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
[0]; /* /proc/sys/kernel/perf_event_max_stack */
64 struct perf_callchain_entry_ctx
{
65 struct perf_callchain_entry
*entry
;
72 typedef unsigned long (*perf_copy_f
)(void *dst
, const void *src
,
73 unsigned long off
, unsigned long len
);
75 struct perf_raw_frag
{
77 struct perf_raw_frag
*next
;
85 struct perf_raw_record
{
86 struct perf_raw_frag frag
;
91 * branch stack layout:
92 * nr: number of taken branches stored in entries[]
94 * Note that nr can vary from sample to sample
95 * branches (to, from) are stored from most recent
96 * to least recent, i.e., entries[0] contains the most
99 struct perf_branch_stack
{
101 struct perf_branch_entry entries
[0];
107 * extra PMU register associated with an event
109 struct hw_perf_event_extra
{
110 u64 config
; /* register value */
111 unsigned int reg
; /* register address or index */
112 int alloc
; /* extra register already allocated */
113 int idx
; /* index in shared_regs->regs[] */
117 * struct hw_perf_event - performance event hardware details:
119 struct hw_perf_event
{
120 #ifdef CONFIG_PERF_EVENTS
122 struct { /* hardware */
125 unsigned long config_base
;
126 unsigned long event_base
;
127 int event_base_rdpmc
;
132 struct hw_perf_event_extra extra_reg
;
133 struct hw_perf_event_extra branch_reg
;
135 struct { /* software */
136 struct hrtimer hrtimer
;
138 struct { /* tracepoint */
139 /* for tp_event->class */
140 struct list_head tp_list
;
142 struct { /* intel_cqm */
146 struct list_head cqm_events_entry
;
147 struct list_head cqm_groups_entry
;
148 struct list_head cqm_group_entry
;
150 struct { /* itrace */
153 struct { /* amd_power */
157 #ifdef CONFIG_HAVE_HW_BREAKPOINT
158 struct { /* breakpoint */
160 * Crufty hack to avoid the chicken and egg
161 * problem hw_breakpoint has with context
162 * creation and event initalization.
164 struct arch_hw_breakpoint info
;
165 struct list_head bp_list
;
170 * If the event is a per task event, this will point to the task in
171 * question. See the comment in perf_event_alloc().
173 struct task_struct
*target
;
176 * PMU would store hardware filter configuration
181 /* Last sync'ed generation of filters */
182 unsigned long addr_filters_gen
;
185 * hw_perf_event::state flags; used to track the PERF_EF_* state.
187 #define PERF_HES_STOPPED 0x01 /* the counter is stopped */
188 #define PERF_HES_UPTODATE 0x02 /* event->count up-to-date */
189 #define PERF_HES_ARCH 0x04
194 * The last observed hardware counter value, updated with a
195 * local64_cmpxchg() such that pmu::read() can be called nested.
197 local64_t prev_count
;
200 * The period to start the next sample with.
205 * The period we started this sample with.
210 * However much is left of the current period; note that this is
211 * a full 64bit value and allows for generation of periods longer
212 * than hardware might allow.
214 local64_t period_left
;
217 * State for throttling the event, see __perf_event_overflow() and
218 * perf_adjust_freq_unthr_context().
224 * State for freq target events, see __perf_event_overflow() and
225 * perf_adjust_freq_unthr_context().
228 u64 freq_count_stamp
;
235 * Common implementation detail of pmu::{start,commit,cancel}_txn
237 #define PERF_PMU_TXN_ADD 0x1 /* txn to add/schedule event on PMU */
238 #define PERF_PMU_TXN_READ 0x2 /* txn to read event group from PMU */
241 * pmu::capabilities flags
243 #define PERF_PMU_CAP_NO_INTERRUPT 0x01
244 #define PERF_PMU_CAP_NO_NMI 0x02
245 #define PERF_PMU_CAP_AUX_NO_SG 0x04
246 #define PERF_PMU_CAP_AUX_SW_DOUBLEBUF 0x08
247 #define PERF_PMU_CAP_EXCLUSIVE 0x10
248 #define PERF_PMU_CAP_ITRACE 0x20
249 #define PERF_PMU_CAP_HETEROGENEOUS_CPUS 0x40
252 * struct pmu - generic performance monitoring unit
255 struct list_head entry
;
257 struct module
*module
;
259 const struct attribute_group
**attr_groups
;
264 * various common per-pmu feature flags
268 int * __percpu pmu_disable_count
;
269 struct perf_cpu_context
* __percpu pmu_cpu_context
;
270 atomic_t exclusive_cnt
; /* < 0: cpu; > 0: tsk */
272 int hrtimer_interval_ms
;
274 /* number of address filters this PMU can do */
275 unsigned int nr_addr_filters
;
278 * Fully disable/enable this PMU, can be used to protect from the PMI
279 * as well as for lazy/batch writing of the MSRs.
