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 { /* amd_power */
154 #ifdef CONFIG_HAVE_HW_BREAKPOINT
155 struct { /* breakpoint */
157 * Crufty hack to avoid the chicken and egg
158 * problem hw_breakpoint has with context
159 * creation and event initalization.
161 struct arch_hw_breakpoint info
;
162 struct list_head bp_list
;
165 struct { /* amd_iommu */
174 * If the event is a per task event, this will point to the task in
175 * question. See the comment in perf_event_alloc().
177 struct task_struct
*target
;
180 * PMU would store hardware filter configuration
185 /* Last sync'ed generation of filters */
186 unsigned long addr_filters_gen
;
189 * hw_perf_event::state flags; used to track the PERF_EF_* state.
191 #define PERF_HES_STOPPED 0x01 /* the counter is stopped */
192 #define PERF_HES_UPTODATE 0x02 /* event->count up-to-date */
193 #define PERF_HES_ARCH 0x04
198 * The last observed hardware counter value, updated with a
199 * local64_cmpxchg() such that pmu::read() can be called nested.
201 local64_t prev_count
;
204 * The period to start the next sample with.
209 * The period we started this sample with.
214 * However much is left of the current period; note that this is
215 * a full 64bit value and allows for generation of periods longer
216 * than hardware might allow.
218 local64_t period_left
;
221 * State for throttling the event, see __perf_event_overflow() and
222 * perf_adjust_freq_unthr_context().
228 * State for freq target events, see __perf_event_overflow() and
229 * perf_adjust_freq_unthr_context().
232 u64 freq_count_stamp
;
239 * Common implementation detail of pmu::{start,commit,cancel}_txn
241 #define PERF_PMU_TXN_ADD 0x1 /* txn to add/schedule event on PMU */
242 #define PERF_PMU_TXN_READ 0x2 /* txn to read event group from PMU */
245 * pmu::capabilities flags
247 #define PERF_PMU_CAP_NO_INTERRUPT 0x01
248 #define PERF_PMU_CAP_NO_NMI 0x02
249 #define PERF_PMU_CAP_AUX_NO_SG 0x04
250 #define PERF_PMU_CAP_AUX_SW_DOUBLEBUF 0x08
251 #define PERF_PMU_CAP_EXCLUSIVE 0x10
252 #define PERF_PMU_CAP_ITRACE 0x20
253 #define PERF_PMU_CAP_HETEROGENEOUS_CPUS 0x40
256 * struct pmu - generic performance monitoring unit
259 struct list_head entry
;
261 struct module
*module
;
263 const struct attribute_group
**attr_groups
;
268 * various common per-pmu feature flags
272 int * __percpu pmu_disable_count
;
273 struct perf_cpu_context
* __percpu pmu_cpu_context
;
274 atomic_t exclusive_cnt
; /* < 0: cpu; > 0: tsk */
276 int hrtimer_interval_ms
;
278 /* number of address filters this PMU can do */
279 unsigned int nr_addr_filters
;
282 * Fully disable/enable this PMU, can be used to protect from the PMI
283 * as well as for lazy/batch writing of the MSRs.
285 void (*pmu_enable
) (struct pmu
*pmu
); /* optional */
286 void (*pmu_disable
) (struct pmu
*pmu
); /* optional */
289 * Try and initialize the event for this PMU.
292 * -ENOENT -- @event is not for this PMU
294 * -ENODEV -- @event is for this PMU but PMU not present
295 * -EBUSY -- @event is for this PMU but PMU temporarily unavailable
296 * -EINVAL -- @event is for this PMU but @event is not valid
297 * -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
298 * -EACCESS -- @event is for this PMU, @event is valid, but no privilidges
300 * 0 -- @event is for this PMU and valid
302 * Other error return values are allowed.
304 int (*event_init
) (struct perf_event
*event
);
307 * Notification that the event was mapped or unmapped. Called
308 * in the context of the mapping task.
310 void (*event_mapped
) (struct perf_event
*event
, struct mm_struct
*mm
); /* optional */
311 void (*event_unmapped
) (struct perf_event
*event
, struct mm_struct
*mm
); /* optional */
314 * Flags for ->add()/->del()/ ->start()/->stop(). There are
315 * matching hw_perf_event::state flags.
317 #define PERF_EF_START 0x01 /* start the counter when adding */
318 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
319 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
322 * Adds/Removes a counter to/from the PMU, can be done inside a
323 * transaction, see the ->*_txn() methods.
325 * The add/del callbacks will reserve all hardware resources required
326 * to service the event, this includes any counter constraint
329 * Called with IRQs disabled and the PMU disabled on the CPU the event
332 * ->add() called without PERF_EF_START should result in the same state
333 * as ->add() followed by ->stop().
