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_PMU_TXN_ADD 0x1 /* txn to add/schedule event on PMU */
203 #define PERF_PMU_TXN_READ 0x2 /* txn to read event group from PMU */
206 * pmu::capabilities flags
208 #define PERF_PMU_CAP_NO_INTERRUPT 0x01
209 #define PERF_PMU_CAP_NO_NMI 0x02
210 #define PERF_PMU_CAP_AUX_NO_SG 0x04
211 #define PERF_PMU_CAP_AUX_SW_DOUBLEBUF 0x08
212 #define PERF_PMU_CAP_EXCLUSIVE 0x10
213 #define PERF_PMU_CAP_ITRACE 0x20
216 * struct pmu - generic performance monitoring unit
219 struct list_head entry
;
221 struct module
*module
;
223 const struct attribute_group
**attr_groups
;
228 * various common per-pmu feature flags
232 int * __percpu pmu_disable_count
;
233 struct perf_cpu_context
* __percpu pmu_cpu_context
;
234 atomic_t exclusive_cnt
; /* < 0: cpu; > 0: tsk */
236 int hrtimer_interval_ms
;
239 * Fully disable/enable this PMU, can be used to protect from the PMI
240 * as well as for lazy/batch writing of the MSRs.
242 void (*pmu_enable
) (struct pmu
*pmu
); /* optional */
243 void (*pmu_disable
) (struct pmu
*pmu
); /* optional */
246 * Try and initialize the event for this PMU.
249 * -ENOENT -- @event is not for this PMU
251 * -ENODEV -- @event is for this PMU but PMU not present
252 * -EBUSY -- @event is for this PMU but PMU temporarily unavailable
253 * -EINVAL -- @event is for this PMU but @event is not valid
254 * -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
255 * -EACCESS -- @event is for this PMU, @event is valid, but no privilidges
257 * 0 -- @event is for this PMU and valid
259 * Other error return values are allowed.
261 int (*event_init
) (struct perf_event
*event
);
264 * Notification that the event was mapped or unmapped. Called
265 * in the context of the mapping task.
267 void (*event_mapped
) (struct perf_event
*event
); /*optional*/
268 void (*event_unmapped
) (struct perf_event
*event
); /*optional*/
271 * Flags for ->add()/->del()/ ->start()/->stop(). There are
272 * matching hw_perf_event::state flags.
274 #define PERF_EF_START 0x01 /* start the counter when adding */
275 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
276 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
279 * Adds/Removes a counter to/from the PMU, can be done inside a
280 * transaction, see the ->*_txn() methods.
282 * The add/del callbacks will reserve all hardware resources required
283 * to service the event, this includes any counter constraint
286 * Called with IRQs disabled and the PMU disabled on the CPU the event
289 * ->add() called without PERF_EF_START should result in the same state
290 * as ->add() followed by ->stop().
292 * ->del() must always PERF_EF_UPDATE stop an event. If it calls
293 * ->stop() that must deal with already being stopped without
296 int (*add
) (struct perf_event
*event
, int flags
);
297 void (*del
) (struct perf_event
*event
, int flags
);
300 * Starts/Stops a counter present on the PMU.
302 * The PMI handler should stop the counter when perf_event_overflow()
303 * returns !0. ->start() will be used to continue.
305 * Also used to change the sample period.
307 * Called with IRQs disabled and the PMU disabled on the CPU the event
308 * is on -- will be called from NMI context with the PMU generates
311 * ->stop() with PERF_EF_UPDATE will read the counter and update
312 * period/count values like ->read() would.
314 * ->start() with PERF_EF_RELOAD will reprogram the the counter
315 * value, must be preceded by a ->stop() with PERF_EF_UPDATE.
317 void (*start
) (struct perf_event
*event
, int flags
);
318 void (*stop
) (struct perf_event
*event
, int flags
);
321 * Updates the counter value of the event.