281 void (*pmu_enable
) (struct pmu
*pmu
); /* optional */
282 void (*pmu_disable
) (struct pmu
*pmu
); /* optional */
285 * Try and initialize the event for this PMU.
288 * -ENOENT -- @event is not for this PMU
290 * -ENODEV -- @event is for this PMU but PMU not present
291 * -EBUSY -- @event is for this PMU but PMU temporarily unavailable
292 * -EINVAL -- @event is for this PMU but @event is not valid
293 * -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
294 * -EACCESS -- @event is for this PMU, @event is valid, but no privilidges
296 * 0 -- @event is for this PMU and valid
298 * Other error return values are allowed.
300 int (*event_init
) (struct perf_event
*event
);
303 * Notification that the event was mapped or unmapped. Called
304 * in the context of the mapping task.
306 void (*event_mapped
) (struct perf_event
*event
); /*optional*/
307 void (*event_unmapped
) (struct perf_event
*event
); /*optional*/
310 * Flags for ->add()/->del()/ ->start()/->stop(). There are
311 * matching hw_perf_event::state flags.
313 #define PERF_EF_START 0x01 /* start the counter when adding */
314 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
315 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
318 * Adds/Removes a counter to/from the PMU, can be done inside a
319 * transaction, see the ->*_txn() methods.
321 * The add/del callbacks will reserve all hardware resources required
322 * to service the event, this includes any counter constraint
325 * Called with IRQs disabled and the PMU disabled on the CPU the event
328 * ->add() called without PERF_EF_START should result in the same state
329 * as ->add() followed by ->stop().
331 * ->del() must always PERF_EF_UPDATE stop an event. If it calls
332 * ->stop() that must deal with already being stopped without
335 int (*add
) (struct perf_event
*event
, int flags
);
336 void (*del
) (struct perf_event
*event
, int flags
);
339 * Starts/Stops a counter present on the PMU.
341 * The PMI handler should stop the counter when perf_event_overflow()
342 * returns !0. ->start() will be used to continue.
344 * Also used to change the sample period.
346 * Called with IRQs disabled and the PMU disabled on the CPU the event
347 * is on -- will be called from NMI context with the PMU generates
350 * ->stop() with PERF_EF_UPDATE will read the counter and update
351 * period/count values like ->read() would.
353 * ->start() with PERF_EF_RELOAD will reprogram the the counter
354 * value, must be preceded by a ->stop() with PERF_EF_UPDATE.
356 void (*start
) (struct perf_event
*event
, int flags
);
357 void (*stop
) (struct perf_event
*event
, int flags
);
360 * Updates the counter value of the event.
362 * For sampling capable PMUs this will also update the software period
363 * hw_perf_event::period_left field.
365 void (*read
) (struct perf_event
*event
);
368 * Group events scheduling is treated as a transaction, add
369 * group events as a whole and perform one schedulability test.
370 * If the test fails, roll back the whole group
372 * Start the transaction, after this ->add() doesn't need to
373 * do schedulability tests.
377 void (*start_txn
) (struct pmu
*pmu
, unsigned int txn_flags
);
379 * If ->start_txn() disabled the ->add() schedulability test
380 * then ->commit_txn() is required to perform one. On success
381 * the transaction is closed. On error the transaction is kept
382 * open until ->cancel_txn() is called.
386 int (*commit_txn
) (struct pmu
*pmu
);
388 * Will cancel the transaction, assumes ->del() is called
389 * for each successful ->add() during the transaction.
393 void (*cancel_txn
) (struct pmu
*pmu
);
396 * Will return the value for perf_event_mmap_page::index for this event,
397 * if no implementation is provided it will default to: event->hw.idx + 1.
399 int (*event_idx
) (struct perf_event
*event
); /*optional */
402 * context-switches callback
404 void (*sched_task
) (struct perf_event_context
*ctx
,
407 * PMU specific data size
409 size_t task_ctx_size
;
413 * Return the count value for a counter.
415 u64 (*count
) (struct perf_event
*event
); /*optional*/
418 * Set up pmu-private data structures for an AUX area
420 void *(*setup_aux
) (int cpu
, void **pages
,
421 int nr_pages
, bool overwrite
);
425 * Free pmu-private AUX data structures
427 void (*free_aux
) (void *aux
); /* optional */
430 * Validate address range filters: make sure the HW supports the
431 * requested configuration and number of filters; return 0 if the
432 * supplied filters are valid, -errno otherwise.
434 * Runs in the context of the ioctl()ing process and is not serialized
435 * with the rest of the PMU callbacks.
437 int (*addr_filters_validate
) (struct list_head
*filters
);
441 * Synchronize address range filter configuration:
442 * translate hw-agnostic filters into hardware configuration in
443 * event::hw::addr_filters.
445 * Runs as a part of filter sync sequence that is done in ->start()
446 * callback by calling perf_event_addr_filters_sync().
448 * May (and should) traverse event::addr_filters::list, for which its
449 * caller provides necessary serialization.
