4 * Copyright (C) 2008-2009, Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2011, Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2011, Red Hat, Inc., Peter Zijlstra
8 * Data type definitions, declarations, prototypes.
10 * Started by: Thomas Gleixner and Ingo Molnar
12 * For licencing details see kernel-base/COPYING
14 #ifndef _LINUX_PERF_EVENT_H
15 #define _LINUX_PERF_EVENT_H
17 #include <uapi/linux/perf_event.h>
20 * Kernel-internal data types and definitions:
23 #ifdef CONFIG_PERF_EVENTS
24 # include <asm/perf_event.h>
25 # include <asm/local64.h>
28 struct perf_guest_info_callbacks {
29 int (*is_in_guest)(void);
30 int (*is_user_mode)(void);
31 unsigned long (*get_guest_ip)(void);
34 #ifdef CONFIG_HAVE_HW_BREAKPOINT
35 #include <asm/hw_breakpoint.h>
38 #include <linux/list.h>
39 #include <linux/mutex.h>
40 #include <linux/rculist.h>
41 #include <linux/rcupdate.h>
42 #include <linux/spinlock.h>
43 #include <linux/hrtimer.h>
45 #include <linux/pid_namespace.h>
46 #include <linux/workqueue.h>
47 #include <linux/ftrace.h>
48 #include <linux/cpu.h>
49 #include <linux/irq_work.h>
50 #include <linux/static_key.h>
51 #include <linux/jump_label_ratelimit.h>
52 #include <linux/atomic.h>
53 #include <linux/sysfs.h>
54 #include <linux/perf_regs.h>
55 #include <linux/workqueue.h>
56 #include <linux/cgroup.h>
57 #include <asm/local.h>
59 struct perf_callchain_entry {
61 __u64 ip[0]; /* /proc/sys/kernel/perf_event_max_stack */
64 struct perf_callchain_entry_ctx {
65 struct perf_callchain_entry *entry;
72 typedef unsigned long (*perf_copy_f)(void *dst, const void *src,
73 unsigned long off, unsigned long len);
75 struct perf_raw_frag {
77 struct perf_raw_frag *next;
85 struct perf_raw_record {
86 struct perf_raw_frag frag;
91 * branch stack layout:
92 * nr: number of taken branches stored in entries[]
94 * Note that nr can vary from sample to sample
95 * branches (to, from) are stored from most recent
96 * to least recent, i.e., entries[0] contains the most
99 struct perf_branch_stack {
101 struct perf_branch_entry entries[0];
107 * extra PMU register associated with an event
109 struct hw_perf_event_extra {
110 u64 config; /* register value */
111 unsigned int reg; /* register address or index */
112 int alloc; /* extra register already allocated */
113 int idx; /* index in shared_regs->regs[] */
117 * struct hw_perf_event - performance event hardware details:
119 struct hw_perf_event {
120 #ifdef CONFIG_PERF_EVENTS
122 struct { /* hardware */
125 unsigned long config_base;
126 unsigned long event_base;
127 int event_base_rdpmc;
132 struct hw_perf_event_extra extra_reg;
133 struct hw_perf_event_extra branch_reg;
135 struct { /* software */
136 struct hrtimer hrtimer;
138 struct { /* tracepoint */
139 /* for tp_event->class */
140 struct list_head tp_list;
142 struct { /* intel_cqm */
146 struct list_head cqm_events_entry;
147 struct list_head cqm_groups_entry;
148 struct list_head cqm_group_entry;
150 struct { /* itrace */
153 struct { /* amd_power */
157 #ifdef CONFIG_HAVE_HW_BREAKPOINT
158 struct { /* breakpoint */
160 * Crufty hack to avoid the chicken and egg
161 * problem hw_breakpoint has with context
162 * creation and event initalization.
164 struct arch_hw_breakpoint info;
165 struct list_head bp_list;
168 struct { /* amd_iommu */
177 * If the event is a per task event, this will point to the task in
178 * question. See the comment in perf_event_alloc().
180 struct task_struct *target;
183 * PMU would store hardware filter configuration
188 /* Last sync'ed generation of filters */
189 unsigned long addr_filters_gen;
192 * hw_perf_event::state flags; used to track the PERF_EF_* state.
194 #define PERF_HES_STOPPED 0x01 /* the counter is stopped */
195 #define PERF_HES_UPTODATE 0x02 /* event->count up-to-date */
196 #define PERF_HES_ARCH 0x04
201 * The last observed hardware counter value, updated with a
202 * local64_cmpxchg() such that pmu::read() can be called nested.
204 local64_t prev_count;
207 * The period to start the next sample with.
212 * The period we started this sample with.
217 * However much is left of the current period; note that this is
218 * a full 64bit value and allows for generation of periods longer
219 * than hardware might allow.