335 * ->del() must always PERF_EF_UPDATE stop an event. If it calls
336 * ->stop() that must deal with already being stopped without
339 int (*add
) (struct perf_event
*event
, int flags
);
340 void (*del
) (struct perf_event
*event
, int flags
);
343 * Starts/Stops a counter present on the PMU.
345 * The PMI handler should stop the counter when perf_event_overflow()
346 * returns !0. ->start() will be used to continue.
348 * Also used to change the sample period.
350 * Called with IRQs disabled and the PMU disabled on the CPU the event
351 * is on -- will be called from NMI context with the PMU generates
354 * ->stop() with PERF_EF_UPDATE will read the counter and update
355 * period/count values like ->read() would.
357 * ->start() with PERF_EF_RELOAD will reprogram the the counter
358 * value, must be preceded by a ->stop() with PERF_EF_UPDATE.
360 void (*start
) (struct perf_event
*event
, int flags
);
361 void (*stop
) (struct perf_event
*event
, int flags
);
364 * Updates the counter value of the event.
366 * For sampling capable PMUs this will also update the software period
367 * hw_perf_event::period_left field.
369 void (*read
) (struct perf_event
*event
);
372 * Group events scheduling is treated as a transaction, add
373 * group events as a whole and perform one schedulability test.
374 * If the test fails, roll back the whole group
376 * Start the transaction, after this ->add() doesn't need to
377 * do schedulability tests.
381 void (*start_txn
) (struct pmu
*pmu
, unsigned int txn_flags
);
383 * If ->start_txn() disabled the ->add() schedulability test
384 * then ->commit_txn() is required to perform one. On success
385 * the transaction is closed. On error the transaction is kept
386 * open until ->cancel_txn() is called.
390 int (*commit_txn
) (struct pmu
*pmu
);
392 * Will cancel the transaction, assumes ->del() is called
393 * for each successful ->add() during the transaction.
397 void (*cancel_txn
) (struct pmu
*pmu
);
400 * Will return the value for perf_event_mmap_page::index for this event,
401 * if no implementation is provided it will default to: event->hw.idx + 1.
403 int (*event_idx
) (struct perf_event
*event
); /*optional */
406 * context-switches callback
408 void (*sched_task
) (struct perf_event_context
*ctx
,
411 * PMU specific data size
413 size_t task_ctx_size
;
417 * Return the count value for a counter.
419 u64 (*count
) (struct perf_event
*event
); /*optional*/
422 * Set up pmu-private data structures for an AUX area
424 void *(*setup_aux
) (int cpu
, void **pages
,
425 int nr_pages
, bool overwrite
);
429 * Free pmu-private AUX data structures
431 void (*free_aux
) (void *aux
); /* optional */
434 * Validate address range filters: make sure the HW supports the
435 * requested configuration and number of filters; return 0 if the
436 * supplied filters are valid, -errno otherwise.
438 * Runs in the context of the ioctl()ing process and is not serialized
439 * with the rest of the PMU callbacks.
441 int (*addr_filters_validate
) (struct list_head
*filters
);
445 * Synchronize address range filter configuration:
446 * translate hw-agnostic filters into hardware configuration in
447 * event::hw::addr_filters.
449 * Runs as a part of filter sync sequence that is done in ->start()
450 * callback by calling perf_event_addr_filters_sync().
452 * May (and should) traverse event::addr_filters::list, for which its
453 * caller provides necessary serialization.
455 void (*addr_filters_sync
) (struct perf_event
*event
);
459 * Filter events for PMU-specific reasons.
461 int (*filter_match
) (struct perf_event
*event
); /* optional */
465 * struct perf_addr_filter - address range filter definition
466 * @entry: event's filter list linkage
467 * @inode: object file's inode for file-based filters
468 * @offset: filter range offset
469 * @size: filter range size
470 * @range: 1: range, 0: address
471 * @filter: 1: filter/start, 0: stop
473 * This is a hardware-agnostic filter configuration as specified by the user.
475 struct perf_addr_filter
{
476 struct list_head entry
;
478 unsigned long offset
;
480 unsigned int range
: 1,
485 * struct perf_addr_filters_head - container for address range filters
486 * @list: list of filters for this event
487 * @lock: spinlock that serializes accesses to the @list and event's
488 * (and its children's) filter generations.
489 * @nr_file_filters: number of file-based filters
491 * A child event will use parent's @list (and therefore @lock), so they are
492 * bundled together; see perf_event_addr_filters().
