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
;
143 struct task_struct
*target
;
145 local64_t prev_count
;
148 local64_t period_left
;
153 u64 freq_count_stamp
;
158 * hw_perf_event::state flags
160 #define PERF_HES_STOPPED 0x01 /* the counter is stopped */
161 #define PERF_HES_UPTODATE 0x02 /* event->count up-to-date */
162 #define PERF_HES_ARCH 0x04
167 * Common implementation detail of pmu::{start,commit,cancel}_txn
169 #define PERF_EVENT_TXN 0x1
172 * pmu::capabilities flags
174 #define PERF_PMU_CAP_NO_INTERRUPT 0x01
175 #define PERF_PMU_CAP_NO_NMI 0x02
176 #define PERF_PMU_CAP_AUX_NO_SG 0x04
177 #define PERF_PMU_CAP_AUX_SW_DOUBLEBUF 0x08
178 #define PERF_PMU_CAP_EXCLUSIVE 0x10
179 #define PERF_PMU_CAP_ITRACE 0x20
182 * struct pmu - generic performance monitoring unit
185 struct list_head entry
;
187 struct module
*module
;
189 const struct attribute_group
**attr_groups
;
194 * various common per-pmu feature flags
198 int * __percpu pmu_disable_count
;
199 struct perf_cpu_context
* __percpu pmu_cpu_context
;
200 atomic_t exclusive_cnt
; /* < 0: cpu; > 0: tsk */
202 int hrtimer_interval_ms
;
205 * Fully disable/enable this PMU, can be used to protect from the PMI
206 * as well as for lazy/batch writing of the MSRs.
208 void (*pmu_enable
) (struct pmu
*pmu
); /* optional */
209 void (*pmu_disable
) (struct pmu
*pmu
); /* optional */
212 * Try and initialize the event for this PMU.
213 * Should return -ENOENT when the @event doesn't match this PMU.
215 int (*event_init
) (struct perf_event
*event
);
218 * Notification that the event was mapped or unmapped. Called
219 * in the context of the mapping task.
221 void (*event_mapped
) (struct perf_event
*event
); /*optional*/
222 void (*event_unmapped
) (struct perf_event
*event
); /*optional*/
224 #define PERF_EF_START 0x01 /* start the counter when adding */
225 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
226 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
229 * Adds/Removes a counter to/from the PMU, can be done inside
230 * a transaction, see the ->*_txn() methods.
232 int (*add
) (struct perf_event
*event
, int flags
);
233 void (*del
) (struct perf_event
*event
, int flags
);
236 * Starts/Stops a counter present on the PMU. The PMI handler
237 * should stop the counter when perf_event_overflow() returns
238 * !0. ->start() will be used to continue.
240 void (*start
) (struct perf_event
*event
, int flags
);
241 void (*stop
) (struct perf_event
*event
, int flags
);
244 * Updates the counter value of the event.
246 void (*read
) (struct perf_event
*event
);
249 * Group events scheduling is treated as a transaction, add
250 * group events as a whole and perform one schedulability test.
251 * If the test fails, roll back the whole group
253 * Start the transaction, after this ->add() doesn't need to
254 * do schedulability tests.
256 void (*start_txn
) (struct pmu
*pmu
); /* optional */
258 * If ->start_txn() disabled the ->add() schedulability test
259 * then ->commit_txn() is required to perform one. On success
260 * the transaction is closed. On error the transaction is kept
261 * open until ->cancel_txn() is called.
263 int (*commit_txn
) (struct pmu
*pmu
); /* optional */
265 * Will cancel the transaction, assumes ->del() is called
266 * for each successful ->add() during the transaction.
268 void (*cancel_txn
) (struct pmu
*pmu
); /* optional */
271 * Will return the value for perf_event_mmap_page::index for this event,
272 * if no implementation is provided it will default to: event->hw.idx + 1.
274 int (*event_idx
) (struct perf_event
*event
); /*optional */
277 * context-switches callback
279 void (*sched_task
) (struct perf_event_context
*ctx
,
282 * PMU specific data size
284 size_t task_ctx_size
;
288 * Return the count value for a counter.
