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 { /* itrace */
145 struct { /* amd_power */
149 #ifdef CONFIG_HAVE_HW_BREAKPOINT
150 struct { /* breakpoint */
152 * Crufty hack to avoid the chicken and egg
153 * problem hw_breakpoint has with context
154 * creation and event initalization.
156 struct arch_hw_breakpoint info
;
157 struct list_head bp_list
;
160 struct { /* amd_iommu */
169 * If the event is a per task event, this will point to the task in
170 * question. See the comment in perf_event_alloc().
172 struct task_struct
*target
;
175 * PMU would store hardware filter configuration
180 /* Last sync'ed generation of filters */
181 unsigned long addr_filters_gen
;
184 * hw_perf_event::state flags; used to track the PERF_EF_* state.
186 #define PERF_HES_STOPPED 0x01 /* the counter is stopped */
187 #define PERF_HES_UPTODATE 0x02 /* event->count up-to-date */
188 #define PERF_HES_ARCH 0x04
193 * The last observed hardware counter value, updated with a
194 * local64_cmpxchg() such that pmu::read() can be called nested.
196 local64_t prev_count
;
199 * The period to start the next sample with.
204 * The period we started this sample with.
209 * However much is left of the current period; note that this is
210 * a full 64bit value and allows for generation of periods longer
211 * than hardware might allow.
213 local64_t period_left
;
216 * State for throttling the event, see __perf_event_overflow() and
217 * perf_adjust_freq_unthr_context().
223 * State for freq target events, see __perf_event_overflow() and
224 * perf_adjust_freq_unthr_context().
227 u64 freq_count_stamp
;
234 * Common implementation detail of pmu::{start,commit,cancel}_txn
236 #define PERF_PMU_TXN_ADD 0x1 /* txn to add/schedule event on PMU */
237 #define PERF_PMU_TXN_READ 0x2 /* txn to read event group from PMU */
240 * pmu::capabilities flags
242 #define PERF_PMU_CAP_NO_INTERRUPT 0x01
243 #define PERF_PMU_CAP_NO_NMI 0x02
244 #define PERF_PMU_CAP_AUX_NO_SG 0x04
245 #define PERF_PMU_CAP_AUX_SW_DOUBLEBUF 0x08
246 #define PERF_PMU_CAP_EXCLUSIVE 0x10
247 #define PERF_PMU_CAP_ITRACE 0x20
248 #define PERF_PMU_CAP_HETEROGENEOUS_CPUS 0x40
251 * struct pmu - generic performance monitoring unit
254 struct list_head entry
;
256 struct module
*module
;
258 const struct attribute_group
**attr_groups
;
263 * various common per-pmu feature flags
267 int * __percpu pmu_disable_count
;
268 struct perf_cpu_context
* __percpu pmu_cpu_context
;
269 atomic_t exclusive_cnt
; /* < 0: cpu; > 0: tsk */
271 int hrtimer_interval_ms
;
273 /* number of address filters this PMU can do */
274 unsigned int nr_addr_filters
;
277 * Fully disable/enable this PMU, can be used to protect from the PMI
278 * as well as for lazy/batch writing of the MSRs.
280 void (*pmu_enable
) (struct pmu
*pmu
); /* optional */
281 void (*pmu_disable
) (struct pmu
*pmu
); /* optional */
284 * Try and initialize the event for this PMU.
287 * -ENOENT -- @event is not for this PMU
289 * -ENODEV -- @event is for this PMU but PMU not present
290 * -EBUSY -- @event is for this PMU but PMU temporarily unavailable
291 * -EINVAL -- @event is for this PMU but @event is not valid
292 * -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
293 * -EACCESS -- @event is for this PMU, @event is valid, but no privilidges
295 * 0 -- @event is for this PMU and valid
297 * Other error return values are allowed.
299 int (*event_init
) (struct perf_event
*event
);
302 * Notification that the event was mapped or unmapped. Called
303 * in the context of the mapping task.
305 void (*event_mapped
) (struct perf_event
*event
, struct mm_struct
*mm
); /* optional */
306 void (*event_unmapped
) (struct perf_event
*event
, struct mm_struct
*mm
); /* optional */
309 * Flags for ->add()/->del()/ ->start()/->stop(). There are
310 * matching hw_perf_event::state flags.
