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[]; /* /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[]
96 * hw_idx: The low level index of raw branch records
97 * for the most recent branch.
98 * -1ULL means invalid/unknown.
100 * Note that nr can vary from sample to sample
101 * branches (to, from) are stored from most recent
102 * to least recent, i.e., entries[0] contains the most
104 * The entries[] is an abstraction of raw branch records,
105 * which may not be stored in age order in HW, e.g. Intel LBR.
106 * The hw_idx is to expose the low level index of raw
107 * branch record for the most recent branch aka entries[0].
108 * The hw_idx index is between -1 (unknown) and max depth,
109 * which can be retrieved in /sys/devices/cpu/caps/branches.
110 * For the architectures whose raw branch records are
111 * already stored in age order, the hw_idx should be 0.
113 struct perf_branch_stack {
116 struct perf_branch_entry entries[];
122 * extra PMU register associated with an event
124 struct hw_perf_event_extra {
125 u64 config; /* register value */
126 unsigned int reg; /* register address or index */
127 int alloc; /* extra register already allocated */
128 int idx; /* index in shared_regs->regs[] */
132 * struct hw_perf_event - performance event hardware details:
134 struct hw_perf_event {
135 #ifdef CONFIG_PERF_EVENTS
137 struct { /* hardware */
140 unsigned long config_base;
141 unsigned long event_base;
142 int event_base_rdpmc;
147 struct hw_perf_event_extra extra_reg;
148 struct hw_perf_event_extra branch_reg;
150 struct { /* software */
151 struct hrtimer hrtimer;
153 struct { /* tracepoint */
154 /* for tp_event->class */
155 struct list_head tp_list;
157 struct { /* amd_power */
161 #ifdef CONFIG_HAVE_HW_BREAKPOINT
162 struct { /* breakpoint */
164 * Crufty hack to avoid the chicken and egg
165 * problem hw_breakpoint has with context
166 * creation and event initalization.
168 struct arch_hw_breakpoint info;
169 struct list_head bp_list;
172 struct { /* amd_iommu */
181 * If the event is a per task event, this will point to the task in
182 * question. See the comment in perf_event_alloc().
184 struct task_struct *target;
187 * PMU would store hardware filter configuration
192 /* Last sync'ed generation of filters */
193 unsigned long addr_filters_gen;
196 * hw_perf_event::state flags; used to track the PERF_EF_* state.
198 #define PERF_HES_STOPPED 0x01 /* the counter is stopped */
199 #define PERF_HES_UPTODATE 0x02 /* event->count up-to-date */
200 #define PERF_HES_ARCH 0x04
205 * The last observed hardware counter value, updated with a
206 * local64_cmpxchg() such that pmu::read() can be called nested.
208 local64_t prev_count;
211 * The period to start the next sample with.
216 struct { /* Sampling */
218 * The period we started this sample with.
223 * However much is left of the current period;
224 * note that this is a full 64bit value and
225 * allows for generation of periods longer
226 * than hardware might allow.
228 local64_t period_left;
230 struct { /* Topdown events counting for context switch */
237 * State for throttling the event, see __perf_event_overflow() and
238 * perf_adjust_freq_unthr_context().
244 * State for freq target events, see __perf_event_overflow() and
245 * perf_adjust_freq_unthr_context().
248 u64 freq_count_stamp;
255 * Common implementation detail of pmu::{start,commit,cancel}_txn
257 #define PERF_PMU_TXN_ADD 0x1 /* txn to add/schedule event on PMU */
258 #define PERF_PMU_TXN_READ 0x2 /* txn to read event group from PMU */
261 * pmu::capabilities flags
263 #define PERF_PMU_CAP_NO_INTERRUPT 0x01
264 #define PERF_PMU_CAP_NO_NMI 0x02
265 #define PERF_PMU_CAP_AUX_NO_SG 0x04
266 #define PERF_PMU_CAP_EXTENDED_REGS 0x08
267 #define PERF_PMU_CAP_EXCLUSIVE 0x10
268 #define PERF_PMU_CAP_ITRACE 0x20
269 #define PERF_PMU_CAP_HETEROGENEOUS_CPUS 0x40
270 #define PERF_PMU_CAP_NO_EXCLUDE 0x80
271 #define PERF_PMU_CAP_AUX_OUTPUT 0x100
273 struct perf_output_handle;
276 * struct pmu - generic performance monitoring unit
279 struct list_head entry;
281 struct module *module;
283 const struct attribute_group **attr_groups;
284 const struct attribute_group **attr_update;
289 * various common per-pmu feature flags
293 int __percpu *pmu_disable_count;
294 struct perf_cpu_context __percpu *pmu_cpu_context;
295 atomic_t exclusive_cnt; /* < 0: cpu; > 0: tsk */
297 int hrtimer_interval_ms;
299 /* number of address filters this PMU can do */
300 unsigned int nr_addr_filters;
303 * Fully disable/enable this PMU, can be used to protect from the PMI
304 * as well as for lazy/batch writing of the MSRs.
306 void (*pmu_enable) (struct pmu *pmu); /* optional */
307 void (*pmu_disable) (struct pmu *pmu); /* optional */
310 * Try and initialize the event for this PMU.
