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/atomic.h>
52 #include <linux/sysfs.h>
53 #include <linux/perf_regs.h>
54 #include <asm/local.h>
56 struct perf_callchain_entry {
58 __u64 ip[PERF_MAX_STACK_DEPTH];
61 struct perf_raw_record {
67 * branch stack layout:
68 * nr: number of taken branches stored in entries[]
70 * Note that nr can vary from sample to sample
71 * branches (to, from) are stored from most recent
72 * to least recent, i.e., entries[0] contains the most
75 struct perf_branch_stack {
77 struct perf_branch_entry entries[0];
80 struct perf_regs_user {
88 * extra PMU register associated with an event
90 struct hw_perf_event_extra {
91 u64 config; /* register value */
92 unsigned int reg; /* register address or index */
93 int alloc; /* extra register already allocated */
94 int idx; /* index in shared_regs->regs[] */
97 struct event_constraint;
100 * struct hw_perf_event - performance event hardware details:
102 struct hw_perf_event {
103 #ifdef CONFIG_PERF_EVENTS
105 struct { /* hardware */
108 unsigned long config_base;
109 unsigned long event_base;
110 int event_base_rdpmc;
115 struct hw_perf_event_extra extra_reg;
116 struct hw_perf_event_extra branch_reg;
118 struct event_constraint *constraint;
120 struct { /* software */
121 struct hrtimer hrtimer;
123 struct { /* tracepoint */
124 struct task_struct *tp_target;
125 /* for tp_event->class */
126 struct list_head tp_list;
128 #ifdef CONFIG_HAVE_HW_BREAKPOINT
129 struct { /* breakpoint */
131 * Crufty hack to avoid the chicken and egg
132 * problem hw_breakpoint has with context
133 * creation and event initalization.
135 struct task_struct *bp_target;
136 struct arch_hw_breakpoint info;
137 struct list_head bp_list;
142 local64_t prev_count;
145 local64_t period_left;
150 u64 freq_count_stamp;
155 * hw_perf_event::state flags
157 #define PERF_HES_STOPPED 0x01 /* the counter is stopped */
158 #define PERF_HES_UPTODATE 0x02 /* event->count up-to-date */
159 #define PERF_HES_ARCH 0x04
164 * Common implementation detail of pmu::{start,commit,cancel}_txn
166 #define PERF_EVENT_TXN 0x1
169 * struct pmu - generic performance monitoring unit
172 struct list_head entry;
175 const struct attribute_group **attr_groups;
179 int * __percpu pmu_disable_count;
180 struct perf_cpu_context * __percpu pmu_cpu_context;
182 int hrtimer_interval_ms;
185 * Fully disable/enable this PMU, can be used to protect from the PMI
186 * as well as for lazy/batch writing of the MSRs.
188 void (*pmu_enable) (struct pmu *pmu); /* optional */
189 void (*pmu_disable) (struct pmu *pmu); /* optional */
192 * Try and initialize the event for this PMU.
193 * Should return -ENOENT when the @event doesn't match this PMU.
195 int (*event_init) (struct perf_event *event);
197 #define PERF_EF_START 0x01 /* start the counter when adding */
198 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
199 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
202 * Adds/Removes a counter to/from the PMU, can be done inside
203 * a transaction, see the ->*_txn() methods.
205 int (*add) (struct perf_event *event, int flags);
206 void (*del) (struct perf_event *event, int flags);
209 * Starts/Stops a counter present on the PMU. The PMI handler
210 * should stop the counter when perf_event_overflow() returns
211 * !0. ->start() will be used to continue.
213 void (*start) (struct perf_event *event, int flags);
214 void (*stop) (struct perf_event *event, int flags);
217 * Updates the counter value of the event.
219 void (*read) (struct perf_event *event);
222 * Group events scheduling is treated as a transaction, add
223 * group events as a whole and perform one schedulability test.
224 * If the test fails, roll back the whole group
226 * Start the transaction, after this ->add() doesn't need to
227 * do schedulability tests.