312 #define PERF_EF_START 0x01 /* start the counter when adding */
313 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
314 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
317 * Adds/Removes a counter to/from the PMU, can be done inside a
318 * transaction, see the ->*_txn() methods.
320 * The add/del callbacks will reserve all hardware resources required
321 * to service the event, this includes any counter constraint
324 * Called with IRQs disabled and the PMU disabled on the CPU the event
327 * ->add() called without PERF_EF_START should result in the same state
328 * as ->add() followed by ->stop().
330 * ->del() must always PERF_EF_UPDATE stop an event. If it calls
331 * ->stop() that must deal with already being stopped without
334 int (*add
) (struct perf_event
*event
, int flags
);
335 void (*del
) (struct perf_event
*event
, int flags
);
338 * Starts/Stops a counter present on the PMU.
340 * The PMI handler should stop the counter when perf_event_overflow()
341 * returns !0. ->start() will be used to continue.
343 * Also used to change the sample period.
345 * Called with IRQs disabled and the PMU disabled on the CPU the event
346 * is on -- will be called from NMI context with the PMU generates
349 * ->stop() with PERF_EF_UPDATE will read the counter and update
350 * period/count values like ->read() would.
352 * ->start() with PERF_EF_RELOAD will reprogram the the counter
353 * value, must be preceded by a ->stop() with PERF_EF_UPDATE.
355 void (*start
) (struct perf_event
*event
, int flags
);
356 void (*stop
) (struct perf_event
*event
, int flags
);
359 * Updates the counter value of the event.
361 * For sampling capable PMUs this will also update the software period
362 * hw_perf_event::period_left field.
364 void (*read
) (struct perf_event
*event
);
367 * Group events scheduling is treated as a transaction, add
368 * group events as a whole and perform one schedulability test.
369 * If the test fails, roll back the whole group
371 * Start the transaction, after this ->add() doesn't need to
372 * do schedulability tests.
376 void (*start_txn
) (struct pmu
*pmu
, unsigned int txn_flags
);
378 * If ->start_txn() disabled the ->add() schedulability test
379 * then ->commit_txn() is required to perform one. On success
380 * the transaction is closed. On error the transaction is kept
381 * open until ->cancel_txn() is called.
385 int (*commit_txn
) (struct pmu
*pmu
);
387 * Will cancel the transaction, assumes ->del() is called
388 * for each successful ->add() during the transaction.
392 void (*cancel_txn
) (struct pmu
*pmu
);
395 * Will return the value for perf_event_mmap_page::index for this event,
396 * if no implementation is provided it will default to: event->hw.idx + 1.
398 int (*event_idx
) (struct perf_event
*event
); /*optional */
401 * context-switches callback
403 void (*sched_task
) (struct perf_event_context
*ctx
,
406 * PMU specific data size
408 size_t task_ctx_size
;
412 * Set up pmu-private data structures for an AUX area
414 void *(*setup_aux
) (int cpu
, void **pages
,
415 int nr_pages
, bool overwrite
);
419 * Free pmu-private AUX data structures
421 void (*free_aux
) (void *aux
); /* optional */
424 * Validate address range filters: make sure the HW supports the
425 * requested configuration and number of filters; return 0 if the
426 * supplied filters are valid, -errno otherwise.
428 * Runs in the context of the ioctl()ing process and is not serialized
429 * with the rest of the PMU callbacks.
431 int (*addr_filters_validate
) (struct list_head
*filters
);
435 * Synchronize address range filter configuration:
436 * translate hw-agnostic filters into hardware configuration in
437 * event::hw::addr_filters.
439 * Runs as a part of filter sync sequence that is done in ->start()
440 * callback by calling perf_event_addr_filters_sync().
442 * May (and should) traverse event::addr_filters::list, for which its
443 * caller provides necessary serialization.
445 void (*addr_filters_sync
) (struct perf_event
*event
);
449 * Filter events for PMU-specific reasons.
451 int (*filter_match
) (struct perf_event
*event
); /* optional */
455 * struct perf_addr_filter - address range filter definition
456 * @entry: event's filter list linkage
457 * @inode: object file's inode for file-based filters
458 * @offset: filter range offset
459 * @size: filter range size
460 * @range: 1: range, 0: address
461 * @filter: 1: filter/start, 0: stop
463 * This is a hardware-agnostic filter configuration as specified by the user.