290 u64 (*count
) (struct perf_event
*event
); /*optional*/
293 * Set up pmu-private data structures for an AUX area
295 void *(*setup_aux
) (int cpu
, void **pages
,
296 int nr_pages
, bool overwrite
);
300 * Free pmu-private AUX data structures
302 void (*free_aux
) (void *aux
); /* optional */
306 * enum perf_event_active_state - the states of a event
308 enum perf_event_active_state
{
309 PERF_EVENT_STATE_EXIT
= -3,
310 PERF_EVENT_STATE_ERROR
= -2,
311 PERF_EVENT_STATE_OFF
= -1,
312 PERF_EVENT_STATE_INACTIVE
= 0,
313 PERF_EVENT_STATE_ACTIVE
= 1,
317 struct perf_sample_data
;
319 typedef void (*perf_overflow_handler_t
)(struct perf_event
*,
320 struct perf_sample_data
*,
321 struct pt_regs
*regs
);
323 enum perf_group_flag
{
324 PERF_GROUP_SOFTWARE
= 0x1,
327 #define SWEVENT_HLIST_BITS 8
328 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
330 struct swevent_hlist
{
331 struct hlist_head heads
[SWEVENT_HLIST_SIZE
];
332 struct rcu_head rcu_head
;
335 #define PERF_ATTACH_CONTEXT 0x01
336 #define PERF_ATTACH_GROUP 0x02
337 #define PERF_ATTACH_TASK 0x04
338 #define PERF_ATTACH_TASK_DATA 0x08
344 * struct perf_event - performance event kernel representation:
347 #ifdef CONFIG_PERF_EVENTS
349 * entry onto perf_event_context::event_list;
350 * modifications require ctx->lock
351 * RCU safe iterations.
353 struct list_head event_entry
;
356 * XXX: group_entry and sibling_list should be mutually exclusive;
357 * either you're a sibling on a group, or you're the group leader.
358 * Rework the code to always use the same list element.
360 * Locked for modification by both ctx->mutex and ctx->lock; holding
361 * either sufficies for read.
363 struct list_head group_entry
;
364 struct list_head sibling_list
;
367 * We need storage to track the entries in perf_pmu_migrate_context; we
368 * cannot use the event_entry because of RCU and we want to keep the
369 * group in tact which avoids us using the other two entries.
371 struct list_head migrate_entry
;
373 struct hlist_node hlist_entry
;
374 struct list_head active_entry
;
377 struct perf_event
*group_leader
;
380 enum perf_event_active_state state
;
381 unsigned int attach_state
;
383 atomic64_t child_count
;
386 * These are the total time in nanoseconds that the event
387 * has been enabled (i.e. eligible to run, and the task has
388 * been scheduled in, if this is a per-task event)
389 * and running (scheduled onto the CPU), respectively.
391 * They are computed from tstamp_enabled, tstamp_running and
392 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
394 u64 total_time_enabled
;
395 u64 total_time_running
;
398 * These are timestamps used for computing total_time_enabled
399 * and total_time_running when the event is in INACTIVE or
400 * ACTIVE state, measured in nanoseconds from an arbitrary point
402 * tstamp_enabled: the notional time when the event was enabled
403 * tstamp_running: the notional time when the event was scheduled on
404 * tstamp_stopped: in INACTIVE state, the notional time when the
405 * event was scheduled off.
412 * timestamp shadows the actual context timing but it can
413 * be safely used in NMI interrupt context. It reflects the
414 * context time as it was when the event was last scheduled in.
416 * ctx_time already accounts for ctx->timestamp. Therefore to
417 * compute ctx_time for a sample, simply add perf_clock().
421 struct perf_event_attr attr
;
425 struct hw_perf_event hw
;
427 struct perf_event_context
*ctx
;
428 atomic_long_t refcount
;
431 * These accumulate total time (in nanoseconds) that children
432 * events have been enabled and running, respectively.