451 void (*addr_filters_sync
) (struct perf_event
*event
);
455 * Filter events for PMU-specific reasons.
457 int (*filter_match
) (struct perf_event
*event
); /* optional */
461 * struct perf_addr_filter - address range filter definition
462 * @entry: event's filter list linkage
463 * @inode: object file's inode for file-based filters
464 * @offset: filter range offset
465 * @size: filter range size
466 * @range: 1: range, 0: address
467 * @filter: 1: filter/start, 0: stop
469 * This is a hardware-agnostic filter configuration as specified by the user.
471 struct perf_addr_filter
{
472 struct list_head entry
;
474 unsigned long offset
;
476 unsigned int range
: 1,
481 * struct perf_addr_filters_head - container for address range filters
482 * @list: list of filters for this event
483 * @lock: spinlock that serializes accesses to the @list and event's
484 * (and its children's) filter generations.
486 * A child event will use parent's @list (and therefore @lock), so they are
487 * bundled together; see perf_event_addr_filters().
489 struct perf_addr_filters_head
{
490 struct list_head list
;
495 * enum perf_event_active_state - the states of a event
497 enum perf_event_active_state
{
498 PERF_EVENT_STATE_DEAD
= -4,
499 PERF_EVENT_STATE_EXIT
= -3,
500 PERF_EVENT_STATE_ERROR
= -2,
501 PERF_EVENT_STATE_OFF
= -1,
502 PERF_EVENT_STATE_INACTIVE
= 0,
503 PERF_EVENT_STATE_ACTIVE
= 1,
507 struct perf_sample_data
;
509 typedef void (*perf_overflow_handler_t
)(struct perf_event
*,
510 struct perf_sample_data
*,
511 struct pt_regs
*regs
);
514 * Event capabilities. For event_caps and groups caps.
516 * PERF_EV_CAP_SOFTWARE: Is a software event.
517 * PERF_EV_CAP_READ_ACTIVE_PKG: A CPU event (or cgroup event) that can be read
518 * from any CPU in the package where it is active.
520 #define PERF_EV_CAP_SOFTWARE BIT(0)
521 #define PERF_EV_CAP_READ_ACTIVE_PKG BIT(1)
523 #define SWEVENT_HLIST_BITS 8
524 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
526 struct swevent_hlist
{
527 struct hlist_head heads
[SWEVENT_HLIST_SIZE
];
528 struct rcu_head rcu_head
;
531 #define PERF_ATTACH_CONTEXT 0x01
532 #define PERF_ATTACH_GROUP 0x02
533 #define PERF_ATTACH_TASK 0x04
534 #define PERF_ATTACH_TASK_DATA 0x08
539 struct pmu_event_list
{
541 struct list_head list
;
545 * struct perf_event - performance event kernel representation:
548 #ifdef CONFIG_PERF_EVENTS
550 * entry onto perf_event_context::event_list;
551 * modifications require ctx->lock
552 * RCU safe iterations.
554 struct list_head event_entry
;
557 * XXX: group_entry and sibling_list should be mutually exclusive;
558 * either you're a sibling on a group, or you're the group leader.
559 * Rework the code to always use the same list element.
561 * Locked for modification by both ctx->mutex and ctx->lock; holding
562 * either sufficies for read.
564 struct list_head group_entry
;
565 struct list_head sibling_list
;
568 * We need storage to track the entries in perf_pmu_migrate_context; we
569 * cannot use the event_entry because of RCU and we want to keep the
570 * group in tact which avoids us using the other two entries.
572 struct list_head migrate_entry
;
574 struct hlist_node hlist_entry
;
575 struct list_head active_entry
;
578 /* Not serialized. Only written during event initialization. */
580 /* The cumulative AND of all event_caps for events in this group. */
583 struct perf_event
*group_leader
;
587 enum perf_event_active_state state
;
588 unsigned int attach_state
;
590 atomic64_t child_count
;
593 * These are the total time in nanoseconds that the event
594 * has been enabled (i.e. eligible to run, and the task has
595 * been scheduled in, if this is a per-task event)
596 * and running (scheduled onto the CPU), respectively.
598 * They are computed from tstamp_enabled, tstamp_running and
599 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
601 u64 total_time_enabled
;
602 u64 total_time_running
;
605 * These are timestamps used for computing total_time_enabled
606 * and total_time_running when the event is in INACTIVE or
607 * ACTIVE state, measured in nanoseconds from an arbitrary point
609 * tstamp_enabled: the notional time when the event was enabled
610 * tstamp_running: the notional time when the event was scheduled on
611 * tstamp_stopped: in INACTIVE state, the notional time when the
612 * event was scheduled off.
619 * timestamp shadows the actual context timing but it can
620 * be safely used in NMI interrupt context. It reflects the
621 * context time as it was when the event was last scheduled in.
623 * ctx_time already accounts for ctx->timestamp. Therefore to
624 * compute ctx_time for a sample, simply add perf_clock().