494 struct perf_addr_filters_head
{
495 struct list_head list
;
497 unsigned int nr_file_filters
;
501 * enum perf_event_active_state - the states of a event
503 enum perf_event_active_state
{
504 PERF_EVENT_STATE_DEAD
= -4,
505 PERF_EVENT_STATE_EXIT
= -3,
506 PERF_EVENT_STATE_ERROR
= -2,
507 PERF_EVENT_STATE_OFF
= -1,
508 PERF_EVENT_STATE_INACTIVE
= 0,
509 PERF_EVENT_STATE_ACTIVE
= 1,
513 struct perf_sample_data
;
515 typedef void (*perf_overflow_handler_t
)(struct perf_event
*,
516 struct perf_sample_data
*,
517 struct pt_regs
*regs
);
520 * Event capabilities. For event_caps and groups caps.
522 * PERF_EV_CAP_SOFTWARE: Is a software event.
523 * PERF_EV_CAP_READ_ACTIVE_PKG: A CPU event (or cgroup event) that can be read
524 * from any CPU in the package where it is active.
526 #define PERF_EV_CAP_SOFTWARE BIT(0)
527 #define PERF_EV_CAP_READ_ACTIVE_PKG BIT(1)
529 #define SWEVENT_HLIST_BITS 8
530 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
532 struct swevent_hlist
{
533 struct hlist_head heads
[SWEVENT_HLIST_SIZE
];
534 struct rcu_head rcu_head
;
537 #define PERF_ATTACH_CONTEXT 0x01
538 #define PERF_ATTACH_GROUP 0x02
539 #define PERF_ATTACH_TASK 0x04
540 #define PERF_ATTACH_TASK_DATA 0x08
541 #define PERF_ATTACH_ITRACE 0x10
546 struct pmu_event_list
{
548 struct list_head list
;
552 * struct perf_event - performance event kernel representation:
555 #ifdef CONFIG_PERF_EVENTS
557 * entry onto perf_event_context::event_list;
558 * modifications require ctx->lock
559 * RCU safe iterations.
561 struct list_head event_entry
;
564 * XXX: group_entry and sibling_list should be mutually exclusive;
565 * either you're a sibling on a group, or you're the group leader.
566 * Rework the code to always use the same list element.
568 * Locked for modification by both ctx->mutex and ctx->lock; holding
569 * either sufficies for read.
571 struct list_head group_entry
;
572 struct list_head sibling_list
;
575 * We need storage to track the entries in perf_pmu_migrate_context; we
576 * cannot use the event_entry because of RCU and we want to keep the
577 * group in tact which avoids us using the other two entries.
579 struct list_head migrate_entry
;
581 struct hlist_node hlist_entry
;
582 struct list_head active_entry
;
585 /* Not serialized. Only written during event initialization. */
587 /* The cumulative AND of all event_caps for events in this group. */
590 struct perf_event
*group_leader
;
594 enum perf_event_active_state state
;
595 unsigned int attach_state
;
597 atomic64_t child_count
;
600 * These are the total time in nanoseconds that the event
601 * has been enabled (i.e. eligible to run, and the task has
602 * been scheduled in, if this is a per-task event)
603 * and running (scheduled onto the CPU), respectively.
605 * They are computed from tstamp_enabled, tstamp_running and
606 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
608 u64 total_time_enabled
;
609 u64 total_time_running
;
612 * These are timestamps used for computing total_time_enabled
613 * and total_time_running when the event is in INACTIVE or
614 * ACTIVE state, measured in nanoseconds from an arbitrary point
616 * tstamp_enabled: the notional time when the event was enabled
617 * tstamp_running: the notional time when the event was scheduled on
618 * tstamp_stopped: in INACTIVE state, the notional time when the
619 * event was scheduled off.
626 * timestamp shadows the actual context timing but it can
627 * be safely used in NMI interrupt context. It reflects the
628 * context time as it was when the event was last scheduled in.
630 * ctx_time already accounts for ctx->timestamp. Therefore to
631 * compute ctx_time for a sample, simply add perf_clock().
635 struct perf_event_attr attr
;
639 struct hw_perf_event hw
;
641 struct perf_event_context
*ctx
;
642 atomic_long_t refcount
;
645 * These accumulate total time (in nanoseconds) that children
646 * events have been enabled and running, respectively.