434 atomic64_t child_total_time_enabled
;
435 atomic64_t child_total_time_running
;
438 * Protect attach/detach and child_list:
440 struct mutex child_mutex
;
441 struct list_head child_list
;
442 struct perf_event
*parent
;
447 struct list_head owner_entry
;
448 struct task_struct
*owner
;
451 struct mutex mmap_mutex
;
454 struct ring_buffer
*rb
;
455 struct list_head rb_entry
;
456 unsigned long rcu_batches
;
460 wait_queue_head_t waitq
;
461 struct fasync_struct
*fasync
;
463 /* delayed work for NMIs and such */
467 struct irq_work pending
;
469 atomic_t event_limit
;
471 void (*destroy
)(struct perf_event
*);
472 struct rcu_head rcu_head
;
474 struct pid_namespace
*ns
;
478 perf_overflow_handler_t overflow_handler
;
479 void *overflow_handler_context
;
481 #ifdef CONFIG_EVENT_TRACING
482 struct ftrace_event_call
*tp_event
;
483 struct event_filter
*filter
;
484 #ifdef CONFIG_FUNCTION_TRACER
485 struct ftrace_ops ftrace_ops
;
489 #ifdef CONFIG_CGROUP_PERF
490 struct perf_cgroup
*cgrp
; /* cgroup event is attach to */
491 int cgrp_defer_enabled
;
494 #endif /* CONFIG_PERF_EVENTS */
498 * struct perf_event_context - event context structure
500 * Used as a container for task events and CPU events as well:
502 struct perf_event_context
{
505 * Protect the states of the events in the list,
506 * nr_active, and the list:
510 * Protect the list of events. Locking either mutex or lock
511 * is sufficient to ensure the list doesn't change; to change
512 * the list you need to lock both the mutex and the spinlock.
516 struct list_head active_ctx_list
;
517 struct list_head pinned_groups
;
518 struct list_head flexible_groups
;
519 struct list_head event_list
;
527 struct task_struct
*task
;
530 * Context clock, runs when context enabled.
536 * These fields let us detect when two contexts have both
537 * been cloned (inherited) from a common ancestor.
539 struct perf_event_context
*parent_ctx
;
543 int nr_cgroups
; /* cgroup evts */
544 void *task_ctx_data
; /* pmu specific data */
545 struct rcu_head rcu_head
;
547 struct delayed_work orphans_remove
;
548 bool orphans_remove_sched
;
552 * Number of contexts where an event can trigger:
553 * task, softirq, hardirq, nmi.
555 #define PERF_NR_CONTEXTS 4
558 * struct perf_event_cpu_context - per cpu event context structure
560 struct perf_cpu_context
{
561 struct perf_event_context ctx
;
562 struct perf_event_context
*task_ctx
;
565 struct hrtimer hrtimer
;
566 ktime_t hrtimer_interval
;
567 struct pmu
*unique_pmu
;
568 struct perf_cgroup
*cgrp
;
571 struct perf_output_handle
{
572 struct perf_event
*event
;
573 struct ring_buffer
*rb
;
574 unsigned long wakeup
;
583 #ifdef CONFIG_CGROUP_PERF
586 * perf_cgroup_info keeps track of time_enabled for a cgroup.
587 * This is a per-cpu dynamically allocated data structure.
589 struct perf_cgroup_info
{
595 struct cgroup_subsys_state css
;
596 struct perf_cgroup_info __percpu
*info
;
600 * Must ensure cgroup is pinned (css_get) before calling
601 * this function. In other words, we cannot call this function
602 * if there is no cgroup event for the current CPU context.
