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>
18 #include <uapi/linux/bpf_perf_event.h>
21 * Kernel-internal data types and definitions:
24 #ifdef CONFIG_PERF_EVENTS
25 # include <asm/perf_event.h>
26 # include <asm/local64.h>
29 struct perf_guest_info_callbacks
{
30 int (*is_in_guest
)(void);
31 int (*is_user_mode
)(void);
32 unsigned long (*get_guest_ip
)(void);
33 void (*handle_intel_pt_intr
)(void);
36 #ifdef CONFIG_HAVE_HW_BREAKPOINT
37 #include <asm/hw_breakpoint.h>
40 #include <linux/list.h>
41 #include <linux/mutex.h>
42 #include <linux/rculist.h>
43 #include <linux/rcupdate.h>
44 #include <linux/spinlock.h>
45 #include <linux/hrtimer.h>
47 #include <linux/pid_namespace.h>
48 #include <linux/workqueue.h>
49 #include <linux/ftrace.h>
50 #include <linux/cpu.h>
51 #include <linux/irq_work.h>
52 #include <linux/static_key.h>
53 #include <linux/jump_label_ratelimit.h>
54 #include <linux/atomic.h>
55 #include <linux/sysfs.h>
56 #include <linux/perf_regs.h>
57 #include <linux/cgroup.h>
58 #include <linux/refcount.h>
59 #include <linux/security.h>
60 #include <asm/local.h>
62 struct perf_callchain_entry
{
64 __u64 ip
[0]; /* /proc/sys/kernel/perf_event_max_stack */
67 struct perf_callchain_entry_ctx
{
68 struct perf_callchain_entry
*entry
;
75 typedef unsigned long (*perf_copy_f
)(void *dst
, const void *src
,
76 unsigned long off
, unsigned long len
);
78 struct perf_raw_frag
{
80 struct perf_raw_frag
*next
;
88 struct perf_raw_record
{
89 struct perf_raw_frag frag
;
94 * branch stack layout:
95 * nr: number of taken branches stored in entries[]
97 * Note that nr can vary from sample to sample
98 * branches (to, from) are stored from most recent
99 * to least recent, i.e., entries[0] contains the most
102 struct perf_branch_stack
{
104 struct perf_branch_entry entries
[0];
110 * extra PMU register associated with an event
112 struct hw_perf_event_extra
{
113 u64 config
; /* register value */
114 unsigned int reg
; /* register address or index */
115 int alloc
; /* extra register already allocated */
116 int idx
; /* index in shared_regs->regs[] */
120 * struct hw_perf_event - performance event hardware details:
122 struct hw_perf_event
{
123 #ifdef CONFIG_PERF_EVENTS
125 struct { /* hardware */
128 unsigned long config_base
;
129 unsigned long event_base
;
130 int event_base_rdpmc
;
135 struct hw_perf_event_extra extra_reg
;
136 struct hw_perf_event_extra branch_reg
;
138 struct { /* software */
139 struct hrtimer hrtimer
;
141 struct { /* tracepoint */
142 /* for tp_event->class */
143 struct list_head tp_list
;
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_EXTENDED_REGS 0x08
246 #define PERF_PMU_CAP_EXCLUSIVE 0x10
247 #define PERF_PMU_CAP_ITRACE 0x20
248 #define PERF_PMU_CAP_HETEROGENEOUS_CPUS 0x40
249 #define PERF_PMU_CAP_NO_EXCLUDE 0x80
250 #define PERF_PMU_CAP_AUX_OUTPUT 0x100
253 * struct pmu - generic performance monitoring unit
256 struct list_head entry
;
258 struct module
*module
;
260 const struct attribute_group
**attr_groups
;
261 const struct attribute_group
**attr_update
;
266 * various common per-pmu feature flags
270 int __percpu
*pmu_disable_count
;
271 struct perf_cpu_context __percpu
*pmu_cpu_context
;
272 atomic_t exclusive_cnt
; /* < 0: cpu; > 0: tsk */
274 int hrtimer_interval_ms
;
276 /* number of address filters this PMU can do */
277 unsigned int nr_addr_filters
;
280 * Fully disable/enable this PMU, can be used to protect from the PMI
281 * as well as for lazy/batch writing of the MSRs.
283 void (*pmu_enable
) (struct pmu
*pmu
); /* optional */
284 void (*pmu_disable
) (struct pmu
*pmu
); /* optional */
287 * Try and initialize the event for this PMU.
290 * -ENOENT -- @event is not for this PMU
292 * -ENODEV -- @event is for this PMU but PMU not present
293 * -EBUSY -- @event is for this PMU but PMU temporarily unavailable
294 * -EINVAL -- @event is for this PMU but @event is not valid
295 * -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
296 * -EACCES -- @event is for this PMU, @event is valid, but no privileges
298 * 0 -- @event is for this PMU and valid
300 * Other error return values are allowed.
302 int (*event_init
) (struct perf_event
*event
);
305 * Notification that the event was mapped or unmapped. Called
306 * in the context of the mapping task.
308 void (*event_mapped
) (struct perf_event
*event
, struct mm_struct
*mm
); /* optional */
309 void (*event_unmapped
) (struct perf_event
*event
, struct mm_struct
*mm
); /* optional */
312 * Flags for ->add()/->del()/ ->start()/->stop(). There are
313 * matching hw_perf_event::state flags.