648 atomic64_t child_total_time_enabled
;
649 atomic64_t child_total_time_running
;
652 * Protect attach/detach and child_list:
654 struct mutex child_mutex
;
655 struct list_head child_list
;
656 struct perf_event
*parent
;
661 struct list_head owner_entry
;
662 struct task_struct
*owner
;
665 struct mutex mmap_mutex
;
668 struct ring_buffer
*rb
;
669 struct list_head rb_entry
;
670 unsigned long rcu_batches
;
674 wait_queue_head_t waitq
;
675 struct fasync_struct
*fasync
;
677 /* delayed work for NMIs and such */
681 struct irq_work pending
;
683 atomic_t event_limit
;
685 /* address range filters */
686 struct perf_addr_filters_head addr_filters
;
687 /* vma address array for file-based filders */
688 unsigned long *addr_filters_offs
;
689 unsigned long addr_filters_gen
;
691 void (*destroy
)(struct perf_event
*);
692 struct rcu_head rcu_head
;
694 struct pid_namespace
*ns
;
698 perf_overflow_handler_t overflow_handler
;
699 void *overflow_handler_context
;
700 #ifdef CONFIG_BPF_SYSCALL
701 perf_overflow_handler_t orig_overflow_handler
;
702 struct bpf_prog
*prog
;
705 #ifdef CONFIG_EVENT_TRACING
706 struct trace_event_call
*tp_event
;
707 struct event_filter
*filter
;
708 #ifdef CONFIG_FUNCTION_TRACER
709 struct ftrace_ops ftrace_ops
;
713 #ifdef CONFIG_CGROUP_PERF
714 struct perf_cgroup
*cgrp
; /* cgroup event is attach to */
715 int cgrp_defer_enabled
;
718 struct list_head sb_list
;
719 #endif /* CONFIG_PERF_EVENTS */
723 * struct perf_event_context - event context structure
725 * Used as a container for task events and CPU events as well:
727 struct perf_event_context
{
730 * Protect the states of the events in the list,
731 * nr_active, and the list:
735 * Protect the list of events. Locking either mutex or lock
736 * is sufficient to ensure the list doesn't change; to change
737 * the list you need to lock both the mutex and the spinlock.
741 struct list_head active_ctx_list
;
742 struct list_head pinned_groups
;
743 struct list_head flexible_groups
;
744 struct list_head event_list
;
752 struct task_struct
*task
;
755 * Context clock, runs when context enabled.
761 * These fields let us detect when two contexts have both
762 * been cloned (inherited) from a common ancestor.
764 struct perf_event_context
*parent_ctx
;
768 #ifdef CONFIG_CGROUP_PERF
769 int nr_cgroups
; /* cgroup evts */
771 void *task_ctx_data
; /* pmu specific data */
772 struct rcu_head rcu_head
;
776 * Number of contexts where an event can trigger:
777 * task, softirq, hardirq, nmi.
779 #define PERF_NR_CONTEXTS 4
782 * struct perf_event_cpu_context - per cpu event context structure
784 struct perf_cpu_context
{
785 struct perf_event_context ctx
;
786 struct perf_event_context
*task_ctx
;
790 raw_spinlock_t hrtimer_lock
;
791 struct hrtimer hrtimer
;
792 ktime_t hrtimer_interval
;
793 unsigned int hrtimer_active
;
795 #ifdef CONFIG_CGROUP_PERF
796 struct perf_cgroup
*cgrp
;
797 struct list_head cgrp_cpuctx_entry
;
800 struct list_head sched_cb_entry
;
806 struct perf_output_handle
{
807 struct perf_event
*event
;
808 struct ring_buffer
*rb
;
809 unsigned long wakeup
;
819 struct bpf_perf_event_data_kern
{
820 struct pt_regs
*regs
;
821 struct perf_sample_data
*data
;
824 #ifdef CONFIG_CGROUP_PERF
827 * perf_cgroup_info keeps track of time_enabled for a cgroup.
828 * This is a per-cpu dynamically allocated data structure.
830 struct perf_cgroup_info
{
836 struct cgroup_subsys_state css
;
837 struct perf_cgroup_info __percpu
*info
;
841 * Must ensure cgroup is pinned (css_get) before calling
842 * this function. In other words, we cannot call this function
843 * if there is no cgroup event for the current CPU context.