604 static inline struct perf_cgroup
*
605 perf_cgroup_from_task(struct task_struct
*task
)
607 return container_of(task_css(task
, perf_event_cgrp_id
),
608 struct perf_cgroup
, css
);
610 #endif /* CONFIG_CGROUP_PERF */
612 #ifdef CONFIG_PERF_EVENTS
614 extern void *perf_aux_output_begin(struct perf_output_handle
*handle
,
615 struct perf_event
*event
);
616 extern void perf_aux_output_end(struct perf_output_handle
*handle
,
617 unsigned long size
, bool truncated
);
618 extern int perf_aux_output_skip(struct perf_output_handle
*handle
,
620 extern void *perf_get_aux(struct perf_output_handle
*handle
);
622 extern int perf_pmu_register(struct pmu
*pmu
, const char *name
, int type
);
623 extern void perf_pmu_unregister(struct pmu
*pmu
);
625 extern int perf_num_counters(void);
626 extern const char *perf_pmu_name(void);
627 extern void __perf_event_task_sched_in(struct task_struct
*prev
,
628 struct task_struct
*task
);
629 extern void __perf_event_task_sched_out(struct task_struct
*prev
,
630 struct task_struct
*next
);
631 extern int perf_event_init_task(struct task_struct
*child
);
632 extern void perf_event_exit_task(struct task_struct
*child
);
633 extern void perf_event_free_task(struct task_struct
*task
);
634 extern void perf_event_delayed_put(struct task_struct
*task
);
635 extern void perf_event_print_debug(void);
636 extern void perf_pmu_disable(struct pmu
*pmu
);
637 extern void perf_pmu_enable(struct pmu
*pmu
);
638 extern void perf_sched_cb_dec(struct pmu
*pmu
);
639 extern void perf_sched_cb_inc(struct pmu
*pmu
);
640 extern int perf_event_task_disable(void);
641 extern int perf_event_task_enable(void);
642 extern int perf_event_refresh(struct perf_event
*event
, int refresh
);
643 extern void perf_event_update_userpage(struct perf_event
*event
);
644 extern int perf_event_release_kernel(struct perf_event
*event
);
645 extern struct perf_event
*
646 perf_event_create_kernel_counter(struct perf_event_attr
*attr
,
648 struct task_struct
*task
,
649 perf_overflow_handler_t callback
,
651 extern void perf_pmu_migrate_context(struct pmu
*pmu
,
652 int src_cpu
, int dst_cpu
);
653 extern u64
perf_event_read_value(struct perf_event
*event
,
654 u64
*enabled
, u64
*running
);
657 struct perf_sample_data
{
659 * Fields set by perf_sample_data_init(), group so as to
660 * minimize the cachelines touched.
663 struct perf_raw_record
*raw
;
664 struct perf_branch_stack
*br_stack
;
668 union perf_mem_data_src data_src
;
671 * The other fields, optionally {set,used} by
672 * perf_{prepare,output}_sample().
687 struct perf_callchain_entry
*callchain
;
690 * regs_user may point to task_pt_regs or to regs_user_copy, depending
693 struct perf_regs regs_user
;
694 struct pt_regs regs_user_copy
;
696 struct perf_regs regs_intr
;
698 } ____cacheline_aligned
;
700 /* default value for data source */
701 #define PERF_MEM_NA (PERF_MEM_S(OP, NA) |\
702 PERF_MEM_S(LVL, NA) |\
703 PERF_MEM_S(SNOOP, NA) |\
704 PERF_MEM_S(LOCK, NA) |\
707 static inline void perf_sample_data_init(struct perf_sample_data
*data
,
708 u64 addr
, u64 period
)
710 /* remaining struct members initialized in perf_prepare_sample() */
713 data
->br_stack
= NULL
;
714 data
->period
= period
;
716 data
->data_src
.val
= PERF_MEM_NA
;
720 extern void perf_output_sample(struct perf_output_handle
*handle
,
721 struct perf_event_header
*header
,
722 struct perf_sample_data
*data
,
723 struct perf_event
*event
);
724 extern void perf_prepare_sample(struct perf_event_header
*header
,
725 struct perf_sample_data
*data
,
726 struct perf_event
*event
,
727 struct pt_regs
*regs
);
729 extern int perf_event_overflow(struct perf_event
*event
,
730 struct perf_sample_data
*data
,
731 struct pt_regs
*regs
);
733 extern void perf_event_output(struct perf_event
*event
,
734 struct perf_sample_data
*data
,
735 struct pt_regs