221 local64_t period_left;
224 * State for throttling the event, see __perf_event_overflow() and
225 * perf_adjust_freq_unthr_context().
231 * State for freq target events, see __perf_event_overflow() and
232 * perf_adjust_freq_unthr_context().
235 u64 freq_count_stamp;
242 * Common implementation detail of pmu::{start,commit,cancel}_txn
244 #define PERF_PMU_TXN_ADD 0x1 /* txn to add/schedule event on PMU */
245 #define PERF_PMU_TXN_READ 0x2 /* txn to read event group from PMU */
248 * pmu::capabilities flags
250 #define PERF_PMU_CAP_NO_INTERRUPT 0x01
251 #define PERF_PMU_CAP_NO_NMI 0x02
252 #define PERF_PMU_CAP_AUX_NO_SG 0x04
253 #define PERF_PMU_CAP_AUX_SW_DOUBLEBUF 0x08
254 #define PERF_PMU_CAP_EXCLUSIVE 0x10
255 #define PERF_PMU_CAP_ITRACE 0x20
256 #define PERF_PMU_CAP_HETEROGENEOUS_CPUS 0x40
259 * struct pmu - generic performance monitoring unit
262 struct list_head entry;
264 struct module *module;
266 const struct attribute_group **attr_groups;
271 * various common per-pmu feature flags
275 int * __percpu pmu_disable_count;
276 struct perf_cpu_context * __percpu pmu_cpu_context;
277 atomic_t exclusive_cnt; /* < 0: cpu; > 0: tsk */
279 int hrtimer_interval_ms;
281 /* number of address filters this PMU can do */
282 unsigned int nr_addr_filters;
285 * Fully disable/enable this PMU, can be used to protect from the PMI
286 * as well as for lazy/batch writing of the MSRs.
288 void (*pmu_enable) (struct pmu *pmu); /* optional */
289 void (*pmu_disable) (struct pmu *pmu); /* optional */
292 * Try and initialize the event for this PMU.
295 * -ENOENT -- @event is not for this PMU
297 * -ENODEV -- @event is for this PMU but PMU not present
298 * -EBUSY -- @event is for this PMU but PMU temporarily unavailable
299 * -EINVAL -- @event is for this PMU but @event is not valid
300 * -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
301 * -EACCESS -- @event is for this PMU, @event is valid, but no privilidges
303 * 0 -- @event is for this PMU and valid
305 * Other error return values are allowed.
307 int (*event_init) (struct perf_event *event);
310 * Notification that the event was mapped or unmapped. Called
311 * in the context of the mapping task.
313 void (*event_mapped) (struct perf_event *event); /*optional*/
314 void (*event_unmapped) (struct perf_event *event); /*optional*/
317 * Flags for ->add()/->del()/ ->start()/->stop(). There are
318 * matching hw_perf_event::state flags.
320 #define PERF_EF_START 0x01 /* start the counter when adding */
321 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
322 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
325 * Adds/Removes a counter to/from the PMU, can be done inside a
326 * transaction, see the ->*_txn() methods.
328 * The add/del callbacks will reserve all hardware resources required
329 * to service the event, this includes any counter constraint
332 * Called with IRQs disabled and the PMU disabled on the CPU the event
335 * ->add() called without PERF_EF_START should result in the same state
336 * as ->add() followed by ->stop().
338 * ->del() must always PERF_EF_UPDATE stop an event. If it calls
339 * ->stop() that must deal with already being stopped without
342 int (*add) (struct perf_event *event, int flags);
343 void (*del) (struct perf_event *event, int flags);
346 * Starts/Stops a counter present on the PMU.
348 * The PMI handler should stop the counter when perf_event_overflow()
349 * returns !0. ->start() will be used to continue.
351 * Also used to change the sample period.
353 * Called with IRQs disabled and the PMU disabled on the CPU the event
354 * is on -- will be called from NMI context with the PMU generates
357 * ->stop() with PERF_EF_UPDATE will read the counter and update
358 * period/count values like ->read() would.
360 * ->start() with PERF_EF_RELOAD will reprogram the the counter
361 * value, must be preceded by a ->stop() with PERF_EF_UPDATE.
363 void (*start) (struct perf_event *event, int flags);
364 void (*stop) (struct perf_event *event, int flags);
367 * Updates the counter value of the event.
369 * For sampling capable PMUs this will also update the software period
370 * hw_perf_event::period_left field.
372 void (*read) (struct perf_event *event);
375 * Group events scheduling is treated as a transaction, add
376 * group events as a whole and perform one schedulability test.
377 * If the test fails, roll back the whole group
379 * Start the transaction, after this ->add() doesn't need to
380 * do schedulability tests.
