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.
485 * @nr_file_filters: number of file-based filters
487 * A child event will use parent's @list (and therefore @lock), so they are
488 * bundled together; see perf_event_addr_filters().
490 struct perf_addr_filters_head
{
491 struct list_head list
;
493 unsigned int nr_file_filters
;
497 * enum perf_event_active_state - the states of a event
499 enum perf_event_active_state
{
500 PERF_EVENT_STATE_DEAD
= -4,
501 PERF_EVENT_STATE_EXIT
= -3,
502 PERF_EVENT_STATE_ERROR
= -2,
503 PERF_EVENT_STATE_OFF
= -1,
504 PERF_EVENT_STATE_INACTIVE
= 0,
505 PERF_EVENT_STATE_ACTIVE
= 1,
509 struct perf_sample_data
;
511 typedef void (*perf_overflow_handler_t
)(struct perf_event
*,
512 struct perf_sample_data
*,
513 struct pt_regs
*regs
);
516 * Event capabilities. For event_caps and groups caps.
518 * PERF_EV_CAP_SOFTWARE: Is a software event.
519 * PERF_EV_CAP_READ_ACTIVE_PKG: A CPU event (or cgroup event) that can be read
520 * from any CPU in the package where it is active.
522 #define PERF_EV_CAP_SOFTWARE BIT(0)
523 #define PERF_EV_CAP_READ_ACTIVE_PKG BIT(1)
525 #define SWEVENT_HLIST_BITS 8
526 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
528 struct swevent_hlist
{
529 struct hlist_head heads
[SWEVENT_HLIST_SIZE
];
530 struct rcu_head rcu_head
;
533 #define PERF_ATTACH_CONTEXT 0x01
534 #define PERF_ATTACH_GROUP 0x02
535 #define PERF_ATTACH_TASK 0x04
536 #define PERF_ATTACH_TASK_DATA 0x08
541 struct pmu_event_list
{
543 struct list_head list
;
547 * struct perf_event - performance event kernel representation:
550 #ifdef CONFIG_PERF_EVENTS
552 * entry onto perf_event_context::event_list;
553 * modifications require ctx->lock
554 * RCU safe iterations.
556 struct list_head event_entry
;
559 * XXX: group_entry and sibling_list should be mutually exclusive;
560 * either you're a sibling on a group, or you're the group leader.
561 * Rework the code to always use the same list element.
563 * Locked for modification by both ctx->mutex and ctx->lock; holding
564 * either sufficies for read.
566 struct list_head group_entry
;
567 struct list_head sibling_list
;
570 * We need storage to track the entries in perf_pmu_migrate_context; we
571 * cannot use the event_entry because of RCU and we want to keep the
572 * group in tact which avoids us using the other two entries.
574 struct list_head migrate_entry
;
576 struct hlist_node hlist_entry
;
577 struct list_head active_entry
;
580 /* Not serialized. Only written during event initialization. */
582 /* The cumulative AND of all event_caps for events in this group. */
585 struct perf_event
*group_leader
;
589 enum perf_event_active_state state
;
590 unsigned int attach_state
;
592 atomic64_t child_count
;
595 * These are the total time in nanoseconds that the event
596 * has been enabled (i.e. eligible to run, and the task has
597 * been scheduled in, if this is a per-task event)
598 * and running (scheduled onto the CPU), respectively.
600 * They are computed from tstamp_enabled, tstamp_running and
601 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
603 u64 total_time_enabled
;
604 u64 total_time_running
;
607 * These are timestamps used for computing total_time_enabled
608 * and total_time_running when the event is in INACTIVE or
609 * ACTIVE state, measured in nanoseconds from an arbitrary point
611 * tstamp_enabled: the notional time when the event was enabled
612 * tstamp_running: the notional time when the event was scheduled on
613 * tstamp_stopped: in INACTIVE state, the notional time when the
614 * event was scheduled off.
621 * timestamp shadows the actual context timing but it can
622 * be safely used in NMI interrupt context. It reflects the
623 * context time as it was when the event was last scheduled in.
625 * ctx_time already accounts for ctx->timestamp. Therefore to
626 * compute ctx_time for a sample, simply add perf_clock().