315 #define PERF_EF_START 0x01 /* start the counter when adding */
316 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
317 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
320 * Adds/Removes a counter to/from the PMU, can be done inside a
321 * transaction, see the ->*_txn() methods.
323 * The add/del callbacks will reserve all hardware resources required
324 * to service the event, this includes any counter constraint
327 * Called with IRQs disabled and the PMU disabled on the CPU the event
330 * ->add() called without PERF_EF_START should result in the same state
331 * as ->add() followed by ->stop().
333 * ->del() must always PERF_EF_UPDATE stop an event. If it calls
334 * ->stop() that must deal with already being stopped without
337 int (*add
) (struct perf_event
*event
, int flags
);
338 void (*del
) (struct perf_event
*event
, int flags
);
341 * Starts/Stops a counter present on the PMU.
343 * The PMI handler should stop the counter when perf_event_overflow()
344 * returns !0. ->start() will be used to continue.
346 * Also used to change the sample period.
348 * Called with IRQs disabled and the PMU disabled on the CPU the event
349 * is on -- will be called from NMI context with the PMU generates
352 * ->stop() with PERF_EF_UPDATE will read the counter and update
353 * period/count values like ->read() would.
355 * ->start() with PERF_EF_RELOAD will reprogram the the counter
356 * value, must be preceded by a ->stop() with PERF_EF_UPDATE.
358 void (*start
) (struct perf_event
*event
, int flags
);
359 void (*stop
) (struct perf_event
*event
, int flags
);
362 * Updates the counter value of the event.
364 * For sampling capable PMUs this will also update the software period
365 * hw_perf_event::period_left field.
367 void (*read
) (struct perf_event
*event
);
370 * Group events scheduling is treated as a transaction, add
371 * group events as a whole and perform one schedulability test.
372 * If the test fails, roll back the whole group
374 * Start the transaction, after this ->add() doesn't need to
375 * do schedulability tests.
379 void (*start_txn
) (struct pmu
*pmu
, unsigned int txn_flags
);
381 * If ->start_txn() disabled the ->add() schedulability test
382 * then ->commit_txn() is required to perform one. On success
383 * the transaction is closed. On error the transaction is kept
384 * open until ->cancel_txn() is called.
388 int (*commit_txn
) (struct pmu
*pmu
);
390 * Will cancel the transaction, assumes ->del() is called
391 * for each successful ->add() during the transaction.
395 void (*cancel_txn
) (struct pmu
*pmu
);
398 * Will return the value for perf_event_mmap_page::index for this event,
399 * if no implementation is provided it will default to: event->hw.idx + 1.
401 int (*event_idx
) (struct perf_event
*event
); /*optional */
404 * context-switches callback
406 void (*sched_task
) (struct perf_event_context
*ctx
,
409 * PMU specific data size
411 size_t task_ctx_size
;
415 * Set up pmu-private data structures for an AUX area
417 void *(*setup_aux
) (struct perf_event
*event
, void **pages
,
418 int nr_pages
, bool overwrite
);
422 * Free pmu-private AUX data structures
424 void (*free_aux
) (void *aux
); /* optional */
427 * Validate address range filters: make sure the HW supports the
428 * requested configuration and number of filters; return 0 if the
429 * supplied filters are valid, -errno otherwise.
431 * Runs in the context of the ioctl()ing process and is not serialized
432 * with the rest of the PMU callbacks.
434 int (*addr_filters_validate
) (struct list_head
*filters
);
438 * Synchronize address range filter configuration:
439 * translate hw-agnostic filters into hardware configuration in
440 * event::hw::addr_filters.
442 * Runs as a part of filter sync sequence that is done in ->start()
443 * callback by calling perf_event_addr_filters_sync().
445 * May (and should) traverse event::addr_filters::list, for which its
446 * caller provides necessary serialization.
448 void (*addr_filters_sync
) (struct perf_event
*event
);
452 * Check if event can be used for aux_output purposes for
453 * events of this PMU.
455 * Runs from perf_event_open(). Should return 0 for "no match"
456 * or non-zero for "match".
458 int (*aux_output_match
) (struct perf_event
*event
);
462 * Filter events for PMU-specific reasons.
464 int (*filter_match
) (struct perf_event
*event
); /* optional */
467 * Check period value for PERF_EVENT_IOC_PERIOD ioctl.
469 int (*check_period
) (struct perf_event
*event
, u64 value
); /* optional */
472 enum perf_addr_filter_action_t
{
473 PERF_ADDR_FILTER_ACTION_STOP
= 0,
474 PERF_ADDR_FILTER_ACTION_START
,
475 PERF_ADDR_FILTER_ACTION_FILTER
,
479 * struct perf_addr_filter - address range filter definition
480 * @entry: event's filter list linkage
481 * @path: object file's path for file-based filters
482 * @offset: filter range offset
483 * @size: filter range size (size==0 means single address trigger)
484 * @action: filter/start/stop
486 * This is a hardware-agnostic filter configuration as specified by the user.
488 struct perf_addr_filter
{
489 struct list_head entry
;
491 unsigned long offset
;
493 enum perf_addr_filter_action_t action
;
497 * struct perf_addr_filters_head - container for address range filters
498 * @list: list of filters for this event
499 * @lock: spinlock that serializes accesses to the @list and event's
500 * (and its children's) filter generations.
501 * @nr_file_filters: number of file-based filters
503 * A child event will use parent's @list (and therefore @lock), so they are
504 * bundled together; see perf_event_addr_filters().