313 * -ENOENT -- @event is not for this PMU
315 * -ENODEV -- @event is for this PMU but PMU not present
316 * -EBUSY -- @event is for this PMU but PMU temporarily unavailable
317 * -EINVAL -- @event is for this PMU but @event is not valid
318 * -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
319 * -EACCES -- @event is for this PMU, @event is valid, but no privileges
321 * 0 -- @event is for this PMU and valid
323 * Other error return values are allowed.
325 int (*event_init) (struct perf_event *event);
328 * Notification that the event was mapped or unmapped. Called
329 * in the context of the mapping task.
331 void (*event_mapped) (struct perf_event *event, struct mm_struct *mm); /* optional */
332 void (*event_unmapped) (struct perf_event *event, struct mm_struct *mm); /* optional */
335 * Flags for ->add()/->del()/ ->start()/->stop(). There are
336 * matching hw_perf_event::state flags.
338 #define PERF_EF_START 0x01 /* start the counter when adding */
339 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
340 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
343 * Adds/Removes a counter to/from the PMU, can be done inside a
344 * transaction, see the ->*_txn() methods.
346 * The add/del callbacks will reserve all hardware resources required
347 * to service the event, this includes any counter constraint
350 * Called with IRQs disabled and the PMU disabled on the CPU the event
353 * ->add() called without PERF_EF_START should result in the same state
354 * as ->add() followed by ->stop().
356 * ->del() must always PERF_EF_UPDATE stop an event. If it calls
357 * ->stop() that must deal with already being stopped without
360 int (*add) (struct perf_event *event, int flags);
361 void (*del) (struct perf_event *event, int flags);
364 * Starts/Stops a counter present on the PMU.
366 * The PMI handler should stop the counter when perf_event_overflow()
367 * returns !0. ->start() will be used to continue.
369 * Also used to change the sample period.
371 * Called with IRQs disabled and the PMU disabled on the CPU the event
372 * is on -- will be called from NMI context with the PMU generates
375 * ->stop() with PERF_EF_UPDATE will read the counter and update
376 * period/count values like ->read() would.
378 * ->start() with PERF_EF_RELOAD will reprogram the counter
379 * value, must be preceded by a ->stop() with PERF_EF_UPDATE.
381 void (*start) (struct perf_event *event, int flags);
382 void (*stop) (struct perf_event *event, int flags);
385 * Updates the counter value of the event.
387 * For sampling capable PMUs this will also update the software period
388 * hw_perf_event::period_left field.
390 void (*read) (struct perf_event *event);
393 * Group events scheduling is treated as a transaction, add
394 * group events as a whole and perform one schedulability test.
395 * If the test fails, roll back the whole group
397 * Start the transaction, after this ->add() doesn't need to
398 * do schedulability tests.
402 void (*start_txn) (struct pmu *pmu, unsigned int txn_flags);
404 * If ->start_txn() disabled the ->add() schedulability test
405 * then ->commit_txn() is required to perform one. On success
406 * the transaction is closed. On error the transaction is kept
407 * open until ->cancel_txn() is called.
411 int (*commit_txn) (struct pmu *pmu);
413 * Will cancel the transaction, assumes ->del() is called
414 * for each successful ->add() during the transaction.
418 void (*cancel_txn) (struct pmu *pmu);
421 * Will return the value for perf_event_mmap_page::index for this event,
422 * if no implementation is provided it will default to: event->hw.idx + 1.
424 int (*event_idx) (struct perf_event *event); /*optional */
427 * context-switches callback
429 void (*sched_task) (struct perf_event_context *ctx,
433 * Kmem cache of PMU specific data
435 struct kmem_cache *task_ctx_cache;
438 * PMU specific parts of task perf event context (i.e. ctx->task_ctx_data)
439 * can be synchronized using this function. See Intel LBR callstack support
440 * implementation and Perf core context switch handling callbacks for usage
443 void (*swap_task_ctx) (struct perf_event_context *prev,
444 struct perf_event_context *next);
448 * Set up pmu-private data structures for an AUX area
450 void *(*setup_aux) (struct perf_event *event, void **pages,
451 int nr_pages, bool overwrite);
455 * Free pmu-private AUX data structures
457 void (*free_aux) (void *aux); /* optional */
460 * Take a snapshot of the AUX buffer without touching the event
461 * state, so that preempting ->start()/->stop() callbacks does
462 * not interfere with their logic. Called in PMI context.
464 * Returns the size of AUX data copied to the output handle.
468 long (*snapshot_aux) (struct perf_event *event,
469 struct perf_output_handle *handle,
473 * Validate address range filters: make sure the HW supports the
474 * requested configuration and number of filters; return 0 if the
475 * supplied filters are valid, -errno otherwise.
477 * Runs in the context of the ioctl()ing process and is not serialized
478 * with the rest of the PMU callbacks.
480 int (*addr_filters_validate) (struct list_head *filters);
484 * Synchronize address range filter configuration:
485 * translate hw-agnostic filters into hardware configuration in
486 * event::hw::addr_filters.
488 * Runs as a part of filter sync sequence that is done in ->start()
489 * callback by calling perf_event_addr_filters_sync().
491 * May (and should) traverse event::addr_filters::list, for which its
492 * caller provides necessary serialization.
494 void (*addr_filters_sync) (struct perf_event *event);
498 * Check if event can be used for aux_output purposes for
499 * events of this PMU.
501 * Runs from perf_event_open(). Should return 0 for "no match"
502 * or non-zero for "match".
504 int (*aux_output_match) (struct perf_event *event);
508 * Filter events for PMU-specific reasons.
