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 * The period we started this sample with.
221 * However much is left of the current period; note that this is
222 * a full 64bit value and allows for generation of periods longer
223 * than hardware might allow.
225 local64_t period_left;
228 * State for throttling the event, see __perf_event_overflow() and
229 * perf_adjust_freq_unthr_context().
235 * State for freq target events, see __perf_event_overflow() and
236 * perf_adjust_freq_unthr_context().
239 u64 freq_count_stamp;
246 * Common implementation detail of pmu::{start,commit,cancel}_txn
248 #define PERF_PMU_TXN_ADD 0x1 /* txn to add/schedule event on PMU */
249 #define PERF_PMU_TXN_READ 0x2 /* txn to read event group from PMU */
252 * pmu::capabilities flags
254 #define PERF_PMU_CAP_NO_INTERRUPT 0x01
255 #define PERF_PMU_CAP_NO_NMI 0x02
256 #define PERF_PMU_CAP_AUX_NO_SG 0x04
257 #define PERF_PMU_CAP_EXTENDED_REGS 0x08
258 #define PERF_PMU_CAP_EXCLUSIVE 0x10
259 #define PERF_PMU_CAP_ITRACE 0x20
260 #define PERF_PMU_CAP_HETEROGENEOUS_CPUS 0x40
261 #define PERF_PMU_CAP_NO_EXCLUDE 0x80
262 #define PERF_PMU_CAP_AUX_OUTPUT 0x100
264 struct perf_output_handle;
267 * struct pmu - generic performance monitoring unit
270 struct list_head entry;
272 struct module *module;
274 const struct attribute_group **attr_groups;
275 const struct attribute_group **attr_update;
280 * various common per-pmu feature flags
284 int __percpu *pmu_disable_count;
285 struct perf_cpu_context __percpu *pmu_cpu_context;
286 atomic_t exclusive_cnt; /* < 0: cpu; > 0: tsk */
288 int hrtimer_interval_ms;
290 /* number of address filters this PMU can do */
291 unsigned int nr_addr_filters;
294 * Fully disable/enable this PMU, can be used to protect from the PMI
295 * as well as for lazy/batch writing of the MSRs.
297 void (*pmu_enable) (struct pmu *pmu); /* optional */
298 void (*pmu_disable) (struct pmu *pmu); /* optional */
301 * Try and initialize the event for this PMU.
304 * -ENOENT -- @event is not for this PMU
306 * -ENODEV -- @event is for this PMU but PMU not present
307 * -EBUSY -- @event is for this PMU but PMU temporarily unavailable
308 * -EINVAL -- @event is for this PMU but @event is not valid
309 * -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
310 * -EACCES -- @event is for this PMU, @event is valid, but no privileges
312 * 0 -- @event is for this PMU and valid
314 * Other error return values are allowed.
316 int (*event_init) (struct perf_event *event);
319 * Notification that the event was mapped or unmapped. Called
320 * in the context of the mapping task.
322 void (*event_mapped) (struct perf_event *event, struct mm_struct *mm); /* optional */
323 void (*event_unmapped) (struct perf_event *event, struct mm_struct *mm); /* optional */
326 * Flags for ->add()/->del()/ ->start()/->stop(). There are
327 * matching hw_perf_event::state flags.
329 #define PERF_EF_START 0x01 /* start the counter when adding */
330 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
331 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
334 * Adds/Removes a counter to/from the PMU, can be done inside a
335 * transaction, see the ->*_txn() methods.
337 * The add/del callbacks will reserve all hardware resources required
338 * to service the event, this includes any counter constraint
341 * Called with IRQs disabled and the PMU disabled on the CPU the event
344 * ->add() called without PERF_EF_START should result in the same state
345 * as ->add() followed by ->stop().
347 * ->del() must always PERF_EF_UPDATE stop an event. If it calls
348 * ->stop() that must deal with already being stopped without
351 int (*add) (struct perf_event *event, int flags);
352 void (*del) (struct perf_event *event, int flags);
355 * Starts/Stops a counter present on the PMU.
357 * The PMI handler should stop the counter when perf_event_overflow()
358 * returns !0. ->start() will be used to continue.
360 * Also used to change the sample period.
362 * Called with IRQs disabled and the PMU disabled on the CPU the event
363 * is on -- will be called from NMI context with the PMU generates
366 * ->stop() with PERF_EF_UPDATE will read the counter and update
367 * period/count values like ->read() would.
369 * ->start() with PERF_EF_RELOAD will reprogram the the counter
370 * value, must be preceded by a ->stop() with PERF_EF_UPDATE.
372 void (*start) (struct perf_event *event, int flags);
373 void (*stop) (struct perf_event *event, int flags);
376 * Updates the counter value of the event.
378 * For sampling capable PMUs this will also update the software period
379 * hw_perf_event::period_left field.