465 struct perf_addr_filter
{
466 struct list_head entry
;
468 unsigned long offset
;
470 unsigned int range
: 1,
475 * struct perf_addr_filters_head - container for address range filters
476 * @list: list of filters for this event
477 * @lock: spinlock that serializes accesses to the @list and event's
478 * (and its children's) filter generations.
479 * @nr_file_filters: number of file-based filters
481 * A child event will use parent's @list (and therefore @lock), so they are
482 * bundled together; see perf_event_addr_filters().
484 struct perf_addr_filters_head
{
485 struct list_head list
;
487 unsigned int nr_file_filters
;
491 * enum perf_event_active_state - the states of a event
493 enum perf_event_active_state
{
494 PERF_EVENT_STATE_DEAD
= -4,
495 PERF_EVENT_STATE_EXIT
= -3,
496 PERF_EVENT_STATE_ERROR
= -2,
497 PERF_EVENT_STATE_OFF
= -1,
498 PERF_EVENT_STATE_INACTIVE
= 0,
499 PERF_EVENT_STATE_ACTIVE
= 1,
503 struct perf_sample_data
;
505 typedef void (*perf_overflow_handler_t
)(struct perf_event
*,
506 struct perf_sample_data
*,
507 struct pt_regs
*regs
);
510 * Event capabilities. For event_caps and groups caps.
512 * PERF_EV_CAP_SOFTWARE: Is a software event.
513 * PERF_EV_CAP_READ_ACTIVE_PKG: A CPU event (or cgroup event) that can be read
514 * from any CPU in the package where it is active.
516 #define PERF_EV_CAP_SOFTWARE BIT(0)
517 #define PERF_EV_CAP_READ_ACTIVE_PKG BIT(1)
519 #define SWEVENT_HLIST_BITS 8
520 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
522 struct swevent_hlist
{
523 struct hlist_head heads
[SWEVENT_HLIST_SIZE
];
524 struct rcu_head rcu_head
;
527 #define PERF_ATTACH_CONTEXT 0x01
528 #define PERF_ATTACH_GROUP 0x02
529 #define PERF_ATTACH_TASK 0x04
530 #define PERF_ATTACH_TASK_DATA 0x08
535 struct pmu_event_list
{
537 struct list_head list
;
541 * struct perf_event - performance event kernel representation:
544 #ifdef CONFIG_PERF_EVENTS
546 * entry onto perf_event_context::event_list;
547 * modifications require ctx->lock
548 * RCU safe iterations.
550 struct list_head event_entry
;
553 * XXX: group_entry and sibling_list should be mutually exclusive;
554 * either you're a sibling on a group, or you're the group leader.
555 * Rework the code to always use the same list element.
557 * Locked for modification by both ctx->mutex and ctx->lock; holding
558 * either sufficies for read.
560 struct list_head group_entry
;
561 struct list_head sibling_list
;
564 * We need storage to track the entries in perf_pmu_migrate_context; we
565 * cannot use the event_entry because of RCU and we want to keep the
566 * group in tact which avoids us using the other two entries.
568 struct list_head migrate_entry
;
570 struct hlist_node hlist_entry
;
571 struct list_head active_entry
;
574 /* Not serialized. Only written during event initialization. */
576 /* The cumulative AND of all event_caps for events in this group. */
579 struct perf_event
*group_leader
;
583 enum perf_event_active_state state
;
584 unsigned int attach_state
;
586 atomic64_t child_count
;
589 * These are the total time in nanoseconds that the event
590 * has been enabled (i.e. eligible to run, and the task has
591 * been scheduled in, if this is a per-task event)
592 * and running (scheduled onto the CPU), respectively.
594 * They are computed from tstamp_enabled, tstamp_running and
595 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
597 u64 total_time_enabled
;
598 u64 total_time_running
;
601 * These are timestamps used for computing total_time_enabled
602 * and total_time_running when the event is in INACTIVE or
603 * ACTIVE state, measured in nanoseconds from an arbitrary point
605 * tstamp_enabled: the notional time when the event was enabled
606 * tstamp_running: the notional time when the event was scheduled on
607 * tstamp_stopped: in INACTIVE state, the notional time when the
608 * event was scheduled off.
