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 #define PERF_GUEST_ACTIVE 0x01
30 #define PERF_GUEST_USER 0x02
32 struct perf_guest_info_callbacks {
33 unsigned int (*state)(void);
34 unsigned long (*get_ip)(void);
35 unsigned int (*handle_intel_pt_intr)(void);
38 #ifdef CONFIG_HAVE_HW_BREAKPOINT
39 #include <linux/rhashtable-types.h>
40 #include <asm/hw_breakpoint.h>
43 #include <linux/list.h>
44 #include <linux/mutex.h>
45 #include <linux/rculist.h>
46 #include <linux/rcupdate.h>
47 #include <linux/spinlock.h>
48 #include <linux/hrtimer.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/workqueue.h>
52 #include <linux/ftrace.h>
53 #include <linux/cpu.h>
54 #include <linux/irq_work.h>
55 #include <linux/static_key.h>
56 #include <linux/jump_label_ratelimit.h>
57 #include <linux/atomic.h>
58 #include <linux/sysfs.h>
59 #include <linux/perf_regs.h>
60 #include <linux/cgroup.h>
61 #include <linux/refcount.h>
62 #include <linux/security.h>
63 #include <linux/static_call.h>
64 #include <linux/lockdep.h>
65 #include <asm/local.h>
67 struct perf_callchain_entry {
69 __u64 ip[]; /* /proc/sys/kernel/perf_event_max_stack */
72 struct perf_callchain_entry_ctx {
73 struct perf_callchain_entry *entry;
80 typedef unsigned long (*perf_copy_f)(void *dst, const void *src,
81 unsigned long off, unsigned long len);
83 struct perf_raw_frag {
85 struct perf_raw_frag *next;
93 struct perf_raw_record {
94 struct perf_raw_frag frag;
98 static __always_inline bool perf_raw_frag_last(const struct perf_raw_frag *frag)
100 return frag->pad < sizeof(u64);
104 * branch stack layout:
105 * nr: number of taken branches stored in entries[]
106 * hw_idx: The low level index of raw branch records
107 * for the most recent branch.
108 * -1ULL means invalid/unknown.
110 * Note that nr can vary from sample to sample
111 * branches (to, from) are stored from most recent
112 * to least recent, i.e., entries[0] contains the most
114 * The entries[] is an abstraction of raw branch records,
115 * which may not be stored in age order in HW, e.g. Intel LBR.
116 * The hw_idx is to expose the low level index of raw
117 * branch record for the most recent branch aka entries[0].
118 * The hw_idx index is between -1 (unknown) and max depth,
119 * which can be retrieved in /sys/devices/cpu/caps/branches.
120 * For the architectures whose raw branch records are
121 * already stored in age order, the hw_idx should be 0.
123 struct perf_branch_stack {
126 struct perf_branch_entry entries[];
132 * extra PMU register associated with an event
134 struct hw_perf_event_extra {
135 u64 config; /* register value */
136 unsigned int reg; /* register address or index */
137 int alloc; /* extra register already allocated */
138 int idx; /* index in shared_regs->regs[] */
142 * hw_perf_event::flag values
144 * PERF_EVENT_FLAG_ARCH bits are reserved for architecture-specific
147 #define PERF_EVENT_FLAG_ARCH 0x000fffff
148 #define PERF_EVENT_FLAG_USER_READ_CNT 0x80000000
150 static_assert((PERF_EVENT_FLAG_USER_READ_CNT & PERF_EVENT_FLAG_ARCH) == 0);
153 * struct hw_perf_event - performance event hardware details:
155 struct hw_perf_event {
156 #ifdef CONFIG_PERF_EVENTS
158 struct { /* hardware */
161 unsigned long config_base;
162 unsigned long event_base;
163 int event_base_rdpmc;
168 struct hw_perf_event_extra extra_reg;
169 struct hw_perf_event_extra branch_reg;
171 struct { /* software */
172 struct hrtimer hrtimer;
174 struct { /* tracepoint */
175 /* for tp_event->class */
176 struct list_head tp_list;
178 struct { /* amd_power */
182 #ifdef CONFIG_HAVE_HW_BREAKPOINT
183 struct { /* breakpoint */
185 * Crufty hack to avoid the chicken and egg
186 * problem hw_breakpoint has with context
187 * creation and event initalization.
189 struct arch_hw_breakpoint info;
190 struct rhlist_head bp_list;
193 struct { /* amd_iommu */
202 * If the event is a per task event, this will point to the task in
203 * question. See the comment in perf_event_alloc().
205 struct task_struct *target;
208 * PMU would store hardware filter configuration
213 /* Last sync'ed generation of filters */
214 unsigned long addr_filters_gen;
217 * hw_perf_event::state flags; used to track the PERF_EF_* state.
219 #define PERF_HES_STOPPED 0x01 /* the counter is stopped */
220 #define PERF_HES_UPTODATE 0x02 /* event->count up-to-date */
221 #define PERF_HES_ARCH 0x04
226 * The last observed hardware counter value, updated with a
227 * local64_cmpxchg() such that pmu::read() can be called nested.
229 local64_t prev_count;
232 * The period to start the next sample with.
237 struct { /* Sampling */
239 * The period we started this sample with.
244 * However much is left of the current period;
245 * note that this is a full 64bit value and
246 * allows for generation of periods longer
247 * than hardware might allow.
249 local64_t period_left;
251 struct { /* Topdown events counting for context switch */
258 * State for throttling the event, see __perf_event_overflow() and
259 * perf_adjust_freq_unthr_context().
265 * State for freq target events, see __perf_event_overflow() and
266 * perf_adjust_freq_unthr_context().
269 u64 freq_count_stamp;
274 struct perf_event_pmu_context;
277 * Common implementation detail of pmu::{start,commit,cancel}_txn
279 #define PERF_PMU_TXN_ADD 0x1 /* txn to add/schedule event on PMU */
280 #define PERF_PMU_TXN_READ 0x2 /* txn to read event group from PMU */
283 * pmu::capabilities flags
285 #define PERF_PMU_CAP_NO_INTERRUPT 0x0001
286 #define PERF_PMU_CAP_NO_NMI 0x0002
287 #define PERF_PMU_CAP_AUX_NO_SG 0x0004
288 #define PERF_PMU_CAP_EXTENDED_REGS 0x0008
289 #define PERF_PMU_CAP_EXCLUSIVE 0x0010
290 #define PERF_PMU_CAP_ITRACE 0x0020
291 #define PERF_PMU_CAP_NO_EXCLUDE 0x0040
292 #define PERF_PMU_CAP_AUX_OUTPUT 0x0080
293 #define PERF_PMU_CAP_EXTENDED_HW_TYPE 0x0100
295 struct perf_output_handle;
297 #define PMU_NULL_DEV ((void *)(~0UL))
300 * struct pmu - generic performance monitoring unit
303 struct list_head entry;
305 struct module *module;
307 struct device *parent;
308 const struct attribute_group **attr_groups;
309 const struct attribute_group **attr_update;
314 * various common per-pmu feature flags
318 int __percpu *pmu_disable_count;
319 struct perf_cpu_pmu_context __percpu *cpu_pmu_context;
320 atomic_t exclusive_cnt; /* < 0: cpu; > 0: tsk */
322 int hrtimer_interval_ms;
324 /* number of address filters this PMU can do */
325 unsigned int nr_addr_filters;
328 * Fully disable/enable this PMU, can be used to protect from the PMI
329 * as well as for lazy/batch writing of the MSRs.
