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 struct perf_raw_record {
78 * branch stack layout:
79 * nr: number of taken branches stored in entries[]
81 * Note that nr can vary from sample to sample
82 * branches (to, from) are stored from most recent
83 * to least recent, i.e., entries[0] contains the most
86 struct perf_branch_stack {
88 struct perf_branch_entry entries[0];
94 * extra PMU register associated with an event
96 struct hw_perf_event_extra {
97 u64 config; /* register value */
98 unsigned int reg; /* register address or index */
99 int alloc; /* extra register already allocated */
100 int idx; /* index in shared_regs->regs[] */
104 * struct hw_perf_event - performance event hardware details:
106 struct hw_perf_event {
107 #ifdef CONFIG_PERF_EVENTS
109 struct { /* hardware */
112 unsigned long config_base;
113 unsigned long event_base;
114 int event_base_rdpmc;
119 struct hw_perf_event_extra extra_reg;
120 struct hw_perf_event_extra branch_reg;
122 struct { /* software */
123 struct hrtimer hrtimer;
125 struct { /* tracepoint */
126 /* for tp_event->class */
127 struct list_head tp_list;
129 struct { /* intel_cqm */
133 struct list_head cqm_events_entry;
134 struct list_head cqm_groups_entry;
135 struct list_head cqm_group_entry;
137 struct { /* itrace */
140 struct { /* amd_power */
144 #ifdef CONFIG_HAVE_HW_BREAKPOINT
145 struct { /* breakpoint */
147 * Crufty hack to avoid the chicken and egg
148 * problem hw_breakpoint has with context
149 * creation and event initalization.
151 struct arch_hw_breakpoint info;
152 struct list_head bp_list;
157 * If the event is a per task event, this will point to the task in
158 * question. See the comment in perf_event_alloc().
160 struct task_struct *target;
163 * PMU would store hardware filter configuration
168 /* Last sync'ed generation of filters */
169 unsigned long addr_filters_gen;
172 * hw_perf_event::state flags; used to track the PERF_EF_* state.
174 #define PERF_HES_STOPPED 0x01 /* the counter is stopped */
175 #define PERF_HES_UPTODATE 0x02 /* event->count up-to-date */
176 #define PERF_HES_ARCH 0x04
181 * The last observed hardware counter value, updated with a
182 * local64_cmpxchg() such that pmu::read() can be called nested.
184 local64_t prev_count;
187 * The period to start the next sample with.
192 * The period we started this sample with.
197 * However much is left of the current period; note that this is
198 * a full 64bit value and allows for generation of periods longer
199 * than hardware might allow.
201 local64_t period_left;
204 * State for throttling the event, see __perf_event_overflow() and
205 * perf_adjust_freq_unthr_context().
211 * State for freq target events, see __perf_event_overflow() and
212 * perf_adjust_freq_unthr_context().
215 u64 freq_count_stamp;
222 * Common implementation detail of pmu::{start,commit,cancel}_txn
224 #define PERF_PMU_TXN_ADD 0x1 /* txn to add/schedule event on PMU */
225 #define PERF_PMU_TXN_READ 0x2 /* txn to read event group from PMU */
228 * pmu::capabilities flags
230 #define PERF_PMU_CAP_NO_INTERRUPT 0x01
231 #define PERF_PMU_CAP_NO_NMI 0x02
232 #define PERF_PMU_CAP_AUX_NO_SG 0x04
233 #define PERF_PMU_CAP_AUX_SW_DOUBLEBUF 0x08
234 #define PERF_PMU_CAP_EXCLUSIVE 0x10
235 #define PERF_PMU_CAP_ITRACE 0x20
236 #define PERF_PMU_CAP_HETEROGENEOUS_CPUS 0x40
239 * struct pmu - generic performance monitoring unit
242 struct list_head entry;
244 struct module *module;
246 const struct attribute_group **attr_groups;
251 * various common per-pmu feature flags
255 int * __percpu pmu_disable_count;
256 struct perf_cpu_context * __percpu pmu_cpu_context;
257 atomic_t exclusive_cnt; /* < 0: cpu; > 0: tsk */
259 int hrtimer_interval_ms;
261 /* number of address filters this PMU can do */
262 unsigned int nr_addr_filters;
265 * Fully disable/enable this PMU, can be used to protect from the PMI
266 * as well as for lazy/batch writing of the MSRs.