615 * timestamp shadows the actual context timing but it can
616 * be safely used in NMI interrupt context. It reflects the
617 * context time as it was when the event was last scheduled in.
619 * ctx_time already accounts for ctx->timestamp. Therefore to
620 * compute ctx_time for a sample, simply add perf_clock().
624 struct perf_event_attr attr
;
628 struct hw_perf_event hw
;
630 struct perf_event_context
*ctx
;
631 atomic_long_t refcount
;
634 * These accumulate total time (in nanoseconds) that children
635 * events have been enabled and running, respectively.
637 atomic64_t child_total_time_enabled
;
638 atomic64_t child_total_time_running
;
641 * Protect attach/detach and child_list:
643 struct mutex child_mutex
;
644 struct list_head child_list
;
645 struct perf_event
*parent
;
650 struct list_head owner_entry
;
651 struct task_struct
*owner
;
654 struct mutex mmap_mutex
;
657 struct ring_buffer
*rb
;
658 struct list_head rb_entry
;
659 unsigned long rcu_batches
;
663 wait_queue_head_t waitq
;
664 struct fasync_struct
*fasync
;
666 /* delayed work for NMIs and such */
670 struct irq_work pending
;
672 atomic_t event_limit
;
674 /* address range filters */
675 struct perf_addr_filters_head addr_filters
;
676 /* vma address array for file-based filders */
677 unsigned long *addr_filters_offs
;
678 unsigned long addr_filters_gen
;
680 void (*destroy
)(struct perf_event
*);
681 struct rcu_head rcu_head
;
683 struct pid_namespace
*ns
;
687 perf_overflow_handler_t overflow_handler
;
688 void *overflow_handler_context
;
689 #ifdef CONFIG_BPF_SYSCALL
690 perf_overflow_handler_t orig_overflow_handler
;
691 struct bpf_prog
*prog
;
694 #ifdef CONFIG_EVENT_TRACING
695 struct trace_event_call
*tp_event
;
696 struct event_filter
*filter
;
697 #ifdef CONFIG_FUNCTION_TRACER
698 struct ftrace_ops ftrace_ops
;
702 #ifdef CONFIG_CGROUP_PERF
703 struct perf_cgroup
*cgrp
; /* cgroup event is attach to */
704 int cgrp_defer_enabled
;
707 struct list_head sb_list
;
708 #endif /* CONFIG_PERF_EVENTS */
712 * struct perf_event_context - event context structure
714 * Used as a container for task events and CPU events as well:
716 struct perf_event_context
{
719 * Protect the states of the events in the list,
720 * nr_active, and the list:
724 * Protect the list of events. Locking either mutex or lock
725 * is sufficient to ensure the list doesn't change; to change
726 * the list you need to lock both the mutex and the spinlock.
730 struct list_head active_ctx_list
;
731 struct list_head pinned_groups
;
732 struct list_head flexible_groups
;
733 struct list_head event_list
;
741 struct task_struct
*task
;
744 * Context clock, runs when context enabled.
750 * These fields let us detect when two contexts have both
751 * been cloned (inherited) from a common ancestor.
753 struct perf_event_context
*parent_ctx
;
757 #ifdef CONFIG_CGROUP_PERF
758 int nr_cgroups
; /* cgroup evts */
760 void *task_ctx_data
; /* pmu specific data */
761 struct rcu_head rcu_head
;
765 * Number of contexts where an event can trigger:
766 * task, softirq, hardirq, nmi.
768 #define PERF_NR_CONTEXTS 4
771 * struct perf_event_cpu_context - per cpu event context structure
773 struct perf_cpu_context
{
774 struct perf_event_context ctx
;
775 struct perf_event_context
*task_ctx
;
779 raw_spinlock_t hrtimer_lock
;
780 struct hrtimer hrtimer
;
781 ktime_t hrtimer_interval
;
782 unsigned int hrtimer_active
;
784 #ifdef CONFIG_CGROUP_PERF
785 struct perf_cgroup
*cgrp
;
786 struct list_head cgrp_cpuctx_entry
;
789 struct list_head sched_cb_entry
;
795 struct perf_output_handle
{
796 struct perf_event
*event
;
797 struct ring_buffer
*rb
;
798 unsigned long wakeup
;
808 struct bpf_perf_event_data_kern
{
809 struct pt_regs
*regs
;
810 struct perf_sample_data
*data
;
813 #ifdef CONFIG_CGROUP_PERF
816 * perf_cgroup_info keeps track of time_enabled for a cgroup.