628 struct perf_event_attr attr
;
632 struct hw_perf_event hw
;
634 struct perf_event_context
*ctx
;
635 atomic_long_t refcount
;
638 * These accumulate total time (in nanoseconds) that children
639 * events have been enabled and running, respectively.
641 atomic64_t child_total_time_enabled
;
642 atomic64_t child_total_time_running
;
645 * Protect attach/detach and child_list:
647 struct mutex child_mutex
;
648 struct list_head child_list
;
649 struct perf_event
*parent
;
654 struct list_head owner_entry
;
655 struct task_struct
*owner
;
658 struct mutex mmap_mutex
;
661 struct ring_buffer
*rb
;
662 struct list_head rb_entry
;
663 unsigned long rcu_batches
;
667 wait_queue_head_t waitq
;
668 struct fasync_struct
*fasync
;
670 /* delayed work for NMIs and such */
674 struct irq_work pending
;
676 atomic_t event_limit
;
678 /* address range filters */
679 struct perf_addr_filters_head addr_filters
;
680 /* vma address array for file-based filders */
681 unsigned long *addr_filters_offs
;
682 unsigned long addr_filters_gen
;
684 void (*destroy
)(struct perf_event
*);
685 struct rcu_head rcu_head
;
687 struct pid_namespace
*ns
;
691 perf_overflow_handler_t overflow_handler
;
692 void *overflow_handler_context
;
693 #ifdef CONFIG_BPF_SYSCALL
694 perf_overflow_handler_t orig_overflow_handler
;
695 struct bpf_prog
*prog
;
698 #ifdef CONFIG_EVENT_TRACING
699 struct trace_event_call
*tp_event
;
700 struct event_filter
*filter
;
701 #ifdef CONFIG_FUNCTION_TRACER
702 struct ftrace_ops ftrace_ops
;
706 #ifdef CONFIG_CGROUP_PERF
707 struct perf_cgroup
*cgrp
; /* cgroup event is attach to */
708 int cgrp_defer_enabled
;
711 struct list_head sb_list
;
712 #endif /* CONFIG_PERF_EVENTS */
716 * struct perf_event_context - event context structure
718 * Used as a container for task events and CPU events as well:
720 struct perf_event_context
{
723 * Protect the states of the events in the list,
724 * nr_active, and the list:
728 * Protect the list of events. Locking either mutex or lock
729 * is sufficient to ensure the list doesn't change; to change
730 * the list you need to lock both the mutex and the spinlock.
734 struct list_head active_ctx_list
;
735 struct list_head pinned_groups
;
736 struct list_head flexible_groups
;
737 struct list_head event_list
;
745 struct task_struct
*task
;
748 * Context clock, runs when context enabled.
754 * These fields let us detect when two contexts have both
755 * been cloned (inherited) from a common ancestor.
757 struct perf_event_context
*parent_ctx
;
761 #ifdef CONFIG_CGROUP_PERF
762 int nr_cgroups
; /* cgroup evts */
764 void *task_ctx_data
; /* pmu specific data */
765 struct rcu_head rcu_head
;
769 * Number of contexts where an event can trigger:
770 * task, softirq, hardirq, nmi.
772 #define PERF_NR_CONTEXTS 4
775 * struct perf_event_cpu_context - per cpu event context structure
777 struct perf_cpu_context
{
778 struct perf_event_context ctx
;
779 struct perf_event_context
*task_ctx
;
783 raw_spinlock_t hrtimer_lock
;
784 struct hrtimer hrtimer
;
785 ktime_t hrtimer_interval
;
786 unsigned int hrtimer_active
;
788 struct pmu
*unique_pmu
;
789 #ifdef CONFIG_CGROUP_PERF
790 struct perf_cgroup
*cgrp
;
793 struct list_head sched_cb_entry
;
797 struct perf_output_handle
{
798 struct perf_event
*event
;
799 struct ring_buffer
*rb
;
800 unsigned long wakeup
;
809 struct bpf_perf_event_data_kern
{
810 struct pt_regs
*regs
;
811 struct perf_sample_data
*data
;
814 #ifdef CONFIG_CGROUP_PERF
817 * perf_cgroup_info keeps track of time_enabled for a cgroup.
818 * This is a per-cpu dynamically allocated data structure.
820 struct perf_cgroup_info
{
826 struct cgroup_subsys_state css
;
827 struct perf_cgroup_info __percpu
*info
;
831 * Must ensure cgroup is pinned (css_get) before calling
832 * this function. In other words, we cannot call this function
833 * if there is no cgroup event for the current CPU context.