331 void (*pmu_enable) (struct pmu *pmu); /* optional */
332 void (*pmu_disable) (struct pmu *pmu); /* optional */
335 * Try and initialize the event for this PMU.
338 * -ENOENT -- @event is not for this PMU
340 * -ENODEV -- @event is for this PMU but PMU not present
341 * -EBUSY -- @event is for this PMU but PMU temporarily unavailable
342 * -EINVAL -- @event is for this PMU but @event is not valid
343 * -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
344 * -EACCES -- @event is for this PMU, @event is valid, but no privileges
346 * 0 -- @event is for this PMU and valid
348 * Other error return values are allowed.
350 int (*event_init) (struct perf_event *event);
353 * Notification that the event was mapped or unmapped. Called
354 * in the context of the mapping task.
356 void (*event_mapped) (struct perf_event *event, struct mm_struct *mm); /* optional */
357 void (*event_unmapped) (struct perf_event *event, struct mm_struct *mm); /* optional */
360 * Flags for ->add()/->del()/ ->start()/->stop(). There are
361 * matching hw_perf_event::state flags.
363 #define PERF_EF_START 0x01 /* start the counter when adding */
364 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
365 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
368 * Adds/Removes a counter to/from the PMU, can be done inside a
369 * transaction, see the ->*_txn() methods.
371 * The add/del callbacks will reserve all hardware resources required
372 * to service the event, this includes any counter constraint
375 * Called with IRQs disabled and the PMU disabled on the CPU the event
378 * ->add() called without PERF_EF_START should result in the same state
379 * as ->add() followed by ->stop().
381 * ->del() must always PERF_EF_UPDATE stop an event. If it calls
382 * ->stop() that must deal with already being stopped without
385 int (*add) (struct perf_event *event, int flags);
386 void (*del) (struct perf_event *event, int flags);
389 * Starts/Stops a counter present on the PMU.
391 * The PMI handler should stop the counter when perf_event_overflow()
392 * returns !0. ->start() will be used to continue.
394 * Also used to change the sample period.
396 * Called with IRQs disabled and the PMU disabled on the CPU the event
397 * is on -- will be called from NMI context with the PMU generates
400 * ->stop() with PERF_EF_UPDATE will read the counter and update
401 * period/count values like ->read() would.
403 * ->start() with PERF_EF_RELOAD will reprogram the counter
404 * value, must be preceded by a ->stop() with PERF_EF_UPDATE.
406 void (*start) (struct perf_event *event, int flags);
407 void (*stop) (struct perf_event *event, int flags);
410 * Updates the counter value of the event.
412 * For sampling capable PMUs this will also update the software period
413 * hw_perf_event::period_left field.
415 void (*read) (struct perf_event *event);
418 * Group events scheduling is treated as a transaction, add
419 * group events as a whole and perform one schedulability test.
420 * If the test fails, roll back the whole group
422 * Start the transaction, after this ->add() doesn't need to
423 * do schedulability tests.
427 void (*start_txn) (struct pmu *pmu, unsigned int txn_flags);
429 * If ->start_txn() disabled the ->add() schedulability test
430 * then ->commit_txn() is required to perform one. On success
431 * the transaction is closed. On error the transaction is kept
432 * open until ->cancel_txn() is called.
436 int (*commit_txn) (struct pmu *pmu);
438 * Will cancel the transaction, assumes ->del() is called
439 * for each successful ->add() during the transaction.
443 void (*cancel_txn) (struct pmu *pmu);
446 * Will return the value for perf_event_mmap_page::index for this event,
447 * if no implementation is provided it will default to 0 (see
448 * perf_event_idx_default).
450 int (*event_idx) (struct perf_event *event); /*optional */
453 * context-switches callback
455 void (*sched_task) (struct perf_event_pmu_context *pmu_ctx,
459 * Kmem cache of PMU specific data
461 struct kmem_cache *task_ctx_cache;
464 * PMU specific parts of task perf event context (i.e. ctx->task_ctx_data)
465 * can be synchronized using this function. See Intel LBR callstack support
466 * implementation and Perf core context switch handling callbacks for usage
469 void (*swap_task_ctx) (struct perf_event_pmu_context *prev_epc,
470 struct perf_event_pmu_context *next_epc);
474 * Set up pmu-private data structures for an AUX area
476 void *(*setup_aux) (struct perf_event *event, void **pages,
477 int nr_pages, bool overwrite);
481 * Free pmu-private AUX data structures
483 void (*free_aux) (void *aux); /* optional */
486 * Take a snapshot of the AUX buffer without touching the event
487 * state, so that preempting ->start()/->stop() callbacks does
488 * not interfere with their logic. Called in PMI context.
490 * Returns the size of AUX data copied to the output handle.
494 long (*snapshot_aux) (struct perf_event *event,
495 struct perf_output_handle *handle,
499 * Validate address range filters: make sure the HW supports the
500 * requested configuration and number of filters; return 0 if the
501 * supplied filters are valid, -errno otherwise.
503 * Runs in the context of the ioctl()ing process and is not serialized
504 * with the rest of the PMU callbacks.
506 int (*addr_filters_validate) (struct list_head *filters);
510 * Synchronize address range filter configuration:
511 * translate hw-agnostic filters into hardware configuration in
512 * event::hw::addr_filters.
514 * Runs as a part of filter sync sequence that is done in ->start()
515 * callback by calling perf_event_addr_filters_sync().
517 * May (and should) traverse event::addr_filters::list, for which its
518 * caller provides necessary serialization.
520 void (*addr_filters_sync) (struct perf_event *event);
524 * Check if event can be used for aux_output purposes for
525 * events of this PMU.
527 * Runs from perf_event_open(). Should return 0 for "no match"
528 * or non-zero for "match".
530 int (*aux_output_match) (struct perf_event *event);
534 * Skip programming this PMU on the given CPU. Typically needed for
537 bool (*filter) (struct pmu *pmu, int cpu); /* optional */
540 * Check period value for PERF_EVENT_IOC_PERIOD ioctl.