817 * This is a per-cpu dynamically allocated data structure.
819 struct perf_cgroup_info
{
825 struct cgroup_subsys_state css
;
826 struct perf_cgroup_info __percpu
*info
;
830 * Must ensure cgroup is pinned (css_get) before calling
831 * this function. In other words, we cannot call this function
832 * if there is no cgroup event for the current CPU context.
834 static inline struct perf_cgroup
*
835 perf_cgroup_from_task(struct task_struct
*task
, struct perf_event_context
*ctx
)
837 return container_of(task_css_check(task
, perf_event_cgrp_id
,
838 ctx
? lockdep_is_held(&ctx
->lock
)
840 struct perf_cgroup
, css
);
842 #endif /* CONFIG_CGROUP_PERF */
844 #ifdef CONFIG_PERF_EVENTS
846 extern void *perf_aux_output_begin(struct perf_output_handle
*handle
,
847 struct perf_event
*event
);
848 extern void perf_aux_output_end(struct perf_output_handle
*handle
,
850 extern int perf_aux_output_skip(struct perf_output_handle
*handle
,
852 extern void *perf_get_aux(struct perf_output_handle
*handle
);
853 extern void perf_aux_output_flag(struct perf_output_handle
*handle
, u64 flags
);
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 int perf_event_read_local(struct perf_event
*event
, u64
*value
);
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 extern void perf_event_mmap(struct vm_area_struct
*vma
);
1102 extern struct perf_guest_info_callbacks
*perf_guest_cbs
;
1103 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*callbacks
);
1104 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*callbacks
);
1106 extern void perf_event_exec(void);
1107 extern void perf_event_comm(struct task_struct
*tsk
, bool exec
);
1108 extern void perf_event_namespaces(struct task_struct
*tsk
);
1109 extern void perf_event_fork(struct task_struct
*tsk
);
1112 DECLARE_PER_CPU(struct perf_callchain_entry
, perf_callchain_entry
);
1114 extern void perf_callchain_user(struct perf_callchain_entry_ctx
*entry
, struct pt_regs
*regs
);
1115 extern void perf_callchain_kernel(struct perf_callchain_entry_ctx
*entry
, struct pt_regs
*regs
);
1116 extern struct perf_callchain_entry
*
1117 get_perf_callchain(struct pt_regs
*regs
, u32 init_nr
, bool kernel
, bool user
,
1118 u32 max_stack
, bool crosstask
, bool add_mark
);
1119 extern int get_callchain_buffers(int max_stack
);
1120 extern void put_callchain_buffers(void);
1122 extern int sysctl_perf_event_max_stack
;
1123 extern int sysctl_perf_event_max_contexts_per_stack
;
1125 static inline int perf_callchain_store_context(struct perf_callchain_entry_ctx
*ctx
, u64 ip
)
1127 if (ctx
->contexts
< sysctl_perf_event_max_contexts_per_stack
) {
1128 struct perf_callchain_entry
*entry
= ctx
->entry
;
1129 entry
->ip
[entry
->nr
++] = ip
;
1133 ctx
->contexts_maxed
= true;
1134 return -1; /* no more room, stop walking the stack */
1138 static inline int perf_callchain_store(struct perf_callchain_entry_ctx
*ctx
, u64 ip
)
1140 if (ctx
->nr
< ctx
->max_stack
&& !ctx
->contexts_maxed
) {
1141 struct perf_callchain_entry
*entry
= ctx
->entry
;
1142 entry
->ip
[entry
->nr
++] = ip
;
1146 return -1; /* no more room, stop walking the stack */
1150 extern int sysctl_perf_event_paranoid
;
1151 extern int sysctl_perf_event_mlock
;
1152 extern int sysctl_perf_event_sample_rate
;
1153 extern int sysctl_perf_cpu_time_max_percent
;
1155 extern void perf_sample_event_took(u64 sample_len_ns
);
1157 extern int perf_proc_update_handler(struct ctl_table
*table
, int write
,
1158 void __user
*buffer
, size_t *lenp
,
1160 extern int perf_cpu_time_max_percent_handler(struct ctl_table
*table
, int write
,
1161 void __user
*buffer
, size_t *lenp
,
1164 int perf_event_max_stack_handler(struct ctl_table
*table
, int write
,
1165 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
);
1167 static inline bool perf_paranoid_any(void)