268 void (*pmu_enable) (struct pmu *pmu); /* optional */
269 void (*pmu_disable) (struct pmu *pmu); /* optional */
272 * Try and initialize the event for this PMU.
275 * -ENOENT -- @event is not for this PMU
277 * -ENODEV -- @event is for this PMU but PMU not present
278 * -EBUSY -- @event is for this PMU but PMU temporarily unavailable
279 * -EINVAL -- @event is for this PMU but @event is not valid
280 * -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
281 * -EACCESS -- @event is for this PMU, @event is valid, but no privilidges
283 * 0 -- @event is for this PMU and valid
285 * Other error return values are allowed.
287 int (*event_init) (struct perf_event *event);
290 * Notification that the event was mapped or unmapped. Called
291 * in the context of the mapping task.
293 void (*event_mapped) (struct perf_event *event); /*optional*/
294 void (*event_unmapped) (struct perf_event *event); /*optional*/
297 * Flags for ->add()/->del()/ ->start()/->stop(). There are
298 * matching hw_perf_event::state flags.
300 #define PERF_EF_START 0x01 /* start the counter when adding */
301 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
302 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
305 * Adds/Removes a counter to/from the PMU, can be done inside a
306 * transaction, see the ->*_txn() methods.
308 * The add/del callbacks will reserve all hardware resources required
309 * to service the event, this includes any counter constraint
312 * Called with IRQs disabled and the PMU disabled on the CPU the event
315 * ->add() called without PERF_EF_START should result in the same state
316 * as ->add() followed by ->stop().
318 * ->del() must always PERF_EF_UPDATE stop an event. If it calls
319 * ->stop() that must deal with already being stopped without
322 int (*add) (struct perf_event *event, int flags);
323 void (*del) (struct perf_event *event, int flags);
326 * Starts/Stops a counter present on the PMU.
328 * The PMI handler should stop the counter when perf_event_overflow()
329 * returns !0. ->start() will be used to continue.
331 * Also used to change the sample period.
333 * Called with IRQs disabled and the PMU disabled on the CPU the event
334 * is on -- will be called from NMI context with the PMU generates
337 * ->stop() with PERF_EF_UPDATE will read the counter and update
338 * period/count values like ->read() would.
340 * ->start() with PERF_EF_RELOAD will reprogram the the counter
341 * value, must be preceded by a ->stop() with PERF_EF_UPDATE.
343 void (*start) (struct perf_event *event, int flags);
344 void (*stop) (struct perf_event *event, int flags);
347 * Updates the counter value of the event.
349 * For sampling capable PMUs this will also update the software period
350 * hw_perf_event::period_left field.
352 void (*read) (struct perf_event *event);
355 * Group events scheduling is treated as a transaction, add
356 * group events as a whole and perform one schedulability test.
357 * If the test fails, roll back the whole group
359 * Start the transaction, after this ->add() doesn't need to
360 * do schedulability tests.
364 void (*start_txn) (struct pmu *pmu, unsigned int txn_flags);
366 * If ->start_txn() disabled the ->add() schedulability test
367 * then ->commit_txn() is required to perform one. On success
368 * the transaction is closed. On error the transaction is kept
369 * open until ->cancel_txn() is called.
373 int (*commit_txn) (struct pmu *pmu);
375 * Will cancel the transaction, assumes ->del() is called
376 * for each successful ->add() during the transaction.
380 void (*cancel_txn) (struct pmu *pmu);
383 * Will return the value for perf_event_mmap_page::index for this event,
384 * if no implementation is provided it will default to: event->hw.idx + 1.
386 int (*event_idx) (struct perf_event *event); /*optional */
389 * context-switches callback
391 void (*sched_task) (struct perf_event_context *ctx,
394 * PMU specific data size
396 size_t task_ctx_size;
400 * Return the count value for a counter.
402 u64 (*count) (struct perf_event *event); /*optional*/
405 * Set up pmu-private data structures for an AUX area
407 void *(*setup_aux) (int cpu, void **pages,
408 int nr_pages, bool overwrite);
412 * Free pmu-private AUX data structures
414 void (*free_aux) (void *aux); /* optional */
417 * Validate address range filters: make sure the HW supports the
418 * requested configuration and number of filters; return 0 if the
419 * supplied filters are valid, -errno otherwise.
421 * Runs in the context of the ioctl()ing process and is not serialized
422 * with the rest of the PMU callbacks.