*regs
);
738 perf_event_header__init_id(struct perf_event_header
*header
,
739 struct perf_sample_data
*data
,
740 struct perf_event
*event
);
742 perf_event__output_id_sample(struct perf_event
*event
,
743 struct perf_output_handle
*handle
,
744 struct perf_sample_data
*sample
);
747 perf_log_lost_samples(struct perf_event
*event
, u64 lost
);
749 static inline bool is_sampling_event(struct perf_event
*event
)
751 return event
->attr
.sample_period
!= 0;
755 * Return 1 for a software event, 0 for a hardware event
757 static inline int is_software_event(struct perf_event
*event
)
759 return event
->pmu
->task_ctx_nr
== perf_sw_context
;
762 extern struct static_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
764 extern void ___perf_sw_event(u32
, u64
, struct pt_regs
*, u64
);
765 extern void __perf_sw_event(u32
, u64
, struct pt_regs
*, u64
);
767 #ifndef perf_arch_fetch_caller_regs
768 static inline void perf_arch_fetch_caller_regs(struct pt_regs
*regs
, unsigned long ip
) { }
772 * Take a snapshot of the regs. Skip ip and frame pointer to
773 * the nth caller. We only need a few of the regs:
774 * - ip for PERF_SAMPLE_IP
775 * - cs for user_mode() tests
776 * - bp for callchains
777 * - eflags, for future purposes, just in case
779 static inline void perf_fetch_caller_regs(struct pt_regs
*regs
)
781 memset(regs
, 0, sizeof(*regs
));
783 perf_arch_fetch_caller_regs(regs
, CALLER_ADDR0
);
786 static __always_inline
void
787 perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
789 if (static_key_false(&perf_swevent_enabled
[event_id
]))
790 __perf_sw_event(event_id
, nr
, regs
, addr
);
793 DECLARE_PER_CPU(struct pt_regs
, __perf_regs
[4]);
796 * 'Special' version for the scheduler, it hard assumes no recursion,
797 * which is guaranteed by us not actually scheduling inside other swevents
798 * because those disable preemption.
800 static __always_inline
void
801 perf_sw_event_sched(u32 event_id
, u64 nr
, u64 addr
)
803 if (static_key_false(&perf_swevent_enabled
[event_id
])) {
804 struct pt_regs
*regs
= this_cpu_ptr(&__perf_regs
[0]);
806 perf_fetch_caller_regs(regs
);
807 ___perf_sw_event(event_id
, nr
, regs
, addr
);
811 extern struct static_key_deferred perf_sched_events
;
813 static __always_inline
bool
814 perf_sw_migrate_enabled(void)
816 if (static_key_false(&perf_swevent_enabled
[PERF_COUNT_SW_CPU_MIGRATIONS
]))
821 static inline void perf_event_task_migrate(struct task_struct
*task
)
823 if (perf_sw_migrate_enabled())
824 task
->sched_migrated
= 1;
827 static inline void perf_event_task_sched_in(struct task_struct
*prev
,
828 struct task_struct
*task
)
830 if (static_key_false(&perf_sched_events
.key
))
831 __perf_event_task_sched_in(prev
, task
);
833 if (perf_sw_migrate_enabled() && task
->sched_migrated
) {
834 struct pt_regs
*regs
= this_cpu_ptr(&__perf_regs
[0]);
836 perf_fetch_caller_regs(regs
);
837 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, regs
, 0);
838 task
->sched_migrated
= 0;
842 static inline void perf_event_task_sched_out(struct task_struct
*prev
,
843 struct task_struct
*next
)
845 perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES
, 1, 0);
847 if (static_key_false(&perf_sched_events
.key
))
848 __perf_event_task_sched_out(prev
, next
);
851 static inline u64
__perf_event_count(struct perf_event
*event
)
853 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
856 extern void perf_event_mmap(struct vm_area_struct
*vma
);
857 extern struct perf_guest_info_callbacks
*perf_guest_cbs
;
858 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*callbacks
);
859 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*callbacks
);
861 extern void perf_event_exec(void);
862 extern void perf_event_comm(struct task_struct
*tsk
, bool exec
);
863 extern void perf_event_fork(struct task_struct
*tsk
);