630 struct perf_event_attr attr
;
634 struct hw_perf_event hw
;
636 struct perf_event_context
*ctx
;
637 atomic_long_t refcount
;
640 * These accumulate total time (in nanoseconds) that children
641 * events have been enabled and running, respectively.
643 atomic64_t child_total_time_enabled
;
644 atomic64_t child_total_time_running
;
647 * Protect attach/detach and child_list:
649 struct mutex child_mutex
;
650 struct list_head child_list
;
651 struct perf_event
*parent
;
656 struct list_head owner_entry
;
657 struct task_struct
*owner
;
660 struct mutex mmap_mutex
;
663 struct ring_buffer
*rb
;
664 struct list_head rb_entry
;
665 unsigned long rcu_batches
;
669 wait_queue_head_t waitq
;
670 struct fasync_struct
*fasync
;
672 /* delayed work for NMIs and such */
676 struct irq_work pending
;
678 atomic_t event_limit
;
680 /* address range filters */
681 struct perf_addr_filters_head addr_filters
;
682 /* vma address array for file-based filders */
683 unsigned long *addr_filters_offs
;
684 unsigned long addr_filters_gen
;
686 void (*destroy
)(struct perf_event
*);
687 struct rcu_head rcu_head
;
689 struct pid_namespace
*ns
;
693 perf_overflow_handler_t overflow_handler
;
694 void *overflow_handler_context
;
695 #ifdef CONFIG_BPF_SYSCALL
696 perf_overflow_handler_t orig_overflow_handler
;
697 struct bpf_prog
*prog
;
700 #ifdef CONFIG_EVENT_TRACING
701 struct trace_event_call
*tp_event
;
702 struct event_filter
*filter
;
703 #ifdef CONFIG_FUNCTION_TRACER
704 struct ftrace_ops ftrace_ops
;
708 #ifdef CONFIG_CGROUP_PERF
709 struct perf_cgroup
*cgrp
; /* cgroup event is attach to */
710 int cgrp_defer_enabled
;
713 struct list_head sb_list
;
714 #endif /* CONFIG_PERF_EVENTS */
718 * struct perf_event_context - event context structure
720 * Used as a container for task events and CPU events as well:
722 struct perf_event_context
{
725 * Protect the states of the events in the list,
726 * nr_active, and the list:
730 * Protect the list of events. Locking either mutex or lock
731 * is sufficient to ensure the list doesn't change; to change
732 * the list you need to lock both the mutex and the spinlock.
736 struct list_head active_ctx_list
;
737 struct list_head pinned_groups
;
738 struct list_head flexible_groups
;
739 struct list_head event_list
;
747 struct task_struct
*task
;
750 * Context clock, runs when context enabled.
756 * These fields let us detect when two contexts have both
757 * been cloned (inherited) from a common ancestor.
759 struct perf_event_context
*parent_ctx
;
763 #ifdef CONFIG_CGROUP_PERF
764 int nr_cgroups
; /* cgroup evts */
766 void *task_ctx_data
; /* pmu specific data */
767 struct rcu_head rcu_head
;
771 * Number of contexts where an event can trigger:
772 * task, softirq, hardirq, nmi.
774 #define PERF_NR_CONTEXTS 4
777 * struct perf_event_cpu_context - per cpu event context structure
779 struct perf_cpu_context
{
780 struct perf_event_context ctx
;
781 struct perf_event_context
*task_ctx
;
785 raw_spinlock_t hrtimer_lock
;
786 struct hrtimer hrtimer
;
787 ktime_t hrtimer_interval
;
788 unsigned int hrtimer_active
;
790 #ifdef CONFIG_CGROUP_PERF
791 struct perf_cgroup
*cgrp
;
792 struct list_head cgrp_cpuctx_entry
;
795 struct list_head sched_cb_entry
;
799 struct perf_output_handle
{
800 struct perf_event
*event
;
801 struct ring_buffer
*rb
;
802 unsigned long wakeup
;
811 struct bpf_perf_event_data_kern
{
812 struct pt_regs
*regs
;
813 struct perf_sample_data
*data
;
816 #ifdef CONFIG_CGROUP_PERF
819 * perf_cgroup_info keeps track of time_enabled for a cgroup.