542 int (*check_period) (struct perf_event *event, u64 value); /* optional */
545 enum perf_addr_filter_action_t {
546 PERF_ADDR_FILTER_ACTION_STOP = 0,
547 PERF_ADDR_FILTER_ACTION_START,
548 PERF_ADDR_FILTER_ACTION_FILTER,
552 * struct perf_addr_filter - address range filter definition
553 * @entry: event's filter list linkage
554 * @path: object file's path for file-based filters
555 * @offset: filter range offset
556 * @size: filter range size (size==0 means single address trigger)
557 * @action: filter/start/stop
559 * This is a hardware-agnostic filter configuration as specified by the user.
561 struct perf_addr_filter {
562 struct list_head entry;
564 unsigned long offset;
566 enum perf_addr_filter_action_t action;
570 * struct perf_addr_filters_head - container for address range filters
571 * @list: list of filters for this event
572 * @lock: spinlock that serializes accesses to the @list and event's
573 * (and its children's) filter generations.
574 * @nr_file_filters: number of file-based filters
576 * A child event will use parent's @list (and therefore @lock), so they are
577 * bundled together; see perf_event_addr_filters().
579 struct perf_addr_filters_head {
580 struct list_head list;
582 unsigned int nr_file_filters;
585 struct perf_addr_filter_range {
591 * enum perf_event_state - the states of an event:
593 enum perf_event_state {
594 PERF_EVENT_STATE_DEAD = -4,
595 PERF_EVENT_STATE_EXIT = -3,
596 PERF_EVENT_STATE_ERROR = -2,
597 PERF_EVENT_STATE_OFF = -1,
598 PERF_EVENT_STATE_INACTIVE = 0,
599 PERF_EVENT_STATE_ACTIVE = 1,
603 struct perf_sample_data;
605 typedef void (*perf_overflow_handler_t)(struct perf_event *,
606 struct perf_sample_data *,
607 struct pt_regs *regs);
610 * Event capabilities. For event_caps and groups caps.
612 * PERF_EV_CAP_SOFTWARE: Is a software event.
613 * PERF_EV_CAP_READ_ACTIVE_PKG: A CPU event (or cgroup event) that can be read
614 * from any CPU in the package where it is active.
615 * PERF_EV_CAP_SIBLING: An event with this flag must be a group sibling and
616 * cannot be a group leader. If an event with this flag is detached from the
617 * group it is scheduled out and moved into an unrecoverable ERROR state.
619 #define PERF_EV_CAP_SOFTWARE BIT(0)
620 #define PERF_EV_CAP_READ_ACTIVE_PKG BIT(1)
621 #define PERF_EV_CAP_SIBLING BIT(2)
623 #define SWEVENT_HLIST_BITS 8
624 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
626 struct swevent_hlist {
627 struct hlist_head heads[SWEVENT_HLIST_SIZE];
628 struct rcu_head rcu_head;
631 #define PERF_ATTACH_CONTEXT 0x01
632 #define PERF_ATTACH_GROUP 0x02
633 #define PERF_ATTACH_TASK 0x04
634 #define PERF_ATTACH_TASK_DATA 0x08
635 #define PERF_ATTACH_ITRACE 0x10
636 #define PERF_ATTACH_SCHED_CB 0x20
637 #define PERF_ATTACH_CHILD 0x40
643 struct pmu_event_list {
645 struct list_head list;
649 * event->sibling_list is modified whole holding both ctx->lock and ctx->mutex
650 * as such iteration must hold either lock. However, since ctx->lock is an IRQ
651 * safe lock, and is only held by the CPU doing the modification, having IRQs
652 * disabled is sufficient since it will hold-off the IPIs.
654 #ifdef CONFIG_PROVE_LOCKING
655 #define lockdep_assert_event_ctx(event) \
656 WARN_ON_ONCE(__lockdep_enabled && \
657 (this_cpu_read(hardirqs_enabled) && \
658 lockdep_is_held(&(event)->ctx->mutex) != LOCK_STATE_HELD))
660 #define lockdep_assert_event_ctx(event)
663 #define for_each_sibling_event(sibling, event) \
664 lockdep_assert_event_ctx(event); \
665 if ((event)->group_leader == (event)) \
666 list_for_each_entry((sibling), &(event)->sibling_list, sibling_list)
669 * struct perf_event - performance event kernel representation:
672 #ifdef CONFIG_PERF_EVENTS
674 * entry onto perf_event_context::event_list;
675 * modifications require ctx->lock
676 * RCU safe iterations.
678 struct list_head event_entry;
681 * Locked for modification by both ctx->mutex and ctx->lock; holding
682 * either sufficies for read.
684 struct list_head sibling_list;
685 struct list_head active_list;
687 * Node on the pinned or flexible tree located at the event context;
689 struct rb_node group_node;
692 * We need storage to track the entries in perf_pmu_migrate_context; we
693 * cannot use the event_entry because of RCU and we want to keep the
694 * group in tact which avoids us using the other two entries.
696 struct list_head migrate_entry;
698 struct hlist_node hlist_entry;
699 struct list_head active_entry;
702 /* Not serialized. Only written during event initialization. */
704 /* The cumulative AND of all event_caps for events in this group. */
707 unsigned int group_generation;
708 struct perf_event *group_leader;
710 * event->pmu will always point to pmu in which this event belongs.
711 * Whereas event->pmu_ctx->pmu may point to other pmu when group of
712 * different pmu events is created.
717 enum perf_event_state state;
718 unsigned int attach_state;
720 atomic64_t child_count;
723 * These are the total time in nanoseconds that the event
724 * has been enabled (i.e. eligible to run, and the task has
725 * been scheduled in, if this is a per-task event)
726 * and running (scheduled onto the CPU), respectively.
728 u64 total_time_enabled;
729 u64 total_time_running;
732 struct perf_event_attr attr;
736 struct hw_perf_event hw;
738 struct perf_event_context *ctx;
740 * event->pmu_ctx points to perf_event_pmu_context in which the event
741 * is added. This pmu_ctx can be of other pmu for sw event when that
742 * sw event is part of a group which also contains non-sw events.
744 struct perf_event_pmu_context *pmu_ctx;
745 atomic_long_t refcount;
748 * These accumulate total time (in nanoseconds) that children
749 * events have been enabled and running, respectively.