323 * For sampling capable PMUs this will also update the software period
324 * hw_perf_event::period_left field.
326 void (*read
) (struct perf_event
*event
);
329 * Group events scheduling is treated as a transaction, add
330 * group events as a whole and perform one schedulability test.
331 * If the test fails, roll back the whole group
333 * Start the transaction, after this ->add() doesn't need to
334 * do schedulability tests.
338 void (*start_txn
) (struct pmu
*pmu
, unsigned int txn_flags
);
340 * If ->start_txn() disabled the ->add() schedulability test
341 * then ->commit_txn() is required to perform one. On success
342 * the transaction is closed. On error the transaction is kept
343 * open until ->cancel_txn() is called.
347 int (*commit_txn
) (struct pmu
*pmu
);
349 * Will cancel the transaction, assumes ->del() is called
350 * for each successful ->add() during the transaction.
354 void (*cancel_txn
) (struct pmu
*pmu
);
357 * Will return the value for perf_event_mmap_page::index for this event,
358 * if no implementation is provided it will default to: event->hw.idx + 1.
360 int (*event_idx
) (struct perf_event
*event
); /*optional */
363 * context-switches callback
365 void (*sched_task
) (struct perf_event_context
*ctx
,
368 * PMU specific data size
370 size_t task_ctx_size
;
374 * Return the count value for a counter.
376 u64 (*count
) (struct perf_event
*event
); /*optional*/
379 * Set up pmu-private data structures for an AUX area
381 void *(*setup_aux
) (int cpu
, void **pages
,
382 int nr_pages
, bool overwrite
);
386 * Free pmu-private AUX data structures
388 void (*free_aux
) (void *aux
); /* optional */
391 * Filter events for PMU-specific reasons.
393 int (*filter_match
) (struct perf_event
*event
); /* optional */
397 * enum perf_event_active_state - the states of a event
399 enum perf_event_active_state
{
400 PERF_EVENT_STATE_EXIT
= -3,
401 PERF_EVENT_STATE_ERROR
= -2,
402 PERF_EVENT_STATE_OFF
= -1,
403 PERF_EVENT_STATE_INACTIVE
= 0,
404 PERF_EVENT_STATE_ACTIVE
= 1,
408 struct perf_sample_data
;
410 typedef void (*perf_overflow_handler_t
)(struct perf_event
*,
411 struct perf_sample_data
*,
412 struct pt_regs
*regs
);
414 enum perf_group_flag
{
415 PERF_GROUP_SOFTWARE
= 0x1,
418 #define SWEVENT_HLIST_BITS 8
419 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
421 struct swevent_hlist
{
422 struct hlist_head heads
[SWEVENT_HLIST_SIZE
];
423 struct rcu_head rcu_head
;
426 #define PERF_ATTACH_CONTEXT 0x01
427 #define PERF_ATTACH_GROUP 0x02
428 #define PERF_ATTACH_TASK 0x04
429 #define PERF_ATTACH_TASK_DATA 0x08
435 * struct perf_event - performance event kernel representation:
438 #ifdef CONFIG_PERF_EVENTS
440 * entry onto perf_event_context::event_list;
441 * modifications require ctx->lock
442 * RCU safe iterations.
444 struct list_head event_entry
;
447 * XXX: group_entry and sibling_list should be mutually exclusive;
448 * either you're a sibling on a group, or you're the group leader.
449 * Rework the code to always use the same list element.
451 * Locked for modification by both ctx->mutex and ctx->lock; holding
452 * either sufficies for read.
454 struct list_head group_entry
;
455 struct list_head sibling_list
;
458 * We need storage to track the entries in perf_pmu_migrate_context; we
459 * cannot use the event_entry because of RCU and we want to keep the
460 * group in tact which avoids us using the other two entries.