820 * This is a per-cpu dynamically allocated data structure.
822 struct perf_cgroup_info
{
828 struct cgroup_subsys_state css
;
829 struct perf_cgroup_info __percpu
*info
;
833 * Must ensure cgroup is pinned (css_get) before calling
834 * this function. In other words, we cannot call this function
835 * if there is no cgroup event for the current CPU context.
837 static inline struct perf_cgroup
*
838 perf_cgroup_from_task(struct task_struct
*task
, struct perf_event_context
*ctx
)
840 return container_of(task_css_check(task
, perf_event_cgrp_id
,
841 ctx
? lockdep_is_held(&ctx
->lock
)
843 struct perf_cgroup
, css
);
845 #endif /* CONFIG_CGROUP_PERF */
847 #ifdef CONFIG_PERF_EVENTS
849 extern void *perf_aux_output_begin(struct perf_output_handle
*handle
,
850 struct perf_event
*event
);
851 extern void perf_aux_output_end(struct perf_output_handle
*handle
,
852 unsigned long size
, bool truncated
);
853 extern int perf_aux_output_skip(struct perf_output_handle
*handle
,
855 extern void *perf_get_aux(struct perf_output_handle
*handle
);
857 extern int perf_pmu_register(struct pmu
*pmu
, const char *name
, int type
);
858 extern void perf_pmu_unregister(struct pmu
*pmu
);
860 extern int perf_num_counters(void);
861 extern const char *perf_pmu_name(void);
862 extern void __perf_event_task_sched_in(struct task_struct
*prev
,
863 struct task_struct
*task
);
864 extern void __perf_event_task_sched_out(struct task_struct
*prev
,
865 struct task_struct
*next
);
866 extern int perf_event_init_task(struct task_struct
*child
);
867 extern void perf_event_exit_task(struct task_struct
*child
);
868 extern void perf_event_free_task(struct task_struct
*task
);
869 extern void perf_event_delayed_put(struct task_struct
*task
);
870 extern struct file
*perf_event_get(unsigned int fd
);
871 extern const struct perf_event_attr
*perf_event_attrs(struct perf_event
*event
);
872 extern void perf_event_print_debug(void);
873 extern void perf_pmu_disable(struct pmu
*pmu
);
874 extern void perf_pmu_enable(struct pmu
*pmu
);
875 extern void perf_sched_cb_dec(struct pmu
*pmu
);
876 extern void perf_sched_cb_inc(struct pmu
*pmu
);
877 extern int perf_event_task_disable(void);
878 extern int perf_event_task_enable(void);
879 extern int perf_event_refresh(struct perf_event
*event
, int refresh
);
880 extern void perf_event_update_userpage(struct perf_event
*event
);
881 extern int perf_event_release_kernel(struct perf_event
*event
);
882 extern struct perf_event
*
883 perf_event_create_kernel_counter(struct perf_event_attr
*attr
,
885 struct task_struct
*task
,
886 perf_overflow_handler_t callback
,
888 extern void perf_pmu_migrate_context(struct pmu
*pmu
,
889 int src_cpu
, int dst_cpu
);
890 extern u64
perf_event_read_local(struct perf_event
*event
);
891 extern u64
perf_event_read_value(struct perf_event
*event
,
892 u64
*enabled
, u64
*running
);
895 struct perf_sample_data
{
897 * Fields set by perf_sample_data_init(), group so as to
898 * minimize the cachelines touched.
901 struct perf_raw_record
*raw
;
902 struct perf_branch_stack
*br_stack
;
906 union perf_mem_data_src data_src
;
909 * The other fields, optionally {set,used} by
910 * perf_{prepare,output}_sample().