506 struct perf_addr_filters_head
{
507 struct list_head list
;
509 unsigned int nr_file_filters
;
512 struct perf_addr_filter_range
{
518 * enum perf_event_state - the states of an event:
520 enum perf_event_state
{
521 PERF_EVENT_STATE_DEAD
= -4,
522 PERF_EVENT_STATE_EXIT
= -3,
523 PERF_EVENT_STATE_ERROR
= -2,
524 PERF_EVENT_STATE_OFF
= -1,
525 PERF_EVENT_STATE_INACTIVE
= 0,
526 PERF_EVENT_STATE_ACTIVE
= 1,
530 struct perf_sample_data
;
532 typedef void (*perf_overflow_handler_t
)(struct perf_event
*,
533 struct perf_sample_data
*,
534 struct pt_regs
*regs
);
537 * Event capabilities. For event_caps and groups caps.
539 * PERF_EV_CAP_SOFTWARE: Is a software event.
540 * PERF_EV_CAP_READ_ACTIVE_PKG: A CPU event (or cgroup event) that can be read
541 * from any CPU in the package where it is active.
543 #define PERF_EV_CAP_SOFTWARE BIT(0)
544 #define PERF_EV_CAP_READ_ACTIVE_PKG BIT(1)
546 #define SWEVENT_HLIST_BITS 8
547 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
549 struct swevent_hlist
{
550 struct hlist_head heads
[SWEVENT_HLIST_SIZE
];
551 struct rcu_head rcu_head
;
554 #define PERF_ATTACH_CONTEXT 0x01
555 #define PERF_ATTACH_GROUP 0x02
556 #define PERF_ATTACH_TASK 0x04
557 #define PERF_ATTACH_TASK_DATA 0x08
558 #define PERF_ATTACH_ITRACE 0x10
563 struct pmu_event_list
{
565 struct list_head list
;
568 #define for_each_sibling_event(sibling, event) \
569 if ((event)->group_leader == (event)) \
570 list_for_each_entry((sibling), &(event)->sibling_list, sibling_list)
573 * struct perf_event - performance event kernel representation:
576 #ifdef CONFIG_PERF_EVENTS
578 * entry onto perf_event_context::event_list;
579 * modifications require ctx->lock
580 * RCU safe iterations.
582 struct list_head event_entry
;
585 * Locked for modification by both ctx->mutex and ctx->lock; holding
586 * either sufficies for read.
588 struct list_head sibling_list
;
589 struct list_head active_list
;
591 * Node on the pinned or flexible tree located at the event context;
593 struct rb_node group_node
;
596 * We need storage to track the entries in perf_pmu_migrate_context; we
597 * cannot use the event_entry because of RCU and we want to keep the
598 * group in tact which avoids us using the other two entries.
600 struct list_head migrate_entry
;
602 struct hlist_node hlist_entry
;
603 struct list_head active_entry
;
606 /* Not serialized. Only written during event initialization. */
608 /* The cumulative AND of all event_caps for events in this group. */
611 struct perf_event
*group_leader
;
615 enum perf_event_state state
;
616 unsigned int attach_state
;
618 atomic64_t child_count
;
621 * These are the total time in nanoseconds that the event
622 * has been enabled (i.e. eligible to run, and the task has
623 * been scheduled in, if this is a per-task event)
624 * and running (scheduled onto the CPU), respectively.
626 u64 total_time_enabled
;
627 u64 total_time_running
;
631 * timestamp shadows the actual context timing but it can
632 * be safely used in NMI interrupt context. It reflects the
633 * context time as it was when the event was last scheduled in.
635 * ctx_time already accounts for ctx->timestamp. Therefore to
636 * compute ctx_time for a sample, simply add perf_clock().
640 struct perf_event_attr attr
;
644 struct hw_perf_event hw
;
646 struct perf_event_context
*ctx
;
647 atomic_long_t refcount
;
650 * These accumulate total time (in nanoseconds) that children
651 * events have been enabled and running, respectively.
653 atomic64_t child_total_time_enabled
;
654 atomic64_t child_total_time_running
;
657 * Protect attach/detach and child_list:
659 struct mutex child_mutex
;
660 struct list_head child_list
;
661 struct perf_event
*parent
;
666 struct list_head owner_entry
;
667 struct task_struct
*owner
;
670 struct mutex mmap_mutex
;
673 struct ring_buffer
*rb
;
674 struct list_head rb_entry
;
675 unsigned long rcu_batches
;
679 wait_queue_head_t waitq
;
680 struct fasync_struct
*fasync
;
682 /* delayed work for NMIs and such */
686 struct irq_work pending
;
688 atomic_t event_limit
;
690 /* address range filters */
691 struct perf_addr_filters_head addr_filters
;
692 /* vma address array for file-based filders */
693 struct perf_addr_filter_range
*addr_filter_ranges
;
694 unsigned long addr_filters_gen
;
696 /* for aux_output events */
697 struct perf_event
*aux_event
;
699 void (*destroy
)(struct perf_event
*);
700 struct rcu_head rcu_head
;
702 struct pid_namespace
*ns
;
706 perf_overflow_handler_t overflow_handler
;
707 void *overflow_handler_context
;
708 #ifdef CONFIG_BPF_SYSCALL
709 perf_overflow_handler_t orig_overflow_handler
;
710 struct bpf_prog
*prog
;
713 #ifdef CONFIG_EVENT_TRACING
714 struct trace_event_call
*tp_event
;
715 struct event_filter
*filter
;
716 #ifdef CONFIG_FUNCTION_TRACER
717 struct ftrace_ops ftrace_ops
;
721 #ifdef CONFIG_CGROUP_PERF
722 struct perf_cgroup
*cgrp
; /* cgroup event is attach to */
725 #ifdef CONFIG_SECURITY
728 struct list_head sb_list
;
729 #endif /* CONFIG_PERF_EVENTS */
733 struct perf_event_groups
{
739 * struct perf_event_context - event context structure
741 * Used as a container for task events and CPU events as well:
743 struct perf_event_context
{
746 * Protect the states of the events in the list,
747 * nr_active, and the list:
751 * Protect the list of events. Locking either mutex or lock
752 * is sufficient to ensure the list doesn't change; to change
753 * the list you need to lock both the mutex and the spinlock.