462 struct list_head migrate_entry
;
464 struct hlist_node hlist_entry
;
465 struct list_head active_entry
;
468 struct perf_event
*group_leader
;
471 enum perf_event_active_state state
;
472 unsigned int attach_state
;
474 atomic64_t child_count
;
477 * These are the total time in nanoseconds that the event
478 * has been enabled (i.e. eligible to run, and the task has
479 * been scheduled in, if this is a per-task event)
480 * and running (scheduled onto the CPU), respectively.
482 * They are computed from tstamp_enabled, tstamp_running and
483 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
485 u64 total_time_enabled
;
486 u64 total_time_running
;
489 * These are timestamps used for computing total_time_enabled
490 * and total_time_running when the event is in INACTIVE or
491 * ACTIVE state, measured in nanoseconds from an arbitrary point
493 * tstamp_enabled: the notional time when the event was enabled
494 * tstamp_running: the notional time when the event was scheduled on
495 * tstamp_stopped: in INACTIVE state, the notional time when the
496 * event was scheduled off.
503 * timestamp shadows the actual context timing but it can
504 * be safely used in NMI interrupt context. It reflects the
505 * context time as it was when the event was last scheduled in.
507 * ctx_time already accounts for ctx->timestamp. Therefore to
508 * compute ctx_time for a sample, simply add perf_clock().
512 struct perf_event_attr attr
;
516 struct hw_perf_event hw
;
518 struct perf_event_context
*ctx
;
519 atomic_long_t refcount
;
522 * These accumulate total time (in nanoseconds) that children
523 * events have been enabled and running, respectively.
525 atomic64_t child_total_time_enabled
;
526 atomic64_t child_total_time_running
;
529 * Protect attach/detach and child_list:
531 struct mutex child_mutex
;
532 struct list_head child_list
;
533 struct perf_event
*parent
;
538 struct list_head owner_entry
;
539 struct task_struct
*owner
;
542 struct mutex mmap_mutex
;
545 struct ring_buffer
*rb
;
546 struct list_head rb_entry
;
547 unsigned long rcu_batches
;
551 wait_queue_head_t waitq
;
552 struct fasync_struct
*fasync
;
554 /* delayed work for NMIs and such */
558 struct irq_work pending
;
560 atomic_t event_limit
;
562 void (*destroy
)(struct perf_event
*);
563 struct rcu_head rcu_head
;
565 struct pid_namespace
*ns
;
569 perf_overflow_handler_t overflow_handler
;
570 void *overflow_handler_context
;
572 #ifdef CONFIG_EVENT_TRACING
573 struct trace_event_call
*tp_event
;
574 struct event_filter
*filter
;
575 #ifdef CONFIG_FUNCTION_TRACER
576 struct ftrace_ops ftrace_ops
;
580 #ifdef CONFIG_CGROUP_PERF
581 struct perf_cgroup
*cgrp
; /* cgroup event is attach to */
582 int cgrp_defer_enabled
;
585 #endif /* CONFIG_PERF_EVENTS */
589 * struct perf_event_context - event context structure
591 * Used as a container for task events and CPU events as well:
593 struct perf_event_context
{
596 * Protect the states of the events in the list,
597 * nr_active, and the list:
601 * Protect the list of events. Locking either mutex or lock
602 * is sufficient to ensure the list doesn't change; to change
603 * the list you need to lock both the mutex and the spinlock.
607 struct list_head active_ctx_list
;
608 struct list_head pinned_groups
;
609 struct list_head flexible_groups
;
610 struct list_head event_list
;
618 struct task_struct
*task
;
621 * Context clock, runs when context enabled.
627 * These fields let us detect when two contexts have both
628 * been cloned (inherited) from a common ancestor.
630 struct perf_event_context
*parent_ctx
;
634 int nr_cgroups
; /* cgroup evts */
635 void *task_ctx_data
; /* pmu specific data */
636 struct rcu_head rcu_head
;
638 struct delayed_work orphans_remove
;
639 bool orphans_remove_sched
;
643 * Number of contexts where an event can trigger:
644 * task, softirq, hardirq, nmi.
