64 sample_len_ns
);
975 extern int perf_proc_update_handler(struct ctl_table
*table
, int write
,
976 void __user
*buffer
, size_t *lenp
,
978 extern int perf_cpu_time_max_percent_handler(struct ctl_table
*table
, int write
,
979 void __user
*buffer
, size_t *lenp
,
983 static inline bool perf_paranoid_tracepoint_raw(void)
985 return sysctl_perf_event_paranoid
> -1;
988 static inline bool perf_paranoid_cpu(void)
990 return sysctl_perf_event_paranoid
> 0;
993 static inline bool perf_paranoid_kernel(void)
995 return sysctl_perf_event_paranoid
> 1;
998 extern void perf_event_init(void);
999 extern void perf_tp_event(u64 addr
, u64 count
, void *record
,
1000 int entry_size
, struct pt_regs
*regs
,
1001 struct hlist_head
*head
, int rctx
,
1002 struct task_struct
*task
);
1003 extern void perf_bp_event(struct perf_event
*event
, void *data
);
1005 #ifndef perf_misc_flags
1006 # define perf_misc_flags(regs) \
1007 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1008 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1011 static inline bool has_branch_stack(struct perf_event
*event
)
1013 return event
->attr
.sample_type
& PERF_SAMPLE_BRANCH_STACK
;
1016 static inline bool needs_branch_stack(struct perf_event
*event
)
1018 return event
->attr
.branch_sample_type
!= 0;
1021 static inline bool has_aux(struct perf_event
*event
)
1023 return event
->pmu
->setup_aux
;
1026 extern int perf_output_begin(struct perf_output_handle
*handle
,
1027 struct perf_event
*event
, unsigned int size
);
1028 extern void perf_output_end(struct perf_output_handle
*handle
);
1029 extern unsigned int perf_output_copy(struct perf_output_handle
*handle
,
1030 const void *buf
, unsigned int len
);
1031 extern unsigned int perf_output_skip(struct perf_output_handle
*handle
,
1033 extern int perf_swevent_get_recursion_context(void);
1034 extern void perf_swevent_put_recursion_context(int rctx
);
1035 extern u64
perf_swevent_set_period(struct perf_event
*event
);
1036 extern void perf_event_enable(struct perf_event
*event
);
1037 extern void perf_event_disable(struct perf_event
*event
);
1038 extern int __perf_event_disable(void *info
);
1039 extern void perf_event_task_tick(void);
1040 #else /* !CONFIG_PERF_EVENTS: */
1041 static inline void *
1042 perf_aux_output_begin(struct perf_output_handle
*handle
,
1043 struct perf_event
*event
) { return NULL
; }
1045 perf_aux_output_end(struct perf_output_handle
*handle
, unsigned long size
,
1048 perf_aux_output_skip(struct perf_output_handle
*handle
,
1049 unsigned long size
) { return -EINVAL
; }
1050 static inline void *
1051 perf_get_aux(struct perf_output_handle
*handle
) { return NULL
; }
1053 perf_event_task_migrate(struct task_struct
*task
) { }
1055 perf_event_task_sched_in(struct task_struct
*prev
,
1056 struct task_struct
*task
) { }
1058 perf_event_task_sched_out(struct task_struct
*prev
,
1059 struct task_struct
*next
) { }
1060 static inline int perf_event_init_task(struct task_struct
*child
) { return 0; }
1061 static inline void perf_event_exit_task(struct task_struct
*child
) { }
1062 static inline void perf_event_free_task(struct task_struct
*task
) { }
1063 static inline void perf_event_delayed_put(struct task_struct
*task
) { }
1064 static inline struct perf_event
*perf_event_get(unsigned int fd
) { return ERR_PTR(-EINVAL
); }
1065 static inline const struct perf_event_attr
*perf_event_attrs(struct perf_event
*event
)
1067 return ERR_PTR(-EINVAL
);
1069 static inline u64
perf_event_read_local(struct perf_event
*event
) { return -EINVAL
; }
1070 static inline void perf_event_print_debug(void) { }
1071 static inline int perf_event_task_disable(void) { return -EINVAL
; }
1072 static inline int perf_event_task_enable(void) { return -EINVAL
; }