757 struct list_head active_ctx_list
;
758 struct perf_event_groups pinned_groups
;
759 struct perf_event_groups flexible_groups
;
760 struct list_head event_list
;
762 struct list_head pinned_active
;
763 struct list_head flexible_active
;
772 * Set when nr_events != nr_active, except tolerant to events not
773 * necessary to be active due to scheduling constraints, such as cgroups.
775 int rotate_necessary
;
777 struct task_struct
*task
;
780 * Context clock, runs when context enabled.
786 * These fields let us detect when two contexts have both
787 * been cloned (inherited) from a common ancestor.
789 struct perf_event_context
*parent_ctx
;
793 #ifdef CONFIG_CGROUP_PERF
794 int nr_cgroups
; /* cgroup evts */
796 void *task_ctx_data
; /* pmu specific data */
797 struct rcu_head rcu_head
;
801 * Number of contexts where an event can trigger:
802 * task, softirq, hardirq, nmi.
804 #define PERF_NR_CONTEXTS 4
807 * struct perf_event_cpu_context - per cpu event context structure
809 struct perf_cpu_context
{
810 struct perf_event_context ctx
;
811 struct perf_event_context
*task_ctx
;
815 raw_spinlock_t hrtimer_lock
;
816 struct hrtimer hrtimer
;
817 ktime_t hrtimer_interval
;
818 unsigned int hrtimer_active
;
820 #ifdef CONFIG_CGROUP_PERF
821 struct perf_cgroup
*cgrp
;
822 struct list_head cgrp_cpuctx_entry
;
825 struct list_head sched_cb_entry
;
831 struct perf_output_handle
{
832 struct perf_event
*event
;
833 struct ring_buffer
*rb
;
834 unsigned long wakeup
;
844 struct bpf_perf_event_data_kern
{
845 bpf_user_pt_regs_t
*regs
;
846 struct perf_sample_data
*data
;
847 struct perf_event
*event
;
850 #ifdef CONFIG_CGROUP_PERF
853 * perf_cgroup_info keeps track of time_enabled for a cgroup.
854 * This is a per-cpu dynamically allocated data structure.
856 struct perf_cgroup_info
{
862 struct cgroup_subsys_state css
;
863 struct perf_cgroup_info __percpu
*info
;
867 * Must ensure cgroup is pinned (css_get) before calling
868 * this function. In other words, we cannot call this function
869 * if there is no cgroup event for the current CPU context.
871 static inline struct perf_cgroup
*
872 perf_cgroup_from_task(struct task_struct
*task
, struct perf_event_context
*ctx
)
874 return container_of(task_css_check(task
, perf_event_cgrp_id
,
875 ctx
? lockdep_is_held(&ctx
->lock
)
877 struct perf_cgroup
, css
);
879 #endif /* CONFIG_CGROUP_PERF */
881 #ifdef CONFIG_PERF_EVENTS
883 extern void *perf_aux_output_begin(struct perf_output_handle
*handle
,
884 struct perf_event
*event
);
885 extern void perf_aux_output_end(struct perf_output_handle
*handle
,
887 extern int perf_aux_output_skip(struct perf_output_handle
*handle
,
889 extern void *perf_get_aux(struct perf_output_handle
*handle
);
890 extern void perf_aux_output_flag(struct perf_output_handle
*handle
, u64 flags
);
891 extern void perf_event_itrace_started(struct perf_event
*event
);
893 extern int perf_pmu_register(struct pmu
*pmu
, const char *name
, int type
);
894 extern void perf_pmu_unregister(struct pmu
*pmu
);
896 extern int perf_num_counters(void);
897 extern const char *perf_pmu_name(void);
898 extern void __perf_event_task_sched_in(struct task_struct
*prev
,
899 struct task_struct
*task
);
900 extern void __perf_event_task_sched_out(struct task_struct
*prev
,
901 struct task_struct
*next
);
902 extern int perf_event_init_task(struct task_struct
*child
);
903 extern void perf_event_exit_task(struct task_struct
*child
);
904 extern void perf_event_free_task(struct task_struct
*task
);
905 extern void perf_event_delayed_put(struct task_struct
*task
);
906 extern struct file
*perf_event_get(unsigned int fd
);
907 extern const struct perf_event
*perf_get_event(struct file
*file
);
908 extern const struct perf_event_attr
*perf_event_attrs(struct perf_event
*event
);
909 extern void perf_event_print_debug(void);
910 extern void perf_pmu_disable(struct pmu
*pmu
);
911 extern void perf_pmu_enable(struct pmu
*pmu
);
912 extern void perf_sched_cb_dec(struct pmu
*pmu
);
913 extern void perf_sched_cb_inc(struct pmu
*pmu
);
914 extern int perf_event_task_disable(void);
915 extern int perf_event_task_enable(void);
917 extern void perf_pmu_resched(struct pmu
*pmu
);
919 extern int perf_event_refresh(struct perf_event
*event
, int refresh
);
920 extern void perf_event_update_userpage(struct perf_event
*event
);
921 extern int perf_event_release_kernel(struct perf_event
*event
);
922 extern struct perf_event
*
923 perf_event_create_kernel_counter(struct perf_event_attr
*attr
,
925 struct task_struct
*task
,
926 perf_overflow_handler_t callback
,
928 extern void perf_pmu_migrate_context(struct pmu
*pmu
,
929 int src_cpu
, int dst_cpu
);
930 int perf_event_read_local(struct perf_event
*event
, u64
*value
,
931 u64
*enabled
, u64
*running
);
932 extern u64
perf_event_read_value(struct perf_event
*event
,
933 u64
*enabled
, u64
*running
);
936 struct perf_sample_data
{
938 * Fields set by perf_sample_data_init(), group so as to
939 * minimize the cachelines touched.