381 void (*read) (struct perf_event *event);
384 * Group events scheduling is treated as a transaction, add
385 * group events as a whole and perform one schedulability test.
386 * If the test fails, roll back the whole group
388 * Start the transaction, after this ->add() doesn't need to
389 * do schedulability tests.
393 void (*start_txn) (struct pmu *pmu, unsigned int txn_flags);
395 * If ->start_txn() disabled the ->add() schedulability test
396 * then ->commit_txn() is required to perform one. On success
397 * the transaction is closed. On error the transaction is kept
398 * open until ->cancel_txn() is called.
402 int (*commit_txn) (struct pmu *pmu);
404 * Will cancel the transaction, assumes ->del() is called
405 * for each successful ->add() during the transaction.
409 void (*cancel_txn) (struct pmu *pmu);
412 * Will return the value for perf_event_mmap_page::index for this event,
413 * if no implementation is provided it will default to: event->hw.idx + 1.
415 int (*event_idx) (struct perf_event *event); /*optional */
418 * context-switches callback
420 void (*sched_task) (struct perf_event_context *ctx,
423 * PMU specific data size
425 size_t task_ctx_size;
428 * PMU specific parts of task perf event context (i.e. ctx->task_ctx_data)
429 * can be synchronized using this function. See Intel LBR callstack support
430 * implementation and Perf core context switch handling callbacks for usage
433 void (*swap_task_ctx) (struct perf_event_context *prev,
434 struct perf_event_context *next);
438 * Set up pmu-private data structures for an AUX area
440 void *(*setup_aux) (struct perf_event *event, void **pages,
441 int nr_pages, bool overwrite);
445 * Free pmu-private AUX data structures
447 void (*free_aux) (void *aux); /* optional */
450 * Take a snapshot of the AUX buffer without touching the event
451 * state, so that preempting ->start()/->stop() callbacks does
452 * not interfere with their logic. Called in PMI context.
454 * Returns the size of AUX data copied to the output handle.
458 long (*snapshot_aux) (struct perf_event *event,
459 struct perf_output_handle *handle,
463 * Validate address range filters: make sure the HW supports the
464 * requested configuration and number of filters; return 0 if the
465 * supplied filters are valid, -errno otherwise.
467 * Runs in the context of the ioctl()ing process and is not serialized
468 * with the rest of the PMU callbacks.
470 int (*addr_filters_validate) (struct list_head *filters);
474 * Synchronize address range filter configuration:
475 * translate hw-agnostic filters into hardware configuration in
476 * event::hw::addr_filters.
478 * Runs as a part of filter sync sequence that is done in ->start()
479 * callback by calling perf_event_addr_filters_sync().
481 * May (and should) traverse event::addr_filters::list, for which its
482 * caller provides necessary serialization.
484 void (*addr_filters_sync) (struct perf_event *event);
488 * Check if event can be used for aux_output purposes for
489 * events of this PMU.
491 * Runs from perf_event_open(). Should return 0 for "no match"
492 * or non-zero for "match".
494 int (*aux_output_match) (struct perf_event *event);
498 * Filter events for PMU-specific reasons.
500 int (*filter_match) (struct perf_event *event); /* optional */
503 * Check period value for PERF_EVENT_IOC_PERIOD ioctl.
505 int (*check_period) (struct perf_event *event, u64 value); /* optional */
508 enum perf_addr_filter_action_t {
509 PERF_ADDR_FILTER_ACTION_STOP = 0,
510 PERF_ADDR_FILTER_ACTION_START,
511 PERF_ADDR_FILTER_ACTION_FILTER,
515 * struct perf_addr_filter - address range filter definition
516 * @entry: event's filter list linkage
517 * @path: object file's path for file-based filters
518 * @offset: filter range offset
519 * @size: filter range size (size==0 means single address trigger)
520 * @action: filter/start/stop
522 * This is a hardware-agnostic filter configuration as specified by the user.
524 struct perf_addr_filter {
525 struct list_head entry;
527 unsigned long offset;
529 enum perf_addr_filter_action_t action;
533 * struct perf_addr_filters_head - container for address range filters
534 * @list: list of filters for this event
535 * @lock: spinlock that serializes accesses to the @list and event's
536 * (and its children's) filter generations.
537 * @nr_file_filters: number of file-based filters
539 * A child event will use parent's @list (and therefore @lock), so they are
540 * bundled together; see perf_event_addr_filters().
542 struct perf_addr_filters_head {
543 struct list_head list;
545 unsigned int nr_file_filters;
548 struct perf_addr_filter_range {
554 * enum perf_event_state - the states of an event:
556 enum perf_event_state {
557 PERF_EVENT_STATE_DEAD = -4,
558 PERF_EVENT_STATE_EXIT = -3,
559 PERF_EVENT_STATE_ERROR = -2,
560 PERF_EVENT_STATE_OFF = -1,
561 PERF_EVENT_STATE_INACTIVE = 0,
562 PERF_EVENT_STATE_ACTIVE = 1,
566 struct perf_sample_data;
568 typedef void (*perf_overflow_handler_t)(struct perf_event *,
569 struct perf_sample_data *,
570 struct pt_regs *regs);
573 * Event capabilities. For event_caps and groups caps.