1169 return sysctl_perf_event_paranoid
> 2;
1172 static inline bool perf_paranoid_tracepoint_raw(void)
1174 return sysctl_perf_event_paranoid
> -1;
1177 static inline bool perf_paranoid_cpu(void)
1179 return sysctl_perf_event_paranoid
> 0;
1182 static inline bool perf_paranoid_kernel(void)
1184 return sysctl_perf_event_paranoid
> 1;
1187 extern void perf_event_init(void);
1188 extern void perf_tp_event(u16 event_type
, u64 count
, void *record
,
1189 int entry_size
, struct pt_regs
*regs
,
1190 struct hlist_head
*head
, int rctx
,
1191 struct task_struct
*task
, struct perf_event
*event
);
1192 extern void perf_bp_event(struct perf_event
*event
, void *data
);
1194 #ifndef perf_misc_flags
1195 # define perf_misc_flags(regs) \
1196 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1197 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1200 static inline bool has_branch_stack(struct perf_event
*event
)
1202 return event
->attr
.sample_type
& PERF_SAMPLE_BRANCH_STACK
;
1205 static inline bool needs_branch_stack(struct perf_event
*event
)
1207 return event
->attr
.branch_sample_type
!= 0;
1210 static inline bool has_aux(struct perf_event
*event
)
1212 return event
->pmu
->setup_aux
;
1215 static inline bool is_write_backward(struct perf_event
*event
)
1217 return !!event
->attr
.write_backward
;
1220 static inline bool has_addr_filter(struct perf_event
*event
)
1222 return event
->pmu
->nr_addr_filters
;
1226 * An inherited event uses parent's filters
1228 static inline struct perf_addr_filters_head
*
1229 perf_event_addr_filters(struct perf_event
*event
)
1231 struct perf_addr_filters_head
*ifh
= &event
->addr_filters
;
1234 ifh
= &event
->parent
->addr_filters
;
1239 extern void perf_event_addr_filters_sync(struct perf_event
*event
);
1241 extern int perf_output_begin(struct perf_output_handle
*handle
,
1242 struct perf_event
*event
, unsigned int size
);
1243 extern int perf_output_begin_forward(struct perf_output_handle
*handle
,
1244 struct perf_event
*event
,
1246 extern int perf_output_begin_backward(struct perf_output_handle
*handle
,
1247 struct perf_event
*event
,
1250 extern void perf_output_end(struct perf_output_handle
*handle
);
1251 extern unsigned int perf_output_copy(struct perf_output_handle
*handle
,
1252 const void *buf
, unsigned int len
);
1253 extern unsigned int perf_output_skip(struct perf_output_handle
*handle
,
1255 extern int perf_swevent_get_recursion_context(void);
1256 extern void perf_swevent_put_recursion_context(int rctx
);
1257 extern u64
perf_swevent_set_period(struct perf_event
*event
);
1258 extern void perf_event_enable(struct perf_event
*event
);
1259 extern void perf_event_disable(struct perf_event
*event
);
1260 extern void perf_event_disable_local(struct perf_event
*event
);
1261 extern void perf_event_disable_inatomic(struct perf_event
*event
);
1262 extern void perf_event_task_tick(void);
1263 extern int perf_event_account_interrupt(struct perf_event
*event
);
1264 #else /* !CONFIG_PERF_EVENTS: */
1265 static inline void *
1266 perf_aux_output_begin(struct perf_output_handle
*handle
,
1267 struct perf_event
*event
) { return NULL
; }
1269 perf_aux_output_end(struct perf_output_handle
*handle
, unsigned long size
)
1272 perf_aux_output_skip(struct perf_output_handle
*handle
,
1273 unsigned long size
) { return -EINVAL
; }
1274 static inline void *
1275 perf_get_aux(struct perf_output_handle
*handle
) { return NULL
; }
1277 perf_event_task_migrate(struct task_struct
*task
) { }
1279 perf_event_task_sched_in(struct task_struct
*prev
,
1280 struct task_struct
*task
) { }
1282 perf_event_task_sched_out(struct task_struct
*prev
,
1283 struct task_struct
*next
) { }
1284 static inline int perf_event_init_task(struct task_struct
*child
) { return 0; }
1285 static inline void perf_event_exit_task(struct task_struct