845 static inline struct perf_cgroup
*
846 perf_cgroup_from_task(struct task_struct
*task
, struct perf_event_context
*ctx
)
848 return container_of(task_css_check(task
, perf_event_cgrp_id
,
849 ctx
? lockdep_is_held(&ctx
->lock
)
851 struct perf_cgroup
, css
);
853 #endif /* CONFIG_CGROUP_PERF */
855 #ifdef CONFIG_PERF_EVENTS
857 extern void *perf_aux_output_begin(struct perf_output_handle
*handle
,
858 struct perf_event
*event
);
859 extern void perf_aux_output_end(struct perf_output_handle
*handle
,
861 extern int perf_aux_output_skip(struct perf_output_handle
*handle
,
863 extern void *perf_get_aux(struct perf_output_handle
*handle
);
864 extern void perf_aux_output_flag(struct perf_output_handle
*handle
, u64 flags
);
865 extern void perf_event_itrace_started(struct perf_event
*event
);
867 extern int perf_pmu_register(struct pmu
*pmu
, const char *name
, int type
);
868 extern void perf_pmu_unregister(struct pmu
*pmu
);
870 extern int perf_num_counters(void);
871 extern const char *perf_pmu_name(void);
872 extern void __perf_event_task_sched_in(struct task_struct
*prev
,
873 struct task_struct
*task
);
874 extern void __perf_event_task_sched_out(struct task_struct
*prev
,
875 struct task_struct
*next
);
876 extern int perf_event_init_task(struct task_struct
*child
);
877 extern void perf_event_exit_task(struct task_struct
*child
);
878 extern void perf_event_free_task(struct task_struct
*task
);
879 extern void perf_event_delayed_put(struct task_struct
*task
);
880 extern struct file
*perf_event_get(unsigned int fd
);
881 extern const struct perf_event_attr
*perf_event_attrs(struct perf_event
*event
);
882 extern void perf_event_print_debug(void);
883 extern void perf_pmu_disable(struct pmu
*pmu
);
884 extern void perf_pmu_enable(struct pmu
*pmu
);
885 extern void perf_sched_cb_dec(struct pmu
*pmu
);
886 extern void perf_sched_cb_inc(struct pmu
*pmu
);
887 extern int perf_event_task_disable(void);
888 extern int perf_event_task_enable(void);
889 extern int perf_event_refresh(struct perf_event
*event
, int refresh
);
890 extern void perf_event_update_userpage(struct perf_event
*event
);
891 extern int perf_event_release_kernel(struct perf_event
*event
);
892 extern struct perf_event
*
893 perf_event_create_kernel_counter(struct perf_event_attr
*attr
,
895 struct task_struct
*task
,
896 perf_overflow_handler_t callback
,
898 extern void perf_pmu_migrate_context(struct pmu
*pmu
,
899 int src_cpu
, int dst_cpu
);
900 int perf_event_read_local(struct perf_event
*event
, u64
*value
);
901 extern u64
perf_event_read_value(struct perf_event
*event
,
902 u64
*enabled
, u64
*running
);
905 struct perf_sample_data
{
907 * Fields set by perf_sample_data_init(), group so as to
908 * minimize the cachelines touched.
911 struct perf_raw_record
*raw
;
912 struct perf_branch_stack
*br_stack
;
916 union perf_mem_data_src data_src
;
919 * The other fields, optionally {set,used} by
920 * perf_{prepare,output}_sample().
935 struct perf_callchain_entry
*callchain
;
938 * regs_user may point to task_pt_regs or to regs_user_copy, depending
941 struct perf_regs regs_user
;
942 struct pt_regs regs_user_copy
;
944 struct perf_regs regs_intr
;
948 } ____cacheline_aligned
;
950 /* default value for data source */
951 #define PERF_MEM_NA (PERF_MEM_S(OP, NA) |\
952 PERF_MEM_S(LVL, NA) |\
953 PERF_MEM_S(SNOOP, NA) |\
954 PERF_MEM_S(LOCK, NA) |\
957 static inline void perf_sample_data_init(struct perf_sample_data
*data
,
958 u64 addr
, u64 period
)
960 /* remaining struct members initialized in perf_prepare_sample() */
963 data
->br_stack
= NULL
;
964 data
->period
= period
;
966 data
->data_src
.val
= PERF_MEM_NA
;
970 extern void perf_output_sample(struct perf_output_handle
*handle
,
971 struct perf_event_header
*header
,
972 struct perf_sample_data
*data
,
973 struct perf_event
*event
);
974 extern void perf_prepare_sample(struct perf_event_header
*header
,
975 struct perf_sample_data
*data
,
976 struct perf_event
*event
,
977 struct pt_regs
*regs
);
979 extern int perf_event_overflow(struct perf_event
*event
,
980 struct perf_sample_data
*data
,
981 struct pt_regs
*regs
);
983 extern void perf_event_output_forward(struct perf_event
*event
,
984 struct perf_sample_data
*data
,
985 struct pt_regs
*regs
);
986 extern void perf_event_output_backward(struct perf_event
*event
,
987 struct perf_sample_data
*data
,
988 struct pt_regs
*regs
);
989 extern void perf_event_output(struct perf_event
*event
,
990 struct perf_sample_data
*data
,
991 struct pt_regs
*regs
);
994 is_default_overflow_handler(struct perf_event
*event
)
996 if (likely(event
->overflow_handler
== perf_event_output_forward
))
998 if (unlikely(event
->overflow_handler
== perf_event_output_backward
))
1004 perf_event_header__init_id(struct perf_event_header
*header
,
1005 struct perf_sample_data
*data
,
1006 struct perf_event
*event
);
1008 perf_event__output_id_sample(struct perf_event
*event
,
1009 struct perf_output_handle
*handle
,
1010 struct perf_sample_data
*sample
);
1013 perf_log_lost_samples(struct perf_event
*event
, u64 lost
);
1015 static inline bool is_sampling_event(struct perf_event
*event
)
1017 return event
->attr
.sample_period
!= 0;
1021 * Return 1 for a software event, 0 for a hardware event
1023 static inline int is_software_event(struct perf_event
*event
)
1025 return event
->event_caps
& PERF_EV_CAP_SOFTWARE
;
1028 extern struct static_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
1030 extern void ___perf_sw_event(u32
, u64
, struct pt_regs
*, u64
);
1031 extern void __perf_sw_event(u32
, u64
, struct pt_regs
*, u64
);
1033 #ifndef perf_arch_fetch_caller_regs
1034 static inline void perf_arch_fetch_caller_regs(struct pt_regs
*regs
, unsigned long ip
) { }
1038 * Take a snapshot of the regs. Skip ip and frame pointer to
1039 * the nth caller. We only need a few of the regs:
1040 * - ip for PERF_SAMPLE_IP
1041 * - cs for user_mode() tests
1042 * - bp for callchains
1043 * - eflags, for future purposes, just in case
1045 static inline void perf_fetch_caller_regs(struct pt_regs
*regs
)
1047 perf_arch_fetch_caller_regs(regs
, CALLER_ADDR0
);
1050 static __always_inline
void
1051 perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
1053 if (static_key_false(&perf_swevent_enabled
[event_id
]))
1054 __perf_sw_event(event_id
, nr
, regs
, addr
);
1057 DECLARE_PER_CPU(struct pt_regs
, __perf_regs
[4]);
1060 * 'Special' version for the scheduler, it hard assumes no recursion,
1061 * which is guaranteed by us not actually scheduling inside other swevents
1062 * because those disable preemption.