575 * PERF_EV_CAP_SOFTWARE: Is a software event.
576 * PERF_EV_CAP_READ_ACTIVE_PKG: A CPU event (or cgroup event) that can be read
577 * from any CPU in the package where it is active.
579 #define PERF_EV_CAP_SOFTWARE BIT(0)
580 #define PERF_EV_CAP_READ_ACTIVE_PKG BIT(1)
582 #define SWEVENT_HLIST_BITS 8
583 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
585 struct swevent_hlist {
586 struct hlist_head heads[SWEVENT_HLIST_SIZE];
587 struct rcu_head rcu_head;
590 #define PERF_ATTACH_CONTEXT 0x01
591 #define PERF_ATTACH_GROUP 0x02
592 #define PERF_ATTACH_TASK 0x04
593 #define PERF_ATTACH_TASK_DATA 0x08
594 #define PERF_ATTACH_ITRACE 0x10
599 struct pmu_event_list {
601 struct list_head list;
604 #define for_each_sibling_event(sibling, event) \
605 if ((event)->group_leader == (event)) \
606 list_for_each_entry((sibling), &(event)->sibling_list, sibling_list)
609 * struct perf_event - performance event kernel representation:
612 #ifdef CONFIG_PERF_EVENTS
614 * entry onto perf_event_context::event_list;
615 * modifications require ctx->lock
616 * RCU safe iterations.
618 struct list_head event_entry;
621 * Locked for modification by both ctx->mutex and ctx->lock; holding
622 * either sufficies for read.
624 struct list_head sibling_list;
625 struct list_head active_list;
627 * Node on the pinned or flexible tree located at the event context;
629 struct rb_node group_node;
632 * We need storage to track the entries in perf_pmu_migrate_context; we
633 * cannot use the event_entry because of RCU and we want to keep the
634 * group in tact which avoids us using the other two entries.
636 struct list_head migrate_entry;
638 struct hlist_node hlist_entry;
639 struct list_head active_entry;
642 /* Not serialized. Only written during event initialization. */
644 /* The cumulative AND of all event_caps for events in this group. */
647 struct perf_event *group_leader;
651 enum perf_event_state state;
652 unsigned int attach_state;
654 atomic64_t child_count;
657 * These are the total time in nanoseconds that the event
658 * has been enabled (i.e. eligible to run, and the task has
659 * been scheduled in, if this is a per-task event)
660 * and running (scheduled onto the CPU), respectively.
662 u64 total_time_enabled;
663 u64 total_time_running;
667 * timestamp shadows the actual context timing but it can
668 * be safely used in NMI interrupt context. It reflects the
669 * context time as it was when the event was last scheduled in.
671 * ctx_time already accounts for ctx->timestamp. Therefore to
672 * compute ctx_time for a sample, simply add perf_clock().
676 struct perf_event_attr attr;
680 struct hw_perf_event hw;
682 struct perf_event_context *ctx;
683 atomic_long_t refcount;
686 * These accumulate total time (in nanoseconds) that children
687 * events have been enabled and running, respectively.
689 atomic64_t child_total_time_enabled;
690 atomic64_t child_total_time_running;
693 * Protect attach/detach and child_list:
695 struct mutex child_mutex;
696 struct list_head child_list;
697 struct perf_event *parent;
702 struct list_head owner_entry;
703 struct task_struct *owner;
706 struct mutex mmap_mutex;
709 struct perf_buffer *rb;
710 struct list_head rb_entry;
711 unsigned long rcu_batches;
715 wait_queue_head_t waitq;
716 struct fasync_struct *fasync;
718 /* delayed work for NMIs and such */
722 struct irq_work pending;
724 atomic_t event_limit;
726 /* address range filters */
727 struct perf_addr_filters_head addr_filters;
728 /* vma address array for file-based filders */
729 struct perf_addr_filter_range *addr_filter_ranges;
730 unsigned long addr_filters_gen;
732 /* for aux_output events */
733 struct perf_event *aux_event;
735 void (*destroy)(struct perf_event *);
736 struct rcu_head rcu_head;
738 struct pid_namespace *ns;
742 perf_overflow_handler_t overflow_handler;
743 void *overflow_handler_context;
744 #ifdef CONFIG_BPF_SYSCALL
745 perf_overflow_handler_t orig_overflow_handler;
746 struct bpf_prog *prog;
749 #ifdef CONFIG_EVENT_TRACING
750 struct trace_event_call *tp_event;
751 struct event_filter *filter;
752 #ifdef CONFIG_FUNCTION_TRACER
753 struct ftrace_ops ftrace_ops;
757 #ifdef CONFIG_CGROUP_PERF
758 struct perf_cgroup *cgrp; /* cgroup event is attach to */
761 #ifdef CONFIG_SECURITY
764 struct list_head sb_list;
765 #endif /* CONFIG_PERF_EVENTS */
769 struct perf_event_groups {
775 * struct perf_event_context - event context structure
777 * Used as a container for task events and CPU events as well:
779 struct perf_event_context {
782 * Protect the states of the events in the list,
783 * nr_active, and the list:
787 * Protect the list of events. Locking either mutex or lock
788 * is sufficient to ensure the list doesn't change; to change
789 * the list you need to lock both the mutex and the spinlock.