835 static inline struct perf_cgroup
*
836 perf_cgroup_from_task(struct task_struct
*task
, struct perf_event_context
*ctx
)
838 return container_of(task_css_check(task
, perf_event_cgrp_id
,
839 ctx
? lockdep_is_held(&ctx
->lock
)
841 struct perf_cgroup
, css
);
843 #endif /* CONFIG_CGROUP_PERF */
845 #ifdef CONFIG_PERF_EVENTS
847 extern void *perf_aux_output_begin(struct perf_output_handle
*handle
,
848 struct perf_event
*event
);
849 extern void perf_aux_output_end(struct perf_output_handle
*handle
,
850 unsigned long size
, bool truncated
);
851 extern int perf_aux_output_skip(struct perf_output_handle
*handle
,
853 extern void *perf_get_aux(struct perf_output_handle
*handle
);
855 extern int perf_pmu_register(struct pmu
*pmu
, const char *name
, int type
);
856 extern void perf_pmu_unregister(struct pmu
*pmu
);
858 extern int perf_num_counters(void);
859 extern const char *perf_pmu_name(void);
860 extern void __perf_event_task_sched_in(struct task_struct
*prev
,
861 struct task_struct
*task
);
862 extern void __perf_event_task_sched_out(struct task_struct
*prev
,
863 struct task_struct
*next
);
864 extern int perf_event_init_task(struct task_struct
*child
);
865 extern void perf_event_exit_task(struct task_struct
*child
);
866 extern void perf_event_free_task(struct task_struct
*task
);
867 extern void perf_event_delayed_put(struct task_struct
*task
);
868 extern struct file
*perf_event_get(unsigned int fd
);
869 extern const struct perf_event_attr
*perf_event_attrs(struct perf_event
*event
);
870 extern void perf_event_print_debug(void);
871 extern void perf_pmu_disable(struct pmu
*pmu
);
872 extern void perf_pmu_enable(struct pmu
*pmu
);
873 extern void perf_sched_cb_dec(struct pmu
*pmu
);
874 extern void perf_sched_cb_inc(struct pmu
*pmu
);
875 extern int perf_event_task_disable(void);
876 extern int perf_event_task_enable(void);
877 extern int perf_event_refresh(struct perf_event
*event
, int refresh
);
878 extern void perf_event_update_userpage(struct perf_event
*event
);
879 extern int perf_event_release_kernel(struct perf_event
*event
);
880 extern struct perf_event
*
881 perf_event_create_kernel_counter(struct perf_event_attr
*attr
,
883 struct task_struct
*task
,
884 perf_overflow_handler_t callback
,
886 extern void perf_pmu_migrate_context(struct pmu
*pmu
,
887 int src_cpu
, int dst_cpu
);
888 extern u64
perf_event_read_local(struct perf_event
*event
);
889 extern u64
perf_event_read_value(struct perf_event
*event
,
890 u64
*enabled
, u64
*running
);
893 struct perf_sample_data
{
895 * Fields set by perf_sample_data_init(), group so as to
896 * minimize the cachelines touched.
899 struct perf_raw_record
*raw
;
900 struct perf_branch_stack
*br_stack
;
904 union perf_mem_data_src data_src
;
907 * The other fields, optionally {set,used} by
908 * perf_{prepare,output}_sample().
923 struct perf_callchain_entry
*callchain
;
926 * regs_user may point to task_pt_regs or to regs_user_copy, depending
929 struct perf_regs regs_user
;
930 struct pt_regs regs_user_copy
;
932 struct perf_regs regs_intr
;
934 } ____cacheline_aligned
;
936 /* default value for data source */
937 #define PERF_MEM_NA (PERF_MEM_S(OP, NA) |\
938 PERF_MEM_S(LVL, NA) |\
939 PERF_MEM_S(SNOOP, NA) |\
940 PERF_MEM_S(LOCK, NA) |\
943 static inline void perf_sample_data_init(struct perf_sample_data
*data
,
944 u64 addr
, u64 period
)
946 /* remaining struct members initialized in perf_prepare_sample() */
949 data
->br_stack
= NULL
;
950 data
->period
= period
;
952 data
->data_src
.val
= PERF_MEM_NA
;
956 extern void perf_output_sample(struct perf_output_handle
*handle
,
957 struct perf_event_header
*header
,
958 struct perf_sample_data
*data
,
959 struct perf_event
*event
);
960 extern void perf_prepare_sample(struct perf_event_header
*header
,
961 struct perf_sample_data
*data
,
962 struct perf_event
*event
,
963 struct pt_regs
*regs
);
965 extern int perf_event_overflow(struct perf_event
*event
,
966 struct perf_sample_data
*data
,
967 struct pt_regs
*regs
);
969 extern void perf_event_output_forward(struct perf_event
*event
,
970 struct perf_sample_data
*data
,
971 struct pt_regs
*regs
);
972 extern void perf_event_output_backward(struct perf_event
*event
,
973 struct perf_sample_data
*data
,
974 struct pt_regs
*regs
);
975 extern void perf_event_output(struct perf_event
*event
,
976 struct perf_sample_data
*data
,
977 struct pt_regs
*regs
);
980 is_default_overflow_handler(struct perf_event
*event