925 struct perf_callchain_entry
*callchain
;
928 * regs_user may point to task_pt_regs or to regs_user_copy, depending
931 struct perf_regs regs_user
;
932 struct pt_regs regs_user_copy
;
934 struct perf_regs regs_intr
;
936 } ____cacheline_aligned
;
938 /* default value for data source */
939 #define PERF_MEM_NA (PERF_MEM_S(OP, NA) |\
940 PERF_MEM_S(LVL, NA) |\
941 PERF_MEM_S(SNOOP, NA) |\
942 PERF_MEM_S(LOCK, NA) |\
945 static inline void perf_sample_data_init(struct perf_sample_data
*data
,
946 u64 addr
, u64 period
)
948 /* remaining struct members initialized in perf_prepare_sample() */
951 data
->br_stack
= NULL
;
952 data
->period
= period
;
954 data
->data_src
.val
= PERF_MEM_NA
;
958 extern void perf_output_sample(struct perf_output_handle
*handle
,
959 struct perf_event_header
*header
,
960 struct perf_sample_data
*data
,
961 struct perf_event
*event
);
962 extern void perf_prepare_sample(struct perf_event_header
*header
,
963 struct perf_sample_data
*data
,
964 struct perf_event
*event
,
965 struct pt_regs
*regs
);
967 extern int perf_event_overflow(struct perf_event
*event
,
968 struct perf_sample_data
*data
,
969 struct pt_regs
*regs
);
971 extern void perf_event_output_forward(struct perf_event
*event
,
972 struct perf_sample_data
*data
,
973 struct pt_regs
*regs
);
974 extern void perf_event_output_backward(struct perf_event
*event
,
975 struct perf_sample_data
*data
,
976 struct pt_regs
*regs
);
977 extern void perf_event_output(struct perf_event
*event
,
978 struct perf_sample_data
*data
,
979 struct pt_regs
*regs
);
982 is_default_overflow_handler(struct perf_event
*event
)
984 if (likely(event
->overflow_handler
== perf_event_output_forward
))
986 if (unlikely(event
->overflow_handler
== perf_event_output_backward
))
992 perf_event_header__init_id(struct perf_event_header
*header
,
993 struct perf_sample_data
*data
,
994 struct perf_event
*event
);
996 perf_event__output_id_sample(struct perf_event
*event
,
997 struct perf_output_handle
*handle
,
998 struct perf_sample_data
*sample
);
1001 perf_log_lost_samples(struct perf_event
*event
, u64 lost
);
1003 static inline bool is_sampling_event(struct perf_event
*event
)
1005 return event
->attr
.sample_period
!= 0;
1009 * Return 1 for a software event, 0 for a hardware event
1011 static inline int is_software_event(struct perf_event
*event
)
1013 return event
->event_caps
& PERF_EV_CAP_SOFTWARE
;
1016 extern struct static_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
1018 extern void ___perf_sw_event(u32
, u64
, struct pt_regs
*, u64
);
1019 extern void __perf_sw_event(u32
, u64
, struct pt_regs
*, u64
);
1021 #ifndef perf_arch_fetch_caller_regs
1022 static inline void perf_arch_fetch_caller_regs(struct pt_regs
*regs
, unsigned long ip
) { }
1026 * Take a snapshot of the regs. Skip ip and frame pointer to
1027 * the nth caller. We only need a few of the regs:
1028 * - ip for PERF_SAMPLE_IP
1029 * - cs for user_mode() tests
1030 * - bp for callchains
1031 * - eflags, for future purposes, just in case
1033 static inline void perf_fetch_caller_regs(struct pt_regs
*regs
)
1035 perf_arch_fetch_caller_regs(regs
, CALLER_ADDR0
);
1038 static __always_inline
void
1039 perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
1041 if (static_key_false(&perf_swevent_enabled
[event_id
]))
1042 __perf_sw_event(event_id
, nr
, regs
, addr
);
1045 DECLARE_PER_CPU(struct pt_regs
, __perf_regs
[4]);
1048 * 'Special' version for the scheduler, it hard assumes no recursion,
1049 * which is guaranteed by us not actually scheduling inside other swevents
1050 * because those disable preemption.