751 atomic64_t child_total_time_enabled;
752 atomic64_t child_total_time_running;
755 * Protect attach/detach and child_list:
757 struct mutex child_mutex;
758 struct list_head child_list;
759 struct perf_event *parent;
764 struct list_head owner_entry;
765 struct task_struct *owner;
768 struct mutex mmap_mutex;
771 struct perf_buffer *rb;
772 struct list_head rb_entry;
773 unsigned long rcu_batches;
777 wait_queue_head_t waitq;
778 struct fasync_struct *fasync;
780 /* delayed work for NMIs and such */
781 unsigned int pending_wakeup;
782 unsigned int pending_kill;
783 unsigned int pending_disable;
784 unsigned int pending_sigtrap;
785 unsigned long pending_addr; /* SIGTRAP */
786 struct irq_work pending_irq;
787 struct callback_head pending_task;
788 unsigned int pending_work;
790 atomic_t event_limit;
792 /* address range filters */
793 struct perf_addr_filters_head addr_filters;
794 /* vma address array for file-based filders */
795 struct perf_addr_filter_range *addr_filter_ranges;
796 unsigned long addr_filters_gen;
798 /* for aux_output events */
799 struct perf_event *aux_event;
801 void (*destroy)(struct perf_event *);
802 struct rcu_head rcu_head;
804 struct pid_namespace *ns;
807 atomic64_t lost_samples;
810 perf_overflow_handler_t overflow_handler;
811 void *overflow_handler_context;
812 #ifdef CONFIG_BPF_SYSCALL
813 perf_overflow_handler_t orig_overflow_handler;
814 struct bpf_prog *prog;
818 #ifdef CONFIG_EVENT_TRACING
819 struct trace_event_call *tp_event;
820 struct event_filter *filter;
821 #ifdef CONFIG_FUNCTION_TRACER
822 struct ftrace_ops ftrace_ops;
826 #ifdef CONFIG_CGROUP_PERF
827 struct perf_cgroup *cgrp; /* cgroup event is attach to */
830 #ifdef CONFIG_SECURITY
833 struct list_head sb_list;
836 * Certain events gets forwarded to another pmu internally by over-
837 * writing kernel copy of event->attr.type without user being aware
838 * of it. event->orig_type contains original 'type' requested by
842 #endif /* CONFIG_PERF_EVENTS */
846 * ,-----------------------[1:n]----------------------.
848 * perf_event_context <-[1:n]-> perf_event_pmu_context <--- perf_event
850 * `--------[1:n]---------' `-[n:1]-> pmu <-[1:n]-'
853 * struct perf_event_pmu_context lifetime is refcount based and RCU freed
854 * (similar to perf_event_context). Locking is as if it were a member of
855 * perf_event_context; specifically:
857 * modification, both: ctx->mutex && ctx->lock
858 * reading, either: ctx->mutex || ctx->lock
860 * There is one exception to this; namely put_pmu_ctx() isn't always called
861 * with ctx->mutex held; this means that as long as we can guarantee the epc
862 * has events the above rules hold.
864 * Specificially, sys_perf_event_open()'s group_leader case depends on
865 * ctx->mutex pinning the configuration. Since we hold a reference on
866 * group_leader (through the filedesc) it can't go away, therefore it's
867 * associated pmu_ctx must exist and cannot change due to ctx->mutex.
869 struct perf_event_pmu_context {
871 struct perf_event_context *ctx;
873 struct list_head pmu_ctx_entry;
875 struct list_head pinned_active;
876 struct list_head flexible_active;
878 /* Used to avoid freeing per-cpu perf_event_pmu_context */
879 unsigned int embedded : 1;
881 unsigned int nr_events;
883 atomic_t refcount; /* event <-> epc */
884 struct rcu_head rcu_head;
886 void *task_ctx_data; /* pmu specific data */
888 * Set when one or more (plausibly active) event can't be scheduled
889 * due to pmu overcommit or pmu constraints, except tolerant to
890 * events not necessary to be active due to scheduling constraints,
893 int rotate_necessary;
896 struct perf_event_groups {
903 * struct perf_event_context - event context structure
905 * Used as a container for task events and CPU events as well:
907 struct perf_event_context {
909 * Protect the states of the events in the list,
910 * nr_active, and the list:
914 * Protect the list of events. Locking either mutex or lock
915 * is sufficient to ensure the list doesn't change; to change
916 * the list you need to lock both the mutex and the spinlock.
920 struct list_head pmu_ctx_list;
921 struct perf_event_groups pinned_groups;
922 struct perf_event_groups flexible_groups;
923 struct list_head event_list;
934 refcount_t refcount; /* event <-> ctx */
935 struct task_struct *task;
938 * Context clock, runs when context enabled.
945 * These fields let us detect when two contexts have both
946 * been cloned (inherited) from a common ancestor.
948 struct perf_event_context *parent_ctx;
952 #ifdef CONFIG_CGROUP_PERF
953 int nr_cgroups; /* cgroup evts */
955 struct rcu_head rcu_head;
958 * Sum (event->pending_sigtrap + event->pending_work)
960 * The SIGTRAP is targeted at ctx->task, as such it won't do changing
961 * that until the signal is delivered.
967 * Number of contexts where an event can trigger:
968 * task, softirq, hardirq, nmi.
970 #define PERF_NR_CONTEXTS 4
972 struct perf_cpu_pmu_context {
973 struct perf_event_pmu_context epc;
974 struct perf_event_pmu_context *task_epc;
976 struct list_head sched_cb_entry;
982 raw_spinlock_t hrtimer_lock;
983 struct hrtimer hrtimer;
984 ktime_t hrtimer_interval;
985 unsigned int hrtimer_active;
989 * struct perf_event_cpu_context - per cpu event context structure
991 struct perf_cpu_context {
992 struct perf_event_context ctx;
993 struct perf_event_context *task_ctx;
996 #ifdef CONFIG_CGROUP_PERF
997 struct perf_cgroup *cgrp;
1001 * Per-CPU storage for iterators used in visit_groups_merge. The default
1002 * storage is of size 2 to hold the CPU and any CPU event iterators.
1005 struct perf_event **heap;
1006 struct perf_event *heap_default[2];
1009 struct perf_output_handle {
1010 struct perf_event *event;
1011 struct perf_buffer *rb;
1012 unsigned long wakeup;
1022 struct bpf_perf_event_data_kern {
1023 bpf_user_pt_regs_t *regs;
1024 struct perf_sample_data *data;
1025 struct perf_event *event;
1028 #ifdef CONFIG_CGROUP_PERF
1031 * perf_cgroup_info keeps track of time_enabled for a cgroup.
1032 * This is a per-cpu dynamically allocated data structure.
1034 struct perf_cgroup_info {
1041 struct perf_cgroup {
1042 struct cgroup_subsys_state css;
1043 struct perf_cgroup_info __percpu *info;
1047 * Must ensure cgroup is pinned (css_get) before calling
1048 * this function. In other words, we cannot call this function
1049 * if there is no cgroup event for the current CPU context.