646 #define PERF_NR_CONTEXTS 4
649 * struct perf_event_cpu_context - per cpu event context structure
651 struct perf_cpu_context
{
652 struct perf_event_context ctx
;
653 struct perf_event_context
*task_ctx
;
657 raw_spinlock_t hrtimer_lock
;
658 struct hrtimer hrtimer
;
659 ktime_t hrtimer_interval
;
660 unsigned int hrtimer_active
;
662 struct pmu
*unique_pmu
;
663 struct perf_cgroup
*cgrp
;
666 struct perf_output_handle
{
667 struct perf_event
*event
;
668 struct ring_buffer
*rb
;
669 unsigned long wakeup
;
678 #ifdef CONFIG_CGROUP_PERF
681 * perf_cgroup_info keeps track of time_enabled for a cgroup.
682 * This is a per-cpu dynamically allocated data structure.
684 struct perf_cgroup_info
{
690 struct cgroup_subsys_state css
;
691 struct perf_cgroup_info __percpu
*info
;
695 * Must ensure cgroup is pinned (css_get) before calling
696 * this function. In other words, we cannot call this function
697 * if there is no cgroup event for the current CPU context.
699 static inline struct perf_cgroup
*
700 perf_cgroup_from_task(struct task_struct
*task
, struct perf_event_context
*ctx
)
702 return container_of(task_css_check(task
, perf_event_cgrp_id
,
703 ctx
? lockdep_is_held(&ctx
->lock
)
705 struct perf_cgroup
, css
);
707 #endif /* CONFIG_CGROUP_PERF */
709 #ifdef CONFIG_PERF_EVENTS
711 extern void *perf_aux_output_begin(struct perf_output_handle
*handle
,
712 struct perf_event
*event
);
713 extern void perf_aux_output_end(struct perf_output_handle
*handle
,
714 unsigned long size
, bool truncated
);
715 extern int perf_aux_output_skip(struct perf_output_handle
*handle
,
717 extern void *perf_get_aux(struct perf_output_handle
*handle
);
719 extern int perf_pmu_register(struct pmu
*pmu
, const char *name
, int type
);
720 extern void perf_pmu_unregister(struct pmu
*pmu
);
722 extern int perf_num_counters(void);
723 extern const char *perf_pmu_name(void);
724 extern void __perf_event_task_sched_in(struct task_struct
*prev
,
725 struct task_struct
*task
);
726 extern void __perf_event_task_sched_out(struct task_struct
*prev
,
727 struct task_struct
*next
);
728 extern int perf_event_init_task(struct task_struct
*child
);
729 extern void perf_event_exit_task(struct task_struct
*child
);
730 extern void perf_event_free_task(struct task_struct
*task
);
731 extern void perf_event_delayed_put(struct task_struct
*task
);
732 extern struct perf_event
*perf_event_get(unsigned int fd
);
733 extern const struct perf_event_attr
*perf_event_attrs(struct perf_event
*event
);
734 extern void perf_event_print_debug(void);
735 extern void perf_pmu_disable(struct pmu
*pmu
);
736 extern void perf_pmu_enable(struct pmu
*pmu
);
737 extern void perf_sched_cb_dec(struct pmu
*pmu
);
738 extern void perf_sched_cb_inc(struct pmu
*pmu
);
739 extern int perf_event_task_disable(void);
740 extern int perf_event_task_enable(void);
741 extern int perf_event_refresh(struct perf_event
*event
, int refresh
);
742 extern void perf_event_update_userpage(struct perf_event
*event
);
743 extern int perf_event_release_kernel(struct perf_event
*event
);
744 extern struct perf_event
*
745 perf_event_create_kernel_counter(struct perf_event_attr
*attr
,
747 struct task_struct
*task
,
748 perf_overflow_handler_t callback
,
750 extern void perf_pmu_migrate_context(struct pmu
*pmu
,
751 int src_cpu
, int dst_cpu
);
752 extern u64
perf_event_read_local(struct perf_event
*event
);
753 extern u64
perf_event_read_value(struct perf_event
*event
,
754 u64
*enabled
, u64
*running
);
757 struct perf_sample_data
{
759 * Fields set by perf_sample_data_init(), group so as to
760 * minimize the cachelines touched.