384 void (*start_txn) (struct pmu *pmu, unsigned int txn_flags);
386 * If ->start_txn() disabled the ->add() schedulability test
387 * then ->commit_txn() is required to perform one. On success
388 * the transaction is closed. On error the transaction is kept
389 * open until ->cancel_txn() is called.
393 int (*commit_txn) (struct pmu *pmu);
395 * Will cancel the transaction, assumes ->del() is called
396 * for each successful ->add() during the transaction.
400 void (*cancel_txn) (struct pmu *pmu);
403 * Will return the value for perf_event_mmap_page::index for this event,
404 * if no implementation is provided it will default to: event->hw.idx + 1.
406 int (*event_idx) (struct perf_event *event); /*optional */
409 * context-switches callback
411 void (*sched_task) (struct perf_event_context *ctx,
414 * PMU specific data size
416 size_t task_ctx_size;
420 * Return the count value for a counter.
422 u64 (*count) (struct perf_event *event); /*optional*/
425 * Set up pmu-private data structures for an AUX area
427 void *(*setup_aux) (int cpu, void **pages,
428 int nr_pages, bool overwrite);
432 * Free pmu-private AUX data structures
434 void (*free_aux) (void *aux); /* optional */
437 * Validate address range filters: make sure the HW supports the
438 * requested configuration and number of filters; return 0 if the
439 * supplied filters are valid, -errno otherwise.
441 * Runs in the context of the ioctl()ing process and is not serialized
442 * with the rest of the PMU callbacks.
444 int (*addr_filters_validate) (struct list_head *filters);
448 * Synchronize address range filter configuration:
449 * translate hw-agnostic filters into hardware configuration in
450 * event::hw::addr_filters.
452 * Runs as a part of filter sync sequence that is done in ->start()
453 * callback by calling perf_event_addr_filters_sync().
455 * May (and should) traverse event::addr_filters::list, for which its
456 * caller provides necessary serialization.
458 void (*addr_filters_sync) (struct perf_event *event);
462 * Filter events for PMU-specific reasons.
464 int (*filter_match) (struct perf_event *event); /* optional */
468 * struct perf_addr_filter - address range filter definition
469 * @entry: event's filter list linkage
470 * @inode: object file's inode for file-based filters
471 * @offset: filter range offset
472 * @size: filter range size
473 * @range: 1: range, 0: address
474 * @filter: 1: filter/start, 0: stop
476 * This is a hardware-agnostic filter configuration as specified by the user.
478 struct perf_addr_filter {
479 struct list_head entry;
481 unsigned long offset;
483 unsigned int range : 1,
488 * struct perf_addr_filters_head - container for address range filters
489 * @list: list of filters for this event
490 * @lock: spinlock that serializes accesses to the @list and event's
491 * (and its children's) filter generations.
492 * @nr_file_filters: number of file-based filters
494 * A child event will use parent's @list (and therefore @lock), so they are
495 * bundled together; see perf_event_addr_filters().
497 struct perf_addr_filters_head {
498 struct list_head list;
500 unsigned int nr_file_filters;
504 * enum perf_event_active_state - the states of a event
506 enum perf_event_active_state {
507 PERF_EVENT_STATE_DEAD = -4,
508 PERF_EVENT_STATE_EXIT = -3,
509 PERF_EVENT_STATE_ERROR = -2,
510 PERF_EVENT_STATE_OFF = -1,
511 PERF_EVENT_STATE_INACTIVE = 0,
512 PERF_EVENT_STATE_ACTIVE = 1,
516 struct perf_sample_data;
518 typedef void (*perf_overflow_handler_t)(struct perf_event *,
519 struct perf_sample_data *,
520 struct pt_regs *regs);
523 * Event capabilities. For event_caps and groups caps.
525 * PERF_EV_CAP_SOFTWARE: Is a software event.
526 * PERF_EV_CAP_READ_ACTIVE_PKG: A CPU event (or cgroup event) that can be read
527 * from any CPU in the package where it is active.
529 #define PERF_EV_CAP_SOFTWARE BIT(0)
530 #define PERF_EV_CAP_READ_ACTIVE_PKG BIT(1)
532 #define SWEVENT_HLIST_BITS 8
533 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
535 struct swevent_hlist {
536 struct hlist_head heads[SWEVENT_HLIST_SIZE];
537 struct rcu_head rcu_head;
540 #define PERF_ATTACH_CONTEXT 0x01
541 #define PERF_ATTACH_GROUP 0x02
542 #define PERF_ATTACH_TASK 0x04
543 #define PERF_ATTACH_TASK_DATA 0x08
548 struct pmu_event_list {
550 struct list_head list;
554 * struct perf_event - performance event kernel representation:
557 #ifdef CONFIG_PERF_EVENTS
559 * entry onto perf_event_context::event_list;
560 * modifications require ctx->lock
561 * RCU safe iterations.