510 int (*filter_match) (struct perf_event *event); /* optional */
513 * Check period value for PERF_EVENT_IOC_PERIOD ioctl.
515 int (*check_period) (struct perf_event *event, u64 value); /* optional */
518 enum perf_addr_filter_action_t {
519 PERF_ADDR_FILTER_ACTION_STOP = 0,
520 PERF_ADDR_FILTER_ACTION_START,
521 PERF_ADDR_FILTER_ACTION_FILTER,
525 * struct perf_addr_filter - address range filter definition
526 * @entry: event's filter list linkage
527 * @path: object file's path for file-based filters
528 * @offset: filter range offset
529 * @size: filter range size (size==0 means single address trigger)
530 * @action: filter/start/stop
532 * This is a hardware-agnostic filter configuration as specified by the user.
534 struct perf_addr_filter {
535 struct list_head entry;
537 unsigned long offset;
539 enum perf_addr_filter_action_t action;
543 * struct perf_addr_filters_head - container for address range filters
544 * @list: list of filters for this event
545 * @lock: spinlock that serializes accesses to the @list and event's
546 * (and its children's) filter generations.
547 * @nr_file_filters: number of file-based filters
549 * A child event will use parent's @list (and therefore @lock), so they are
550 * bundled together; see perf_event_addr_filters().
552 struct perf_addr_filters_head {
553 struct list_head list;
555 unsigned int nr_file_filters;
558 struct perf_addr_filter_range {
564 * enum perf_event_state - the states of an event:
566 enum perf_event_state {
567 PERF_EVENT_STATE_DEAD = -4,
568 PERF_EVENT_STATE_EXIT = -3,
569 PERF_EVENT_STATE_ERROR = -2,
570 PERF_EVENT_STATE_OFF = -1,
571 PERF_EVENT_STATE_INACTIVE = 0,
572 PERF_EVENT_STATE_ACTIVE = 1,
576 struct perf_sample_data;
578 typedef void (*perf_overflow_handler_t)(struct perf_event *,
579 struct perf_sample_data *,
580 struct pt_regs *regs);
583 * Event capabilities. For event_caps and groups caps.
585 * PERF_EV_CAP_SOFTWARE: Is a software event.
586 * PERF_EV_CAP_READ_ACTIVE_PKG: A CPU event (or cgroup event) that can be read
587 * from any CPU in the package where it is active.
588 * PERF_EV_CAP_SIBLING: An event with this flag must be a group sibling and
589 * cannot be a group leader. If an event with this flag is detached from the
590 * group it is scheduled out and moved into an unrecoverable ERROR state.
592 #define PERF_EV_CAP_SOFTWARE BIT(0)
593 #define PERF_EV_CAP_READ_ACTIVE_PKG BIT(1)
594 #define PERF_EV_CAP_SIBLING BIT(2)
596 #define SWEVENT_HLIST_BITS 8
597 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
599 struct swevent_hlist {
600 struct hlist_head heads[SWEVENT_HLIST_SIZE];
601 struct rcu_head rcu_head;
604 #define PERF_ATTACH_CONTEXT 0x01
605 #define PERF_ATTACH_GROUP 0x02
606 #define PERF_ATTACH_TASK 0x04
607 #define PERF_ATTACH_TASK_DATA 0x08
608 #define PERF_ATTACH_ITRACE 0x10
609 #define PERF_ATTACH_SCHED_CB 0x20
614 struct pmu_event_list {
616 struct list_head list;
619 #define for_each_sibling_event(sibling, event) \
620 if ((event)->group_leader == (event)) \
621 list_for_each_entry((sibling), &(event)->sibling_list, sibling_list)
624 * struct perf_event - performance event kernel representation:
627 #ifdef CONFIG_PERF_EVENTS
629 * entry onto perf_event_context::event_list;
630 * modifications require ctx->lock
631 * RCU safe iterations.
633 struct list_head event_entry;
636 * Locked for modification by both ctx->mutex and ctx->lock; holding
637 * either sufficies for read.
639 struct list_head sibling_list;
640 struct list_head active_list;
642 * Node on the pinned or flexible tree located at the event context;
644 struct rb_node group_node;
647 * We need storage to track the entries in perf_pmu_migrate_context; we
648 * cannot use the event_entry because of RCU and we want to keep the
649 * group in tact which avoids us using the other two entries.
651 struct list_head migrate_entry;
653 struct hlist_node hlist_entry;
654 struct list_head active_entry;
657 /* Not serialized. Only written during event initialization. */
659 /* The cumulative AND of all event_caps for events in this group. */
662 unsigned int group_generation;
663 struct perf_event *group_leader;
667 enum perf_event_state state;
668 unsigned int attach_state;
670 atomic64_t child_count;
673 * These are the total time in nanoseconds that the event
674 * has been enabled (i.e. eligible to run, and the task has
675 * been scheduled in, if this is a per-task event)
676 * and running (scheduled onto the CPU), respectively.
678 u64 total_time_enabled;
679 u64 total_time_running;
682 struct perf_event_attr attr;
686 struct hw_perf_event hw;
688 struct perf_event_context *ctx;
689 atomic_long_t refcount;
692 * These accumulate total time (in nanoseconds) that children
693 * events have been enabled and running, respectively.