)
982 if (likely(event
->overflow_handler
== perf_event_output_forward
))
984 if (unlikely(event
->overflow_handler
== perf_event_output_backward
))
990 perf_event_header__init_id(struct perf_event_header
*header
,
991 struct perf_sample_data
*data
,
992 struct perf_event
*event
);
994 perf_event__output_id_sample(struct perf_event
*event
,
995 struct perf_output_handle
*handle
,
996 struct perf_sample_data
*sample
);
999 perf_log_lost_samples(struct perf_event
*event
, u64 lost
);
1001 static inline bool is_sampling_event(struct perf_event
*event
)
1003 return event
->attr
.sample_period
!= 0;
1007 * Return 1 for a software event, 0 for a hardware event
1009 static inline int is_software_event(struct perf_event
*event
)
1011 return event
->event_caps
& PERF_EV_CAP_SOFTWARE
;
1014 extern struct static_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
1016 extern void ___perf_sw_event(u32
, u64
, struct pt_regs
*, u64
);
1017 extern void __perf_sw_event(u32
, u64
, struct pt_regs
*, u64
);
1019 #ifndef perf_arch_fetch_caller_regs
1020 static inline void perf_arch_fetch_caller_regs(struct pt_regs
*regs
, unsigned long ip
) { }
1024 * Take a snapshot of the regs. Skip ip and frame pointer to
1025 * the nth caller. We only need a few of the regs:
1026 * - ip for PERF_SAMPLE_IP
1027 * - cs for user_mode() tests
1028 * - bp for callchains
1029 * - eflags, for future purposes, just in case
1031 static inline void perf_fetch_caller_regs(struct pt_regs
*regs
)
1033 perf_arch_fetch_caller_regs(regs
, CALLER_ADDR0
);
1036 static __always_inline
void
1037 perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
1039 if (static_key_false(&perf_swevent_enabled
[event_id
]))
1040 __perf_sw_event(event_id
, nr
, regs
, addr
);
1043 DECLARE_PER_CPU(struct pt_regs
, __perf_regs
[4]);
1046 * 'Special' version for the scheduler, it hard assumes no recursion,
1047 * which is guaranteed by us not actually scheduling inside other swevents
1048 * because those disable preemption.
1050 static __always_inline
void
1051 perf_sw_event_sched(u32 event_id
, u64 nr
, u64 addr
)
1053 if (static_key_false(&perf_swevent_enabled
[event_id
])) {
1054 struct pt_regs
*regs
= this_cpu_ptr(&__perf_regs
[0]);
1056 perf_fetch_caller_regs(regs
);
1057 ___perf_sw_event(event_id
, nr
, regs
, addr
);
1061 extern struct static_key_false perf_sched_events
;
1063 static __always_inline
bool
1064 perf_sw_migrate_enabled(void)
1066 if (static_key_false(&perf_swevent_enabled
[PERF_COUNT_SW_CPU_MIGRATIONS
]))
1071 static inline void perf_event_task_migrate(struct task_struct
*task
)
1073 if (perf_sw_migrate_enabled())
1074 task
->sched_migrated
= 1;
1077 static inline void perf_event_task_sched_in(struct task_struct
*prev
,
1078 struct task_struct
*task
)
1080 if (static_branch_unlikely(&perf_sched_events
))
1081 __perf_event_task_sched_in(prev
, task
);
1083 if (perf_sw_migrate_enabled() && task
->sched_migrated
) {
1084 struct pt_regs
*regs
= this_cpu_ptr(&__perf_regs
[0]);
1086 perf_fetch_caller_regs(regs
);
1087 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, regs
, 0);
1088 task
->sched_migrated
= 0;
1092 static inline void perf_event_task_sched_out(struct task_struct
*prev
,
1093 struct task_struct
*next
)
1095 perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES
, 1, 0);
1097 if (static_branch_unlikely(&perf_sched_events
))
1098 __perf_event_task_sched_out(prev
, next
);
1101 static inline u64
__perf_event_count(struct perf_event
*event
)
1103 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
1106 extern void perf_event_mmap(struct vm_area_struct
*vma
);
1107 extern struct perf_guest_info_callbacks
*perf_guest_cbs
;
1108 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*callbacks
);
1109 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*callbacks
);
1111 extern void perf_event_exec(void);
1112 extern void perf_event_comm(struct task_struct
*tsk
, bool exec
);
1113 extern void perf_event_fork(struct task_struct
*tsk
);
1116 DECLARE_PER_CPU(struct perf_callchain_entry
, perf_callchain_entry
);
1118 extern void perf_callchain_user(struct perf_callchain_entry_ctx
*entry
, struct pt_regs
*regs
);
1119 extern void perf_callchain_kernel(struct perf_callchain_entry_ctx
*entry
, struct pt_regs
*regs
);
1120 extern struct perf_callchain_entry
*
1121 get_perf_callchain(struct pt_regs
*regs
, u32 init_nr
, bool kernel
, bool user
,
1122 u32 max_stack
, bool crosstask
, bool add_mark
);