563 struct list_head event_entry;
566 * XXX: group_entry and sibling_list should be mutually exclusive;
567 * either you're a sibling on a group, or you're the group leader.
568 * Rework the code to always use the same list element.
570 * Locked for modification by both ctx->mutex and ctx->lock; holding
571 * either sufficies for read.
573 struct list_head group_entry;
574 struct list_head sibling_list;
577 * We need storage to track the entries in perf_pmu_migrate_context; we
578 * cannot use the event_entry because of RCU and we want to keep the
579 * group in tact which avoids us using the other two entries.
581 struct list_head migrate_entry;
583 struct hlist_node hlist_entry;
584 struct list_head active_entry;
587 /* Not serialized. Only written during event initialization. */
589 /* The cumulative AND of all event_caps for events in this group. */
592 struct perf_event *group_leader;
596 enum perf_event_active_state state;
597 unsigned int attach_state;
599 atomic64_t child_count;
602 * These are the total time in nanoseconds that the event
603 * has been enabled (i.e. eligible to run, and the task has
604 * been scheduled in, if this is a per-task event)
605 * and running (scheduled onto the CPU), respectively.
607 * They are computed from tstamp_enabled, tstamp_running and
608 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
610 u64 total_time_enabled;
611 u64 total_time_running;
614 * These are timestamps used for computing total_time_enabled
615 * and total_time_running when the event is in INACTIVE or
616 * ACTIVE state, measured in nanoseconds from an arbitrary point
618 * tstamp_enabled: the notional time when the event was enabled
619 * tstamp_running: the notional time when the event was scheduled on
620 * tstamp_stopped: in INACTIVE state, the notional time when the
621 * event was scheduled off.
628 * timestamp shadows the actual context timing but it can
629 * be safely used in NMI interrupt context. It reflects the
630 * context time as it was when the event was last scheduled in.
632 * ctx_time already accounts for ctx->timestamp. Therefore to
633 * compute ctx_time for a sample, simply add perf_clock().
637 struct perf_event_attr attr;
641 struct hw_perf_event hw;
643 struct perf_event_context *ctx;
644 atomic_long_t refcount;
647 * These accumulate total time (in nanoseconds) that children
648 * events have been enabled and running, respectively.
650 atomic64_t child_total_time_enabled;
651 atomic64_t child_total_time_running;
654 * Protect attach/detach and child_list:
656 struct mutex child_mutex;
657 struct list_head child_list;
658 struct perf_event *parent;
663 struct list_head owner_entry;
664 struct task_struct *owner;
667 struct mutex mmap_mutex;
670 struct ring_buffer *rb;
671 struct list_head rb_entry;
672 unsigned long rcu_batches;
676 wait_queue_head_t waitq;
677 struct fasync_struct *fasync;
679 /* delayed work for NMIs and such */
683 struct irq_work pending;
685 atomic_t event_limit;
687 /* address range filters */
688 struct perf_addr_filters_head addr_filters;
689 /* vma address array for file-based filders */
690 unsigned long *addr_filters_offs;
691 unsigned long addr_filters_gen;
693 void (*destroy)(struct perf_event *);
694 struct rcu_head rcu_head;
696 struct pid_namespace *ns;
700 perf_overflow_handler_t overflow_handler;
701 void *overflow_handler_context;
702 #ifdef CONFIG_BPF_SYSCALL
703 perf_overflow_handler_t orig_overflow_handler;
704 struct bpf_prog *prog;
707 #ifdef CONFIG_EVENT_TRACING
708 struct trace_event_call *tp_event;
709 struct event_filter *filter;
710 #ifdef CONFIG_FUNCTION_TRACER
711 struct ftrace_ops ftrace_ops;
715 #ifdef CONFIG_CGROUP_PERF
716 struct perf_cgroup *cgrp; /* cgroup event is attach to */
717 int cgrp_defer_enabled;
720 struct list_head sb_list;
721 #endif /* CONFIG_PERF_EVENTS */
725 * struct perf_event_context - event context structure
727 * Used as a container for task events and CPU events as well:
729 struct perf_event_context {
732 * Protect the states of the events in the list,
733 * nr_active, and the list:
737 * Protect the list of events. Locking either mutex or lock
738 * is sufficient to ensure the list doesn't change; to change
739 * the list you need to lock both the mutex and the spinlock.
743 struct list_head active_ctx_list;
744 struct list_head pinned_groups;
745 struct list_head flexible_groups;
746 struct list_head event_list;
754 struct task_struct *task;
757 * Context clock, runs when context enabled.