793 struct list_head active_ctx_list;
794 struct perf_event_groups pinned_groups;
795 struct perf_event_groups flexible_groups;
796 struct list_head event_list;
798 struct list_head pinned_active;
799 struct list_head flexible_active;
808 * Set when nr_events != nr_active, except tolerant to events not
809 * necessary to be active due to scheduling constraints, such as cgroups.
811 int rotate_necessary;
813 struct task_struct *task;
816 * Context clock, runs when context enabled.
822 * These fields let us detect when two contexts have both
823 * been cloned (inherited) from a common ancestor.
825 struct perf_event_context *parent_ctx;
829 #ifdef CONFIG_CGROUP_PERF
830 int nr_cgroups; /* cgroup evts */
832 void *task_ctx_data; /* pmu specific data */
833 struct rcu_head rcu_head;
837 * Number of contexts where an event can trigger:
838 * task, softirq, hardirq, nmi.
840 #define PERF_NR_CONTEXTS 4
843 * struct perf_event_cpu_context - per cpu event context structure
845 struct perf_cpu_context {
846 struct perf_event_context ctx;
847 struct perf_event_context *task_ctx;
851 raw_spinlock_t hrtimer_lock;
852 struct hrtimer hrtimer;
853 ktime_t hrtimer_interval;
854 unsigned int hrtimer_active;
856 #ifdef CONFIG_CGROUP_PERF
857 struct perf_cgroup *cgrp;
858 struct list_head cgrp_cpuctx_entry;
861 struct list_head sched_cb_entry;
866 * Per-CPU storage for iterators used in visit_groups_merge. The default
867 * storage is of size 2 to hold the CPU and any CPU event iterators.
870 struct perf_event **heap;
871 struct perf_event *heap_default[2];
874 struct perf_output_handle {
875 struct perf_event *event;
876 struct perf_buffer *rb;
877 unsigned long wakeup;
887 struct bpf_perf_event_data_kern {
888 bpf_user_pt_regs_t *regs;
889 struct perf_sample_data *data;
890 struct perf_event *event;
893 #ifdef CONFIG_CGROUP_PERF
896 * perf_cgroup_info keeps track of time_enabled for a cgroup.
897 * This is a per-cpu dynamically allocated data structure.
899 struct perf_cgroup_info {
905 struct cgroup_subsys_state css;
906 struct perf_cgroup_info __percpu *info;
910 * Must ensure cgroup is pinned (css_get) before calling
911 * this function. In other words, we cannot call this function
912 * if there is no cgroup event for the current CPU context.
914 static inline struct perf_cgroup *
915 perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
917 return container_of(task_css_check(task, perf_event_cgrp_id,
918 ctx ? lockdep_is_held(&ctx->lock)
920 struct perf_cgroup, css);
922 #endif /* CONFIG_CGROUP_PERF */
924 #ifdef CONFIG_PERF_EVENTS
926 extern void *perf_aux_output_begin(struct perf_output_handle *handle,
927 struct perf_event *event);
928 extern void perf_aux_output_end(struct perf_output_handle *handle,
930 extern int perf_aux_output_skip(struct perf_output_handle *handle,
932 extern void *perf_get_aux(struct perf_output_handle *handle);
933 extern void perf_aux_output_flag(struct perf_output_handle *handle, u64 flags);
934 extern void perf_event_itrace_started(struct perf_event *event);
936 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
937 extern void perf_pmu_unregister(struct pmu *pmu);
939 extern int perf_num_counters(void);
940 extern const char *perf_pmu_name(void);
941 extern void __perf_event_task_sched_in(struct task_struct *prev,
942 struct task_struct *task);
943 extern void __perf_event_task_sched_out(struct task_struct *prev,
944 struct task_struct *next);
945 extern int perf_event_init_task(struct task_struct *child);
946 extern void perf_event_exit_task(struct task_struct *child);
947 extern void perf_event_free_task(struct task_struct *task);
948 extern void perf_event_delayed_put(struct task_struct *task);
949 extern struct file *perf_event_get(unsigned int fd);
950 extern const struct perf_event *perf_get_event(struct file *file);
951 extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
952 extern void perf_event_print_debug(void);
953 extern void perf_pmu_disable(struct pmu *pmu);
954 extern void perf_pmu_enable(struct pmu *pmu);
955 extern void perf_sched_cb_dec(struct pmu *pmu);
956 extern void perf_sched_cb_inc(struct pmu *pmu);
957 extern int perf_event_task_disable(void);
958 extern int perf_event_task_enable(void);
960 extern void perf_pmu_resched(struct pmu *pmu);
962 extern int perf_event_refresh(struct perf_event *event, int refresh);
963 extern void perf_event_update_userpage(struct perf_event *event);
964 extern int perf_event_release_kernel(struct perf_event *event);
965 extern struct perf_event *
966 perf_event_create_kernel_counter(struct perf_event_attr *attr,
968 struct task_struct *task,
969 perf_overflow_handler_t callback,
971 extern void perf_pmu_migrate_context(struct pmu *pmu,
972 int src_cpu, int dst_cpu);
973 int perf_event_read_local(struct perf_event *event, u64 *value,
974 u64 *enabled, u64 *running);
975 extern u64 perf_event_read_value(struct perf_event *event,
976 u64 *enabled, u64 *running);
979 struct perf_sample_data {
981 * Fields set by perf_sample_data_init(), group so as to
982 * minimize the cachelines touched.