695 atomic64_t child_total_time_enabled;
696 atomic64_t child_total_time_running;
699 * Protect attach/detach and child_list:
701 struct mutex child_mutex;
702 struct list_head child_list;
703 struct perf_event *parent;
708 struct list_head owner_entry;
709 struct task_struct *owner;
712 struct mutex mmap_mutex;
715 struct perf_buffer *rb;
716 struct list_head rb_entry;
717 unsigned long rcu_batches;
721 wait_queue_head_t waitq;
722 struct fasync_struct *fasync;
724 /* delayed work for NMIs and such */
728 struct irq_work pending;
730 atomic_t event_limit;
732 /* address range filters */
733 struct perf_addr_filters_head addr_filters;
734 /* vma address array for file-based filders */
735 struct perf_addr_filter_range *addr_filter_ranges;
736 unsigned long addr_filters_gen;
738 /* for aux_output events */
739 struct perf_event *aux_event;
741 void (*destroy)(struct perf_event *);
742 struct rcu_head rcu_head;
744 struct pid_namespace *ns;
747 atomic64_t lost_samples;
750 perf_overflow_handler_t overflow_handler;
751 void *overflow_handler_context;
752 #ifdef CONFIG_BPF_SYSCALL
753 perf_overflow_handler_t orig_overflow_handler;
754 struct bpf_prog *prog;
757 #ifdef CONFIG_EVENT_TRACING
758 struct trace_event_call *tp_event;
759 struct event_filter *filter;
760 #ifdef CONFIG_FUNCTION_TRACER
761 struct ftrace_ops ftrace_ops;
765 #ifdef CONFIG_CGROUP_PERF
766 struct perf_cgroup *cgrp; /* cgroup event is attach to */
769 #ifdef CONFIG_SECURITY
772 struct list_head sb_list;
773 #endif /* CONFIG_PERF_EVENTS */
777 struct perf_event_groups {
783 * struct perf_event_context - event context structure
785 * Used as a container for task events and CPU events as well:
787 struct perf_event_context {
790 * Protect the states of the events in the list,
791 * nr_active, and the list:
795 * Protect the list of events. Locking either mutex or lock
796 * is sufficient to ensure the list doesn't change; to change
797 * the list you need to lock both the mutex and the spinlock.
801 struct list_head active_ctx_list;
802 struct perf_event_groups pinned_groups;
803 struct perf_event_groups flexible_groups;
804 struct list_head event_list;
806 struct list_head pinned_active;
807 struct list_head flexible_active;
816 * Set when nr_events != nr_active, except tolerant to events not
817 * necessary to be active due to scheduling constraints, such as cgroups.
819 int rotate_necessary;
821 struct task_struct *task;
824 * Context clock, runs when context enabled.
831 * These fields let us detect when two contexts have both
832 * been cloned (inherited) from a common ancestor.
834 struct perf_event_context *parent_ctx;
838 #ifdef CONFIG_CGROUP_PERF
839 int nr_cgroups; /* cgroup evts */
841 void *task_ctx_data; /* pmu specific data */
842 struct rcu_head rcu_head;
846 * Number of contexts where an event can trigger:
847 * task, softirq, hardirq, nmi.
849 #define PERF_NR_CONTEXTS 4
852 * struct perf_event_cpu_context - per cpu event context structure
854 struct perf_cpu_context {
855 struct perf_event_context ctx;
856 struct perf_event_context *task_ctx;
860 raw_spinlock_t hrtimer_lock;
861 struct hrtimer hrtimer;
862 ktime_t hrtimer_interval;
863 unsigned int hrtimer_active;
865 #ifdef CONFIG_CGROUP_PERF
866 struct perf_cgroup *cgrp;
867 struct list_head cgrp_cpuctx_entry;
870 struct list_head sched_cb_entry;
875 * Per-CPU storage for iterators used in visit_groups_merge. The default
876 * storage is of size 2 to hold the CPU and any CPU event iterators.
879 struct perf_event **heap;
880 struct perf_event *heap_default[2];
883 struct perf_output_handle {
884 struct perf_event *event;
885 struct perf_buffer *rb;
886 unsigned long wakeup;
896 struct bpf_perf_event_data_kern {
897 bpf_user_pt_regs_t *regs;
898 struct perf_sample_data *data;
899 struct perf_event *event;
902 #ifdef CONFIG_CGROUP_PERF
905 * perf_cgroup_info keeps track of time_enabled for a cgroup.
906 * This is a per-cpu dynamically allocated data structure.
908 struct perf_cgroup_info {
916 struct cgroup_subsys_state css;
917 struct perf_cgroup_info __percpu *info;
921 * Must ensure cgroup is pinned (css_get) before calling
922 * this function. In other words, we cannot call this function
923 * if there is no cgroup event for the current CPU context.