763 * These fields let us detect when two contexts have both
764 * been cloned (inherited) from a common ancestor.
766 struct perf_event_context *parent_ctx;
770 #ifdef CONFIG_CGROUP_PERF
771 int nr_cgroups; /* cgroup evts */
773 void *task_ctx_data; /* pmu specific data */
774 struct rcu_head rcu_head;
778 * Number of contexts where an event can trigger:
779 * task, softirq, hardirq, nmi.
781 #define PERF_NR_CONTEXTS 4
784 * struct perf_event_cpu_context - per cpu event context structure
786 struct perf_cpu_context {
787 struct perf_event_context ctx;
788 struct perf_event_context *task_ctx;
792 raw_spinlock_t hrtimer_lock;
793 struct hrtimer hrtimer;
794 ktime_t hrtimer_interval;
795 unsigned int hrtimer_active;
797 #ifdef CONFIG_CGROUP_PERF
798 struct perf_cgroup *cgrp;
799 struct list_head cgrp_cpuctx_entry;
802 struct list_head sched_cb_entry;
806 struct perf_output_handle {
807 struct perf_event *event;
808 struct ring_buffer *rb;
809 unsigned long wakeup;
819 struct bpf_perf_event_data_kern {
820 struct pt_regs *regs;
821 struct perf_sample_data *data;
824 #ifdef CONFIG_CGROUP_PERF
827 * perf_cgroup_info keeps track of time_enabled for a cgroup.
828 * This is a per-cpu dynamically allocated data structure.
830 struct perf_cgroup_info {
836 struct cgroup_subsys_state css;
837 struct perf_cgroup_info __percpu *info;
841 * Must ensure cgroup is pinned (css_get) before calling
842 * this function. In other words, we cannot call this function
843 * if there is no cgroup event for the current CPU context.
845 static inline struct perf_cgroup *
846 perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
848 return container_of(task_css_check(task, perf_event_cgrp_id,
849 ctx ? lockdep_is_held(&ctx->lock)
851 struct perf_cgroup, css);
853 #endif /* CONFIG_CGROUP_PERF */
855 #ifdef CONFIG_PERF_EVENTS
857 extern void *perf_aux_output_begin(struct perf_output_handle *handle,
858 struct perf_event *event);
859 extern void perf_aux_output_end(struct perf_output_handle *handle,
861 extern int perf_aux_output_skip(struct perf_output_handle *handle,
863 extern void *perf_get_aux(struct perf_output_handle *handle);
864 extern void perf_aux_output_flag(struct perf_output_handle *handle, u64 flags);
866 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
867 extern void perf_pmu_unregister(struct pmu *pmu);
869 extern int perf_num_counters(void);
870 extern const char *perf_pmu_name(void);
871 extern void __perf_event_task_sched_in(struct task_struct *prev,
872 struct task_struct *task);
873 extern void __perf_event_task_sched_out(struct task_struct *prev,
874 struct task_struct *next);
875 extern int perf_event_init_task(struct task_struct *child);
876 extern void perf_event_exit_task(struct task_struct *child);
877 extern void perf_event_free_task(struct task_struct *task);
878 extern void perf_event_delayed_put(struct task_struct *task);
879 extern struct file *perf_event_get(unsigned int fd);
880 extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
881 extern void perf_event_print_debug(void);
882 extern void perf_pmu_disable(struct pmu *pmu);
883 extern void perf_pmu_enable(struct pmu *pmu);
884 extern void perf_sched_cb_dec(struct pmu *pmu);
885 extern void perf_sched_cb_inc(struct pmu *pmu);
886 extern int perf_event_task_disable(void);
887 extern int perf_event_task_enable(void);
888 extern int perf_event_refresh(struct perf_event *event, int refresh);
889 extern void perf_event_update_userpage(struct perf_event *event);
890 extern int perf_event_release_kernel(struct perf_event *event);
891 extern struct perf_event *
892 perf_event_create_kernel_counter(struct perf_event_attr *attr,
894 struct task_struct *task,
895 perf_overflow_handler_t callback,
897 extern void perf_pmu_migrate_context(struct pmu *pmu,
898 int src_cpu, int dst_cpu);
899 extern u64 perf_event_read_local(struct perf_event *event);
900 extern u64 perf_event_read_value(struct perf_event *event,
901 u64 *enabled, u64 *running);
904 struct perf_sample_data {
906 * Fields set by perf_sample_data_init(), group so as to
907 * minimize the cachelines touched.
910 struct perf_raw_record *raw;
911 struct perf_branch_stack *br_stack;
915 union perf_mem_data_src data_src;
918 * The other fields, optionally {set,used} by
919 * perf_{prepare,output}_sample().