1064 static __always_inline
void
1065 perf_sw_event_sched(u32 event_id
, u64 nr
, u64 addr
)
1067 if (static_key_false(&perf_swevent_enabled
[event_id
])) {
1068 struct pt_regs
*regs
= this_cpu_ptr(&__perf_regs
[0]);
1070 perf_fetch_caller_regs(regs
);
1071 ___perf_sw_event(event_id
, nr
, regs
, addr
);
1075 extern struct static_key_false perf_sched_events
;
1077 static __always_inline
bool
1078 perf_sw_migrate_enabled(void)
1080 if (static_key_false(&perf_swevent_enabled
[PERF_COUNT_SW_CPU_MIGRATIONS
]))
1085 static inline void perf_event_task_migrate(struct task_struct
*task
)
1087 if (perf_sw_migrate_enabled())
1088 task
->sched_migrated
= 1;
1091 static inline void perf_event_task_sched_in(struct task_struct
*prev
,
1092 struct task_struct
*task
)
1094 if (static_branch_unlikely(&perf_sched_events
))
1095 __perf_event_task_sched_in(prev
, task
);
1097 if (perf_sw_migrate_enabled() && task
->sched_migrated
) {
1098 struct pt_regs
*regs
= this_cpu_ptr(&__perf_regs
[0]);
1100 perf_fetch_caller_regs(regs
);
1101 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, regs
, 0);
1102 task
->sched_migrated
= 0;
1106 static inline void perf_event_task_sched_out(struct task_struct
*prev
,
1107 struct task_struct
*next
)
1109 perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES
, 1, 0);
1111 if (static_branch_unlikely(&perf_sched_events
))
1112 __perf_event_task_sched_out(prev
, next
);
1115 static inline u64
__perf_event_count(struct perf_event
*event
)
1117 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
1120 extern void perf_event_mmap(struct vm_area_struct
*vma
);
1121 extern struct perf_guest_info_callbacks
*perf_guest_cbs
;
1122 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*callbacks
);
1123 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*callbacks
);
1125 extern void perf_event_exec(void);
1126 extern void perf_event_comm(struct task_struct
*tsk
, bool exec
);
1127 extern void perf_event_namespaces(struct task_struct
*tsk
);
1128 extern void perf_event_fork(struct task_struct
*tsk
);
1131 DECLARE_PER_CPU(struct perf_callchain_entry
, perf_callchain_entry
);
1133 extern void perf_callchain_user(struct perf_callchain_entry_ctx
*entry
, struct pt_regs
*regs
);
1134 extern void perf_callchain_kernel(struct perf_callchain_entry_ctx
*entry
, struct pt_regs
*regs
);
1135 extern struct perf_callchain_entry
*
1136 get_perf_callchain(struct pt_regs
*regs
, u32 init_nr
, bool kernel
, bool user
,
1137 u32 max_stack
, bool crosstask
, bool add_mark
);
1138 extern int get_callchain_buffers(int max_stack
);
1139 extern void put_callchain_buffers(void);
1141 extern int sysctl_perf_event_max_stack
;
1142 extern int sysctl_perf_event_max_contexts_per_stack
;
1144 static inline int perf_callchain_store_context(struct perf_callchain_entry_ctx
*ctx
, u64 ip
)
1146 if (ctx
->contexts
< sysctl_perf_event_max_contexts_per_stack
) {
1147 struct perf_callchain_entry
*entry
= ctx
->entry
;
1148 entry
->ip
[entry
->nr
++] = ip
;
1152 ctx
->contexts_maxed
= true;
1153 return -1; /* no more room, stop walking the stack */
1157 static inline int perf_callchain_store(struct perf_callchain_entry_ctx
*ctx
, u64 ip
)
1159 if (ctx
->nr
< ctx
->max_stack
&& !ctx
->contexts_maxed
) {
1160 struct perf_callchain_entry
*entry
= ctx
->entry
;
1161 entry
->ip
[entry
->nr
++] = ip
;
1165 return -1; /* no more room, stop walking the stack */
1169 extern int sysctl_perf_event_paranoid
;
1170 extern int sysctl_perf_event_mlock
;
1171 extern int sysctl_perf_event_sample_rate
;
1172 extern int sysctl_perf_cpu_time_max_percent
;
1174 extern void perf_sample_event_took(u64 sample_len_ns
);
1176 extern int perf_proc_update_handler(struct ctl_table
*table
, int write
,
1177 void __user
*buffer
, size_t *lenp
,
1179 extern int perf_cpu_time_max_percent_handler(struct ctl_table
*table
, int write
,
1180 void __user
*buffer