1123 extern int get_callchain_buffers(int max_stack
);
1124 extern void put_callchain_buffers(void);
1126 extern int sysctl_perf_event_max_stack
;
1127 extern int sysctl_perf_event_max_contexts_per_stack
;
1129 static inline int perf_callchain_store_context(struct perf_callchain_entry_ctx
*ctx
, u64 ip
)
1131 if (ctx
->contexts
< sysctl_perf_event_max_contexts_per_stack
) {
1132 struct perf_callchain_entry
*entry
= ctx
->entry
;
1133 entry
->ip
[entry
->nr
++] = ip
;
1137 ctx
->contexts_maxed
= true;
1138 return -1; /* no more room, stop walking the stack */
1142 static inline int perf_callchain_store(struct perf_callchain_entry_ctx
*ctx
, u64 ip
)
1144 if (ctx
->nr
< ctx
->max_stack
&& !ctx
->contexts_maxed
) {
1145 struct perf_callchain_entry
*entry
= ctx
->entry
;
1146 entry
->ip
[entry
->nr
++] = ip
;
1150 return -1; /* no more room, stop walking the stack */
1154 extern int sysctl_perf_event_paranoid
;
1155 extern int sysctl_perf_event_mlock
;
1156 extern int sysctl_perf_event_sample_rate
;
1157 extern int sysctl_perf_cpu_time_max_percent
;
1159 extern void perf_sample_event_took(u64 sample_len_ns
);
1161 extern int perf_proc_update_handler(struct ctl_table
*table
, int write
,
1162 void __user
*buffer
, size_t *lenp
,
1164 extern int perf_cpu_time_max_percent_handler(struct ctl_table
*table
, int write
,
1165 void __user
*buffer
, size_t *lenp
,
1168 int perf_event_max_stack_handler(struct ctl_table
*table
, int write
,
1169 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
);
1171 static inline bool perf_paranoid_tracepoint_raw(void)
1173 return sysctl_perf_event_paranoid
> -1;
1176 static inline bool perf_paranoid_cpu(void)
1178 return sysctl_perf_event_paranoid
> 0;
1181 static inline bool perf_paranoid_kernel(void)
1183 return sysctl_perf_event_paranoid
> 1;
1186 extern void perf_event_init(void);
1187 extern void perf_tp_event(u16 event_type
, u64 count
, void *record
,
1188 int entry_size
, struct pt_regs
*regs
,
1189 struct hlist_head
*head
, int rctx
,
1190 struct task_struct
*task
);
1191 extern void perf_bp_event(struct perf_event
*event
, void *data
);
1193 #ifndef perf_misc_flags
1194 # define perf_misc_flags(regs) \
1195 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1196 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1199 static inline bool has_branch_stack(struct perf_event
*event
)
1201 return event
->attr
.sample_type
& PERF_SAMPLE_BRANCH_STACK
;
1204 static inline bool needs_branch_stack(struct perf_event
*event
)
1206 return event
->attr
.branch_sample_type
!= 0;
1209 static inline bool has_aux(struct perf_event
*event
)
1211 return event
->pmu
->setup_aux
;
1214 static inline bool is_write_backward(struct perf_event
*event
)
1216 return !!event
->attr
.write_backward
;
1219 static inline bool has_addr_filter(struct perf_event
*event
)
1221 return event
->pmu
->nr_addr_filters
;
1225 * An inherited event uses parent's filters
1227 static inline struct perf_addr_filters_head
*
1228 perf_event_addr_filters(struct perf_event
*event
)
1230 struct perf_addr_filters_head
*ifh
= &event
->addr_filters
;
1233 ifh
= &event
->parent
->addr_filters
;
1238 extern void perf_event_addr_filters_sync(struct perf_event
*event
);
1240 extern int perf_output_begin(struct perf_output_handle
*handle
,
1241 struct perf_event
*event
, unsigned int size
);
1242 extern int perf_output_begin_forward(struct perf_output_handle
*handle
,
1243 struct perf_event
*event
,
1245 extern int perf_output_begin_backward(struct perf_output_handle
*handle
,
1246 struct perf_event
*event
,
1249 extern void perf_output_end(struct perf_output_handle
*handle
);
1250 extern unsigned int perf_output_copy(struct perf_output_handle
*handle
,
1251 const void *buf
, unsigned int len
);
1252 extern unsigned int perf_output_skip(struct perf_output_handle
*handle
,
1254 extern int perf_swevent_get_recursion_context(void);
1255 extern void perf_swevent_put_recursion_context(int rctx
);
1256 extern u64
perf_swevent_set_period(struct perf_event
*event
);
1257 extern void perf_event_enable(struct perf_event
*event
);
1258 extern void perf_event_disable(struct perf_event
*event
);
1259 extern void perf_event_disable_local(struct perf_event
*event
);
1260 extern void perf_event_disable_inatomic(struct perf_event
*event
);
1261 extern void perf_event_task_tick(void);
1262 #else /* !CONFIG_PERF_EVENTS: */
1263 static inline void *
1264 perf_aux_output_begin(struct perf_output_handle
*handle
,
1265 struct perf_event