1051 static inline struct perf_cgroup *
1052 perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
1054 return container_of(task_css_check(task, perf_event_cgrp_id,
1055 ctx ? lockdep_is_held(&ctx->lock)
1057 struct perf_cgroup, css);
1059 #endif /* CONFIG_CGROUP_PERF */
1061 #ifdef CONFIG_PERF_EVENTS
1063 extern struct perf_event_context *perf_cpu_task_ctx(void);
1065 extern void *perf_aux_output_begin(struct perf_output_handle *handle,
1066 struct perf_event *event);
1067 extern void perf_aux_output_end(struct perf_output_handle *handle,
1068 unsigned long size);
1069 extern int perf_aux_output_skip(struct perf_output_handle *handle,
1070 unsigned long size);
1071 extern void *perf_get_aux(struct perf_output_handle *handle);
1072 extern void perf_aux_output_flag(struct perf_output_handle *handle, u64 flags);
1073 extern void perf_event_itrace_started(struct perf_event *event);
1075 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
1076 extern void perf_pmu_unregister(struct pmu *pmu);
1078 extern void __perf_event_task_sched_in(struct task_struct *prev,
1079 struct task_struct *task);
1080 extern void __perf_event_task_sched_out(struct task_struct *prev,
1081 struct task_struct *next);
1082 extern int perf_event_init_task(struct task_struct *child, u64 clone_flags);
1083 extern void perf_event_exit_task(struct task_struct *child);
1084 extern void perf_event_free_task(struct task_struct *task);
1085 extern void perf_event_delayed_put(struct task_struct *task);
1086 extern struct file *perf_event_get(unsigned int fd);
1087 extern const struct perf_event *perf_get_event(struct file *file);
1088 extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
1089 extern void perf_event_print_debug(void);
1090 extern void perf_pmu_disable(struct pmu *pmu);
1091 extern void perf_pmu_enable(struct pmu *pmu);
1092 extern void perf_sched_cb_dec(struct pmu *pmu);
1093 extern void perf_sched_cb_inc(struct pmu *pmu);
1094 extern int perf_event_task_disable(void);
1095 extern int perf_event_task_enable(void);
1097 extern void perf_pmu_resched(struct pmu *pmu);
1099 extern int perf_event_refresh(struct perf_event *event, int refresh);
1100 extern void perf_event_update_userpage(struct perf_event *event);
1101 extern int perf_event_release_kernel(struct perf_event *event);
1102 extern struct perf_event *
1103 perf_event_create_kernel_counter(struct perf_event_attr *attr,
1105 struct task_struct *task,
1106 perf_overflow_handler_t callback,
1108 extern void perf_pmu_migrate_context(struct pmu *pmu,
1109 int src_cpu, int dst_cpu);
1110 int perf_event_read_local(struct perf_event *event, u64 *value,
1111 u64 *enabled, u64 *running);
1112 extern u64 perf_event_read_value(struct perf_event *event,
1113 u64 *enabled, u64 *running);
1115 extern struct perf_callchain_entry *perf_callchain(struct perf_event *event, struct pt_regs *regs);
1117 static inline bool branch_sample_no_flags(const struct perf_event *event)
1119 return event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_NO_FLAGS;
1122 static inline bool branch_sample_no_cycles(const struct perf_event *event)
1124 return event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_NO_CYCLES;
1127 static inline bool branch_sample_type(const struct perf_event *event)
1129 return event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_TYPE_SAVE;
1132 static inline bool branch_sample_hw_index(const struct perf_event *event)
1134 return event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX;
1137 static inline bool branch_sample_priv(const struct perf_event *event)
1139 return event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_PRIV_SAVE;
1143 struct perf_sample_data {
1145 * Fields set by perf_sample_data_init() unconditionally,
1146 * group so as to minimize the cachelines touched.
1153 * Fields commonly set by __perf_event_header__init_id(),
1154 * group so as to minimize the cachelines touched.
1169 * The other fields, optionally {set,used} by
1170 * perf_{prepare,output}_sample().
1173 struct perf_callchain_entry *callchain;
1174 struct perf_raw_record *raw;
1175 struct perf_branch_stack *br_stack;
1176 union perf_sample_weight weight;
1177 union perf_mem_data_src data_src;
1180 struct perf_regs regs_user;
1181 struct perf_regs regs_intr;
1182 u64 stack_user_size;
1191 } ____cacheline_aligned;
1193 /* default value for data source */
1194 #define PERF_MEM_NA (PERF_MEM_S(OP, NA) |\
1195 PERF_MEM_S(LVL, NA) |\
1196 PERF_MEM_S(SNOOP, NA) |\
1197 PERF_MEM_S(LOCK, NA) |\
1198 PERF_MEM_S(TLB, NA) |\
1199 PERF_MEM_S(LVLNUM, NA))
1201 static inline void perf_sample_data_init(struct perf_sample_data *data,
1202 u64 addr, u64 period)
1204 /* remaining struct members initialized in perf_prepare_sample() */
1205 data->sample_flags = PERF_SAMPLE_PERIOD;
1206 data->period = period;
1211 data->sample_flags |= PERF_SAMPLE_ADDR;
1215 static inline void perf_sample_save_callchain(struct perf_sample_data *data,
1216 struct perf_event *event,
1217 struct pt_regs *regs)
1221 data->callchain = perf_callchain(event, regs);
1222 size += data->callchain->nr;
1224 data->dyn_size += size * sizeof(u64);
1225 data->sample_flags |= PERF_SAMPLE_CALLCHAIN;
1228 static inline void perf_sample_save_raw_data(struct perf_sample_data *data,
1229 struct perf_raw_record *raw)
1231 struct perf_raw_frag *frag = &raw->frag;
1237 if (perf_raw_frag_last(frag))
1242 size = round_up(sum + sizeof(u32), sizeof(u64));
1243 raw->size = size - sizeof(u32);
1244 frag->pad = raw->size - sum;
1247 data->dyn_size += size;
1248 data->sample_flags |= PERF_SAMPLE_RAW;
1251 static inline void perf_sample_save_brstack(struct perf_sample_data *data,
1252 struct perf_event *event,
1253 struct perf_branch_stack *brs)
1255 int size = sizeof(u64); /* nr */
1257 if (branch_sample_hw_index(event))
1258 size += sizeof(u64);
1259 size += brs->nr * sizeof(struct perf_branch_entry);
1261 data->br_stack = brs;
1262 data->dyn_size += size;
1263 data->sample_flags |= PERF_SAMPLE_BRANCH_STACK;
1266 static inline u32 perf_sample_data_size(struct perf_sample_data *data,
1267 struct perf_event *event)
1269 u32 size = sizeof(struct perf_event_header);
1271 size += event->header_size + event->id_header_size;
1272 size += data->dyn_size;
1278 * Clear all bitfields in the perf_branch_entry.
1279 * The to and from fields are not cleared because they are
1280 * systematically modified by caller.