925 static inline struct perf_cgroup *
926 perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
928 return container_of(task_css_check(task, perf_event_cgrp_id,
929 ctx ? lockdep_is_held(&ctx->lock)
931 struct perf_cgroup, css);
933 #endif /* CONFIG_CGROUP_PERF */
935 #ifdef CONFIG_PERF_EVENTS
937 extern void *perf_aux_output_begin(struct perf_output_handle *handle,
938 struct perf_event *event);
939 extern void perf_aux_output_end(struct perf_output_handle *handle,
941 extern int perf_aux_output_skip(struct perf_output_handle *handle,
943 extern void *perf_get_aux(struct perf_output_handle *handle);
944 extern void perf_aux_output_flag(struct perf_output_handle *handle, u64 flags);
945 extern void perf_event_itrace_started(struct perf_event *event);
947 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
948 extern void perf_pmu_unregister(struct pmu *pmu);
950 extern int perf_num_counters(void);
951 extern const char *perf_pmu_name(void);
952 extern void __perf_event_task_sched_in(struct task_struct *prev,
953 struct task_struct *task);
954 extern void __perf_event_task_sched_out(struct task_struct *prev,
955 struct task_struct *next);
956 extern int perf_event_init_task(struct task_struct *child);
957 extern void perf_event_exit_task(struct task_struct *child);
958 extern void perf_event_free_task(struct task_struct *task);
959 extern void perf_event_delayed_put(struct task_struct *task);
960 extern struct file *perf_event_get(unsigned int fd);
961 extern const struct perf_event *perf_get_event(struct file *file);
962 extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
963 extern void perf_event_print_debug(void);
964 extern void perf_pmu_disable(struct pmu *pmu);
965 extern void perf_pmu_enable(struct pmu *pmu);
966 extern void perf_sched_cb_dec(struct pmu *pmu);
967 extern void perf_sched_cb_inc(struct pmu *pmu);
968 extern int perf_event_task_disable(void);
969 extern int perf_event_task_enable(void);
971 extern void perf_pmu_resched(struct pmu *pmu);
973 extern int perf_event_refresh(struct perf_event *event, int refresh);
974 extern void perf_event_update_userpage(struct perf_event *event);
975 extern int perf_event_release_kernel(struct perf_event *event);
976 extern struct perf_event *
977 perf_event_create_kernel_counter(struct perf_event_attr *attr,
979 struct task_struct *task,
980 perf_overflow_handler_t callback,
982 extern void perf_pmu_migrate_context(struct pmu *pmu,
983 int src_cpu, int dst_cpu);
984 int perf_event_read_local(struct perf_event *event, u64 *value,
985 u64 *enabled, u64 *running);
986 extern u64 perf_event_read_value(struct perf_event *event,
987 u64 *enabled, u64 *running);
990 struct perf_sample_data {
992 * Fields set by perf_sample_data_init(), group so as to
993 * minimize the cachelines touched.
996 struct perf_raw_record *raw;
997 struct perf_branch_stack *br_stack;
1001 union perf_mem_data_src data_src;
1004 * The other fields, optionally {set,used} by
1005 * perf_{prepare,output}_sample().
1020 struct perf_callchain_entry *callchain;
1023 struct perf_regs regs_user;
1024 struct perf_regs regs_intr;
1025 u64 stack_user_size;
1029 } ____cacheline_aligned;
1031 /* default value for data source */
1032 #define PERF_MEM_NA (PERF_MEM_S(OP, NA) |\
1033 PERF_MEM_S(LVL, NA) |\
1034 PERF_MEM_S(SNOOP, NA) |\
1035 PERF_MEM_S(LOCK, NA) |\
1036 PERF_MEM_S(TLB, NA))
1038 static inline void perf_sample_data_init(struct perf_sample_data *data,
1039 u64 addr, u64 period)
1041 /* remaining struct members initialized in perf_prepare_sample() */
1044 data->br_stack = NULL;
1045 data->period = period;
1047 data->data_src.val = PERF_MEM_NA;
1051 extern void perf_output_sample(struct perf_output_handle *handle,
1052 struct perf_event_header *header,
1053 struct perf_sample_data *data,
1054 struct perf_event *event);
1055 extern void perf_prepare_sample(struct perf_event_header *header,
1056 struct perf_sample_data *data,
1057 struct perf_event *event,
1058 struct pt_regs *regs);
1060 extern int perf_event_overflow(struct perf_event *event,
1061 struct perf_sample_data *data,
1062 struct pt_regs *regs);
1064 extern void perf_event_output_forward(struct perf_event *event,
1065 struct perf_sample_data *data,
1066 struct pt_regs *regs);
1067 extern void perf_event_output_backward(struct perf_event *event,
1068 struct perf_sample_data *data,
1069 struct pt_regs *regs);
1070 extern int perf_event_output(struct perf_event *event,
1071 struct perf_sample_data *data,
1072 struct pt_regs *regs);
1075 __is_default_overflow_handler(perf_overflow_handler_t overflow_handler)
1077 if (likely(overflow_handler == perf_event_output_forward))
1079 if (unlikely(overflow_handler == perf_event_output_backward))
1084 #define is_default_overflow_handler(event) \
1085 __is_default_overflow_handler((event)->overflow_handler)
1087 #ifdef CONFIG_BPF_SYSCALL
1088 static inline bool uses_default_overflow_handler(struct perf_event *event)
1090 if (likely(is_default_overflow_handler(event)))
1093 return __is_default_overflow_handler(event->orig_overflow_handler);
1096 #define uses_default_overflow_handler(event) \
1097 is_default_overflow_handler(event)
1101 perf_event_header__init_id(struct perf_event_header *header,
1102 struct perf_sample_data *data,
1103 struct perf_event *event);
1105 perf_event__output_id_sample(struct perf_event *event,
1106 struct perf_output_handle *handle,
1107 struct perf_sample_data *sample);
1110 perf_log_lost_samples(struct perf_event *event, u64 lost);
1112 static inline bool event_has_any_exclude_flag(struct perf_event *event)
1114 struct perf_event_attr *attr = &event->attr;
1116 return attr->exclude_idle || attr->exclude_user ||
1117 attr->exclude_kernel || attr->exclude_hv ||
1118 attr->exclude_guest || attr->exclude_host;
1121 static inline bool is_sampling_event(struct perf_event *event)
1123 return event->attr.sample_period != 0;
1127 * Return 1 for a software event, 0 for a hardware event
1129 static inline int is_software_event(struct perf_event *event)
1131 return event->event_caps & PERF_EV_CAP_SOFTWARE;
1135 * Return 1 for event in sw context, 0 for event in hw context
1137 static inline int in_software_context(struct perf_event *event)
1139 return event->ctx->pmu->task_ctx_nr == perf_sw_context;
1142 static inline int is_exclusive_pmu(struct pmu *pmu)
1144 return pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE;
1147 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
1149 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
1150 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
1152 #ifndef perf_arch_fetch_caller_regs
1153 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
1157 * When generating a perf sample in-line, instead of from an interrupt /
1158 * exception, we lack a pt_regs. This is typically used from software events
1159 * like: SW_CONTEXT_SWITCHES, SW_MIGRATIONS and the tie-in with tracepoints.
