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 <asm/local.h>
57 struct perf_callchain_entry {
59 __u64 ip[PERF_MAX_STACK_DEPTH];
62 struct perf_raw_record {
68 * branch stack layout:
69 * nr: number of taken branches stored in entries[]
71 * Note that nr can vary from sample to sample
72 * branches (to, from) are stored from most recent
73 * to least recent, i.e., entries[0] contains the most
76 struct perf_branch_stack {
78 struct perf_branch_entry entries[0];
81 struct perf_regs_user {
89 * extra PMU register associated with an event
91 struct hw_perf_event_extra {
92 u64 config; /* register value */
93 unsigned int reg; /* register address or index */
94 int alloc; /* extra register already allocated */
95 int idx; /* index in shared_regs->regs[] */
98 struct event_constraint;
101 * struct hw_perf_event - performance event hardware details:
103 struct hw_perf_event {
104 #ifdef CONFIG_PERF_EVENTS
106 struct { /* hardware */
109 unsigned long config_base;
110 unsigned long event_base;
111 int event_base_rdpmc;
116 struct hw_perf_event_extra extra_reg;
117 struct hw_perf_event_extra branch_reg;
119 struct event_constraint *constraint;
121 struct { /* software */
122 struct hrtimer hrtimer;
124 struct { /* tracepoint */
125 struct task_struct *tp_target;
126 /* for tp_event->class */
127 struct list_head tp_list;
129 #ifdef CONFIG_HAVE_HW_BREAKPOINT
130 struct { /* breakpoint */
132 * Crufty hack to avoid the chicken and egg
133 * problem hw_breakpoint has with context
134 * creation and event initalization.
136 struct task_struct *bp_target;
137 struct arch_hw_breakpoint info;
138 struct list_head bp_list;
143 local64_t prev_count;
146 local64_t period_left;
151 u64 freq_count_stamp;
156 * hw_perf_event::state flags
158 #define PERF_HES_STOPPED 0x01 /* the counter is stopped */
159 #define PERF_HES_UPTODATE 0x02 /* event->count up-to-date */
160 #define PERF_HES_ARCH 0x04
165 * Common implementation detail of pmu::{start,commit,cancel}_txn
167 #define PERF_EVENT_TXN 0x1
170 * struct pmu - generic performance monitoring unit
173 struct list_head entry;
176 const struct attribute_group **attr_groups;
180 int * __percpu pmu_disable_count;
181 struct perf_cpu_context * __percpu pmu_cpu_context;
183 int hrtimer_interval_ms;
186 * Fully disable/enable this PMU, can be used to protect from the PMI
187 * as well as for lazy/batch writing of the MSRs.
189 void (*pmu_enable) (struct pmu *pmu); /* optional */
190 void (*pmu_disable) (struct pmu *pmu); /* optional */
193 * Try and initialize the event for this PMU.
194 * Should return -ENOENT when the @event doesn't match this PMU.
196 int (*event_init) (struct perf_event *event);
198 #define PERF_EF_START 0x01 /* start the counter when adding */
199 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
200 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
203 * Adds/Removes a counter to/from the PMU, can be done inside
204 * a transaction, see the ->*_txn() methods.
206 int (*add) (struct perf_event *event, int flags);
207 void (*del) (struct perf_event *event, int flags);
210 * Starts/Stops a counter present on the PMU. The PMI handler
211 * should stop the counter when perf_event_overflow() returns
212 * !0. ->start() will be used to continue.
214 void (*start) (struct perf_event *event, int flags);
215 void (*stop) (struct perf_event *event, int flags);
218 * Updates the counter value of the event.
220 void (*read) (struct perf_event *event);
223 * Group events scheduling is treated as a transaction, add
224 * group events as a whole and perform one schedulability test.
225 * If the test fails, roll back the whole group
227 * Start the transaction, after this ->add() doesn't need to
228 * do schedulability tests.
230 void (*start_txn) (struct pmu *pmu); /* optional */
232 * If ->start_txn() disabled the ->add() schedulability test
233 * then ->commit_txn() is required to perform one. On success
234 * the transaction is closed. On error the transaction is kept
235 * open until ->cancel_txn() is called.