1282 static inline void perf_clear_branch_entry_bitfields(struct perf_branch_entry *br)
1290 br->spec = PERF_BR_SPEC_NA;
1294 extern void perf_output_sample(struct perf_output_handle *handle,
1295 struct perf_event_header *header,
1296 struct perf_sample_data *data,
1297 struct perf_event *event);
1298 extern void perf_prepare_sample(struct perf_sample_data *data,
1299 struct perf_event *event,
1300 struct pt_regs *regs);
1301 extern void perf_prepare_header(struct perf_event_header *header,
1302 struct perf_sample_data *data,
1303 struct perf_event *event,
1304 struct pt_regs *regs);
1306 extern int perf_event_overflow(struct perf_event *event,
1307 struct perf_sample_data *data,
1308 struct pt_regs *regs);
1310 extern void perf_event_output_forward(struct perf_event *event,
1311 struct perf_sample_data *data,
1312 struct pt_regs *regs);
1313 extern void perf_event_output_backward(struct perf_event *event,
1314 struct perf_sample_data *data,
1315 struct pt_regs *regs);
1316 extern int perf_event_output(struct perf_event *event,
1317 struct perf_sample_data *data,
1318 struct pt_regs *regs);
1321 __is_default_overflow_handler(perf_overflow_handler_t overflow_handler)
1323 if (likely(overflow_handler == perf_event_output_forward))
1325 if (unlikely(overflow_handler == perf_event_output_backward))
1330 #define is_default_overflow_handler(event) \
1331 __is_default_overflow_handler((event)->overflow_handler)
1333 #ifdef CONFIG_BPF_SYSCALL
1334 static inline bool uses_default_overflow_handler(struct perf_event *event)
1336 if (likely(is_default_overflow_handler(event)))
1339 return __is_default_overflow_handler(event->orig_overflow_handler);
1342 #define uses_default_overflow_handler(event) \
1343 is_default_overflow_handler(event)
1347 perf_event_header__init_id(struct perf_event_header *header,
1348 struct perf_sample_data *data,
1349 struct perf_event *event);
1351 perf_event__output_id_sample(struct perf_event *event,
1352 struct perf_output_handle *handle,
1353 struct perf_sample_data *sample);
1356 perf_log_lost_samples(struct perf_event *event, u64 lost);
1358 static inline bool event_has_any_exclude_flag(struct perf_event *event)
1360 struct perf_event_attr *attr = &event->attr;
1362 return attr->exclude_idle || attr->exclude_user ||
1363 attr->exclude_kernel || attr->exclude_hv ||
1364 attr->exclude_guest || attr->exclude_host;
1367 static inline bool is_sampling_event(struct perf_event *event)
1369 return event->attr.sample_period != 0;
1373 * Return 1 for a software event, 0 for a hardware event
1375 static inline int is_software_event(struct perf_event *event)
1377 return event->event_caps & PERF_EV_CAP_SOFTWARE;
1381 * Return 1 for event in sw context, 0 for event in hw context
1383 static inline int in_software_context(struct perf_event *event)
1385 return event->pmu_ctx->pmu->task_ctx_nr == perf_sw_context;
1388 static inline int is_exclusive_pmu(struct pmu *pmu)
1390 return pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE;
1393 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
1395 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
1396 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
1398 #ifndef perf_arch_fetch_caller_regs
1399 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
1403 * When generating a perf sample in-line, instead of from an interrupt /
1404 * exception, we lack a pt_regs. This is typically used from software events
1405 * like: SW_CONTEXT_SWITCHES, SW_MIGRATIONS and the tie-in with tracepoints.
1407 * We typically don't need a full set, but (for x86) do require:
1408 * - ip for PERF_SAMPLE_IP
1409 * - cs for user_mode() tests
1410 * - sp for PERF_SAMPLE_CALLCHAIN
1411 * - eflags for MISC bits and CALLCHAIN (see: perf_hw_regs())
1413 * NOTE: assumes @regs is otherwise already 0 filled; this is important for
1414 * things like PERF_SAMPLE_REGS_INTR.
1416 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1418 perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1421 static __always_inline void
1422 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1424 if (static_key_false(&perf_swevent_enabled[event_id]))
1425 __perf_sw_event(event_id, nr, regs, addr);
1428 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
1431 * 'Special' version for the scheduler, it hard assumes no recursion,
1432 * which is guaranteed by us not actually scheduling inside other swevents
1433 * because those disable preemption.
1435 static __always_inline void __perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
1437 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1439 perf_fetch_caller_regs(regs);
1440 ___perf_sw_event(event_id, nr, regs, addr);
1443 extern struct static_key_false perf_sched_events;
1445 static __always_inline bool __perf_sw_enabled(int swevt)
1447 return static_key_false(&perf_swevent_enabled[swevt]);
1450 static inline void perf_event_task_migrate(struct task_struct *task)
1452 if (__perf_sw_enabled(PERF_COUNT_SW_CPU_MIGRATIONS))
1453 task->sched_migrated = 1;
1456 static inline void perf_event_task_sched_in(struct task_struct *prev,
1457 struct task_struct *task)
1459 if (static_branch_unlikely(&perf_sched_events))
1460 __perf_event_task_sched_in(prev, task);
1462 if (__perf_sw_enabled(PERF_COUNT_SW_CPU_MIGRATIONS) &&
1463 task->sched_migrated) {
1464 __perf_sw_event_sched(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 0);
1465 task->sched_migrated = 0;
1469 static inline void perf_event_task_sched_out(struct task_struct *prev,
1470 struct task_struct *next)
1472 if (__perf_sw_enabled(PERF_COUNT_SW_CONTEXT_SWITCHES))
1473 __perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
1475 #ifdef CONFIG_CGROUP_PERF
1476 if (__perf_sw_enabled(PERF_COUNT_SW_CGROUP_SWITCHES) &&
1477 perf_cgroup_from_task(prev, NULL) !=
1478 perf_cgroup_from_task(next, NULL))
1479 __perf_sw_event_sched(PERF_COUNT_SW_CGROUP_SWITCHES, 1, 0);
1482 if (static_branch_unlikely(&perf_sched_events))
1483 __perf_event_task_sched_out(prev, next);
1486 extern void perf_event_mmap(struct vm_area_struct *vma);
1488 extern void perf_event_ksymbol(u16 ksym_type, u64 addr, u32 len,
1489 bool unregister, const char *sym);
1490 extern void perf_event_bpf_event(struct bpf_prog *prog,
1491 enum perf_bpf_event_type type,
1494 #ifdef CONFIG_GUEST_PERF_EVENTS
1495 extern struct perf_guest_info_callbacks __rcu *perf_guest_cbs;
1497 DECLARE_STATIC_CALL(__perf_guest_state, *perf_guest_cbs->state);
1498 DECLARE_STATIC_CALL(__perf_guest_get_ip, *perf_guest_cbs->get_ip);
1499 DECLARE_STATIC_CALL(__perf_guest_handle_intel_pt_intr, *perf_guest_cbs->handle_intel_pt_intr);
1501 static inline unsigned int perf_guest_state(void)