866 DECLARE_PER_CPU(struct perf_callchain_entry
, perf_callchain_entry
);
868 extern void perf_callchain_user(struct perf_callchain_entry
*entry
, struct pt_regs
*regs
);
869 extern void perf_callchain_kernel(struct perf_callchain_entry
*entry
, struct pt_regs
*regs
);
871 static inline void perf_callchain_store(struct perf_callchain_entry
*entry
, u64 ip
)
873 if (entry
->nr
< PERF_MAX_STACK_DEPTH
)
874 entry
->ip
[entry
->nr
++] = ip
;
877 extern int sysctl_perf_event_paranoid
;
878 extern int sysctl_perf_event_mlock
;
879 extern int sysctl_perf_event_sample_rate
;
880 extern int sysctl_perf_cpu_time_max_percent
;
882 extern void perf_sample_event_took(u64 sample_len_ns
);
884 extern int perf_proc_update_handler(struct ctl_table
*table
, int write
,
885 void __user
*buffer
, size_t *lenp
,
887 extern int perf_cpu_time_max_percent_handler(struct ctl_table
*table
, int write
,
888 void __user
*buffer
, size_t *lenp
,
892 static inline bool perf_paranoid_tracepoint_raw(void)
894 return sysctl_perf_event_paranoid
> -1;
897 static inline bool perf_paranoid_cpu(void)
899 return sysctl_perf_event_paranoid
> 0;
902 static inline bool perf_paranoid_kernel(void)
904 return sysctl_perf_event_paranoid
> 1;
907 extern void perf_event_init(void);
908 extern void perf_tp_event(u64 addr
, u64 count
, void *record
,
909 int entry_size
, struct pt_regs
*regs
,
910 struct hlist_head
*head
, int rctx
,
911 struct task_struct
*task
);
912 extern void perf_bp_event(struct perf_event
*event
, void *data
);
914 #ifndef perf_misc_flags
915 # define perf_misc_flags(regs) \
916 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
917 # define perf_instruction_pointer(regs) instruction_pointer(regs)
920 static inline bool has_branch_stack(struct perf_event
*event
)
922 return event
->attr
.sample_type
& PERF_SAMPLE_BRANCH_STACK
;
925 static inline bool needs_branch_stack(struct perf_event
*event
)
927 return event
->attr
.branch_sample_type
!= 0;
930 static inline bool has_aux(struct perf_event
*event
)
932 return event
->pmu
->setup_aux
;
935 extern int perf_output_begin(struct perf_output_handle
*handle
,
936 struct perf_event
*event
, unsigned int size
);
937 extern void perf_output_end(struct perf_output_handle
*handle
);
938 extern unsigned int perf_output_copy(struct perf_output_handle
*handle
,
939 const void *buf
, unsigned int len
);
940 extern unsigned int perf_output_skip(struct perf_output_handle
*handle
,
942 extern int perf_swevent_get_recursion_context(void);
943 extern void perf_swevent_put_recursion_context(int rctx
);
944 extern u64
perf_swevent_set_period(struct perf_event
*event
);
945 extern void perf_event_enable(struct perf_event
*event
);
946 extern void perf_event_disable(struct perf_event
*event
);
947 extern int __perf_event_disable(void *info
);
948 extern void perf_event_task_tick(void);
949 #else /* !CONFIG_PERF_EVENTS: */
951 perf_aux_output_begin(struct perf_output_handle
*handle
,
952 struct perf_event
*event
) { return NULL
; }
954 perf_aux_output_end(struct perf_output_handle
*handle
, unsigned long size
,
957 perf_aux_output_skip(struct perf_output_handle
*handle
,
958 unsigned long size
) { return -EINVAL
; }
960 perf_get_aux(struct perf_output_handle
*handle
) { return NULL
; }
962 perf_event_task_migrate(struct task_struct
*task
) { }
964 perf_event_task_sched_in(struct task_struct
*prev
,
965 struct task_struct
*task
) { }
967 perf_event_task_sched_out(struct task_struct
*prev
,
968 struct task_struct
*next
) { }
969 static inline int perf_event_init_task(struct task_struct
*child
) { return 0; }
970 static inline void perf_event_exit_task(struct task_struct
*child
) { }
971 static inline void perf_event_free_task(struct task_struct
*task
) { }
972 static inline void perf_event_delayed_put(struct task_struct
*task
) { }
973 static inline void perf_event_print_debug(void) { }