237 int (*commit_txn) (struct pmu *pmu); /* optional */
239 * Will cancel the transaction, assumes ->del() is called
240 * for each successful ->add() during the transaction.
242 void (*cancel_txn) (struct pmu *pmu); /* optional */
245 * Will return the value for perf_event_mmap_page::index for this event,
246 * if no implementation is provided it will default to: event->hw.idx + 1.
248 int (*event_idx) (struct perf_event *event); /*optional */
251 * flush branch stack on context-switches (needed in cpu-wide mode)
253 void (*flush_branch_stack) (void);
257 * enum perf_event_active_state - the states of a event
259 enum perf_event_active_state {
260 PERF_EVENT_STATE_ERROR = -2,
261 PERF_EVENT_STATE_OFF = -1,
262 PERF_EVENT_STATE_INACTIVE = 0,
263 PERF_EVENT_STATE_ACTIVE = 1,
267 struct perf_sample_data;
269 typedef void (*perf_overflow_handler_t)(struct perf_event *,
270 struct perf_sample_data *,
271 struct pt_regs *regs);
273 enum perf_group_flag {
274 PERF_GROUP_SOFTWARE = 0x1,
277 #define SWEVENT_HLIST_BITS 8
278 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
280 struct swevent_hlist {
281 struct hlist_head heads[SWEVENT_HLIST_SIZE];
282 struct rcu_head rcu_head;
285 #define PERF_ATTACH_CONTEXT 0x01
286 #define PERF_ATTACH_GROUP 0x02
287 #define PERF_ATTACH_TASK 0x04
293 * struct perf_event - performance event kernel representation:
296 #ifdef CONFIG_PERF_EVENTS
298 * entry onto perf_event_context::event_list;
299 * modifications require ctx->lock
300 * RCU safe iterations.
302 struct list_head event_entry;
305 * XXX: group_entry and sibling_list should be mutually exclusive;
306 * either you're a sibling on a group, or you're the group leader.
307 * Rework the code to always use the same list element.
309 * Locked for modification by both ctx->mutex and ctx->lock; holding
310 * either sufficies for read.
312 struct list_head group_entry;
313 struct list_head sibling_list;
316 * We need storage to track the entries in perf_pmu_migrate_context; we
317 * cannot use the event_entry because of RCU and we want to keep the
318 * group in tact which avoids us using the other two entries.
320 struct list_head migrate_entry;
322 struct hlist_node hlist_entry;
323 struct list_head active_entry;
326 struct perf_event *group_leader;
329 enum perf_event_active_state state;
330 unsigned int attach_state;
332 atomic64_t child_count;
335 * These are the total time in nanoseconds that the event
336 * has been enabled (i.e. eligible to run, and the task has
337 * been scheduled in, if this is a per-task event)
338 * and running (scheduled onto the CPU), respectively.
340 * They are computed from tstamp_enabled, tstamp_running and
341 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
343 u64 total_time_enabled;
344 u64 total_time_running;
347 * These are timestamps used for computing total_time_enabled
348 * and total_time_running when the event is in INACTIVE or
349 * ACTIVE state, measured in nanoseconds from an arbitrary point
351 * tstamp_enabled: the notional time when the event was enabled
352 * tstamp_running: the notional time when the event was scheduled on
353 * tstamp_stopped: in INACTIVE state, the notional time when the
354 * event was scheduled off.
361 * timestamp shadows the actual context timing but it can
362 * be safely used in NMI interrupt context. It reflects the
363 * context time as it was when the event was last scheduled in.
365 * ctx_time already accounts for ctx->timestamp. Therefore to
366 * compute ctx_time for a sample, simply add perf_clock().
370 struct perf_event_attr attr;
374 struct hw_perf_event hw;
376 struct perf_event_context *ctx;
377 atomic_long_t refcount;
380 * These accumulate total time (in nanoseconds) that children
381 * events have been enabled and running, respectively.