942 struct perf_raw_record
*raw
;
943 struct perf_branch_stack
*br_stack
;
947 union perf_mem_data_src data_src
;
950 * The other fields, optionally {set,used} by
951 * perf_{prepare,output}_sample().
966 struct perf_callchain_entry
*callchain
;
969 * regs_user may point to task_pt_regs or to regs_user_copy, depending
972 struct perf_regs regs_user
;
973 struct pt_regs regs_user_copy
;
975 struct perf_regs regs_intr
;
979 } ____cacheline_aligned
;
981 /* default value for data source */
982 #define PERF_MEM_NA (PERF_MEM_S(OP, NA) |\
983 PERF_MEM_S(LVL, NA) |\
984 PERF_MEM_S(SNOOP, NA) |\
985 PERF_MEM_S(LOCK, NA) |\
988 static inline void perf_sample_data_init(struct perf_sample_data
*data
,
989 u64 addr
, u64 period
)
991 /* remaining struct members initialized in perf_prepare_sample() */
994 data
->br_stack
= NULL
;
995 data
->period
= period
;
997 data
->data_src
.val
= PERF_MEM_NA
;
1001 extern void perf_output_sample(struct perf_output_handle
*handle
,
1002 struct perf_event_header
*header
,
1003 struct perf_sample_data
*data
,
1004 struct perf_event
*event
);
1005 extern void perf_prepare_sample(struct perf_event_header
*header
,
1006 struct perf_sample_data
*data
,
1007 struct perf_event
*event
,
1008 struct pt_regs
*regs
);
1010 extern int perf_event_overflow(struct perf_event
*event
,
1011 struct perf_sample_data
*data
,
1012 struct pt_regs
*regs
);
1014 extern void perf_event_output_forward(struct perf_event
*event
,
1015 struct perf_sample_data
*data
,
1016 struct pt_regs
*regs
);
1017 extern void perf_event_output_backward(struct perf_event
*event
,
1018 struct perf_sample_data
*data
,
1019 struct pt_regs
*regs
);
1020 extern int perf_event_output(struct perf_event
*event
,
1021 struct perf_sample_data
*data
,
1022 struct pt_regs
*regs
);
1025 is_default_overflow_handler(struct perf_event
*event
)
1027 if (likely(event
->overflow_handler
== perf_event_output_forward
))
1029 if (unlikely(event
->overflow_handler
== perf_event_output_backward
))
1035 perf_event_header__init_id(struct perf_event_header
*header
,
1036 struct perf_sample_data
*data
,
1037 struct perf_event
*event
);
1039 perf_event__output_id_sample(struct perf_event
*event
,
1040 struct perf_output_handle
*handle
,
1041 struct perf_sample_data
*sample
);
1044 perf_log_lost_samples(struct perf_event
*event
, u64 lost
);
1046 static inline bool event_has_any_exclude_flag(struct perf_event
*event
)
1048 struct perf_event_attr
*attr
= &event
->attr
;
1050 return attr
->exclude_idle
|| attr
->exclude_user
||
1051 attr
->exclude_kernel
|| attr
->exclude_hv
||
1052 attr
->exclude_guest
|| attr
->exclude_host
;
1055 static inline bool is_sampling_event(struct perf_event
*event
)
1057 return event
->attr
.sample_period
!= 0;
1061 * Return 1 for a software event, 0 for a hardware event
1063 static inline int is_software_event(struct perf_event
*event
)
1065 return event
->event_caps
& PERF_EV_CAP_SOFTWARE
;
1069 * Return 1 for event in sw context, 0 for event in hw context
1071 static inline int in_software_context(struct perf_event
*event
)
1073 return event
->ctx
->pmu
->task_ctx_nr
== perf_sw_context
;
1076 static inline int is_exclusive_pmu(struct pmu
*pmu
)
1078 return pmu
->capabilities
& PERF_PMU_CAP_EXCLUSIVE
;
1081 extern struct static_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
1083 extern void ___perf_sw_event(u32
, u64
, struct pt_regs
*, u64
);
1084 extern void __perf_sw_event(u32
, u64
, struct pt_regs
*, u64
);
1086 #ifndef perf_arch_fetch_caller_regs
1087 static inline void perf_arch_fetch_caller_regs(struct pt_regs
*regs
, unsigned long ip
) { }
1091 * When generating a perf sample in-line, instead of from an interrupt /
1092 * exception, we lack a pt_regs. This is typically used from software events
1093 * like: SW_CONTEXT_SWITCHES, SW_MIGRATIONS and the tie-in with tracepoints.