974 static inline int perf_event_task_disable(void) { return -EINVAL
; }
975 static inline int perf_event_task_enable(void) { return -EINVAL
; }
976 static inline int perf_event_refresh(struct perf_event
*event
, int refresh
)
982 perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
) { }
984 perf_sw_event_sched(u32 event_id
, u64 nr
, u64 addr
) { }
986 perf_bp_event(struct perf_event
*event
, void *data
) { }
988 static inline int perf_register_guest_info_callbacks
989 (struct perf_guest_info_callbacks
*callbacks
) { return 0; }
990 static inline int perf_unregister_guest_info_callbacks
991 (struct perf_guest_info_callbacks
*callbacks
) { return 0; }
993 static inline void perf_event_mmap(struct vm_area_struct
*vma
) { }
994 static inline void perf_event_exec(void) { }
995 static inline void perf_event_comm(struct task_struct
*tsk
, bool exec
) { }
996 static inline void perf_event_fork(struct task_struct
*tsk
) { }
997 static inline void perf_event_init(void) { }
998 static inline int perf_swevent_get_recursion_context(void) { return -1; }
999 static inline void perf_swevent_put_recursion_context(int rctx
) { }
1000 static inline u64
perf_swevent_set_period(struct perf_event
*event
) { return 0; }
1001 static inline void perf_event_enable(struct perf_event
*event
) { }
1002 static inline void perf_event_disable(struct perf_event
*event
) { }
1003 static inline int __perf_event_disable(void *info
) { return -1; }
1004 static inline void perf_event_task_tick(void) { }
1007 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_NO_HZ_FULL)
1008 extern bool perf_event_can_stop_tick(void);
1010 static inline bool perf_event_can_stop_tick(void) { return true; }
1013 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1014 extern void perf_restore_debug_store(void);
1016 static inline void perf_restore_debug_store(void) { }
1019 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1022 * This has to have a higher priority than migration_notifier in sched/core.c.
1024 #define perf_cpu_notifier(fn) \
1026 static struct notifier_block fn##_nb = \
1027 { .notifier_call = fn, .priority = CPU_PRI_PERF }; \
1028 unsigned long cpu = smp_processor_id(); \
1029 unsigned long flags; \
1031 cpu_notifier_register_begin(); \
1032 fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE, \
1033 (void *)(unsigned long)cpu); \
1034 local_irq_save(flags); \
1035 fn(&fn##_nb, (unsigned long)CPU_STARTING, \
1036 (void *)(unsigned long)cpu); \
1037 local_irq_restore(flags); \
1038 fn(&fn##_nb, (unsigned long)CPU_ONLINE, \
1039 (void *)(unsigned long)cpu); \
1040 __register_cpu_notifier(&fn##_nb); \
1041 cpu_notifier_register_done(); \
1045 * Bare-bones version of perf_cpu_notifier(), which doesn't invoke the
1046 * callback for already online CPUs.
1048 #define __perf_cpu_notifier(fn) \
1050 static struct notifier_block fn##_nb = \
1051 { .notifier_call = fn, .priority = CPU_PRI_PERF }; \
1053 __register_cpu_notifier(&fn##_nb); \
1056 struct perf_pmu_events_attr
{
1057 struct device_attribute attr
;
1059 const char *event_str
;
1062 ssize_t
perf_event_sysfs_show(struct device
*dev
, struct device_attribute
*attr
,
1065 #define PMU_EVENT_ATTR(_name, _var, _id, _show) \
1066 static struct perf_pmu_events_attr _var = { \
1067 .attr = __ATTR(_name, 0444, _show, NULL), \
1071 #define PMU_EVENT_ATTR_STRING(_name, _var, _str) \
1072 static struct perf_pmu_events_attr _var = { \
1073 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1075 .event_str = _str, \
1078 #define PMU_FORMAT_ATTR(_name, _format) \
1080 _name##_show(struct device *dev, \
1081 struct device_attribute *attr, \
1084 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
1085 return sprintf(page, _format "\n"); \
1088 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1090 #endif /* _LINUX_PERF_EVENT_H */