424 int (*addr_filters_validate) (struct list_head *filters);
428 * Synchronize address range filter configuration:
429 * translate hw-agnostic filters into hardware configuration in
430 * event::hw::addr_filters.
432 * Runs as a part of filter sync sequence that is done in ->start()
433 * callback by calling perf_event_addr_filters_sync().
435 * May (and should) traverse event::addr_filters::list, for which its
436 * caller provides necessary serialization.
438 void (*addr_filters_sync) (struct perf_event *event);
442 * Filter events for PMU-specific reasons.
444 int (*filter_match) (struct perf_event *event); /* optional */
448 * struct perf_addr_filter - address range filter definition
449 * @entry: event's filter list linkage
450 * @inode: object file's inode for file-based filters
451 * @offset: filter range offset
452 * @size: filter range size
453 * @range: 1: range, 0: address
454 * @filter: 1: filter/start, 0: stop
456 * This is a hardware-agnostic filter configuration as specified by the user.
458 struct perf_addr_filter {
459 struct list_head entry;
461 unsigned long offset;
463 unsigned int range : 1,
468 * struct perf_addr_filters_head - container for address range filters
469 * @list: list of filters for this event
470 * @lock: spinlock that serializes accesses to the @list and event's
471 * (and its children's) filter generations.
473 * A child event will use parent's @list (and therefore @lock), so they are
474 * bundled together; see perf_event_addr_filters().
476 struct perf_addr_filters_head {
477 struct list_head list;
482 * enum perf_event_active_state - the states of a event
484 enum perf_event_active_state {
485 PERF_EVENT_STATE_DEAD = -4,
486 PERF_EVENT_STATE_EXIT = -3,
487 PERF_EVENT_STATE_ERROR = -2,
488 PERF_EVENT_STATE_OFF = -1,
489 PERF_EVENT_STATE_INACTIVE = 0,
490 PERF_EVENT_STATE_ACTIVE = 1,
494 struct perf_sample_data;
496 typedef void (*perf_overflow_handler_t)(struct perf_event *,
497 struct perf_sample_data *,
498 struct pt_regs *regs);
500 enum perf_group_flag {
501 PERF_GROUP_SOFTWARE = 0x1,
504 #define SWEVENT_HLIST_BITS 8
505 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
507 struct swevent_hlist {
508 struct hlist_head heads[SWEVENT_HLIST_SIZE];
509 struct rcu_head rcu_head;
512 #define PERF_ATTACH_CONTEXT 0x01
513 #define PERF_ATTACH_GROUP 0x02
514 #define PERF_ATTACH_TASK 0x04
515 #define PERF_ATTACH_TASK_DATA 0x08
520 struct pmu_event_list {
522 struct list_head list;
526 * struct perf_event - performance event kernel representation:
529 #ifdef CONFIG_PERF_EVENTS
531 * entry onto perf_event_context::event_list;
532 * modifications require ctx->lock
533 * RCU safe iterations.
535 struct list_head event_entry;
538 * XXX: group_entry and sibling_list should be mutually exclusive;
539 * either you're a sibling on a group, or you're the group leader.
540 * Rework the code to always use the same list element.
542 * Locked for modification by both ctx->mutex and ctx->lock; holding
543 * either sufficies for read.
545 struct list_head group_entry;
546 struct list_head sibling_list;
549 * We need storage to track the entries in perf_pmu_migrate_context; we
550 * cannot use the event_entry because of RCU and we want to keep the
551 * group in tact which avoids us using the other two entries.
553 struct list_head migrate_entry;
555 struct hlist_node hlist_entry;
556 struct list_head active_entry;
559 struct perf_event *group_leader;
563 enum perf_event_active_state state;
564 unsigned int attach_state;
566 atomic64_t child_count;
569 * These are the total time in nanoseconds that the event
570 * has been enabled (i.e. eligible to run, and the task has
571 * been scheduled in, if this is a per-task event)
572 * and running (scheduled onto the CPU), respectively.
574 * They are computed from tstamp_enabled, tstamp_running and
575 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
577 u64 total_time_enabled;
578 u64 total_time_running;
581 * These are timestamps used for computing total_time_enabled
582 * and total_time_running when the event is in INACTIVE or
583 * ACTIVE state, measured in nanoseconds from an arbitrary point
585 * tstamp_enabled: the notional time when the event was enabled
586 * tstamp_running: the notional time when the event was scheduled on
587 * tstamp_stopped: in INACTIVE state, the notional time when the
588 * event was scheduled off.