763 struct perf_raw_record
*raw
;
764 struct perf_branch_stack
*br_stack
;
768 union perf_mem_data_src data_src
;
771 * The other fields, optionally {set,used} by
772 * perf_{prepare,output}_sample().
787 struct perf_callchain_entry
*callchain
;
790 * regs_user may point to task_pt_regs or to regs_user_copy, depending
793 struct perf_regs regs_user
;
794 struct pt_regs regs_user_copy
;
796 struct perf_regs regs_intr
;
798 } ____cacheline_aligned
;
800 /* default value for data source */
801 #define PERF_MEM_NA (PERF_MEM_S(OP, NA) |\
802 PERF_MEM_S(LVL, NA) |\
803 PERF_MEM_S(SNOOP, NA) |\
804 PERF_MEM_S(LOCK, NA) |\
807 static inline void perf_sample_data_init(struct perf_sample_data
*data
,
808 u64 addr
, u64 period
)
810 /* remaining struct members initialized in perf_prepare_sample() */
813 data
->br_stack
= NULL
;
814 data
->period
= period
;
816 data
->data_src
.val
= PERF_MEM_NA
;
820 extern void perf_output_sample(struct perf_output_handle
*handle
,
821 struct perf_event_header
*header
,
822 struct perf_sample_data
*data
,
823 struct perf_event
*event
);
824 extern void perf_prepare_sample(struct perf_event_header
*header
,
825 struct perf_sample_data
*data
,
826 struct perf_event
*event
,
827 struct pt_regs
*regs
);
829 extern int perf_event_overflow(struct perf_event
*event
,
830 struct perf_sample_data
*data
,
831 struct pt_regs
*regs
);
833 extern void perf_event_output(struct perf_event
*event
,
834 struct perf_sample_data
*data
,
835 struct pt_regs
*regs
);
838 perf_event_header__init_id(struct perf_event_header
*header
,
839 struct perf_sample_data
*data
,
840 struct perf_event
*event
);
842 perf_event__output_id_sample(struct perf_event
*event
,
843 struct perf_output_handle
*handle
,
844 struct perf_sample_data
*sample
);
847 perf_log_lost_samples(struct perf_event
*event
, u64 lost
);
849 static inline bool is_sampling_event(struct perf_event
*event
)
851 return event
->attr
.sample_period
!= 0;
855 * Return 1 for a software event, 0 for a hardware event
857 static inline int is_software_event(struct perf_event
*event
)
859 return event
->pmu
->task_ctx_nr
== perf_sw_context
;
862 extern struct static_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
864 extern void ___perf_sw_event(u32
, u64
, struct pt_regs
*, u64
);
865 extern void __perf_sw_event(u32
, u64
, struct pt_regs
*, u64
);
867 #ifndef perf_arch_fetch_caller_regs
868 static inline void perf_arch_fetch_caller_regs(struct pt_regs
*regs
, unsigned long ip
) { }
872 * Take a snapshot of the regs. Skip ip and frame pointer to
873 * the nth caller. We only need a few of the regs:
874 * - ip for PERF_SAMPLE_IP
875 * - cs for user_mode() tests
876 * - bp for callchains
877 * - eflags, for future purposes, just in case
879 static inline void perf_fetch_caller_regs(struct pt_regs
*regs
)
881 memset(regs
, 0, sizeof(*regs
));
883 perf_arch_fetch_caller_regs(regs
, CALLER_ADDR0
);
886 static __always_inline
void
887 perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
889 if (static_key_false(&perf_swevent_enabled
[event_id
]))
890 __perf_sw_event(event_id
, nr
, regs
, addr
);
893 DECLARE_PER_CPU(struct pt_regs
, __perf_regs
[4]);
896 * 'Special' version for the scheduler, it hard assumes no recursion,
897 * which is guaranteed by us not actually scheduling inside other swevents
898 * because those disable preemption.