985 struct perf_raw_record *raw;
986 struct perf_branch_stack *br_stack;
990 union perf_mem_data_src data_src;
993 * The other fields, optionally {set,used} by
994 * perf_{prepare,output}_sample().
1009 struct perf_callchain_entry *callchain;
1013 * regs_user may point to task_pt_regs or to regs_user_copy, depending
1016 struct perf_regs regs_user;
1017 struct pt_regs regs_user_copy;
1019 struct perf_regs regs_intr;
1020 u64 stack_user_size;
1024 } ____cacheline_aligned;
1026 /* default value for data source */
1027 #define PERF_MEM_NA (PERF_MEM_S(OP, NA) |\
1028 PERF_MEM_S(LVL, NA) |\
1029 PERF_MEM_S(SNOOP, NA) |\
1030 PERF_MEM_S(LOCK, NA) |\
1031 PERF_MEM_S(TLB, NA))
1033 static inline void perf_sample_data_init(struct perf_sample_data *data,
1034 u64 addr, u64 period)
1036 /* remaining struct members initialized in perf_prepare_sample() */
1039 data->br_stack = NULL;
1040 data->period = period;
1042 data->data_src.val = PERF_MEM_NA;
1046 extern void perf_output_sample(struct perf_output_handle *handle,
1047 struct perf_event_header *header,
1048 struct perf_sample_data *data,
1049 struct perf_event *event);
1050 extern void perf_prepare_sample(struct perf_event_header *header,
1051 struct perf_sample_data *data,
1052 struct perf_event *event,
1053 struct pt_regs *regs);
1055 extern int perf_event_overflow(struct perf_event *event,
1056 struct perf_sample_data *data,
1057 struct pt_regs *regs);
1059 extern void perf_event_output_forward(struct perf_event *event,
1060 struct perf_sample_data *data,
1061 struct pt_regs *regs);
1062 extern void perf_event_output_backward(struct perf_event *event,
1063 struct perf_sample_data *data,
1064 struct pt_regs *regs);
1065 extern int perf_event_output(struct perf_event *event,
1066 struct perf_sample_data *data,
1067 struct pt_regs *regs);
1070 is_default_overflow_handler(struct perf_event *event)
1072 if (likely(event->overflow_handler == perf_event_output_forward))
1074 if (unlikely(event->overflow_handler == perf_event_output_backward))
1080 perf_event_header__init_id(struct perf_event_header *header,
1081 struct perf_sample_data *data,
1082 struct perf_event *event);
1084 perf_event__output_id_sample(struct perf_event *event,
1085 struct perf_output_handle *handle,
1086 struct perf_sample_data *sample);
1089 perf_log_lost_samples(struct perf_event *event, u64 lost);
1091 static inline bool event_has_any_exclude_flag(struct perf_event *event)
1093 struct perf_event_attr *attr = &event->attr;
1095 return attr->exclude_idle || attr->exclude_user ||
1096 attr->exclude_kernel || attr->exclude_hv ||
1097 attr->exclude_guest || attr->exclude_host;
1100 static inline bool is_sampling_event(struct perf_event *event)
1102 return event->attr.sample_period != 0;
1106 * Return 1 for a software event, 0 for a hardware event
1108 static inline int is_software_event(struct perf_event *event)
1110 return event->event_caps & PERF_EV_CAP_SOFTWARE;
1114 * Return 1 for event in sw context, 0 for event in hw context
1116 static inline int in_software_context(struct perf_event *event)
1118 return event->ctx->pmu->task_ctx_nr == perf_sw_context;
1121 static inline int is_exclusive_pmu(struct pmu *pmu)
1123 return pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE;
1126 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
1128 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
1129 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
1131 #ifndef perf_arch_fetch_caller_regs
1132 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
1136 * When generating a perf sample in-line, instead of from an interrupt /
1137 * exception, we lack a pt_regs. This is typically used from software events
1138 * like: SW_CONTEXT_SWITCHES, SW_MIGRATIONS and the tie-in with tracepoints.