1161 * We typically don't need a full set, but (for x86) do require:
1162 * - ip for PERF_SAMPLE_IP
1163 * - cs for user_mode() tests
1164 * - sp for PERF_SAMPLE_CALLCHAIN
1165 * - eflags for MISC bits and CALLCHAIN (see: perf_hw_regs())
1167 * NOTE: assumes @regs is otherwise already 0 filled; this is important for
1168 * things like PERF_SAMPLE_REGS_INTR.
1170 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1172 perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1175 static __always_inline void
1176 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1178 if (static_key_false(&perf_swevent_enabled[event_id]))
1179 __perf_sw_event(event_id, nr, regs, addr);
1182 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
1185 * 'Special' version for the scheduler, it hard assumes no recursion,
1186 * which is guaranteed by us not actually scheduling inside other swevents
1187 * because those disable preemption.
1189 static __always_inline void
1190 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
1192 if (static_key_false(&perf_swevent_enabled[event_id])) {
1193 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1195 perf_fetch_caller_regs(regs);
1196 ___perf_sw_event(event_id, nr, regs, addr);
1200 extern struct static_key_false perf_sched_events;
1202 static __always_inline bool
1203 perf_sw_migrate_enabled(void)
1205 if (static_key_false(&perf_swevent_enabled[PERF_COUNT_SW_CPU_MIGRATIONS]))
1210 static inline void perf_event_task_migrate(struct task_struct *task)
1212 if (perf_sw_migrate_enabled())
1213 task->sched_migrated = 1;
1216 static inline void perf_event_task_sched_in(struct task_struct *prev,
1217 struct task_struct *task)
1219 if (static_branch_unlikely(&perf_sched_events))
1220 __perf_event_task_sched_in(prev, task);
1222 if (perf_sw_migrate_enabled() && task->sched_migrated) {
1223 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1225 perf_fetch_caller_regs(regs);
1226 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, regs, 0);
1227 task->sched_migrated = 0;
1231 static inline void perf_event_task_sched_out(struct task_struct *prev,
1232 struct task_struct *next)
1234 perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
1236 if (static_branch_unlikely(&perf_sched_events))
1237 __perf_event_task_sched_out(prev, next);
1240 extern void perf_event_mmap(struct vm_area_struct *vma);
1242 extern void perf_event_ksymbol(u16 ksym_type, u64 addr, u32 len,
1243 bool unregister, const char *sym);
1244 extern void perf_event_bpf_event(struct bpf_prog *prog,
1245 enum perf_bpf_event_type type,
1248 extern struct perf_guest_info_callbacks __rcu *perf_guest_cbs;
1249 static inline struct perf_guest_info_callbacks *perf_get_guest_cbs(void)
1252 * Callbacks are RCU-protected and must be READ_ONCE to avoid reloading
1253 * the callbacks between a !NULL check and dereferences, to ensure
1254 * pending stores/changes to the callback pointers are visible before a
1255 * non-NULL perf_guest_cbs is visible to readers, and to prevent a
1256 * module from unloading callbacks while readers are active.