1073 static inline int perf_event_refresh(struct perf_event
*event
, int refresh
)
1079 perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
) { }
1081 perf_sw_event_sched(u32 event_id
, u64 nr
, u64 addr
) { }
1083 perf_bp_event(struct perf_event
*event
, void *data
) { }
1085 static inline int perf_register_guest_info_callbacks
1086 (struct perf_guest_info_callbacks
*callbacks
) { return 0; }
1087 static inline int perf_unregister_guest_info_callbacks
1088 (struct perf_guest_info_callbacks
*callbacks
) { return 0; }
1090 static inline void perf_event_mmap(struct vm_area_struct
*vma
) { }
1091 static inline void perf_event_exec(void) { }
1092 static inline void perf_event_comm(struct task_struct
*tsk
, bool exec
) { }
1093 static inline void perf_event_fork(struct task_struct
*tsk
) { }
1094 static inline void perf_event_init(void) { }
1095 static inline int perf_swevent_get_recursion_context(void) { return -1; }
1096 static inline void perf_swevent_put_recursion_context(int rctx
) { }
1097 static inline u64
perf_swevent_set_period(struct perf_event
*event
) { return 0; }
1098 static inline void perf_event_enable(struct perf_event
*event
) { }
1099 static inline void perf_event_disable(struct perf_event
*event
) { }
1100 static inline int __perf_event_disable(void *info
) { return -1; }
1101 static inline void perf_event_task_tick(void) { }
1102 static inline int perf_event_release_kernel(struct perf_event
*event
) { return 0; }
1105 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_NO_HZ_FULL)
1106 extern bool perf_event_can_stop_tick(void);
1108 static inline bool perf_event_can_stop_tick(void) { return true; }
1111 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1112 extern void perf_restore_debug_store(void);
1114 static inline void perf_restore_debug_store(void) { }
1117 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1120 * This has to have a higher priority than migration_notifier in sched/core.c.
1122 #define perf_cpu_notifier(fn) \
1124 static struct notifier_block fn##_nb = \
1125 { .notifier_call = fn, .priority = CPU_PRI_PERF }; \
1126 unsigned long cpu = smp_processor_id(); \
1127 unsigned long flags; \
1129 cpu_notifier_register_begin(); \
1130 fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE, \
1131 (void *)(unsigned long)cpu); \
1132 local_irq_save(flags); \
1133 fn(&fn##_nb, (unsigned long)CPU_STARTING, \
1134 (void *)(unsigned long)cpu); \
1135 local_irq_restore(flags); \
1136 fn(&fn##_nb, (unsigned long)CPU_ONLINE, \
1137 (void *)(unsigned long)cpu); \
1138 __register_cpu_notifier(&fn##_nb); \
1139 cpu_notifier_register_done(); \
1143 * Bare-bones version of perf_cpu_notifier(), which doesn't invoke the
1144 * callback for already online CPUs.
1146 #define __perf_cpu_notifier(fn) \
1148 static struct notifier_block fn##_nb = \
1149 { .notifier_call = fn, .priority = CPU_PRI_PERF }; \
1151 __register_cpu_notifier(&fn##_nb); \
1154 struct perf_pmu_events_attr
{
1155 struct device_attribute attr
;
1157 const char *event_str
;
1160 ssize_t
perf_event_sysfs_show(struct device
*dev
, struct device_attribute
*attr
,
1163 #define PMU_EVENT_ATTR(_name, _var, _id, _show) \
1164 static struct perf_pmu_events_attr _var = { \
1165 .attr = __ATTR(_name, 0444, _show, NULL), \
1169 #define PMU_EVENT_ATTR_STRING(_name, _var, _str) \
1170 static struct perf_pmu_events_attr _var = { \
1171 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1173 .event_str = _str, \
1176 #define PMU_FORMAT_ATTR(_name, _format) \
1178 _name##_show(struct device *dev, \
1179 struct device_attribute *attr, \
1182 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
1183 return sprintf(page, _format "\n"); \
1186 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1188 #endif /* _LINUX_PERF_EVENT_H */