1140 * We typically don't need a full set, but (for x86) do require:
1141 * - ip for PERF_SAMPLE_IP
1142 * - cs for user_mode() tests
1143 * - sp for PERF_SAMPLE_CALLCHAIN
1144 * - eflags for MISC bits and CALLCHAIN (see: perf_hw_regs())
1146 * NOTE: assumes @regs is otherwise already 0 filled; this is important for
1147 * things like PERF_SAMPLE_REGS_INTR.
1149 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1151 perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1154 static __always_inline void
1155 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1157 if (static_key_false(&perf_swevent_enabled[event_id]))
1158 __perf_sw_event(event_id, nr, regs, addr);
1161 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
1164 * 'Special' version for the scheduler, it hard assumes no recursion,
1165 * which is guaranteed by us not actually scheduling inside other swevents
1166 * because those disable preemption.
1168 static __always_inline void
1169 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
1171 if (static_key_false(&perf_swevent_enabled[event_id])) {
1172 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1174 perf_fetch_caller_regs(regs);
1175 ___perf_sw_event(event_id, nr, regs, addr);
1179 extern struct static_key_false perf_sched_events;
1181 static __always_inline bool
1182 perf_sw_migrate_enabled(void)
1184 if (static_key_false(&perf_swevent_enabled[PERF_COUNT_SW_CPU_MIGRATIONS]))
1189 static inline void perf_event_task_migrate(struct task_struct *task)
1191 if (perf_sw_migrate_enabled())
1192 task->sched_migrated = 1;
1195 static inline void perf_event_task_sched_in(struct task_struct *prev,
1196 struct task_struct *task)
1198 if (static_branch_unlikely(&perf_sched_events))
1199 __perf_event_task_sched_in(prev, task);
1201 if (perf_sw_migrate_enabled() && task->sched_migrated) {
1202 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1204 perf_fetch_caller_regs(regs);
1205 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, regs, 0);
1206 task->sched_migrated = 0;
1210 static inline void perf_event_task_sched_out(struct task_struct *prev,
1211 struct task_struct *next)
1213 perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
1215 if (static_branch_unlikely(&perf_sched_events))
1216 __perf_event_task_sched_out(prev, next);
1219 extern void perf_event_mmap(struct vm_area_struct *vma);
1221 extern void perf_event_ksymbol(u16 ksym_type, u64 addr, u32 len,
1222 bool unregister, const char *sym);
1223 extern void perf_event_bpf_event(struct bpf_prog *prog,
1224 enum perf_bpf_event_type type,
1227 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1228 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1229 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1231 extern void perf_event_exec(void);
1232 extern void perf_event_comm(struct task_struct *tsk, bool exec);
1233 extern void perf_event_namespaces(struct task_struct *tsk);
1234 extern void perf_event_fork(struct task_struct *tsk);
1237 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1239 extern void perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1240 extern void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1241 extern struct perf_callchain_entry *
1242 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
1243 u32 max_stack, bool crosstask, bool add_mark);
1244 extern struct perf_callchain_entry *perf_callchain(struct perf_event *event, struct pt_regs *regs);
1245 extern int get_callchain_buffers(int max_stack);
1246 extern void put_callchain_buffers(void);
1248 extern int sysctl_perf_event_max_stack;
1249 extern int sysctl_perf_event_max_contexts_per_stack;
1251 static inline int perf_callchain_store_context(struct perf_callchain_entry_ctx *ctx, u64 ip)
1253 if (ctx->contexts < sysctl_perf_event_max_contexts_per_stack) {
1254 struct perf_callchain_entry *entry = ctx->entry;
1255 entry->ip[entry->nr++] = ip;
1259 ctx->contexts_maxed = true;
1260 return -1; /* no more room, stop walking the stack */
1264 static inline int perf_callchain_store(struct perf_callchain_entry_ctx *ctx, u64 ip)
1266 if (ctx->nr < ctx->max_stack && !ctx->contexts_maxed) {
1267 struct perf_callchain_entry *entry = ctx->entry;
1268 entry->ip[entry->nr++] = ip;
1272 return -1; /* no more room, stop walking the stack */
1276 extern int sysctl_perf_event_paranoid;
1277 extern int sysctl_perf_event_mlock;
1278 extern int sysctl_perf_event_sample_rate;
1279 extern int sysctl_perf_cpu_time_max_percent;
1281 extern void perf_sample_event_took(u64 sample_len_ns);
1283 int perf_proc_update_handler(struct ctl_table *table, int write,
1284 void *buffer, size_t *lenp, loff_t *ppos);