900 static __always_inline
void
901 perf_sw_event_sched(u32 event_id
, u64 nr
, u64 addr
)
903 if (static_key_false(&perf_swevent_enabled
[event_id
])) {
904 struct pt_regs
*regs
= this_cpu_ptr(&__perf_regs
[0]);
906 perf_fetch_caller_regs(regs
);
907 ___perf_sw_event(event_id
, nr
, regs
, addr
);
911 extern struct static_key_deferred perf_sched_events
;
913 static __always_inline
bool
914 perf_sw_migrate_enabled(void)
916 if (static_key_false(&perf_swevent_enabled
[PERF_COUNT_SW_CPU_MIGRATIONS
]))
921 static inline void perf_event_task_migrate(struct task_struct
*task
)
923 if (perf_sw_migrate_enabled())
924 task
->sched_migrated
= 1;
927 static inline void perf_event_task_sched_in(struct task_struct
*prev
,
928 struct task_struct
*task
)
930 if (static_key_false(&perf_sched_events
.key
))
931 __perf_event_task_sched_in(prev
, task
);
933 if (perf_sw_migrate_enabled() && task
->sched_migrated
) {
934 struct pt_regs
*regs
= this_cpu_ptr(&__perf_regs
[0]);
936 perf_fetch_caller_regs(regs
);
937 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, regs
, 0);
938 task
->sched_migrated
= 0;
942 static inline void perf_event_task_sched_out(struct task_struct
*prev
,
943 struct task_struct
*next
)
945 perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES
, 1, 0);
947 if (static_key_false(&perf_sched_events
.key
))
948 __perf_event_task_sched_out(prev
, next
);
951 static inline u64
__perf_event_count(struct perf_event
*event
)
953 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
956 extern void perf_event_mmap(struct vm_area_struct
*vma
);
957 extern struct perf_guest_info_callbacks
*perf_guest_cbs
;
958 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*callbacks
);
959 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*callbacks
);
961 extern void perf_event_exec(void);
962 extern void perf_event_comm(struct task_struct
*tsk
, bool exec
);
963 extern void perf_event_fork(struct task_struct
*tsk
);
966 DECLARE_PER_CPU(struct perf_callchain_entry
, perf_callchain_entry
);
968 extern void perf_callchain_user(struct perf_callchain_entry
*entry
, struct pt_regs
*regs
);
969 extern void perf_callchain_kernel(struct perf_callchain_entry
*entry
, struct pt_regs
*regs
);
971 static inline void perf_callchain_store(struct perf_callchain_entry
*entry
, u64 ip
)
973 if (entry
->nr
< PERF_MAX_STACK_DEPTH
)
974 entry
->ip
[entry
->nr
++] = ip
;
977 extern int sysctl_perf_event_paranoid
;
978 extern int sysctl_perf_event_mlock
;
979 extern int sysctl_perf_event_sample_rate
;
980 extern int sysctl_perf_cpu_time_max_percent
;
982 extern void perf_sample_event_took(u64 sample_len_ns
);
984 extern int perf_proc_update_handler(struct ctl_table
*table
, int write
,
985 void __user
*buffer
, size_t *lenp
,
987 extern int perf_cpu_time_max_percent_handler(struct ctl_table
*table
, int write
,
988 void __user
*buffer
, size_t *lenp
,
992 static inline bool perf_paranoid_tracepoint_raw(void)
994 return sysctl_perf_event_paranoid
> -1;
997 static inline bool perf_paranoid_cpu(void)
999 return sysctl_perf_event_paranoid
> 0;
1002 static inline bool perf_paranoid_kernel(void)
1004 return sysctl_perf_event_paranoid
> 1;
1007 extern void perf_event_init(void);
1008 extern void perf_tp_event(u64 addr
, u64 count
, void *record
,
1009 int entry_size
, struct pt_regs
*regs
,
1010 struct hlist_head
*head
, int rctx
,
1011 struct task_struct