1095 * We typically don't need a full set, but (for x86) do require:
1096 * - ip for PERF_SAMPLE_IP
1097 * - cs for user_mode() tests
1098 * - sp for PERF_SAMPLE_CALLCHAIN
1099 * - eflags for MISC bits and CALLCHAIN (see: perf_hw_regs())
1101 * NOTE: assumes @regs is otherwise already 0 filled; this is important for
1102 * things like PERF_SAMPLE_REGS_INTR.
1104 static inline void perf_fetch_caller_regs(struct pt_regs
*regs
)
1106 perf_arch_fetch_caller_regs(regs
, CALLER_ADDR0
);
1109 static __always_inline
void
1110 perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
1112 if (static_key_false(&perf_swevent_enabled
[event_id
]))
1113 __perf_sw_event(event_id
, nr
, regs
, addr
);
1116 DECLARE_PER_CPU(struct pt_regs
, __perf_regs
[4]);
1119 * 'Special' version for the scheduler, it hard assumes no recursion,
1120 * which is guaranteed by us not actually scheduling inside other swevents
1121 * because those disable preemption.
1123 static __always_inline
void
1124 perf_sw_event_sched(u32 event_id
, u64 nr
, u64 addr
)
1126 if (static_key_false(&perf_swevent_enabled
[event_id
])) {
1127 struct pt_regs
*regs
= this_cpu_ptr(&__perf_regs
[0]);
1129 perf_fetch_caller_regs(regs
);
1130 ___perf_sw_event(event_id
, nr
, regs
, addr
);
1134 extern struct static_key_false perf_sched_events
;
1136 static __always_inline
bool
1137 perf_sw_migrate_enabled(void)
1139 if (static_key_false(&perf_swevent_enabled
[PERF_COUNT_SW_CPU_MIGRATIONS
]))
1144 static inline void perf_event_task_migrate(struct task_struct
*task
)
1146 if (perf_sw_migrate_enabled())
1147 task
->sched_migrated
= 1;
1150 static inline void perf_event_task_sched_in(struct task_struct
*prev
,
1151 struct task_struct
*task
)
1153 if (static_branch_unlikely(&perf_sched_events
))
1154 __perf_event_task_sched_in(prev
, task
);
1156 if (perf_sw_migrate_enabled() && task
->sched_migrated
) {
1157 struct pt_regs
*regs
= this_cpu_ptr(&__perf_regs
[0]);
1159 perf_fetch_caller_regs(regs
);
1160 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, regs
, 0);
1161 task
->sched_migrated
= 0;
1165 static inline void perf_event_task_sched_out(struct task_struct
*prev
,
1166 struct task_struct
*next
)
1168 perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES
, 1, 0);
1170 if (static_branch_unlikely(&perf_sched_events
))
1171 __perf_event_task_sched_out(prev
, next
);
1174 extern void perf_event_mmap(struct vm_area_struct
*vma
);
1176 extern void perf_event_ksymbol(u16 ksym_type
, u64 addr
, u32 len
,
1177 bool unregister
, const char *sym
);
1178 extern void perf_event_bpf_event(struct bpf_prog
*prog
,
1179 enum perf_bpf_event_type type
,
1182 extern struct perf_guest_info_callbacks
*perf_guest_cbs
;
1183 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*callbacks
);
1184 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*callbacks
);
1186 extern void perf_event_exec(void);
1187 extern void perf_event_comm(struct task_struct
*tsk
, bool exec
);
1188 extern void perf_event_namespaces(struct task_struct
*tsk
);
1189 extern void perf_event_fork(struct task_struct
*tsk
);
1192 DECLARE_PER_CPU(struct perf_callchain_entry
, perf_callchain_entry
);
1194 extern void perf_callchain_user(struct perf_callchain_entry_ctx
*entry
, struct pt_regs
*regs
);
1195 extern void perf_callchain_kernel(struct perf_callchain_entry_ctx
*entry
, struct pt_regs
*regs
);
1196 extern struct perf_callchain_entry
*
1197 get_perf_callchain(struct pt_regs
*regs
, u32 init_nr
, bool kernel
, bool user
,
1198 u32 max_stack
, bool crosstask
, bool add_mark
);
1199 extern struct perf_callchain_entry
*perf_callchain(struct perf_event
*event
, struct pt_regs
*regs
);
1200 extern int get_callchain_buffers(int max_stack
);
1201 extern void put_callchain_buffers(void);
1203 extern int sysctl_perf_event_max_stack
;
1204 extern int sysctl_perf_event_max_contexts_per_stack
;
1206 static inline int perf_callchain_store_context(struct perf_callchain_entry_ctx
*ctx
, u64 ip
)
1208 if (ctx
->contexts
< sysctl_perf_event_max_contexts_per_stack
) {
1209 struct perf_callchain_entry
*entry
= ctx
->entry
;
1210 entry
->ip
[entry
->nr
++] = ip
;
1214 ctx
->contexts_maxed
= true;
1215 return -1; /* no more room, stop walking the stack */
1219 static inline int perf_callchain_store(struct perf_callchain_entry_ctx
*ctx
, u64 ip
)
1221 if (ctx
->nr
< ctx
->max_stack
&& !ctx
->contexts_maxed
) {
1222 struct perf_callchain_entry
*entry
= ctx
->entry
;
1223 entry
->ip
[entry
->nr
++] = ip
;
1227 return -1; /* no more room, stop walking the stack */
1231 extern int sysctl_perf_event_paranoid
;
1232 extern int sysctl_perf_event_mlock
;
1233 extern int sysctl_perf_event_sample_rate