595 * timestamp shadows the actual context timing but it can
596 * be safely used in NMI interrupt context. It reflects the
597 * context time as it was when the event was last scheduled in.
599 * ctx_time already accounts for ctx->timestamp. Therefore to
600 * compute ctx_time for a sample, simply add perf_clock().
604 struct perf_event_attr attr;
608 struct hw_perf_event hw;
610 struct perf_event_context *ctx;
611 atomic_long_t refcount;
614 * These accumulate total time (in nanoseconds) that children
615 * events have been enabled and running, respectively.
617 atomic64_t child_total_time_enabled;
618 atomic64_t child_total_time_running;
621 * Protect attach/detach and child_list:
623 struct mutex child_mutex;
624 struct list_head child_list;
625 struct perf_event *parent;
630 struct list_head owner_entry;
631 struct task_struct *owner;
634 struct mutex mmap_mutex;
637 struct ring_buffer *rb;
638 struct list_head rb_entry;
639 unsigned long rcu_batches;
643 wait_queue_head_t waitq;
644 struct fasync_struct *fasync;
646 /* delayed work for NMIs and such */
650 struct irq_work pending;
652 atomic_t event_limit;
654 /* address range filters */
655 struct perf_addr_filters_head addr_filters;
656 /* vma address array for file-based filders */
657 unsigned long *addr_filters_offs;
658 unsigned long addr_filters_gen;
660 void (*destroy)(struct perf_event *);
661 struct rcu_head rcu_head;
663 struct pid_namespace *ns;
667 perf_overflow_handler_t overflow_handler;
668 void *overflow_handler_context;
670 #ifdef CONFIG_EVENT_TRACING
671 struct trace_event_call *tp_event;
672 struct event_filter *filter;
673 #ifdef CONFIG_FUNCTION_TRACER
674 struct ftrace_ops ftrace_ops;
678 #ifdef CONFIG_CGROUP_PERF
679 struct perf_cgroup *cgrp; /* cgroup event is attach to */
680 int cgrp_defer_enabled;
683 struct list_head sb_list;
684 #endif /* CONFIG_PERF_EVENTS */
688 * struct perf_event_context - event context structure
690 * Used as a container for task events and CPU events as well:
692 struct perf_event_context {
695 * Protect the states of the events in the list,
696 * nr_active, and the list:
700 * Protect the list of events. Locking either mutex or lock
701 * is sufficient to ensure the list doesn't change; to change
702 * the list you need to lock both the mutex and the spinlock.
706 struct list_head active_ctx_list;
707 struct list_head pinned_groups;
708 struct list_head flexible_groups;
709 struct list_head event_list;
717 struct task_struct *task;
720 * Context clock, runs when context enabled.
726 * These fields let us detect when two contexts have both
727 * been cloned (inherited) from a common ancestor.
729 struct perf_event_context *parent_ctx;
733 int nr_cgroups; /* cgroup evts */
734 void *task_ctx_data; /* pmu specific data */
735 struct rcu_head rcu_head;
739 * Number of contexts where an event can trigger:
740 * task, softirq, hardirq, nmi.
742 #define PERF_NR_CONTEXTS 4
745 * struct perf_event_cpu_context - per cpu event context structure
747 struct perf_cpu_context {
748 struct perf_event_context ctx;
749 struct perf_event_context *task_ctx;
753 raw_spinlock_t hrtimer_lock;
754 struct hrtimer hrtimer;
755 ktime_t hrtimer_interval;
756 unsigned int hrtimer_active;
758 struct pmu *unique_pmu;
759 struct perf_cgroup *cgrp;
762 struct perf_output_handle {
763 struct perf_event *event;
764 struct ring_buffer *rb;
765 unsigned long wakeup;
774 #ifdef CONFIG_CGROUP_PERF
777 * perf_cgroup_info keeps track of time_enabled for a cgroup.
778 * This is a per-cpu dynamically allocated data structure.
780 struct perf_cgroup_info {
786 struct cgroup_subsys_state css;
787 struct perf_cgroup_info __percpu *info;
791 * Must ensure cgroup is pinned (css_get) before calling
792 * this function. In other words, we cannot call this function
793 * if there is no cgroup event for the current CPU context.