*task
);
1012 extern void perf_bp_event(struct perf_event
*event
, void *data
);
1014 #ifndef perf_misc_flags
1015 # define perf_misc_flags(regs) \
1016 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1017 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1020 static inline bool has_branch_stack(struct perf_event
*event
)
1022 return event
->attr
.sample_type
& PERF_SAMPLE_BRANCH_STACK
;
1025 static inline bool needs_branch_stack(struct perf_event
*event
)
1027 return event
->attr
.branch_sample_type
!= 0;
1030 static inline bool has_aux(struct perf_event
*event
)
1032 return event
->pmu
->setup_aux
;
1035 extern int perf_output_begin(struct perf_output_handle
*handle
,
1036 struct perf_event
*event
, unsigned int size
);
1037 extern void perf_output_end(struct perf_output_handle
*handle
);
1038 extern unsigned int perf_output_copy(struct perf_output_handle
*handle
,
1039 const void *buf
, unsigned int len
);
1040 extern unsigned int perf_output_skip(struct perf_output_handle
*handle
,
1042 extern int perf_swevent_get_recursion_context(void);
1043 extern void perf_swevent_put_recursion_context(int rctx
);
1044 extern u64
perf_swevent_set_period(struct perf_event
*event
);
1045 extern void perf_event_enable(struct perf_event
*event
);
1046 extern void perf_event_disable(struct perf_event
*event
);
1047 extern int __perf_event_disable(void *info
);
1048 extern void perf_event_task_tick(void);
1049 #else /* !CONFIG_PERF_EVENTS: */
1050 static inline void *
1051 perf_aux_output_begin(struct perf_output_handle
*handle
,
1052 struct perf_event
*event
) { return NULL
; }
1054 perf_aux_output_end(struct perf_output_handle
*handle
, unsigned long size
,
1057 perf_aux_output_skip(struct perf_output_handle
*handle
,
1058 unsigned long size
) { return -EINVAL
; }
1059 static inline void *
1060 perf_get_aux(struct perf_output_handle
*handle
) { return NULL
; }
1062 perf_event_task_migrate(struct task_struct
*task
) { }
1064 perf_event_task_sched_in(struct task_struct
*prev
,
1065 struct task_struct
*task
) { }
1067 perf_event_task_sched_out(struct task_struct
*prev
,
1068 struct task_struct
*next
) { }
1069 static inline int perf_event_init_task(struct task_struct
*child
) { return 0; }
1070 static inline void perf_event_exit_task(struct task_struct
*child
) { }
1071 static inline void perf_event_free_task(struct task_struct
*task
) { }
1072 static inline void perf_event_delayed_put(struct task_struct
*task
) { }
1073 static inline struct perf_event
*perf_event_get(unsigned int fd
) { return ERR_PTR(-EINVAL
); }
1074 static inline const struct perf_event_attr
*perf_event_attrs(struct perf_event
*event
)
1076 return ERR_PTR(-EINVAL
);
1078 static inline u64
perf_event_read_local(struct perf_event
*event
) { return -EINVAL
; }
1079 static inline void perf_event_print_debug(void) { }
1080 static inline int perf_event_task_disable(void) { return -EINVAL
; }
1081 static inline int perf_event_task_enable(void) { return -EINVAL
; }
1082 static inline int perf_event_refresh(struct perf_event
*event
, int refresh
)
1088 perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
) { }
1090 perf_sw_event_sched(u32 event_id
, u64 nr
, u64 addr
) { }
1092 perf_bp_event(struct perf_event
*event
, void *data
) { }
1094 static inline int perf_register_guest_info_callbacks
1095 (struct perf_guest_info_callbacks
*callbacks
) { return 0; }