1052 static __always_inline
void
1053 perf_sw_event_sched(u32 event_id
, u64 nr
, u64 addr
)
1055 if (static_key_false(&perf_swevent_enabled
[event_id
])) {
1056 struct pt_regs
*regs
= this_cpu_ptr(&__perf_regs
[0]);
1058 perf_fetch_caller_regs(regs
);
1059 ___perf_sw_event(event_id
, nr
, regs
, addr
);
1063 extern struct static_key_false perf_sched_events
;
1065 static __always_inline
bool
1066 perf_sw_migrate_enabled(void)
1068 if (static_key_false(&perf_swevent_enabled
[PERF_COUNT_SW_CPU_MIGRATIONS
]))
1073 static inline void perf_event_task_migrate(struct task_struct
*task
)
1075 if (perf_sw_migrate_enabled())
1076 task
->sched_migrated
= 1;
1079 static inline void perf_event_task_sched_in(struct task_struct
*prev
,
1080 struct task_struct
*task
)
1082 if (static_branch_unlikely(&perf_sched_events
))
1083 __perf_event_task_sched_in(prev
, task
);
1085 if (perf_sw_migrate_enabled() && task
->sched_migrated
) {
1086 struct pt_regs
*regs
= this_cpu_ptr(&__perf_regs
[0]);
1088 perf_fetch_caller_regs(regs
);
1089 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, regs
, 0);
1090 task
->sched_migrated
= 0;
1094 static inline void perf_event_task_sched_out(struct task_struct
*prev
,
1095 struct task_struct
*next
)
1097 perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES
, 1, 0);
1099 if (static_branch_unlikely(&perf_sched_events
))
1100 __perf_event_task_sched_out(prev
, next
);
1103 static inline u64
__perf_event_count(struct perf_event
*event
)
1105 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
1108 extern void perf_event_mmap(struct vm_area_struct
*vma
);
1109 extern struct perf_guest_info_callbacks
*perf_guest_cbs
;
1110 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*callbacks
);
1111 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*callbacks
);
1113 extern void perf_event_exec(void);
1114 extern void perf_event_comm(struct task_struct
*tsk
, bool exec
);
1115 extern void perf_event_fork(struct task_struct
*tsk
);
1118 DECLARE_PER_CPU(struct perf_callchain_entry
, perf_callchain_entry
);
1120 extern void perf_callchain_user(struct perf_callchain_entry_ctx
*entry
, struct pt_regs
*regs
);
1121 extern void perf_callchain_kernel(struct perf_callchain_entry_ctx
*entry
, struct pt_regs
*regs
);
1122 extern struct perf_callchain_entry
*
1123 get_perf_callchain(struct pt_regs
*regs
, u32 init_nr
, bool kernel
, bool user
,
1124 u32 max_stack
, bool crosstask
, bool add_mark
);
1125 extern int get_callchain_buffers(int max_stack
);
1126 extern void put_callchain_buffers(void);
1128 extern int sysctl_perf_event_max_stack
;
1129 extern int sysctl_perf_event_max_contexts_per_stack
;
1131 static inline int perf_callchain_store_context(struct perf_callchain_entry_ctx
*ctx
, u64 ip
)
1133 if (ctx
->contexts
< sysctl_perf_event_max_contexts_per_stack
) {
1134 struct perf_callchain_entry
*entry
= ctx
->entry
;
1135 entry
->ip
[entry
->nr
++] = ip
;
1139 ctx
->contexts_maxed
= true;
1140 return -1; /* no more room, stop walking the stack */
1144 static inline int perf_callchain_store(struct perf_callchain_entry_ctx
*ctx
, u64 ip
)
1146 if (ctx
->nr
< ctx
->max_stack
&& !ctx
->contexts_maxed
) {
1147 struct perf_callchain_entry
*entry
= ctx
->entry
;
1148 entry
->ip
[entry
->nr
++] = ip
;
1152 return -1; /* no more room, stop walking the stack */
1156 extern int sysctl_perf_event_paranoid
;
1157 extern int sysctl_perf_event_mlock
;
1158 extern int sysctl_perf_event_sample_rate
;
1159 extern int sysctl_perf_cpu_time_max_percent
;
1161 extern void perf_sample_event_took(u64 sample_len_ns
);
1163 extern int perf_proc_update_handler(struct ctl_table
*table
, int write
,
1164 void __user
*buffer
, size_t *lenp
,
1166 extern int perf_cpu_time_max_percent_handler(struct ctl_table
*table
, int write
,
1167 void __user
*buffer
, size_t *lenp