1258 return rcu_dereference(perf_guest_cbs);
1260 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1261 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1263 extern void perf_event_exec(void);
1264 extern void perf_event_comm(struct task_struct *tsk, bool exec);
1265 extern void perf_event_namespaces(struct task_struct *tsk);
1266 extern void perf_event_fork(struct task_struct *tsk);
1267 extern void perf_event_text_poke(const void *addr,
1268 const void *old_bytes, size_t old_len,
1269 const void *new_bytes, size_t new_len);
1272 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1274 extern void perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1275 extern void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1276 extern struct perf_callchain_entry *
1277 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
1278 u32 max_stack, bool crosstask, bool add_mark);
1279 extern struct perf_callchain_entry *perf_callchain(struct perf_event *event, struct pt_regs *regs);
1280 extern int get_callchain_buffers(int max_stack);
1281 extern void put_callchain_buffers(void);
1282 extern struct perf_callchain_entry *get_callchain_entry(int *rctx);
1283 extern void put_callchain_entry(int rctx);
1285 extern int sysctl_perf_event_max_stack;
1286 extern int sysctl_perf_event_max_contexts_per_stack;
1288 static inline int perf_callchain_store_context(struct perf_callchain_entry_ctx *ctx, u64 ip)
1290 if (ctx->contexts < sysctl_perf_event_max_contexts_per_stack) {
1291 struct perf_callchain_entry *entry = ctx->entry;
1292 entry->ip[entry->nr++] = ip;
1296 ctx->contexts_maxed = true;
1297 return -1; /* no more room, stop walking the stack */
1301 static inline int perf_callchain_store(struct perf_callchain_entry_ctx *ctx, u64 ip)
1303 if (ctx->nr < ctx->max_stack && !ctx->contexts_maxed) {
1304 struct perf_callchain_entry *entry = ctx->entry;
1305 entry->ip[entry->nr++] = ip;
1309 return -1; /* no more room, stop walking the stack */
1313 extern int sysctl_perf_event_paranoid;
1314 extern int sysctl_perf_event_mlock;
1315 extern int sysctl_perf_event_sample_rate;
1316 extern int sysctl_perf_cpu_time_max_percent;
1318 extern void perf_sample_event_took(u64 sample_len_ns);
1320 int perf_proc_update_handler(struct ctl_table *table, int write,
1321 void *buffer, size_t *lenp, loff_t *ppos);
1322 int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
1323 void *buffer, size_t *lenp, loff_t *ppos);
1324 int perf_event_max_stack_handler(struct ctl_table *table, int write,
1325 void *buffer, size_t *lenp, loff_t *ppos);
1327 /* Access to perf_event_open(2) syscall. */
1328 #define PERF_SECURITY_OPEN 0
1330 /* Finer grained perf_event_open(2) access control. */
1331 #define PERF_SECURITY_CPU 1
1332 #define PERF_SECURITY_KERNEL 2
1333 #define PERF_SECURITY_TRACEPOINT 3
1335 static inline int perf_is_paranoid(void)
1337 return sysctl_perf_event_paranoid > -1;
1340 static inline int perf_allow_kernel(struct perf_event_attr *attr)
1342 if (sysctl_perf_event_paranoid > 1 && !perfmon_capable())
1345 return security_perf_event_open(attr, PERF_SECURITY_KERNEL);
1348 static inline int perf_allow_cpu(struct perf_event_attr *attr)
1350 if (sysctl_perf_event_paranoid > 0 && !perfmon_capable())
1353 return security_perf_event_open(attr, PERF_SECURITY_CPU);
1356 static inline int perf_allow_tracepoint(struct perf_event_attr *attr)
1358 if (sysctl_perf_event_paranoid > -1 && !perfmon_capable())
1361 return security_perf_event_open(attr, PERF_SECURITY_TRACEPOINT);
1364 extern void perf_event_init(void);
1365 extern void perf_tp_event(u16 event_type, u64 count, void *record,
1366 int entry_size, struct pt_regs *regs,
1367 struct hlist_head *head, int rctx,
1368 struct task_struct *task);
1369 extern void perf_bp_event(struct perf_event *event, void *data);
1371 #ifndef perf_misc_flags
1372 # define perf_misc_flags(regs) \
1373 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1374 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1376 #ifndef perf_arch_bpf_user_pt_regs
1377 # define perf_arch_bpf_user_pt_regs(regs) regs
1380 static inline bool has_branch_stack(struct perf_event *event)
1382 return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1385 static inline bool needs_branch_stack(struct perf_event *event)
1387 return event->attr.branch_sample_type != 0;
1390 static inline bool has_aux(struct perf_event *event)
1392 return event->pmu->setup_aux;
1395 static inline bool is_write_backward(struct perf_event *event)
1397 return !!event->attr.write_backward;
1400 static inline bool has_addr_filter(struct perf_event *event)
1402 return event->pmu->nr_addr_filters;
1406 * An inherited event uses parent's filters
1408 static inline struct perf_addr_filters_head *
1409 perf_event_addr_filters(struct perf_event *event)
1411 struct perf_addr_filters_head *ifh = &event->addr_filters;
1414 ifh = &event->parent->addr_filters;
1419 extern void perf_event_addr_filters_sync(struct perf_event *event);
1421 extern int perf_output_begin(struct perf_output_handle *handle,
1422 struct perf_sample_data *data,
1423 struct perf_event *event, unsigned int size);
1424 extern int perf_output_begin_forward(struct perf_output_handle *handle,
1425 struct perf_sample_data *data,
1426 struct perf_event *event,
1428 extern int perf_output_begin_backward(struct perf_output_handle *handle,
1429 struct perf_sample_data *data,
1430 struct perf_event *event,
1433 extern void perf_output_end(struct perf_output_handle *handle);
1434 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
1435 const void *buf, unsigned int len);
1436 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
1438 extern long perf_output_copy_aux(struct perf_output_handle *aux_handle,
1439 struct perf_output_handle *handle,
1440 unsigned long from, unsigned long to);
1441 extern int perf_swevent_get_recursion_context(void);
1442 extern void perf_swevent_put_recursion_context(int rctx);