383 atomic64_t child_total_time_enabled;
384 atomic64_t child_total_time_running;
387 * Protect attach/detach and child_list:
389 struct mutex child_mutex;
390 struct list_head child_list;
391 struct perf_event *parent;
396 struct list_head owner_entry;
397 struct task_struct *owner;
400 struct mutex mmap_mutex;
403 struct ring_buffer *rb;
404 struct list_head rb_entry;
407 wait_queue_head_t waitq;
408 struct fasync_struct *fasync;
410 /* delayed work for NMIs and such */
414 struct irq_work pending;
416 atomic_t event_limit;
418 void (*destroy)(struct perf_event *);
419 struct rcu_head rcu_head;
421 struct pid_namespace *ns;
424 perf_overflow_handler_t overflow_handler;
425 void *overflow_handler_context;
427 #ifdef CONFIG_EVENT_TRACING
428 struct ftrace_event_call *tp_event;
429 struct event_filter *filter;
430 #ifdef CONFIG_FUNCTION_TRACER
431 struct ftrace_ops ftrace_ops;
435 #ifdef CONFIG_CGROUP_PERF
436 struct perf_cgroup *cgrp; /* cgroup event is attach to */
437 int cgrp_defer_enabled;
440 #endif /* CONFIG_PERF_EVENTS */
443 enum perf_event_context_type {
449 * struct perf_event_context - event context structure
451 * Used as a container for task events and CPU events as well:
453 struct perf_event_context {
455 enum perf_event_context_type type;
457 * Protect the states of the events in the list,
458 * nr_active, and the list:
462 * Protect the list of events. Locking either mutex or lock
463 * is sufficient to ensure the list doesn't change; to change
464 * the list you need to lock both the mutex and the spinlock.
468 struct list_head pinned_groups;
469 struct list_head flexible_groups;
470 struct list_head event_list;
478 struct task_struct *task;
481 * Context clock, runs when context enabled.
487 * These fields let us detect when two contexts have both
488 * been cloned (inherited) from a common ancestor.
490 struct perf_event_context *parent_ctx;
494 int nr_cgroups; /* cgroup evts */
495 int nr_branch_stack; /* branch_stack evt */
496 struct rcu_head rcu_head;
500 * Number of contexts where an event can trigger:
501 * task, softirq, hardirq, nmi.
503 #define PERF_NR_CONTEXTS 4
506 * struct perf_event_cpu_context - per cpu event context structure
508 struct perf_cpu_context {
509 struct perf_event_context ctx;
510 struct perf_event_context *task_ctx;
513 struct hrtimer hrtimer;
514 ktime_t hrtimer_interval;
515 struct list_head rotation_list;
516 struct pmu *unique_pmu;
517 struct perf_cgroup *cgrp;
520 struct perf_output_handle {
521 struct perf_event *event;
522 struct ring_buffer *rb;
523 unsigned long wakeup;
529 #ifdef CONFIG_PERF_EVENTS
531 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
532 extern void perf_pmu_unregister(struct pmu *pmu);
534 extern int perf_num_counters(void);
535 extern const char *perf_pmu_name(void);
536 extern void __perf_event_task_sched_in(struct task_struct *prev,
537 struct task_struct *task);
538 extern void __perf_event_task_sched_out(struct task_struct *prev,
539 struct task_struct *next);
540 extern int perf_event_init_task(struct task_struct *child);
541 extern void perf_event_exit_task(struct task_struct *child);
542 extern void perf_event_free_task(struct task_struct *task);
543 extern void perf_event_delayed_put(struct task_struct *task);
544 extern void perf_event_print_debug(void);
545 extern void perf_pmu_disable(struct pmu *pmu);
546 extern void perf_pmu_enable(struct pmu *pmu);
547 extern int perf_event_task_disable(void);
548 extern int perf_event_task_enable(void);
549 extern int perf_event_refresh(struct perf_event *event, int refresh);
550 extern void perf_event_update_userpage(struct perf_event *event);
551 extern int perf_event_release_kernel(struct perf_event *event);
552 extern struct perf_event *
553 perf_event_create_kernel_counter(struct perf_event_attr *attr,
555 struct task_struct *task,
556 perf_overflow_handler_t callback,
558 extern void perf_pmu_migrate_context(struct pmu *pmu,
559 int src_cpu, int dst_cpu);
560 extern u64 perf_event_read_value(struct perf_event *event,
561 u64 *enabled, u64 *running);
564 struct perf_sample_data {
581 union perf_mem_data_src data_src;
582 struct perf_callchain_entry *callchain;
583 struct perf_raw_record *raw;
584 struct perf_branch_stack *br_stack;
585 struct perf_regs_user regs_user;
589 * Transaction flags for abort events:
594 static inline void perf_sample_data_init(struct perf_sample_data *data,