795 static inline struct perf_cgroup *
796 perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
798 return container_of(task_css_check(task, perf_event_cgrp_id,
799 ctx ? lockdep_is_held(&ctx->lock)
801 struct perf_cgroup, css);
803 #endif /* CONFIG_CGROUP_PERF */
805 #ifdef CONFIG_PERF_EVENTS
807 extern void *perf_aux_output_begin(struct perf_output_handle *handle,
808 struct perf_event *event);
809 extern void perf_aux_output_end(struct perf_output_handle *handle,
810 unsigned long size, bool truncated);
811 extern int perf_aux_output_skip(struct perf_output_handle *handle,
813 extern void *perf_get_aux(struct perf_output_handle *handle);
815 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
816 extern void perf_pmu_unregister(struct pmu *pmu);
818 extern int perf_num_counters(void);
819 extern const char *perf_pmu_name(void);
820 extern void __perf_event_task_sched_in(struct task_struct *prev,
821 struct task_struct *task);
822 extern void __perf_event_task_sched_out(struct task_struct *prev,
823 struct task_struct *next);
824 extern int perf_event_init_task(struct task_struct *child);
825 extern void perf_event_exit_task(struct task_struct *child);
826 extern void perf_event_free_task(struct task_struct *task);
827 extern void perf_event_delayed_put(struct task_struct *task);
828 extern struct file *perf_event_get(unsigned int fd);
829 extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
830 extern void perf_event_print_debug(void);
831 extern void perf_pmu_disable(struct pmu *pmu);
832 extern void perf_pmu_enable(struct pmu *pmu);
833 extern void perf_sched_cb_dec(struct pmu *pmu);
834 extern void perf_sched_cb_inc(struct pmu *pmu);
835 extern int perf_event_task_disable(void);
836 extern int perf_event_task_enable(void);
837 extern int perf_event_refresh(struct perf_event *event, int refresh);
838 extern void perf_event_update_userpage(struct perf_event *event);
839 extern int perf_event_release_kernel(struct perf_event *event);
840 extern struct perf_event *
841 perf_event_create_kernel_counter(struct perf_event_attr *attr,
843 struct task_struct *task,
844 perf_overflow_handler_t callback,
846 extern void perf_pmu_migrate_context(struct pmu *pmu,
847 int src_cpu, int dst_cpu);
848 extern u64 perf_event_read_local(struct perf_event *event);
849 extern u64 perf_event_read_value(struct perf_event *event,
850 u64 *enabled, u64 *running);
853 struct perf_sample_data {
855 * Fields set by perf_sample_data_init(), group so as to
856 * minimize the cachelines touched.
859 struct perf_raw_record *raw;
860 struct perf_branch_stack *br_stack;
864 union perf_mem_data_src data_src;
867 * The other fields, optionally {set,used} by
868 * perf_{prepare,output}_sample().
883 struct perf_callchain_entry *callchain;
886 * regs_user may point to task_pt_regs or to regs_user_copy, depending
889 struct perf_regs regs_user;
890 struct pt_regs regs_user_copy;
892 struct perf_regs regs_intr;
894 } ____cacheline_aligned;
896 /* default value for data source */
897 #define PERF_MEM_NA (PERF_MEM_S(OP, NA) |\
898 PERF_MEM_S(LVL, NA) |\
899 PERF_MEM_S(SNOOP, NA) |\
900 PERF_MEM_S(LOCK, NA) |\
903 static inline void perf_sample_data_init(struct perf_sample_data *data,
904 u64 addr, u64 period)
906 /* remaining struct members initialized in perf_prepare_sample() */
909 data->br_stack = NULL;
910 data->period = period;
912 data->data_src.val = PERF_MEM_NA;
916 extern void perf_output_sample(struct perf_output_handle *handle,
917 struct perf_event_header *header,
918 struct perf_sample_data *data,
919 struct perf_event *event);
920 extern void perf_prepare_sample(struct perf_event_header *header,
921 struct perf_sample_data *data,
922 struct perf_event *event,
923 struct pt_regs *regs);
925 extern int perf_event_overflow(struct perf_event *event,
926 struct perf_sample_data *data,
927 struct pt_regs *regs);
929 extern void perf_event_output_forward(struct perf_event *event,
930 struct perf_sample_data *data,
931 struct pt_regs *regs);
932 extern void perf_event_output_backward(struct perf_event *event,
933 struct perf_sample_data *data,
934 struct pt_regs *regs);
935 extern void perf_event_output(struct perf_event *event,
936 struct perf_sample_data *data,
937 struct pt_regs *regs);
940 is_default_overflow_handler(struct perf_event *event)
942 if (likely(event->overflow_handler == perf_event_output_forward))
944 if (unlikely(event->overflow_handler == perf_event_output_backward))