1443 extern u64 perf_swevent_set_period(struct perf_event *event);
1444 extern void perf_event_enable(struct perf_event *event);
1445 extern void perf_event_disable(struct perf_event *event);
1446 extern void perf_event_disable_local(struct perf_event *event);
1447 extern void perf_event_disable_inatomic(struct perf_event *event);
1448 extern void perf_event_task_tick(void);
1449 extern int perf_event_account_interrupt(struct perf_event *event);
1450 extern int perf_event_period(struct perf_event *event, u64 value);
1451 extern u64 perf_event_pause(struct perf_event *event, bool reset);
1452 #else /* !CONFIG_PERF_EVENTS: */
1453 static inline void *
1454 perf_aux_output_begin(struct perf_output_handle *handle,
1455 struct perf_event *event) { return NULL; }
1457 perf_aux_output_end(struct perf_output_handle *handle, unsigned long size)
1460 perf_aux_output_skip(struct perf_output_handle *handle,
1461 unsigned long size) { return -EINVAL; }
1462 static inline void *
1463 perf_get_aux(struct perf_output_handle *handle) { return NULL; }
1465 perf_event_task_migrate(struct task_struct *task) { }
1467 perf_event_task_sched_in(struct task_struct *prev,
1468 struct task_struct *task) { }
1470 perf_event_task_sched_out(struct task_struct *prev,
1471 struct task_struct *next) { }
1472 static inline int perf_event_init_task(struct task_struct *child) { return 0; }
1473 static inline void perf_event_exit_task(struct task_struct *child) { }
1474 static inline void perf_event_free_task(struct task_struct *task) { }
1475 static inline void perf_event_delayed_put(struct task_struct *task) { }
1476 static inline struct file *perf_event_get(unsigned int fd) { return ERR_PTR(-EINVAL); }
1477 static inline const struct perf_event *perf_get_event(struct file *file)
1479 return ERR_PTR(-EINVAL);
1481 static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
1483 return ERR_PTR(-EINVAL);
1485 static inline int perf_event_read_local(struct perf_event *event, u64 *value,
1486 u64 *enabled, u64 *running)
1490 static inline void perf_event_print_debug(void) { }
1491 static inline int perf_event_task_disable(void) { return -EINVAL; }
1492 static inline int perf_event_task_enable(void) { return -EINVAL; }
1493 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1499 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) { }
1501 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr) { }
1503 perf_bp_event(struct perf_event *event, void *data) { }
1505 static inline int perf_register_guest_info_callbacks
1506 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1507 static inline int perf_unregister_guest_info_callbacks
1508 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1510 static inline void perf_event_mmap(struct vm_area_struct *vma) { }
1512 typedef int (perf_ksymbol_get_name_f)(char *name, int name_len, void *data);
1513 static inline void perf_event_ksymbol(u16 ksym_type, u64 addr, u32 len,
1514 bool unregister, const char *sym) { }
1515 static inline void perf_event_bpf_event(struct bpf_prog *prog,
1516 enum perf_bpf_event_type type,
1518 static inline void perf_event_exec(void) { }
1519 static inline void perf_event_comm(struct task_struct *tsk, bool exec) { }
1520 static inline void perf_event_namespaces(struct task_struct *tsk) { }
1521 static inline void perf_event_fork(struct task_struct *tsk) { }
1522 static inline void perf_event_text_poke(const void *addr,
1523 const void *old_bytes,
1525 const void *new_bytes,
1527 static inline void perf_event_init(void) { }
1528 static inline int perf_swevent_get_recursion_context(void) { return -1; }
1529 static inline void perf_swevent_put_recursion_context(int rctx) { }
1530 static inline u64 perf_swevent_set_period(struct perf_event *event) { return 0; }
1531 static inline void perf_event_enable(struct perf_event *event) { }
1532 static inline void perf_event_disable(struct perf_event *event) { }
1533 static inline int __perf_event_disable(void *info) { return -1; }
1534 static inline void perf_event_task_tick(void) { }
1535 static inline int perf_event_release_kernel(struct perf_event *event) { return 0; }
1536 static inline int perf_event_period(struct perf_event *event, u64 value)
1540 static inline u64 perf_event_pause(struct perf_event *event, bool reset)
1546 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1547 extern void perf_restore_debug_store(void);
1549 static inline void perf_restore_debug_store(void) { }
1552 static __always_inline bool perf_raw_frag_last(const struct perf_raw_frag *frag)
1554 return frag->pad < sizeof(u64);
1557 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1559 struct perf_pmu_events_attr {
1560 struct device_attribute attr;
1562 const char *event_str;
1565 struct perf_pmu_events_ht_attr {
1566 struct device_attribute attr;
1568 const char *event_str_ht;
1569 const char *event_str_noht;
1572 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1575 #define PMU_EVENT_ATTR(_name, _var, _id, _show) \
1576 static struct perf_pmu_events_attr _var = { \
1577 .attr = __ATTR(_name, 0444, _show, NULL), \
1581 #define PMU_EVENT_ATTR_STRING(_name, _var, _str) \
1582 static struct perf_pmu_events_attr _var = { \
1583 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1585 .event_str = _str, \
1588 #define PMU_FORMAT_ATTR(_name, _format) \
1590 _name##_show(struct device *dev, \
1591 struct device_attribute *attr, \
1594 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
1595 return sprintf(page, _format "\n"); \
1598 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1600 /* Performance counter hotplug functions */
1601 #ifdef CONFIG_PERF_EVENTS
1602 int perf_event_init_cpu(unsigned int cpu);
1603 int perf_event_exit_cpu(unsigned int cpu);
1605 #define perf_event_init_cpu NULL
1606 #define perf_event_exit_cpu NULL
1609 extern void __weak arch_perf_update_userpage(struct perf_event *event,
1610 struct perf_event_mmap_page *userpg,
1613 #endif /* _LINUX_PERF_EVENT_H */