;
1234 extern int sysctl_perf_cpu_time_max_percent
;
1236 extern void perf_sample_event_took(u64 sample_len_ns
);
1238 extern int perf_proc_update_handler(struct ctl_table
*table
, int write
,
1239 void __user
*buffer
, size_t *lenp
,
1241 extern int perf_cpu_time_max_percent_handler(struct ctl_table
*table
, int write
,
1242 void __user
*buffer
, size_t *lenp
,
1245 int perf_event_max_stack_handler(struct ctl_table
*table
, int write
,
1246 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
);
1248 /* Access to perf_event_open(2) syscall. */
1249 #define PERF_SECURITY_OPEN 0
1251 /* Finer grained perf_event_open(2) access control. */
1252 #define PERF_SECURITY_CPU 1
1253 #define PERF_SECURITY_KERNEL 2
1254 #define PERF_SECURITY_TRACEPOINT 3
1256 static inline int perf_is_paranoid(void)
1258 return sysctl_perf_event_paranoid
> -1;
1261 static inline int perf_allow_kernel(struct perf_event_attr
*attr
)
1263 if (sysctl_perf_event_paranoid
> 1 && !capable(CAP_SYS_ADMIN
))
1266 return security_perf_event_open(attr
, PERF_SECURITY_KERNEL
);
1269 static inline int perf_allow_cpu(struct perf_event_attr
*attr
)
1271 if (sysctl_perf_event_paranoid
> 0 && !capable(CAP_SYS_ADMIN
))
1274 return security_perf_event_open(attr
, PERF_SECURITY_CPU
);
1277 static inline int perf_allow_tracepoint(struct perf_event_attr
*attr
)
1279 if (sysctl_perf_event_paranoid
> -1 && !capable(CAP_SYS_ADMIN
))
1282 return security_perf_event_open(attr
, PERF_SECURITY_TRACEPOINT
);
1285 extern void perf_event_init(void);
1286 extern void perf_tp_event(u16 event_type
, u64 count
, void *record
,
1287 int entry_size
, struct pt_regs
*regs
,
1288 struct hlist_head
*head
, int rctx
,
1289 struct task_struct
*task
);
1290 extern void perf_bp_event(struct perf_event
*event
, void *data
);
1292 #ifndef perf_misc_flags
1293 # define perf_misc_flags(regs) \
1294 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1295 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1297 #ifndef perf_arch_bpf_user_pt_regs
1298 # define perf_arch_bpf_user_pt_regs(regs) regs
1301 static inline bool has_branch_stack(struct perf_event
*event
)
1303 return event
->attr
.sample_type
& PERF_SAMPLE_BRANCH_STACK
;
1306 static inline bool needs_branch_stack(struct perf_event
*event
)
1308 return event
->attr
.branch_sample_type
!= 0;
1311 static inline bool has_aux(struct perf_event
*event
)
1313 return event
->pmu
->setup_aux
;
1316 static inline bool is_write_backward(struct perf_event
*event
)
1318 return !!event
->attr
.write_backward
;
1321 static inline bool has_addr_filter(struct perf_event
*event
)
1323 return event
->pmu
->nr_addr_filters
;
1327 * An inherited event uses parent's filters
1329 static inline struct perf_addr_filters_head
*
1330 perf_event_addr_filters(struct perf_event
*event
)
1332 struct perf_addr_filters_head
*ifh
= &event
->addr_filters
;
1335 ifh
= &event
->parent
->addr_filters
;
1340 extern void perf_event_addr_filters_sync(struct perf_event
*event
);
1342 extern int perf_output_begin(struct perf_output_handle
*handle
,
1343 struct perf_event
*event
, unsigned int size
);
1344 extern int perf_output_begin_forward(struct perf_output_handle
*handle
,
1345 struct perf_event
*event
,
1347 extern int perf_output_begin_backward(struct perf_output_handle
*handle
,
1348 struct perf_event
*event
,
1351 extern void perf_output_end(struct perf_output_handle
*handle
);
1352 extern unsigned int perf_output_copy(struct perf_output_handle
*handle
,
1353 const void *buf
, unsigned int len
);
1354 extern unsigned int perf_output_skip(struct perf_output_handle
*handle
,
1356 extern int perf_swevent_get_recursion_context(void);
1357 extern void perf_swevent_put_recursion_context(int rctx
);
1358 extern u64
perf_swevent_set_period(struct perf_event
*event
);
1359 extern void perf_event_enable(struct perf_event
*event
);
1360 extern void perf_event_disable(struct perf_event
*event
);
1361 extern void perf_event_disable_local(struct perf_event
*event
);
1362 extern void perf_event_disable_inatomic(struct perf_event
*event
);
1363 extern void perf_event_task_tick(void);
1364 extern int perf_event_account_interrupt(struct perf_event
*event
);
1365 #else /* !CONFIG_PERF_EVENTS: */
1366 static inline void *
1367 perf_aux_output_begin(struct perf_output_handle
*handle
,
1368 struct perf_event
*event
) { return NULL
; }
1370 perf_aux_output_end(struct perf_output_handle
*handle
, unsigned long size
)
1373 perf_aux_output_skip(struct perf_output_handle
*handle
,
1374 unsigned long size
) { return -EINVAL
; }
1375 static inline void *
1376 perf_get_aux(struct perf_output_handle
*handle
) { return NULL
; }