950 perf_event_header__init_id(struct perf_event_header *header,
951 struct perf_sample_data *data,
952 struct perf_event *event);
954 perf_event__output_id_sample(struct perf_event *event,
955 struct perf_output_handle *handle,
956 struct perf_sample_data *sample);
959 perf_log_lost_samples(struct perf_event *event, u64 lost);
961 static inline bool is_sampling_event(struct perf_event *event)
963 return event->attr.sample_period != 0;
967 * Return 1 for a software event, 0 for a hardware event
969 static inline int is_software_event(struct perf_event *event)
971 return event->pmu->task_ctx_nr == perf_sw_context;
974 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
976 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
977 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
979 #ifndef perf_arch_fetch_caller_regs
980 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
984 * Take a snapshot of the regs. Skip ip and frame pointer to
985 * the nth caller. We only need a few of the regs:
986 * - ip for PERF_SAMPLE_IP
987 * - cs for user_mode() tests
988 * - bp for callchains
989 * - eflags, for future purposes, just in case
991 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
993 memset(regs, 0, sizeof(*regs));
995 perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
998 static __always_inline void
999 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1001 if (static_key_false(&perf_swevent_enabled[event_id]))
1002 __perf_sw_event(event_id, nr, regs, addr);
1005 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
1008 * 'Special' version for the scheduler, it hard assumes no recursion,
1009 * which is guaranteed by us not actually scheduling inside other swevents
1010 * because those disable preemption.
1012 static __always_inline void
1013 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
1015 if (static_key_false(&perf_swevent_enabled[event_id])) {
1016 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1018 perf_fetch_caller_regs(regs);
1019 ___perf_sw_event(event_id, nr, regs, addr);
1023 extern struct static_key_false perf_sched_events;
1025 static __always_inline bool
1026 perf_sw_migrate_enabled(void)
1028 if (static_key_false(&perf_swevent_enabled[PERF_COUNT_SW_CPU_MIGRATIONS]))
1033 static inline void perf_event_task_migrate(struct task_struct *task)
1035 if (perf_sw_migrate_enabled())
1036 task->sched_migrated = 1;
1039 static inline void perf_event_task_sched_in(struct task_struct *prev,
1040 struct task_struct *task)
1042 if (static_branch_unlikely(&perf_sched_events))
1043 __perf_event_task_sched_in(prev, task);
1045 if (perf_sw_migrate_enabled() && task->sched_migrated) {
1046 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1048 perf_fetch_caller_regs(regs);
1049 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, regs, 0);
1050 task->sched_migrated = 0;
1054 static inline void perf_event_task_sched_out(struct task_struct *prev,
1055 struct task_struct *next)
1057 perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
1059 if (static_branch_unlikely(&perf_sched_events))
1060 __perf_event_task_sched_out(prev, next);
1063 static inline u64 __perf_event_count(struct perf_event *event)
1065 return local64_read(&event->count) + atomic64_read(&event->child_count);
1068 extern void perf_event_mmap(struct vm_area_struct *vma);
1069 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1070 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1071 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1073 extern void perf_event_exec(void);
1074 extern void perf_event_comm(struct task_struct *tsk, bool exec);
1075 extern void perf_event_fork(struct task_struct *tsk);
1078 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1080 extern void perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1081 extern void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1082 extern struct perf_callchain_entry *
1083 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
1084 u32 max_stack, bool crosstask, bool add_mark);
1085 extern int get_callchain_buffers(int max_stack);
1086 extern void put_callchain_buffers(void);
1088 extern int sysctl_perf_event_max_stack;
1089 extern int sysctl_perf_event_max_contexts_per_stack;
1091 static inline int perf_callchain_store_context(struct perf_callchain_entry_ctx *ctx, u64 ip)
1093 if (ctx->contexts < sysctl_perf_event_max_contexts_per_stack) {