595 u64 addr, u64 period)
597 /* remaining struct members initialized in perf_prepare_sample() */
600 data->br_stack = NULL;
601 data->period = period;
602 data->regs_user.abi = PERF_SAMPLE_REGS_ABI_NONE;
603 data->regs_user.regs = NULL;
604 data->stack_user_size = 0;
606 data->data_src.val = 0;
610 extern void perf_output_sample(struct perf_output_handle *handle,
611 struct perf_event_header *header,
612 struct perf_sample_data *data,
613 struct perf_event *event);
614 extern void perf_prepare_sample(struct perf_event_header *header,
615 struct perf_sample_data *data,
616 struct perf_event *event,
617 struct pt_regs *regs);
619 extern int perf_event_overflow(struct perf_event *event,
620 struct perf_sample_data *data,
621 struct pt_regs *regs);
623 static inline bool is_sampling_event(struct perf_event *event)
625 return event->attr.sample_period != 0;
629 * Return 1 for a software event, 0 for a hardware event
631 static inline int is_software_event(struct perf_event *event)
633 return event->pmu->task_ctx_nr == perf_sw_context;
636 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
638 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
640 #ifndef perf_arch_fetch_caller_regs
641 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
645 * Take a snapshot of the regs. Skip ip and frame pointer to
646 * the nth caller. We only need a few of the regs:
647 * - ip for PERF_SAMPLE_IP
648 * - cs for user_mode() tests
649 * - bp for callchains
650 * - eflags, for future purposes, just in case
652 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
654 memset(regs, 0, sizeof(*regs));
656 perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
659 static __always_inline void
660 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
662 struct pt_regs hot_regs;
664 if (static_key_false(&perf_swevent_enabled[event_id])) {
666 perf_fetch_caller_regs(&hot_regs);
669 __perf_sw_event(event_id, nr, regs, addr);
673 extern struct static_key_deferred perf_sched_events;
675 static inline void perf_event_task_sched_in(struct task_struct *prev,
676 struct task_struct *task)
678 if (static_key_false(&perf_sched_events.key))
679 __perf_event_task_sched_in(prev, task);
682 static inline void perf_event_task_sched_out(struct task_struct *prev,
683 struct task_struct *next)
685 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, NULL, 0);
687 if (static_key_false(&perf_sched_events.key))
688 __perf_event_task_sched_out(prev, next);
691 extern void perf_event_mmap(struct vm_area_struct *vma);
692 extern struct perf_guest_info_callbacks *perf_guest_cbs;
693 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
694 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
696 extern void perf_event_comm(struct task_struct *tsk);
697 extern void perf_event_fork(struct task_struct *tsk);
700 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
702 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
703 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
705 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
707 if (entry->nr < PERF_MAX_STACK_DEPTH)
708 entry->ip[entry->nr++] = ip;
711 extern int sysctl_perf_event_paranoid;
712 extern int sysctl_perf_event_mlock;
713 extern int sysctl_perf_event_sample_rate;
714 extern int sysctl_perf_cpu_time_max_percent;
716 extern void perf_sample_event_took(u64 sample_len_ns);
718 extern int perf_proc_update_handler(struct ctl_table *table, int write,
719 void __user *buffer, size_t *lenp,
721 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
722 void __user *buffer, size_t *lenp,
726 static inline bool perf_paranoid_tracepoint_raw(void)
728 return sysctl_perf_event_paranoid > -1;
731 static inline bool perf_paranoid_cpu(void)
733 return sysctl_perf_event_paranoid > 0;
736 static inline bool perf_paranoid_kernel(void)
738 return sysctl_perf_event_paranoid > 1;
741 extern void perf_event_init(void);
742 extern void perf_tp_event(u64 addr, u64 count, void *record,
743 int entry_size, struct pt_regs *regs,
744 struct hlist_head *head, int rctx,
745 struct task_struct *task);
746 extern void perf_bp_event(struct perf_event *event, void *data);
748 #ifndef perf_misc_flags
749 # define perf_misc_flags(regs) \
750 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
751 # define perf_instruction_pointer(regs) instruction_pointer(regs)