1378 perf_event_task_migrate(struct task_struct
*task
) { }
1380 perf_event_task_sched_in(struct task_struct
*prev
,
1381 struct task_struct
*task
) { }
1383 perf_event_task_sched_out(struct task_struct
*prev
,
1384 struct task_struct
*next
) { }
1385 static inline int perf_event_init_task(struct task_struct
*child
) { return 0; }
1386 static inline void perf_event_exit_task(struct task_struct
*child
) { }
1387 static inline void perf_event_free_task(struct task_struct
*task
) { }
1388 static inline void perf_event_delayed_put(struct task_struct
*task
) { }
1389 static inline struct file
*perf_event_get(unsigned int fd
) { return ERR_PTR(-EINVAL
); }
1390 static inline const struct perf_event
*perf_get_event(struct file
*file
)
1392 return ERR_PTR(-EINVAL
);
1394 static inline const struct perf_event_attr
*perf_event_attrs(struct perf_event
*event
)
1396 return ERR_PTR(-EINVAL
);
1398 static inline int perf_event_read_local(struct perf_event
*event
, u64
*value
,
1399 u64
*enabled
, u64
*running
)
1403 static inline void perf_event_print_debug(void) { }
1404 static inline int perf_event_task_disable(void) { return -EINVAL
; }
1405 static inline int perf_event_task_enable(void) { return -EINVAL
; }
1406 static inline int perf_event_refresh(struct perf_event
*event
, int refresh
)
1412 perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
) { }
1414 perf_sw_event_sched(u32 event_id
, u64 nr
, u64 addr
) { }
1416 perf_bp_event(struct perf_event
*event
, void *data
) { }
1418 static inline int perf_register_guest_info_callbacks
1419 (struct perf_guest_info_callbacks
*callbacks
) { return 0; }
1420 static inline int perf_unregister_guest_info_callbacks
1421 (struct perf_guest_info_callbacks
*callbacks
) { return 0; }
1423 static inline void perf_event_mmap(struct vm_area_struct
*vma
) { }
1425 typedef int (perf_ksymbol_get_name_f
)(char *name
, int name_len
, void *data
);
1426 static inline void perf_event_ksymbol(u16 ksym_type
, u64 addr
, u32 len
,
1427 bool unregister
, const char *sym
) { }
1428 static inline void perf_event_bpf_event(struct bpf_prog
*prog
,
1429 enum perf_bpf_event_type type
,
1431 static inline void perf_event_exec(void) { }
1432 static inline void perf_event_comm(struct task_struct
*tsk
, bool exec
) { }
1433 static inline void perf_event_namespaces(struct task_struct
*tsk
) { }
1434 static inline void perf_event_fork(struct task_struct
*tsk
) { }
1435 static inline void perf_event_init(void) { }
1436 static inline int perf_swevent_get_recursion_context(void) { return -1; }
1437 static inline void perf_swevent_put_recursion_context(int rctx
) { }
1438 static inline u64
perf_swevent_set_period(struct perf_event
*event
) { return 0; }
1439 static inline void perf_event_enable(struct perf_event
*event
) { }
1440 static inline void perf_event_disable(struct perf_event
*event
) { }
1441 static inline int __perf_event_disable(void *info
) { return -1; }
1442 static inline void perf_event_task_tick(void) { }
1443 static inline int perf_event_release_kernel(struct perf_event
*event
) { return 0; }
1446 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1447 extern void perf_restore_debug_store(void);
1449 static inline void perf_restore_debug_store(void) { }
1452 static __always_inline
bool perf_raw_frag_last(const struct perf_raw_frag
*frag
)
1454 return frag
->pad
< sizeof(u64
);
1457 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1459 struct perf_pmu_events_attr
{
1460 struct device_attribute attr
;
1462 const char *event_str
;
1465 struct perf_pmu_events_ht_attr
{
1466 struct device_attribute attr
;
1468 const char *event_str_ht
;
1469 const char *event_str_noht
;
1472 ssize_t
perf_event_sysfs_show(struct device
*dev
, struct device_attribute
*attr
,
1475 #define PMU_EVENT_ATTR(_name, _var, _id, _show) \
1476 static struct perf_pmu_events_attr _var = { \
1477 .attr = __ATTR(_name, 0444, _show, NULL), \
1481 #define PMU_EVENT_ATTR_STRING(_name, _var, _str) \
1482 static struct perf_pmu_events_attr _var = { \
1483 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1485 .event_str = _str, \
1488 #define PMU_FORMAT_ATTR(_name, _format) \
1490 _name##_show(struct device *dev, \
1491 struct device_attribute *attr, \
1494 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
1495 return sprintf(page, _format "\n"); \
1498 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1500 /* Performance counter hotplug functions */
1501 #ifdef CONFIG_PERF_EVENTS
1502 int perf_event_init_cpu(unsigned int cpu
);
1503 int perf_event_exit_cpu(unsigned int cpu
);
1505 #define perf_event_init_cpu NULL
1506 #define perf_event_exit_cpu NULL
1509 #endif /* _LINUX_PERF_EVENT_H */