934 struct perf_callchain_entry *callchain;
937 * regs_user may point to task_pt_regs or to regs_user_copy, depending
940 struct perf_regs regs_user;
941 struct pt_regs regs_user_copy;
943 struct perf_regs regs_intr;
945 } ____cacheline_aligned;
947 /* default value for data source */
948 #define PERF_MEM_NA (PERF_MEM_S(OP, NA) |\
949 PERF_MEM_S(LVL, NA) |\
950 PERF_MEM_S(SNOOP, NA) |\
951 PERF_MEM_S(LOCK, NA) |\
954 static inline void perf_sample_data_init(struct perf_sample_data *data,
955 u64 addr, u64 period)
957 /* remaining struct members initialized in perf_prepare_sample() */
960 data->br_stack = NULL;
961 data->period = period;
963 data->data_src.val = PERF_MEM_NA;
967 extern void perf_output_sample(struct perf_output_handle *handle,
968 struct perf_event_header *header,
969 struct perf_sample_data *data,
970 struct perf_event *event);
971 extern void perf_prepare_sample(struct perf_event_header *header,
972 struct perf_sample_data *data,
973 struct perf_event *event,
974 struct pt_regs *regs);
976 extern int perf_event_overflow(struct perf_event *event,
977 struct perf_sample_data *data,
978 struct pt_regs *regs);
980 extern void perf_event_output_forward(struct perf_event *event,
981 struct perf_sample_data *data,
982 struct pt_regs *regs);
983 extern void perf_event_output_backward(struct perf_event *event,
984 struct perf_sample_data *data,
985 struct pt_regs *regs);
986 extern void perf_event_output(struct perf_event *event,
987 struct perf_sample_data *data,
988 struct pt_regs *regs);
991 is_default_overflow_handler(struct perf_event *event)
993 if (likely(event->overflow_handler == perf_event_output_forward))
995 if (unlikely(event->overflow_handler == perf_event_output_backward))
1001 perf_event_header__init_id(struct perf_event_header *header,
1002 struct perf_sample_data *data,
1003 struct perf_event *event);
1005 perf_event__output_id_sample(struct perf_event *event,
1006 struct perf_output_handle *handle,
1007 struct perf_sample_data *sample);
1010 perf_log_lost_samples(struct perf_event *event, u64 lost);
1012 static inline bool is_sampling_event(struct perf_event *event)
1014 return event->attr.sample_period != 0;
1018 * Return 1 for a software event, 0 for a hardware event
1020 static inline int is_software_event(struct perf_event *event)
1022 return event->event_caps & PERF_EV_CAP_SOFTWARE;
1025 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
1027 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
1028 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
1030 #ifndef perf_arch_fetch_caller_regs
1031 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
1035 * Take a snapshot of the regs. Skip ip and frame pointer to
1036 * the nth caller. We only need a few of the regs:
1037 * - ip for PERF_SAMPLE_IP
1038 * - cs for user_mode() tests
1039 * - bp for callchains
1040 * - eflags, for future purposes, just in case
1042 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1044 perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1047 static __always_inline void
1048 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1050 if (static_key_false(&perf_swevent_enabled[event_id]))
1051 __perf_sw_event(event_id, nr, regs, addr);
1054 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
1057 * 'Special' version for the scheduler, it hard assumes no recursion,
1058 * which is guaranteed by us not actually scheduling inside other swevents
1059 * because those disable preemption.