1503 return static_call(__perf_guest_state)();
1505 static inline unsigned long perf_guest_get_ip(void)
1507 return static_call(__perf_guest_get_ip)();
1509 static inline unsigned int perf_guest_handle_intel_pt_intr(void)
1511 return static_call(__perf_guest_handle_intel_pt_intr)();
1513 extern void perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs);
1514 extern void perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs);
1516 static inline unsigned int perf_guest_state(void) { return 0; }
1517 static inline unsigned long perf_guest_get_ip(void) { return 0; }
1518 static inline unsigned int perf_guest_handle_intel_pt_intr(void) { return 0; }
1519 #endif /* CONFIG_GUEST_PERF_EVENTS */
1521 extern void perf_event_exec(void);
1522 extern void perf_event_comm(struct task_struct *tsk, bool exec);
1523 extern void perf_event_namespaces(struct task_struct *tsk);
1524 extern void perf_event_fork(struct task_struct *tsk);
1525 extern void perf_event_text_poke(const void *addr,
1526 const void *old_bytes, size_t old_len,
1527 const void *new_bytes, size_t new_len);
1530 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1532 extern void perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1533 extern void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1534 extern struct perf_callchain_entry *
1535 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
1536 u32 max_stack, bool crosstask, bool add_mark);
1537 extern int get_callchain_buffers(int max_stack);
1538 extern void put_callchain_buffers(void);
1539 extern struct perf_callchain_entry *get_callchain_entry(int *rctx);
1540 extern void put_callchain_entry(int rctx);
1542 extern int sysctl_perf_event_max_stack;
1543 extern int sysctl_perf_event_max_contexts_per_stack;
1545 static inline int perf_callchain_store_context(struct perf_callchain_entry_ctx *ctx, u64 ip)
1547 if (ctx->contexts < sysctl_perf_event_max_contexts_per_stack) {
1548 struct perf_callchain_entry *entry = ctx->entry;
1549 entry->ip[entry->nr++] = ip;
1553 ctx->contexts_maxed = true;
1554 return -1; /* no more room, stop walking the stack */
1558 static inline int perf_callchain_store(struct perf_callchain_entry_ctx *ctx, u64 ip)
1560 if (ctx->nr < ctx->max_stack && !ctx->contexts_maxed) {
1561 struct perf_callchain_entry *entry = ctx->entry;
1562 entry->ip[entry->nr++] = ip;
1566 return -1; /* no more room, stop walking the stack */
1570 extern int sysctl_perf_event_paranoid;
1571 extern int sysctl_perf_event_mlock;
1572 extern int sysctl_perf_event_sample_rate;
1573 extern int sysctl_perf_cpu_time_max_percent;
1575 extern void perf_sample_event_took(u64 sample_len_ns);
1577 int perf_proc_update_handler(struct ctl_table *table, int write,
1578 void *buffer, size_t *lenp, loff_t *ppos);
1579 int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
1580 void *buffer, size_t *lenp, loff_t *ppos);
1581 int perf_event_max_stack_handler(struct ctl_table *table, int write,
1582 void *buffer, size_t *lenp, loff_t *ppos);
1584 /* Access to perf_event_open(2) syscall. */
1585 #define PERF_SECURITY_OPEN 0
1587 /* Finer grained perf_event_open(2) access control. */
1588 #define PERF_SECURITY_CPU 1
1589 #define PERF_SECURITY_KERNEL 2
1590 #define PERF_SECURITY_TRACEPOINT 3
1592 static inline int perf_is_paranoid(void)
1594 return sysctl_perf_event_paranoid > -1;
1597 static inline int perf_allow_kernel(struct perf_event_attr *attr)
1599 if (sysctl_perf_event_paranoid > 1 && !perfmon_capable())
1602 return security_perf_event_open(attr, PERF_SECURITY_KERNEL);
1605 static inline int perf_allow_cpu(struct perf_event_attr *attr)
1607 if (sysctl_perf_event_paranoid > 0 && !perfmon_capable())
1610 return security_perf_event_open(attr, PERF_SECURITY_CPU);
1613 static inline int perf_allow_tracepoint(struct perf_event_attr *attr)
1615 if (sysctl_perf_event_paranoid > -1 && !perfmon_capable())
1618 return security_perf_event_open(attr, PERF_SECURITY_TRACEPOINT);
1621 extern void perf_event_init(void);
1622 extern void perf_tp_event(u16 event_type, u64 count, void *record,
1623 int entry_size, struct pt_regs *regs,
1624 struct hlist_head *head, int rctx,
1625 struct task_struct *task);
1626 extern void perf_bp_event(struct perf_event *event, void *data);
1628 #ifndef perf_misc_flags
1629 # define perf_misc_flags(regs) \
1630 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1631 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1633 #ifndef perf_arch_bpf_user_pt_regs
1634 # define perf_arch_bpf_user_pt_regs(regs) regs
1637 static inline bool has_branch_stack(struct perf_event *event)
1639 return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1642 static inline bool needs_branch_stack(struct perf_event *event)
1644 return event->attr.branch_sample_type != 0;
1647 static inline bool has_aux(struct perf_event *event)
1649 return event->pmu->setup_aux;
1652 static inline bool is_write_backward(struct perf_event *event)
1654 return !!event->attr.write_backward;
1657 static inline bool has_addr_filter(struct perf_event *event)
1659 return event->pmu->nr_addr_filters;
1663 * An inherited event uses parent's filters
1665 static inline struct perf_addr_filters_head *
1666 perf_event_addr_filters(struct perf_event *event)
1668 struct perf_addr_filters_head *ifh = &event->addr_filters;
1671 ifh = &event->parent->addr_filters;
1676 extern void perf_event_addr_filters_sync(struct perf_event *event);
1677 extern void perf_report_aux_output_id(struct perf_event *event, u64 hw_id);
1679 extern int perf_output_begin(struct perf_output_handle *handle,
1680 struct perf_sample_data *data,
1681 struct perf_event *event, unsigned int size);
1682 extern int perf_output_begin_forward(struct perf_output_handle *handle,
1683 struct perf_sample_data *data,
1684 struct perf_event *event,
1686 extern int perf_output_begin_backward(struct perf_output_handle *handle,
1687 struct perf_sample_data *data,
1688 struct perf_event *event,
1691 extern void perf_output_end(struct perf_output_handle *handle);
1692 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
1693 const void *buf, unsigned int len);
1694 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
1696 extern long perf_output_copy_aux(struct perf_output_handle *aux_handle,
1697 struct perf_output_handle *handle,
1698 unsigned long from, unsigned long to);
1699 extern int perf_swevent_get_recursion_context(void);
1700 extern void perf_swevent_put_recursion_context(int rctx);
1701 extern u64 perf_swevent_set_period(struct perf_event *event);
1702 extern void perf_event_enable(struct perf_event *event);
1703 extern void perf_event_disable(struct perf_event *event);
1704 extern void perf_event_disable_local(struct perf_event *event);