1096 static inline int perf_unregister_guest_info_callbacks
1097 (struct perf_guest_info_callbacks
*callbacks
) { return 0; }
1099 static inline void perf_event_mmap(struct vm_area_struct
*vma
) { }
1100 static inline void perf_event_exec(void) { }
1101 static inline void perf_event_comm(struct task_struct
*tsk
, bool exec
) { }
1102 static inline void perf_event_fork(struct task_struct
*tsk
) { }
1103 static inline void perf_event_init(void) { }
1104 static inline int perf_swevent_get_recursion_context(void) { return -1; }
1105 static inline void perf_swevent_put_recursion_context(int rctx
) { }
1106 static inline u64
perf_swevent_set_period(struct perf_event
*event
) { return 0; }
1107 static inline void perf_event_enable(struct perf_event
*event
) { }
1108 static inline void perf_event_disable(struct perf_event
*event
) { }
1109 static inline int __perf_event_disable(void *info
) { return -1; }
1110 static inline void perf_event_task_tick(void) { }
1111 static inline int perf_event_release_kernel(struct perf_event
*event
) { return 0; }
1114 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_NO_HZ_FULL)
1115 extern bool perf_event_can_stop_tick(void);
1117 static inline bool perf_event_can_stop_tick(void) { return true; }
1120 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1121 extern void perf_restore_debug_store(void);
1123 static inline void perf_restore_debug_store(void) { }
1126 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1129 * This has to have a higher priority than migration_notifier in sched/core.c.
1131 #define perf_cpu_notifier(fn) \
1133 static struct notifier_block fn##_nb = \
1134 { .notifier_call = fn, .priority = CPU_PRI_PERF }; \
1135 unsigned long cpu = smp_processor_id(); \
1136 unsigned long flags; \
1138 cpu_notifier_register_begin(); \
1139 fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE, \
1140 (void *)(unsigned long)cpu); \
1141 local_irq_save(flags); \
1142 fn(&fn##_nb, (unsigned long)CPU_STARTING, \
1143 (void *)(unsigned long)cpu); \
1144 local_irq_restore(flags); \
1145 fn(&fn##_nb, (unsigned long)CPU_ONLINE, \
1146 (void *)(unsigned long)cpu); \
1147 __register_cpu_notifier(&fn##_nb); \
1148 cpu_notifier_register_done(); \
1152 * Bare-bones version of perf_cpu_notifier(), which doesn't invoke the
1153 * callback for already online CPUs.
1155 #define __perf_cpu_notifier(fn) \
1157 static struct notifier_block fn##_nb = \
1158 { .notifier_call = fn, .priority = CPU_PRI_PERF }; \
1160 __register_cpu_notifier(&fn##_nb); \
1163 struct perf_pmu_events_attr
{
1164 struct device_attribute attr
;
1166 const char *event_str
;
1169 ssize_t
perf_event_sysfs_show(struct device
*dev
, struct device_attribute
*attr
,
1172 #define PMU_EVENT_ATTR(_name, _var, _id, _show) \
1173 static struct perf_pmu_events_attr _var = { \
1174 .attr = __ATTR(_name, 0444, _show, NULL), \
1178 #define PMU_EVENT_ATTR_STRING(_name, _var, _str) \
1179 static struct perf_pmu_events_attr _var = { \
1180 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1182 .event_str = _str, \
1185 #define PMU_FORMAT_ATTR(_name, _format) \
1187 _name##_show(struct device *dev, \
1188 struct device_attribute *attr, \
1191 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
1192 return sprintf(page, _format "\n"); \
1195 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1197 #endif /* _LINUX_PERF_EVENT_H */