1285 int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
1286 void *buffer, size_t *lenp, loff_t *ppos);
1287 int perf_event_max_stack_handler(struct ctl_table *table, int write,
1288 void *buffer, size_t *lenp, loff_t *ppos);
1290 /* Access to perf_event_open(2) syscall. */
1291 #define PERF_SECURITY_OPEN 0
1293 /* Finer grained perf_event_open(2) access control. */
1294 #define PERF_SECURITY_CPU 1
1295 #define PERF_SECURITY_KERNEL 2
1296 #define PERF_SECURITY_TRACEPOINT 3
1298 static inline int perf_is_paranoid(void)
1300 return sysctl_perf_event_paranoid > -1;
1303 static inline int perf_allow_kernel(struct perf_event_attr *attr)
1305 if (sysctl_perf_event_paranoid > 1 && !perfmon_capable())
1308 return security_perf_event_open(attr, PERF_SECURITY_KERNEL);
1311 static inline int perf_allow_cpu(struct perf_event_attr *attr)
1313 if (sysctl_perf_event_paranoid > 0 && !perfmon_capable())
1316 return security_perf_event_open(attr, PERF_SECURITY_CPU);
1319 static inline int perf_allow_tracepoint(struct perf_event_attr *attr)
1321 if (sysctl_perf_event_paranoid > -1 && !perfmon_capable())
1324 return security_perf_event_open(attr, PERF_SECURITY_TRACEPOINT);
1327 extern void perf_event_init(void);
1328 extern void perf_tp_event(u16 event_type, u64 count, void *record,
1329 int entry_size, struct pt_regs *regs,
1330 struct hlist_head *head, int rctx,
1331 struct task_struct *task);
1332 extern void perf_bp_event(struct perf_event *event, void *data);
1334 #ifndef perf_misc_flags
1335 # define perf_misc_flags(regs) \
1336 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1337 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1339 #ifndef perf_arch_bpf_user_pt_regs
1340 # define perf_arch_bpf_user_pt_regs(regs) regs
1343 static inline bool has_branch_stack(struct perf_event *event)
1345 return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1348 static inline bool needs_branch_stack(struct perf_event *event)
1350 return event->attr.branch_sample_type != 0;
1353 static inline bool has_aux(struct perf_event *event)
1355 return event->pmu->setup_aux;
1358 static inline bool is_write_backward(struct perf_event *event)
1360 return !!event->attr.write_backward;
1363 static inline bool has_addr_filter(struct perf_event *event)
1365 return event->pmu->nr_addr_filters;
1369 * An inherited event uses parent's filters
1371 static inline struct perf_addr_filters_head *
1372 perf_event_addr_filters(struct perf_event *event)
1374 struct perf_addr_filters_head *ifh = &event->addr_filters;
1377 ifh = &event->parent->addr_filters;
1382 extern void perf_event_addr_filters_sync(struct perf_event *event);
1384 extern int perf_output_begin(struct perf_output_handle *handle,
1385 struct perf_event *event, unsigned int size);
1386 extern int perf_output_begin_forward(struct perf_output_handle *handle,
1387 struct perf_event *event,
1389 extern int perf_output_begin_backward(struct perf_output_handle *handle,
1390 struct perf_event *event,
1393 extern void perf_output_end(struct perf_output_handle *handle);
1394 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
1395 const void *buf, unsigned int len);
1396 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
1398 extern long perf_output_copy_aux(struct perf_output_handle *aux_handle,
1399 struct perf_output_handle *handle,
1400 unsigned long from, unsigned long to);
1401 extern int perf_swevent_get_recursion_context(void);
1402 extern void perf_swevent_put_recursion_context(int rctx);
1403 extern u64 perf_swevent_set_period(struct perf_event *event);
1404 extern void perf_event_enable(struct perf_event *event);
1405 extern void perf_event_disable(struct perf_event *event);
1406 extern void perf_event_disable_local(struct perf_event *event);
1407 extern void perf_event_disable_inatomic(struct perf_event *event);
1408 extern void perf_event_task_tick(void);
1409 extern int perf_event_account_interrupt(struct perf_event *event);
1410 extern int perf_event_period(struct perf_event *event, u64 value);
1411 extern u64 perf_event_pause(struct perf_event *event, bool reset);
1412 #else /* !CONFIG_PERF_EVENTS: */
1413 static inline void *
1414 perf_aux_output_begin(struct perf_output_handle *handle,
1415 struct perf_event *event) { return NULL; }
1417 perf_aux_output_end(struct perf_output_handle *handle, unsigned long size)
1420 perf_aux_output_skip(struct perf_output_handle *handle,
1421 unsigned long size) { return -EINVAL; }
1422 static inline void *
1423 perf_get_aux(struct perf_output_handle *handle) { return NULL; }
1425 perf_event_task_migrate(struct task_struct *task) { }
1427 perf_event_task_sched_in(struct task_struct *prev,