1705 extern void perf_event_disable_inatomic(struct perf_event *event);
1706 extern void perf_event_task_tick(void);
1707 extern int perf_event_account_interrupt(struct perf_event *event);
1708 extern int perf_event_period(struct perf_event *event, u64 value);
1709 extern u64 perf_event_pause(struct perf_event *event, bool reset);
1710 #else /* !CONFIG_PERF_EVENTS: */
1711 static inline void *
1712 perf_aux_output_begin(struct perf_output_handle *handle,
1713 struct perf_event *event) { return NULL; }
1715 perf_aux_output_end(struct perf_output_handle *handle, unsigned long size)
1718 perf_aux_output_skip(struct perf_output_handle *handle,
1719 unsigned long size) { return -EINVAL; }
1720 static inline void *
1721 perf_get_aux(struct perf_output_handle *handle) { return NULL; }
1723 perf_event_task_migrate(struct task_struct *task) { }
1725 perf_event_task_sched_in(struct task_struct *prev,
1726 struct task_struct *task) { }
1728 perf_event_task_sched_out(struct task_struct *prev,
1729 struct task_struct *next) { }
1730 static inline int perf_event_init_task(struct task_struct *child,
1731 u64 clone_flags) { return 0; }
1732 static inline void perf_event_exit_task(struct task_struct *child) { }
1733 static inline void perf_event_free_task(struct task_struct *task) { }
1734 static inline void perf_event_delayed_put(struct task_struct *task) { }
1735 static inline struct file *perf_event_get(unsigned int fd) { return ERR_PTR(-EINVAL); }
1736 static inline const struct perf_event *perf_get_event(struct file *file)
1738 return ERR_PTR(-EINVAL);
1740 static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
1742 return ERR_PTR(-EINVAL);
1744 static inline int perf_event_read_local(struct perf_event *event, u64 *value,
1745 u64 *enabled, u64 *running)
1749 static inline void perf_event_print_debug(void) { }
1750 static inline int perf_event_task_disable(void) { return -EINVAL; }
1751 static inline int perf_event_task_enable(void) { return -EINVAL; }
1752 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1758 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) { }
1760 perf_bp_event(struct perf_event *event, void *data) { }
1762 static inline void perf_event_mmap(struct vm_area_struct *vma) { }
1764 typedef int (perf_ksymbol_get_name_f)(char *name, int name_len, void *data);
1765 static inline void perf_event_ksymbol(u16 ksym_type, u64 addr, u32 len,
1766 bool unregister, const char *sym) { }
1767 static inline void perf_event_bpf_event(struct bpf_prog *prog,
1768 enum perf_bpf_event_type type,
1770 static inline void perf_event_exec(void) { }
1771 static inline void perf_event_comm(struct task_struct *tsk, bool exec) { }
1772 static inline void perf_event_namespaces(struct task_struct *tsk) { }
1773 static inline void perf_event_fork(struct task_struct *tsk) { }
1774 static inline void perf_event_text_poke(const void *addr,
1775 const void *old_bytes,
1777 const void *new_bytes,
1779 static inline void perf_event_init(void) { }
1780 static inline int perf_swevent_get_recursion_context(void) { return -1; }
1781 static inline void perf_swevent_put_recursion_context(int rctx) { }
1782 static inline u64 perf_swevent_set_period(struct perf_event *event) { return 0; }
1783 static inline void perf_event_enable(struct perf_event *event) { }
1784 static inline void perf_event_disable(struct perf_event *event) { }
1785 static inline int __perf_event_disable(void *info) { return -1; }
1786 static inline void perf_event_task_tick(void) { }
1787 static inline int perf_event_release_kernel(struct perf_event *event) { return 0; }
1788 static inline int perf_event_period(struct perf_event *event, u64 value)
1792 static inline u64 perf_event_pause(struct perf_event *event, bool reset)
1798 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1799 extern void perf_restore_debug_store(void);
1801 static inline void perf_restore_debug_store(void) { }
1804 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1806 struct perf_pmu_events_attr {
1807 struct device_attribute attr;
1809 const char *event_str;
1812 struct perf_pmu_events_ht_attr {
1813 struct device_attribute attr;
1815 const char *event_str_ht;
1816 const char *event_str_noht;
1819 struct perf_pmu_events_hybrid_attr {
1820 struct device_attribute attr;
1822 const char *event_str;
1826 struct perf_pmu_format_hybrid_attr {
1827 struct device_attribute attr;
1831 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1834 #define PMU_EVENT_ATTR(_name, _var, _id, _show) \
1835 static struct perf_pmu_events_attr _var = { \
1836 .attr = __ATTR(_name, 0444, _show, NULL), \
1840 #define PMU_EVENT_ATTR_STRING(_name, _var, _str) \
1841 static struct perf_pmu_events_attr _var = { \
1842 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1844 .event_str = _str, \
1847 #define PMU_EVENT_ATTR_ID(_name, _show, _id) \
1848 (&((struct perf_pmu_events_attr[]) { \
1849 { .attr = __ATTR(_name, 0444, _show, NULL), \
1853 #define PMU_FORMAT_ATTR_SHOW(_name, _format) \
1855 _name##_show(struct device *dev, \
1856 struct device_attribute *attr, \
1859 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
1860 return sprintf(page, _format "\n"); \
1863 #define PMU_FORMAT_ATTR(_name, _format) \
1864 PMU_FORMAT_ATTR_SHOW(_name, _format) \
1866 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1868 /* Performance counter hotplug functions */
1869 #ifdef CONFIG_PERF_EVENTS
1870 int perf_event_init_cpu(unsigned int cpu);
1871 int perf_event_exit_cpu(unsigned int cpu);
1873 #define perf_event_init_cpu NULL
1874 #define perf_event_exit_cpu NULL
1877 extern void arch_perf_update_userpage(struct perf_event *event,
1878 struct perf_event_mmap_page *userpg,
1882 * Snapshot branch stack on software events.
1884 * Branch stack can be very useful in understanding software events. For
1885 * example, when a long function, e.g. sys_perf_event_open, returns an
1886 * errno, it is not obvious why the function failed. Branch stack could
1887 * provide very helpful information in this type of scenarios.
1889 * On software event, it is necessary to stop the hardware branch recorder
1890 * fast. Otherwise, the hardware register/buffer will be flushed with
1891 * entries of the triggering event. Therefore, static call is used to
1892 * stop the hardware recorder.
1896 * cnt is the number of entries allocated for entries.
1897 * Return number of entries copied to .
1899 typedef int (perf_snapshot_branch_stack_t)(struct perf_branch_entry *entries,
1901 DECLARE_STATIC_CALL(perf_snapshot_branch_stack, perf_snapshot_branch_stack_t);
1903 #ifndef PERF_NEEDS_LOPWR_CB
1904 static inline void perf_lopwr_cb(bool mode)
1909 #endif /* _LINUX_PERF_EVENT_H */