*child
) { }
1286 static inline void perf_event_free_task(struct task_struct
*task
) { }
1287 static inline void perf_event_delayed_put(struct task_struct
*task
) { }
1288 static inline struct file
*perf_event_get(unsigned int fd
) { return ERR_PTR(-EINVAL
); }
1289 static inline const struct perf_event_attr
*perf_event_attrs(struct perf_event
*event
)
1291 return ERR_PTR(-EINVAL
);
1293 static inline int perf_event_read_local(struct perf_event
*event
, u64
*value
)
1297 static inline void perf_event_print_debug(void) { }
1298 static inline int perf_event_task_disable(void) { return -EINVAL
; }
1299 static inline int perf_event_task_enable(void) { return -EINVAL
; }
1300 static inline int perf_event_refresh(struct perf_event
*event
, int refresh
)
1306 perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
) { }
1308 perf_sw_event_sched(u32 event_id
, u64 nr
, u64 addr
) { }
1310 perf_bp_event(struct perf_event
*event
, void *data
) { }
1312 static inline int perf_register_guest_info_callbacks
1313 (struct perf_guest_info_callbacks
*callbacks
) { return 0; }
1314 static inline int perf_unregister_guest_info_callbacks
1315 (struct perf_guest_info_callbacks
*callbacks
) { return 0; }
1317 static inline void perf_event_mmap(struct vm_area_struct
*vma
) { }
1318 static inline void perf_event_exec(void) { }
1319 static inline void perf_event_comm(struct task_struct
*tsk
, bool exec
) { }
1320 static inline void perf_event_namespaces(struct task_struct
*tsk
) { }
1321 static inline void perf_event_fork(struct task_struct
*tsk
) { }
1322 static inline void perf_event_init(void) { }
1323 static inline int perf_swevent_get_recursion_context(void) { return -1; }
1324 static inline void perf_swevent_put_recursion_context(int rctx
) { }
1325 static inline u64
perf_swevent_set_period(struct perf_event
*event
) { return 0; }
1326 static inline void perf_event_enable(struct perf_event
*event
) { }
1327 static inline void perf_event_disable(struct perf_event
*event
) { }
1328 static inline int __perf_event_disable(void *info
) { return -1; }
1329 static inline void perf_event_task_tick(void) { }
1330 static inline int perf_event_release_kernel(struct perf_event
*event
) { return 0; }
1333 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1334 extern void perf_restore_debug_store(void);
1336 static inline void perf_restore_debug_store(void) { }
1339 static __always_inline
bool perf_raw_frag_last(const struct perf_raw_frag
*frag
)
1341 return frag
->pad
< sizeof(u64
);
1344 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1346 struct perf_pmu_events_attr
{
1347 struct device_attribute attr
;
1349 const char *event_str
;
1352 struct perf_pmu_events_ht_attr
{
1353 struct device_attribute attr
;
1355 const char *event_str_ht
;
1356 const char *event_str_noht
;
1359 ssize_t
perf_event_sysfs_show(struct device
*dev
, struct device_attribute
*attr
,
1362 #define PMU_EVENT_ATTR(_name, _var, _id, _show) \
1363 static struct perf_pmu_events_attr _var = { \
1364 .attr = __ATTR(_name, 0444, _show, NULL), \
1368 #define PMU_EVENT_ATTR_STRING(_name, _var, _str) \
1369 static struct perf_pmu_events_attr _var = { \
1370 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1372 .event_str = _str, \
1375 #define PMU_FORMAT_ATTR(_name, _format) \
1377 _name##_show(struct device *dev, \
1378 struct device_attribute *attr, \
1381 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
1382 return sprintf(page, _format "\n"); \
1385 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1387 /* Performance counter hotplug functions */
1388 #ifdef CONFIG_PERF_EVENTS
1389 int perf_event_init_cpu(unsigned int cpu
);
1390 int perf_event_exit_cpu(unsigned int cpu
);
1392 #define perf_event_init_cpu NULL
1393 #define perf_event_exit_cpu NULL
1396 #endif /* _LINUX_PERF_EVENT_H */