,
1170 int perf_event_max_stack_handler(struct ctl_table
*table
, int write
,
1171 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
);
1173 static inline bool perf_paranoid_tracepoint_raw(void)
1175 return sysctl_perf_event_paranoid
> -1;
1178 static inline bool perf_paranoid_cpu(void)
1180 return sysctl_perf_event_paranoid
> 0;
1183 static inline bool perf_paranoid_kernel(void)
1185 return sysctl_perf_event_paranoid
> 1;
1188 extern void perf_event_init(void);
1189 extern void perf_tp_event(u16 event_type
, u64 count
, void *record
,
1190 int entry_size
, struct pt_regs
*regs
,
1191 struct hlist_head
*head
, int rctx
,
1192 struct task_struct
*task
);
1193 extern void perf_bp_event(struct perf_event
*event
, void *data
);
1195 #ifndef perf_misc_flags
1196 # define perf_misc_flags(regs) \
1197 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1198 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1201 static inline bool has_branch_stack(struct perf_event
*event
)
1203 return event
->attr
.sample_type
& PERF_SAMPLE_BRANCH_STACK
;
1206 static inline bool needs_branch_stack(struct perf_event
*event
)
1208 return event
->attr
.branch_sample_type
!= 0;
1211 static inline bool has_aux(struct perf_event
*event
)
1213 return event
->pmu
->setup_aux
;
1216 static inline bool is_write_backward(struct perf_event
*event
)
1218 return !!event
->attr
.write_backward
;
1221 static inline bool has_addr_filter(struct perf_event
*event
)
1223 return event
->pmu
->nr_addr_filters
;
1227 * An inherited event uses parent's filters
1229 static inline struct perf_addr_filters_head
*
1230 perf_event_addr_filters(struct perf_event
*event
)
1232 struct perf_addr_filters_head
*ifh
= &event
->addr_filters
;
1235 ifh
= &event
->parent
->addr_filters
;
1240 extern void perf_event_addr_filters_sync(struct perf_event
*event
);
1242 extern int perf_output_begin(struct perf_output_handle
*handle
,
1243 struct perf_event
*event
, unsigned int size
);
1244 extern int perf_output_begin_forward(struct perf_output_handle
*handle
,
1245 struct perf_event
*event
,
1247 extern int perf_output_begin_backward(struct perf_output_handle
*handle
,
1248 struct perf_event
*event
,
1251 extern void perf_output_end(struct perf_output_handle
*handle
);
1252 extern unsigned int perf_output_copy(struct perf_output_handle
*handle
,
1253 const void *buf
, unsigned int len
);
1254 extern unsigned int perf_output_skip(struct perf_output_handle
*handle
,
1256 extern int perf_swevent_get_recursion_context(void);
1257 extern void perf_swevent_put_recursion_context(int rctx
);
1258 extern u64
perf_swevent_set_period(struct perf_event
*event
);
1259 extern void perf_event_enable(struct perf_event
*event
);
1260 extern void perf_event_disable(struct perf_event
*event
);
1261 extern void perf_event_disable_local(struct perf_event
*event
);
1262 extern void perf_event_disable_inatomic(struct perf_event
*event
);
1263 extern void perf_event_task_tick(void);
1264 extern int perf_event_account_interrupt(struct perf_event
*event
);
1265 #else /* !CONFIG_PERF_EVENTS: */
1266 static inline void *
1267 perf_aux_output_begin(struct perf_output_handle
*handle
,
1268 struct perf_event
*event
) { return NULL
; }
1270 perf_aux_output_end(struct perf_output_handle
*handle
, unsigned long size
,
1273 perf_aux_output_skip(struct perf_output_handle
*handle
,
1274 unsigned long size
) { return -EINVAL
; }
1275 static inline void *
1276 perf_get_aux(struct perf_output_handle
*handle
) { return NULL
; }
1278 perf_event_task_migrate(struct task_struct
*task
) { }
1280 perf_event_task_sched_in(struct task_struct
*prev
,
1281 struct task_struct
*task
) { }
1283 perf_event_task_sched_out(struct task_struct
*prev
,
1284 struct task_struct
*next
) { }
1285 static inline int perf_event_init_task(struct task_struct
*child