, size_t *lenp
,
1183 int perf_event_max_stack_handler(struct ctl_table
*table
, int write
,
1184 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
);
1186 static inline bool perf_paranoid_tracepoint_raw(void)
1188 return sysctl_perf_event_paranoid
> -1;
1191 static inline bool perf_paranoid_cpu(void)
1193 return sysctl_perf_event_paranoid
> 0;
1196 static inline bool perf_paranoid_kernel(void)
1198 return sysctl_perf_event_paranoid
> 1;
1201 extern void perf_event_init(void);
1202 extern void perf_tp_event(u16 event_type
, u64 count
, void *record
,
1203 int entry_size
, struct pt_regs
*regs
,
1204 struct hlist_head
*head
, int rctx
,
1205 struct task_struct
*task
, struct perf_event
*event
);
1206 extern void perf_bp_event(struct perf_event
*event
, void *data
);
1208 #ifndef perf_misc_flags
1209 # define perf_misc_flags(regs) \
1210 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1211 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1214 static inline bool has_branch_stack(struct perf_event
*event
)
1216 return event
->attr
.sample_type
& PERF_SAMPLE_BRANCH_STACK
;
1219 static inline bool needs_branch_stack(struct perf_event
*event
)
1221 return event
->attr
.branch_sample_type
!= 0;
1224 static inline bool has_aux(struct perf_event
*event
)
1226 return event
->pmu
->setup_aux
;
1229 static inline bool is_write_backward(struct perf_event
*event
)
1231 return !!event
->attr
.write_backward
;
1234 static inline bool has_addr_filter(struct perf_event
*event
)
1236 return event
->pmu
->nr_addr_filters
;
1240 * An inherited event uses parent's filters
1242 static inline struct perf_addr_filters_head
*
1243 perf_event_addr_filters(struct perf_event
*event
)
1245 struct perf_addr_filters_head
*ifh
= &event
->addr_filters
;
1248 ifh
= &event
->parent
->addr_filters
;
1253 extern void perf_event_addr_filters_sync(struct perf_event
*event
);
1255 extern int perf_output_begin(struct perf_output_handle
*handle
,
1256 struct perf_event
*event
, unsigned int size
);
1257 extern int perf_output_begin_forward(struct perf_output_handle
*handle
,
1258 struct perf_event
*event
,
1260 extern int perf_output_begin_backward(struct perf_output_handle
*handle
,
1261 struct perf_event
*event
,
1264 extern void perf_output_end(struct perf_output_handle
*handle
);
1265 extern unsigned int perf_output_copy(struct perf_output_handle
*handle
,
1266 const void *buf
, unsigned int len
);
1267 extern unsigned int perf_output_skip(struct perf_output_handle
*handle
,
1269 extern int perf_swevent_get_recursion_context(void);
1270 extern void perf_swevent_put_recursion_context(int rctx
);
1271 extern u64
perf_swevent_set_period(struct perf_event
*event
);
1272 extern void perf_event_enable(struct perf_event
*event
);
1273 extern void perf_event_disable(struct perf_event
*event
);
1274 extern void perf_event_disable_local(struct perf_event
*event
);
1275 extern void perf_event_disable_inatomic(struct perf_event
*event
);
1276 extern void perf_event_task_tick(void);
1277 extern int perf_event_account_interrupt(struct perf_event
*event
);
1278 #else /* !CONFIG_PERF_EVENTS: */
1279 static inline void *
1280 perf_aux_output_begin(struct perf_output_handle
*handle
,
1281 struct perf_event
*event
) { return NULL
; }
1283 perf_aux_output_end(struct perf_output_handle
*handle
, unsigned long size
)
1286 perf_aux_output_skip(struct perf_output_handle
*handle
,
1287 unsigned long size
) { return -EINVAL
; }
1288 static inline void *
1289 perf_get_aux(struct perf_output_handle
*handle
) { return NULL
; }
1291 perf_event_task_migrate(struct task_struct
*task
) { }
1293 perf_event_task_sched_in(struct task_struct
*prev
,
1294 struct task_struct
*task
) { }
1296 perf_event_task_sched_out(struct task_struct
*prev
,
1297 struct task_struct
*next
) { }
1298 static inline int perf_event_init_task(struct task_struct
*child
) { return 0; }