1061 static __always_inline void
1062 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
1064 if (static_key_false(&perf_swevent_enabled[event_id])) {
1065 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1067 perf_fetch_caller_regs(regs);
1068 ___perf_sw_event(event_id, nr, regs, addr);
1072 extern struct static_key_false perf_sched_events;
1074 static __always_inline bool
1075 perf_sw_migrate_enabled(void)
1077 if (static_key_false(&perf_swevent_enabled[PERF_COUNT_SW_CPU_MIGRATIONS]))
1082 static inline void perf_event_task_migrate(struct task_struct *task)
1084 if (perf_sw_migrate_enabled())
1085 task->sched_migrated = 1;
1088 static inline void perf_event_task_sched_in(struct task_struct *prev,
1089 struct task_struct *task)
1091 if (static_branch_unlikely(&perf_sched_events))
1092 __perf_event_task_sched_in(prev, task);
1094 if (perf_sw_migrate_enabled() && task->sched_migrated) {
1095 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1097 perf_fetch_caller_regs(regs);
1098 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, regs, 0);
1099 task->sched_migrated = 0;
1103 static inline void perf_event_task_sched_out(struct task_struct *prev,
1104 struct task_struct *next)
1106 perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
1108 if (static_branch_unlikely(&perf_sched_events))
1109 __perf_event_task_sched_out(prev, next);
1112 static inline u64 __perf_event_count(struct perf_event *event)
1114 return local64_read(&event->count) + atomic64_read(&event->child_count);
1117 extern void perf_event_mmap(struct vm_area_struct *vma);
1118 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1119 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1120 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1122 extern void perf_event_exec(void);
1123 extern void perf_event_comm(struct task_struct *tsk, bool exec);
1124 extern void perf_event_namespaces(struct task_struct *tsk);
1125 extern void perf_event_fork(struct task_struct *tsk);
1128 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1130 extern void perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1131 extern void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1132 extern struct perf_callchain_entry *
1133 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
1134 u32 max_stack, bool crosstask, bool add_mark);
1135 extern int get_callchain_buffers(int max_stack);
1136 extern void put_callchain_buffers(void);
1138 extern int sysctl_perf_event_max_stack;
1139 extern int sysctl_perf_event_max_contexts_per_stack;
1141 static inline int perf_callchain_store_context(struct perf_callchain_entry_ctx *ctx, u64 ip)
1143 if (ctx->contexts < sysctl_perf_event_max_contexts_per_stack) {
1144 struct perf_callchain_entry *entry = ctx->entry;
1145 entry->ip[entry->nr++] = ip;
1149 ctx->contexts_maxed = true;
1150 return -1; /* no more room, stop walking the stack */
1154 static inline int perf_callchain_store(struct perf_callchain_entry_ctx *ctx, u64 ip)
1156 if (ctx->nr < ctx->max_stack && !ctx->contexts_maxed) {
1157 struct perf_callchain_entry *entry = ctx->entry;
1158 entry->ip[entry->nr++] = ip;
1162 return -1; /* no more room, stop walking the stack */
1166 extern int sysctl_perf_event_paranoid;
1167 extern int sysctl_perf_event_mlock;
1168 extern int sysctl_perf_event_sample_rate;
1169 extern int sysctl_perf_cpu_time_max_percent;
1171 extern void perf_sample_event_took(u64 sample_len_ns);
1173 extern int perf_proc_update_handler(struct ctl_table *table, int write,
1174 void __user *buffer, size_t *lenp,
1176 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
1177 void __user *buffer, size_t *lenp,
1180 int perf_event_max_stack_handler(struct ctl_table *table, int write,
1181 void __user *buffer, size_t *lenp, loff_t *ppos);
1183 static inline bool perf_paranoid_tracepoint_raw(void)
1185 return sysctl_perf_event_paranoid > -1;
1188 static inline bool perf_paranoid_cpu(void)
1190 return sysctl_perf_event_paranoid > 0;
1193 static inline bool perf_paranoid_kernel(void)
1195 return sysctl_perf_event_paranoid > 1;
1198 extern void perf_event_init(void);
1199 extern void perf_tp_event(u16 event_type, u64 count, void *record,
1200 int entry_size, struct pt_regs *regs,
1201 struct hlist_head *head, int rctx,
1202 struct task_struct *task);
1203 extern void perf_bp_event(struct perf_event *event, void *data);
1205 #ifndef perf_misc_flags
1206 # define perf_misc_flags(regs) \
1207 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1208 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1211 static inline bool has_branch_stack(struct perf_event *event)
1213 return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1216 static inline bool needs_branch_stack(struct perf_event *event)
1218 return event->attr.branch_sample_type != 0;
1221 static inline bool has_aux(struct perf_event *event)
1223 return event->pmu->setup_aux;
1226 static inline bool is_write_backward(struct perf_event *event)
1228 return !!event->attr.write_backward;
1231 static inline bool has_addr_filter(struct perf_event *event)
1233 return event->pmu->nr_addr_filters;
1237 * An inherited event uses parent's filters
1239 static inline struct perf_addr_filters_head *
1240 perf_event_addr_filters(struct perf_event *event)
1242 struct perf_addr_filters_head *ifh = &event->addr_filters;
1245 ifh = &event->parent->addr_filters;
1250 extern void perf_event_addr_filters_sync(struct perf_event *event);
1252 extern int perf_output_begin(struct perf_output_handle *handle,
1253 struct perf_event *event, unsigned int size);