*event
) { return NULL
; }
1267 perf_aux_output_end(struct perf_output_handle
*handle
, unsigned long size
,
1270 perf_aux_output_skip(struct perf_output_handle
*handle
,
1271 unsigned long size
) { return -EINVAL
; }
1272 static inline void *
1273 perf_get_aux(struct perf_output_handle
*handle
) { return NULL
; }
1275 perf_event_task_migrate(struct task_struct
*task
) { }
1277 perf_event_task_sched_in(struct task_struct
*prev
,
1278 struct task_struct
*task
) { }
1280 perf_event_task_sched_out(struct task_struct
*prev
,
1281 struct task_struct
*next
) { }
1282 static inline int perf_event_init_task(struct task_struct
*child
) { return 0; }
1283 static inline void perf_event_exit_task(struct task_struct
*child
) { }
1284 static inline void perf_event_free_task(struct task_struct
*task
) { }
1285 static inline void perf_event_delayed_put(struct task_struct
*task
) { }
1286 static inline struct file
*perf_event_get(unsigned int fd
) { return ERR_PTR(-EINVAL
); }
1287 static inline const struct perf_event_attr
*perf_event_attrs(struct perf_event
*event
)
1289 return ERR_PTR(-EINVAL
);
1291 static inline u64
perf_event_read_local(struct perf_event
*event
) { return -EINVAL
; }
1292 static inline void perf_event_print_debug(void) { }
1293 static inline int perf_event_task_disable(void) { return -EINVAL
; }
1294 static inline int perf_event_task_enable(void) { return -EINVAL
; }
1295 static inline int perf_event_refresh(struct perf_event
*event
, int refresh
)
1301 perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
) { }
1303 perf_sw_event_sched(u32 event_id
, u64 nr
, u64 addr
) { }
1305 perf_bp_event(struct perf_event
*event
, void *data
) { }
1307 static inline int perf_register_guest_info_callbacks
1308 (struct perf_guest_info_callbacks
*callbacks
) { return 0; }
1309 static inline int perf_unregister_guest_info_callbacks
1310 (struct perf_guest_info_callbacks
*callbacks
) { return 0; }
1312 static inline void perf_event_mmap(struct vm_area_struct
*vma
) { }
1313 static inline void perf_event_exec(void) { }
1314 static inline void perf_event_comm(struct task_struct
*tsk
, bool exec
) { }
1315 static inline void perf_event_fork(struct task_struct
*tsk
) { }
1316 static inline void perf_event_init(void) { }
1317 static inline int perf_swevent_get_recursion_context(void) { return -1; }
1318 static inline void perf_swevent_put_recursion_context(int rctx
) { }
1319 static inline u64
perf_swevent_set_period(struct perf_event
*event
) { return 0; }
1320 static inline void perf_event_enable(struct perf_event
*event
) { }
1321 static inline void perf_event_disable(struct perf_event
*event
) { }
1322 static inline int __perf_event_disable(void *info
) { return -1; }
1323 static inline void perf_event_task_tick(void) { }
1324 static inline int perf_event_release_kernel(struct perf_event
*event
) { return 0; }
1327 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1328 extern void perf_restore_debug_store(void);
1330 static inline void perf_restore_debug_store(void) { }
1333 static __always_inline
bool perf_raw_frag_last(const struct perf_raw_frag
*frag
)
1335 return frag
->pad
< sizeof(u64
);
1338 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1340 struct perf_pmu_events_attr
{
1341 struct device_attribute attr
;
1343 const char *event_str
;
1346 struct perf_pmu_events_ht_attr
{
1347 struct device_attribute attr
;
1349 const char *event_str_ht
;
1350 const char *event_str_noht
;
1353 ssize_t
perf_event_sysfs_show(struct device
*dev
, struct device_attribute
*attr
,
1356 #define PMU_EVENT_ATTR(_name, _var, _id, _show) \
1357 static struct perf_pmu_events_attr _var = { \
1358 .attr = __ATTR(_name, 0444, _show, NULL), \
1362 #define PMU_EVENT_ATTR_STRING(_name, _var, _str) \
1363 static struct perf_pmu_events_attr _var = { \
1364 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1366 .event_str = _str, \
1369 #define PMU_FORMAT_ATTR(_name, _format) \
1371 _name##_show(struct device *dev, \
1372 struct device_attribute *attr, \
1375 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
1376 return sprintf(page, _format "\n"); \
1379 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1381 /* Performance counter hotplug functions */
1382 #ifdef CONFIG_PERF_EVENTS
1383 int perf_event_init_cpu(unsigned int cpu
);
1384 int perf_event_exit_cpu(unsigned int cpu
);
1386 #define perf_event_init_cpu NULL
1387 #define perf_event_exit_cpu NULL
1390 #endif /* _LINUX_PERF_EVENT_H */