1299 static inline void perf_event_exit_task(struct task_struct
*child
) { }
1300 static inline void perf_event_free_task(struct task_struct
*task
) { }
1301 static inline void perf_event_delayed_put(struct task_struct
*task
) { }
1302 static inline struct file
*perf_event_get(unsigned int fd
) { return ERR_PTR(-EINVAL
); }
1303 static inline const struct perf_event_attr
*perf_event_attrs(struct perf_event
*event
)
1305 return ERR_PTR(-EINVAL
);
1307 static inline int perf_event_read_local(struct perf_event
*event
, u64
*value
)
1311 static inline void perf_event_print_debug(void) { }
1312 static inline int perf_event_task_disable(void) { return -EINVAL
; }
1313 static inline int perf_event_task_enable(void) { return -EINVAL
; }
1314 static inline int perf_event_refresh(struct perf_event
*event
, int refresh
)
1320 perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
) { }
1322 perf_sw_event_sched(u32 event_id
, u64 nr
, u64 addr
) { }
1324 perf_bp_event(struct perf_event
*event
, void *data
) { }
1326 static inline int perf_register_guest_info_callbacks
1327 (struct perf_guest_info_callbacks
*callbacks
) { return 0; }
1328 static inline int perf_unregister_guest_info_callbacks
1329 (struct perf_guest_info_callbacks
*callbacks
) { return 0; }
1331 static inline void perf_event_mmap(struct vm_area_struct
*vma
) { }
1332 static inline void perf_event_exec(void) { }
1333 static inline void perf_event_comm(struct task_struct
*tsk
, bool exec
) { }
1334 static inline void perf_event_namespaces(struct task_struct
*tsk
) { }
1335 static inline void perf_event_fork(struct task_struct
*tsk
) { }
1336 static inline void perf_event_init(void) { }
1337 static inline int perf_swevent_get_recursion_context(void) { return -1; }
1338 static inline void perf_swevent_put_recursion_context(int rctx
) { }
1339 static inline u64
perf_swevent_set_period(struct perf_event
*event
) { return 0; }
1340 static inline void perf_event_enable(struct perf_event
*event
) { }
1341 static inline void perf_event_disable(struct perf_event
*event
) { }
1342 static inline int __perf_event_disable(void *info
) { return -1; }
1343 static inline void perf_event_task_tick(void) { }
1344 static inline int perf_event_release_kernel(struct perf_event
*event
) { return 0; }
1347 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1348 extern void perf_restore_debug_store(void);
1350 static inline void perf_restore_debug_store(void) { }
1353 static __always_inline
bool perf_raw_frag_last(const struct perf_raw_frag
*frag
)
1355 return frag
->pad
< sizeof(u64
);
1358 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1360 struct perf_pmu_events_attr
{
1361 struct device_attribute attr
;
1363 const char *event_str
;
1366 struct perf_pmu_events_ht_attr
{
1367 struct device_attribute attr
;
1369 const char *event_str_ht
;
1370 const char *event_str_noht
;
1373 ssize_t
perf_event_sysfs_show(struct device
*dev
, struct device_attribute
*attr
,
1376 #define PMU_EVENT_ATTR(_name, _var, _id, _show) \
1377 static struct perf_pmu_events_attr _var = { \
1378 .attr = __ATTR(_name, 0444, _show, NULL), \
1382 #define PMU_EVENT_ATTR_STRING(_name, _var, _str) \
1383 static struct perf_pmu_events_attr _var = { \
1384 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1386 .event_str = _str, \
1389 #define PMU_FORMAT_ATTR(_name, _format) \
1391 _name##_show(struct device *dev, \
1392 struct device_attribute *attr, \
1395 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
1396 return sprintf(page, _format "\n"); \
1399 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1401 /* Performance counter hotplug functions */
1402 #ifdef CONFIG_PERF_EVENTS
1403 int perf_event_init_cpu(unsigned int cpu
);
1404 int perf_event_exit_cpu(unsigned int cpu
);
1406 #define perf_event_init_cpu NULL
1407 #define perf_event_exit_cpu NULL
1410 #endif /* _LINUX_PERF_EVENT_H */