1094 struct perf_callchain_entry *entry = ctx->entry;
1095 entry->ip[entry->nr++] = ip;
1099 ctx->contexts_maxed = true;
1100 return -1; /* no more room, stop walking the stack */
1104 static inline int perf_callchain_store(struct perf_callchain_entry_ctx *ctx, u64 ip)
1106 if (ctx->nr < ctx->max_stack && !ctx->contexts_maxed) {
1107 struct perf_callchain_entry *entry = ctx->entry;
1108 entry->ip[entry->nr++] = ip;
1112 return -1; /* no more room, stop walking the stack */
1116 extern int sysctl_perf_event_paranoid;
1117 extern int sysctl_perf_event_mlock;
1118 extern int sysctl_perf_event_sample_rate;
1119 extern int sysctl_perf_cpu_time_max_percent;
1121 extern void perf_sample_event_took(u64 sample_len_ns);
1123 extern int perf_proc_update_handler(struct ctl_table *table, int write,
1124 void __user *buffer, size_t *lenp,
1126 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
1127 void __user *buffer, size_t *lenp,
1130 int perf_event_max_stack_handler(struct ctl_table *table, int write,
1131 void __user *buffer, size_t *lenp, loff_t *ppos);
1133 static inline bool perf_paranoid_tracepoint_raw(void)
1135 return sysctl_perf_event_paranoid > -1;
1138 static inline bool perf_paranoid_cpu(void)
1140 return sysctl_perf_event_paranoid > 0;
1143 static inline bool perf_paranoid_kernel(void)
1145 return sysctl_perf_event_paranoid > 1;
1148 extern void perf_event_init(void);
1149 extern void perf_tp_event(u64 addr, u64 count, void *record,
1150 int entry_size, struct pt_regs *regs,
1151 struct hlist_head *head, int rctx,
1152 struct task_struct *task);
1153 extern void perf_bp_event(struct perf_event *event, void *data);
1155 #ifndef perf_misc_flags
1156 # define perf_misc_flags(regs) \
1157 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1158 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1161 static inline bool has_branch_stack(struct perf_event *event)
1163 return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1166 static inline bool needs_branch_stack(struct perf_event *event)
1168 return event->attr.branch_sample_type != 0;
1171 static inline bool has_aux(struct perf_event *event)
1173 return event->pmu->setup_aux;
1176 static inline bool is_write_backward(struct perf_event *event)
1178 return !!event->attr.write_backward;
1181 static inline bool has_addr_filter(struct perf_event *event)
1183 return event->pmu->nr_addr_filters;
1187 * An inherited event uses parent's filters
1189 static inline struct perf_addr_filters_head *
1190 perf_event_addr_filters(struct perf_event *event)
1192 struct perf_addr_filters_head *ifh = &event->addr_filters;
1195 ifh = &event->parent->addr_filters;
1200 extern void perf_event_addr_filters_sync(struct perf_event *event);
1202 extern int perf_output_begin(struct perf_output_handle *handle,
1203 struct perf_event *event, unsigned int size);
1204 extern int perf_output_begin_forward(struct perf_output_handle *handle,
1205 struct perf_event *event,
1207 extern int perf_output_begin_backward(struct perf_output_handle *handle,
1208 struct perf_event *event,
1211 extern void perf_output_end(struct perf_output_handle *handle);
1212 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
1213 const void *buf, unsigned int len);
1214 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
1216 extern int perf_swevent_get_recursion_context(void);
1217 extern void perf_swevent_put_recursion_context(int rctx);
1218 extern u64 perf_swevent_set_period(struct perf_event *event);
1219 extern void perf_event_enable(struct perf_event *event);
1220 extern void perf_event_disable(struct perf_event *event);
1221 extern void perf_event_disable_local(struct perf_event *event);
1222 extern void perf_event_task_tick(void);
1223 #else /* !CONFIG_PERF_EVENTS: */
1224 static inline void *
1225 perf_aux_output_begin(struct perf_output_handle *handle,
1226 struct perf_event *event) { return NULL; }
1228 perf_aux_output_end(struct perf_output_handle *handle, unsigned long size,
1231 perf_aux_output_skip(struct perf_output_handle *handle,
1232 unsigned long size) { return -EINVAL; }
1233 static inline void *
1234 perf_get_aux(struct perf_output_handle *handle) { return NULL; }
1236 perf_event_task_migrate(struct task_struct *task) { }
1238 perf_event_task_sched_in(struct task_struct *prev,
1239 struct task_struct *task) { }
1241 perf_event_task_sched_out(struct task_struct *prev,
1242 struct task_struct *next) { }