) { return 0; }
1286 static inline void perf_event_exit_task(struct task_struct
*child
) { }
1287 static inline void perf_event_free_task(struct task_struct
*task
) { }
1288 static inline void perf_event_delayed_put(struct task_struct
*task
) { }
1289 static inline struct file
*perf_event_get(unsigned int fd
) { return ERR_PTR(-EINVAL
); }
1290 static inline const struct perf_event_attr
*perf_event_attrs(struct perf_event
*event
)
1292 return ERR_PTR(-EINVAL
);
1294 static inline u64
perf_event_read_local(struct perf_event
*event
) { return -EINVAL
; }
1295 static inline void perf_event_print_debug(void) { }
1296 static inline int perf_event_task_disable(void) { return -EINVAL
; }
1297 static inline int perf_event_task_enable(void) { return -EINVAL
; }
1298 static inline int perf_event_refresh(struct perf_event
*event
, int refresh
)
1304 perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
) { }
1306 perf_sw_event_sched(u32 event_id
, u64 nr
, u64 addr
) { }
1308 perf_bp_event(struct perf_event
*event
, void *data
) { }
1310 static inline int perf_register_guest_info_callbacks
1311 (struct perf_guest_info_callbacks
*callbacks
) { return 0; }
1312 static inline int perf_unregister_guest_info_callbacks
1313 (struct perf_guest_info_callbacks
*callbacks
) { return 0; }
1315 static inline void perf_event_mmap(struct vm_area_struct
*vma
) { }
1316 static inline void perf_event_exec(void) { }
1317 static inline void perf_event_comm(struct task_struct
*tsk
, bool exec
) { }
1318 static inline void perf_event_fork(struct task_struct
*tsk
) { }
1319 static inline void perf_event_init(void) { }
1320 static inline int perf_swevent_get_recursion_context(void) { return -1; }
1321 static inline void perf_swevent_put_recursion_context(int rctx
) { }
1322 static inline u64
perf_swevent_set_period(struct perf_event
*event
) { return 0; }
1323 static inline void perf_event_enable(struct perf_event
*event
) { }
1324 static inline void perf_event_disable(struct perf_event
*event
) { }
1325 static inline int __perf_event_disable(void *info
) { return -1; }
1326 static inline void perf_event_task_tick(void) { }
1327 static inline int perf_event_release_kernel(struct perf_event
*event
) { return 0; }
1330 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1331 extern void perf_restore_debug_store(void);
1333 static inline void perf_restore_debug_store(void) { }
1336 static __always_inline
bool perf_raw_frag_last(const struct perf_raw_frag
*frag
)
1338 return frag
->pad
< sizeof(u64
);
1341 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1343 struct perf_pmu_events_attr
{
1344 struct device_attribute attr
;
1346 const char *event_str
;
1349 struct perf_pmu_events_ht_attr
{
1350 struct device_attribute attr
;
1352 const char *event_str_ht
;
1353 const char *event_str_noht
;
1356 ssize_t
perf_event_sysfs_show(struct device
*dev
, struct device_attribute
*attr
,
1359 #define PMU_EVENT_ATTR(_name, _var, _id, _show) \
1360 static struct perf_pmu_events_attr _var = { \
1361 .attr = __ATTR(_name, 0444, _show, NULL), \
1365 #define PMU_EVENT_ATTR_STRING(_name, _var, _str) \
1366 static struct perf_pmu_events_attr _var = { \
1367 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1369 .event_str = _str, \
1372 #define PMU_FORMAT_ATTR(_name, _format) \
1374 _name##_show(struct device *dev, \
1375 struct device_attribute *attr, \
1378 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
1379 return sprintf(page, _format "\n"); \
1382 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1384 /* Performance counter hotplug functions */
1385 #ifdef CONFIG_PERF_EVENTS
1386 int perf_event_init_cpu(unsigned int cpu
);
1387 int perf_event_exit_cpu(unsigned int cpu
);
1389 #define perf_event_init_cpu NULL
1390 #define perf_event_exit_cpu NULL
1393 #endif /* _LINUX_PERF_EVENT_H */