1254 extern int perf_output_begin_forward(struct perf_output_handle *handle,
1255 struct perf_event *event,
1257 extern int perf_output_begin_backward(struct perf_output_handle *handle,
1258 struct perf_event *event,
1261 extern void perf_output_end(struct perf_output_handle *handle);
1262 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
1263 const void *buf, unsigned int len);
1264 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
1266 extern int perf_swevent_get_recursion_context(void);
1267 extern void perf_swevent_put_recursion_context(int rctx);
1268 extern u64 perf_swevent_set_period(struct perf_event *event);
1269 extern void perf_event_enable(struct perf_event *event);
1270 extern void perf_event_disable(struct perf_event *event);
1271 extern void perf_event_disable_local(struct perf_event *event);
1272 extern void perf_event_disable_inatomic(struct perf_event *event);
1273 extern void perf_event_task_tick(void);
1274 extern int perf_event_account_interrupt(struct perf_event *event);
1275 #else /* !CONFIG_PERF_EVENTS: */
1276 static inline void *
1277 perf_aux_output_begin(struct perf_output_handle *handle,
1278 struct perf_event *event) { return NULL; }
1280 perf_aux_output_end(struct perf_output_handle *handle, unsigned long size)
1283 perf_aux_output_skip(struct perf_output_handle *handle,
1284 unsigned long size) { return -EINVAL; }
1285 static inline void *
1286 perf_get_aux(struct perf_output_handle *handle) { return NULL; }
1288 perf_event_task_migrate(struct task_struct *task) { }
1290 perf_event_task_sched_in(struct task_struct *prev,
1291 struct task_struct *task) { }
1293 perf_event_task_sched_out(struct task_struct *prev,
1294 struct task_struct *next) { }
1295 static inline int perf_event_init_task(struct task_struct *child) { return 0; }
1296 static inline void perf_event_exit_task(struct task_struct *child) { }
1297 static inline void perf_event_free_task(struct task_struct *task) { }
1298 static inline void perf_event_delayed_put(struct task_struct *task) { }
1299 static inline struct file *perf_event_get(unsigned int fd) { return ERR_PTR(-EINVAL); }
1300 static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
1302 return ERR_PTR(-EINVAL);
1304 static inline u64 perf_event_read_local(struct perf_event *event) { return -EINVAL; }
1305 static inline void perf_event_print_debug(void) { }
1306 static inline int perf_event_task_disable(void) { return -EINVAL; }
1307 static inline int perf_event_task_enable(void) { return -EINVAL; }
1308 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1314 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) { }
1316 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr) { }
1318 perf_bp_event(struct perf_event *event, void *data) { }
1320 static inline int perf_register_guest_info_callbacks
1321 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1322 static inline int perf_unregister_guest_info_callbacks
1323 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1325 static inline void perf_event_mmap(struct vm_area_struct *vma) { }
1326 static inline void perf_event_exec(void) { }
1327 static inline void perf_event_comm(struct task_struct *tsk, bool exec) { }
1328 static inline void perf_event_namespaces(struct task_struct *tsk) { }
1329 static inline void perf_event_fork(struct task_struct *tsk) { }
1330 static inline void perf_event_init(void) { }
1331 static inline int perf_swevent_get_recursion_context(void) { return -1; }
1332 static inline void perf_swevent_put_recursion_context(int rctx) { }
1333 static inline u64 perf_swevent_set_period(struct perf_event *event) { return 0; }
1334 static inline void perf_event_enable(struct perf_event *event) { }
1335 static inline void perf_event_disable(struct perf_event *event) { }
1336 static inline int __perf_event_disable(void *info) { return -1; }
1337 static inline void perf_event_task_tick(void) { }
1338 static inline int perf_event_release_kernel(struct perf_event *event) { return 0; }
1341 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1342 extern void perf_restore_debug_store(void);
1344 static inline void perf_restore_debug_store(void) { }
1347 static __always_inline bool perf_raw_frag_last(const struct perf_raw_frag *frag)
1349 return frag->pad < sizeof(u64);
1352 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1354 struct perf_pmu_events_attr {
1355 struct device_attribute attr;
1357 const char *event_str;
1360 struct perf_pmu_events_ht_attr {
1361 struct device_attribute attr;
1363 const char *event_str_ht;
1364 const char *event_str_noht;
1367 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1370 #define PMU_EVENT_ATTR(_name, _var, _id, _show) \
1371 static struct perf_pmu_events_attr _var = { \
1372 .attr = __ATTR(_name, 0444, _show, NULL), \
1376 #define PMU_EVENT_ATTR_STRING(_name, _var, _str) \
1377 static struct perf_pmu_events_attr _var = { \
1378 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1380 .event_str = _str, \
1383 #define PMU_FORMAT_ATTR(_name, _format) \
1385 _name##_show(struct device *dev, \
1386 struct device_attribute *attr, \
1389 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
1390 return sprintf(page, _format "\n"); \
1393 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1395 /* Performance counter hotplug functions */
1396 #ifdef CONFIG_PERF_EVENTS
1397 int perf_event_init_cpu(unsigned int cpu);
1398 int perf_event_exit_cpu(unsigned int cpu);
1400 #define perf_event_init_cpu NULL
1401 #define perf_event_exit_cpu NULL
1404 #endif /* _LINUX_PERF_EVENT_H */