1243 static inline int perf_event_init_task(struct task_struct *child) { return 0; }
1244 static inline void perf_event_exit_task(struct task_struct *child) { }
1245 static inline void perf_event_free_task(struct task_struct *task) { }
1246 static inline void perf_event_delayed_put(struct task_struct *task) { }
1247 static inline struct file *perf_event_get(unsigned int fd) { return ERR_PTR(-EINVAL); }
1248 static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
1250 return ERR_PTR(-EINVAL);
1252 static inline u64 perf_event_read_local(struct perf_event *event) { return -EINVAL; }
1253 static inline void perf_event_print_debug(void) { }
1254 static inline int perf_event_task_disable(void) { return -EINVAL; }
1255 static inline int perf_event_task_enable(void) { return -EINVAL; }
1256 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1262 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) { }
1264 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr) { }
1266 perf_bp_event(struct perf_event *event, void *data) { }
1268 static inline int perf_register_guest_info_callbacks
1269 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1270 static inline int perf_unregister_guest_info_callbacks
1271 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1273 static inline void perf_event_mmap(struct vm_area_struct *vma) { }
1274 static inline void perf_event_exec(void) { }
1275 static inline void perf_event_comm(struct task_struct *tsk, bool exec) { }
1276 static inline void perf_event_fork(struct task_struct *tsk) { }
1277 static inline void perf_event_init(void) { }
1278 static inline int perf_swevent_get_recursion_context(void) { return -1; }
1279 static inline void perf_swevent_put_recursion_context(int rctx) { }
1280 static inline u64 perf_swevent_set_period(struct perf_event *event) { return 0; }
1281 static inline void perf_event_enable(struct perf_event *event) { }
1282 static inline void perf_event_disable(struct perf_event *event) { }
1283 static inline int __perf_event_disable(void *info) { return -1; }
1284 static inline void perf_event_task_tick(void) { }
1285 static inline int perf_event_release_kernel(struct perf_event *event) { return 0; }
1288 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1289 extern void perf_restore_debug_store(void);
1291 static inline void perf_restore_debug_store(void) { }
1294 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1297 * This has to have a higher priority than migration_notifier in sched/core.c.
1299 #define perf_cpu_notifier(fn) \
1301 static struct notifier_block fn##_nb = \
1302 { .notifier_call = fn, .priority = CPU_PRI_PERF }; \
1303 unsigned long cpu = smp_processor_id(); \
1304 unsigned long flags; \
1306 cpu_notifier_register_begin(); \
1307 fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE, \
1308 (void *)(unsigned long)cpu); \
1309 local_irq_save(flags); \
1310 fn(&fn##_nb, (unsigned long)CPU_STARTING, \
1311 (void *)(unsigned long)cpu); \
1312 local_irq_restore(flags); \
1313 fn(&fn##_nb, (unsigned long)CPU_ONLINE, \
1314 (void *)(unsigned long)cpu); \
1315 __register_cpu_notifier(&fn##_nb); \
1316 cpu_notifier_register_done(); \
1320 * Bare-bones version of perf_cpu_notifier(), which doesn't invoke the
1321 * callback for already online CPUs.
1323 #define __perf_cpu_notifier(fn) \
1325 static struct notifier_block fn##_nb = \
1326 { .notifier_call = fn, .priority = CPU_PRI_PERF }; \
1328 __register_cpu_notifier(&fn##_nb); \
1331 struct perf_pmu_events_attr {
1332 struct device_attribute attr;
1334 const char *event_str;
1337 struct perf_pmu_events_ht_attr {
1338 struct device_attribute attr;
1340 const char *event_str_ht;
1341 const char *event_str_noht;
1344 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1347 #define PMU_EVENT_ATTR(_name, _var, _id, _show) \
1348 static struct perf_pmu_events_attr _var = { \
1349 .attr = __ATTR(_name, 0444, _show, NULL), \
1353 #define PMU_EVENT_ATTR_STRING(_name, _var, _str) \
1354 static struct perf_pmu_events_attr _var = { \
1355 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1357 .event_str = _str, \
1360 #define PMU_FORMAT_ATTR(_name, _format) \
1362 _name##_show(struct device *dev, \
1363 struct device_attribute *attr, \
1366 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
1367 return sprintf(page, _format "\n"); \
1370 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1372 #endif /* _LINUX_PERF_EVENT_H */