1428 struct task_struct *task) { }
1430 perf_event_task_sched_out(struct task_struct *prev,
1431 struct task_struct *next) { }
1432 static inline int perf_event_init_task(struct task_struct *child) { return 0; }
1433 static inline void perf_event_exit_task(struct task_struct *child) { }
1434 static inline void perf_event_free_task(struct task_struct *task) { }
1435 static inline void perf_event_delayed_put(struct task_struct *task) { }
1436 static inline struct file *perf_event_get(unsigned int fd) { return ERR_PTR(-EINVAL); }
1437 static inline const struct perf_event *perf_get_event(struct file *file)
1439 return ERR_PTR(-EINVAL);
1441 static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
1443 return ERR_PTR(-EINVAL);
1445 static inline int perf_event_read_local(struct perf_event *event, u64 *value,
1446 u64 *enabled, u64 *running)
1450 static inline void perf_event_print_debug(void) { }
1451 static inline int perf_event_task_disable(void) { return -EINVAL; }
1452 static inline int perf_event_task_enable(void) { return -EINVAL; }
1453 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1459 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) { }
1461 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr) { }
1463 perf_bp_event(struct perf_event *event, void *data) { }
1465 static inline int perf_register_guest_info_callbacks
1466 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1467 static inline int perf_unregister_guest_info_callbacks
1468 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1470 static inline void perf_event_mmap(struct vm_area_struct *vma) { }
1472 typedef int (perf_ksymbol_get_name_f)(char *name, int name_len, void *data);
1473 static inline void perf_event_ksymbol(u16 ksym_type, u64 addr, u32 len,
1474 bool unregister, const char *sym) { }
1475 static inline void perf_event_bpf_event(struct bpf_prog *prog,
1476 enum perf_bpf_event_type type,
1478 static inline void perf_event_exec(void) { }
1479 static inline void perf_event_comm(struct task_struct *tsk, bool exec) { }
1480 static inline void perf_event_namespaces(struct task_struct *tsk) { }
1481 static inline void perf_event_fork(struct task_struct *tsk) { }
1482 static inline void perf_event_init(void) { }
1483 static inline int perf_swevent_get_recursion_context(void) { return -1; }
1484 static inline void perf_swevent_put_recursion_context(int rctx) { }
1485 static inline u64 perf_swevent_set_period(struct perf_event *event) { return 0; }
1486 static inline void perf_event_enable(struct perf_event *event) { }
1487 static inline void perf_event_disable(struct perf_event *event) { }
1488 static inline int __perf_event_disable(void *info) { return -1; }
1489 static inline void perf_event_task_tick(void) { }
1490 static inline int perf_event_release_kernel(struct perf_event *event) { return 0; }
1491 static inline int perf_event_period(struct perf_event *event, u64 value)
1495 static inline u64 perf_event_pause(struct perf_event *event, bool reset)
1501 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1502 extern void perf_restore_debug_store(void);
1504 static inline void perf_restore_debug_store(void) { }
1507 static __always_inline bool perf_raw_frag_last(const struct perf_raw_frag *frag)
1509 return frag->pad < sizeof(u64);
1512 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1514 struct perf_pmu_events_attr {
1515 struct device_attribute attr;
1517 const char *event_str;
1520 struct perf_pmu_events_ht_attr {
1521 struct device_attribute attr;
1523 const char *event_str_ht;
1524 const char *event_str_noht;
1527 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1530 #define PMU_EVENT_ATTR(_name, _var, _id, _show) \
1531 static struct perf_pmu_events_attr _var = { \
1532 .attr = __ATTR(_name, 0444, _show, NULL), \
1536 #define PMU_EVENT_ATTR_STRING(_name, _var, _str) \
1537 static struct perf_pmu_events_attr _var = { \
1538 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1540 .event_str = _str, \
1543 #define PMU_FORMAT_ATTR(_name, _format) \
1545 _name##_show(struct device *dev, \
1546 struct device_attribute *attr, \
1549 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
1550 return sprintf(page, _format "\n"); \
1553 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1555 /* Performance counter hotplug functions */
1556 #ifdef CONFIG_PERF_EVENTS
1557 int perf_event_init_cpu(unsigned int cpu);
1558 int perf_event_exit_cpu(unsigned int cpu);
1560 #define perf_event_init_cpu NULL
1561 #define perf_event_exit_cpu NULL
1564 extern void __weak arch_perf_update_userpage(struct perf_event *event,
1565 struct perf_event_mmap_page *userpg,
1568 #endif /* _LINUX_PERF_EVENT_H */