754 static inline bool has_branch_stack(struct perf_event *event)
756 return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
759 extern int perf_output_begin(struct perf_output_handle *handle,
760 struct perf_event *event, unsigned int size);
761 extern void perf_output_end(struct perf_output_handle *handle);
762 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
763 const void *buf, unsigned int len);
764 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
766 extern int perf_swevent_get_recursion_context(void);
767 extern void perf_swevent_put_recursion_context(int rctx);
768 extern u64 perf_swevent_set_period(struct perf_event *event);
769 extern void perf_event_enable(struct perf_event *event);
770 extern void perf_event_disable(struct perf_event *event);
771 extern int __perf_event_disable(void *info);
772 extern void perf_event_task_tick(void);
775 perf_event_task_sched_in(struct task_struct *prev,
776 struct task_struct *task) { }
778 perf_event_task_sched_out(struct task_struct *prev,
779 struct task_struct *next) { }
780 static inline int perf_event_init_task(struct task_struct *child) { return 0; }
781 static inline void perf_event_exit_task(struct task_struct *child) { }
782 static inline void perf_event_free_task(struct task_struct *task) { }
783 static inline void perf_event_delayed_put(struct task_struct *task) { }
784 static inline void perf_event_print_debug(void) { }
785 static inline int perf_event_task_disable(void) { return -EINVAL; }
786 static inline int perf_event_task_enable(void) { return -EINVAL; }
787 static inline int perf_event_refresh(struct perf_event *event, int refresh)
793 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) { }
795 perf_bp_event(struct perf_event *event, void *data) { }
797 static inline int perf_register_guest_info_callbacks
798 (struct perf_guest_info_callbacks *callbacks) { return 0; }
799 static inline int perf_unregister_guest_info_callbacks
800 (struct perf_guest_info_callbacks *callbacks) { return 0; }
802 static inline void perf_event_mmap(struct vm_area_struct *vma) { }
803 static inline void perf_event_comm(struct task_struct *tsk) { }
804 static inline void perf_event_fork(struct task_struct *tsk) { }
805 static inline void perf_event_init(void) { }
806 static inline int perf_swevent_get_recursion_context(void) { return -1; }
807 static inline void perf_swevent_put_recursion_context(int rctx) { }
808 static inline u64 perf_swevent_set_period(struct perf_event *event) { return 0; }
809 static inline void perf_event_enable(struct perf_event *event) { }
810 static inline void perf_event_disable(struct perf_event *event) { }
811 static inline int __perf_event_disable(void *info) { return -1; }
812 static inline void perf_event_task_tick(void) { }
815 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_NO_HZ_FULL)
816 extern bool perf_event_can_stop_tick(void);
818 static inline bool perf_event_can_stop_tick(void) { return true; }
821 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
822 extern void perf_restore_debug_store(void);
824 static inline void perf_restore_debug_store(void) { }
827 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
830 * This has to have a higher priority than migration_notifier in sched/core.c.
832 #define perf_cpu_notifier(fn) \
834 static struct notifier_block fn##_nb = \
835 { .notifier_call = fn, .priority = CPU_PRI_PERF }; \
836 unsigned long cpu = smp_processor_id(); \
837 unsigned long flags; \
838 fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE, \
839 (void *)(unsigned long)cpu); \
840 local_irq_save(flags); \
841 fn(&fn##_nb, (unsigned long)CPU_STARTING, \
842 (void *)(unsigned long)cpu); \
843 local_irq_restore(flags); \
844 fn(&fn##_nb, (unsigned long)CPU_ONLINE, \
845 (void *)(unsigned long)cpu); \
846 register_cpu_notifier(&fn##_nb); \
850 struct perf_pmu_events_attr {
851 struct device_attribute attr;
853 const char *event_str;
856 #define PMU_EVENT_ATTR(_name, _var, _id, _show) \
857 static struct perf_pmu_events_attr _var = { \
858 .attr = __ATTR(_name, 0444, _show, NULL), \
862 #define PMU_FORMAT_ATTR(_name, _format) \
864 _name##_show(struct device *dev, \
865 struct device_attribute *attr, \
868 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
869 return sprintf(page, _format "\n"); \
872 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
874 #endif /* _LINUX_PERF_EVENT_H */