229 void (*start_txn) (struct pmu *pmu); /* optional */
231 * If ->start_txn() disabled the ->add() schedulability test
232 * then ->commit_txn() is required to perform one. On success
233 * the transaction is closed. On error the transaction is kept
234 * open until ->cancel_txn() is called.
236 int (*commit_txn) (struct pmu *pmu); /* optional */
238 * Will cancel the transaction, assumes ->del() is called
239 * for each successful ->add() during the transaction.
241 void (*cancel_txn) (struct pmu *pmu); /* optional */
244 * Will return the value for perf_event_mmap_page::index for this event,
245 * if no implementation is provided it will default to: event->hw.idx + 1.
247 int (*event_idx) (struct perf_event *event); /*optional */
250 * flush branch stack on context-switches (needed in cpu-wide mode)
252 void (*flush_branch_stack) (void);
256 * enum perf_event_active_state - the states of a event
258 enum perf_event_active_state {
259 PERF_EVENT_STATE_ERROR = -2,
260 PERF_EVENT_STATE_OFF = -1,
261 PERF_EVENT_STATE_INACTIVE = 0,
262 PERF_EVENT_STATE_ACTIVE = 1,
266 struct perf_sample_data;
268 typedef void (*perf_overflow_handler_t)(struct perf_event *,
269 struct perf_sample_data *,
270 struct pt_regs *regs);
272 enum perf_group_flag {
273 PERF_GROUP_SOFTWARE = 0x1,
276 #define SWEVENT_HLIST_BITS 8
277 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
279 struct swevent_hlist {
280 struct hlist_head heads[SWEVENT_HLIST_SIZE];
281 struct rcu_head rcu_head;
284 #define PERF_ATTACH_CONTEXT 0x01
285 #define PERF_ATTACH_GROUP 0x02
286 #define PERF_ATTACH_TASK 0x04
292 * struct perf_event - performance event kernel representation:
295 #ifdef CONFIG_PERF_EVENTS
296 struct list_head group_entry;
297 struct list_head event_entry;
298 struct list_head sibling_list;
299 struct hlist_node hlist_entry;
302 struct perf_event *group_leader;
305 enum perf_event_active_state state;
306 unsigned int attach_state;
308 atomic64_t child_count;
311 * These are the total time in nanoseconds that the event
312 * has been enabled (i.e. eligible to run, and the task has
313 * been scheduled in, if this is a per-task event)
314 * and running (scheduled onto the CPU), respectively.
316 * They are computed from tstamp_enabled, tstamp_running and
317 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
319 u64 total_time_enabled;
320 u64 total_time_running;
323 * These are timestamps used for computing total_time_enabled
324 * and total_time_running when the event is in INACTIVE or
325 * ACTIVE state, measured in nanoseconds from an arbitrary point
327 * tstamp_enabled: the notional time when the event was enabled
328 * tstamp_running: the notional time when the event was scheduled on
329 * tstamp_stopped: in INACTIVE state, the notional time when the
330 * event was scheduled off.
337 * timestamp shadows the actual context timing but it can
338 * be safely used in NMI interrupt context. It reflects the
339 * context time as it was when the event was last scheduled in.
341 * ctx_time already accounts for ctx->timestamp. Therefore to
342 * compute ctx_time for a sample, simply add perf_clock().
346 struct perf_event_attr attr;
350 struct hw_perf_event hw;
352 struct perf_event_context *ctx;
353 atomic_long_t refcount;
356 * These accumulate total time (in nanoseconds) that children
357 * events have been enabled and running, respectively.
359 atomic64_t child_total_time_enabled;
360 atomic64_t child_total_time_running;
363 * Protect attach/detach and child_list:
365 struct mutex child_mutex;
366 struct list_head child_list;
367 struct perf_event *parent;
372 struct list_head owner_entry;
373 struct task_struct *owner;
376 struct mutex mmap_mutex;
379 struct ring_buffer *rb;
380 struct list_head rb_entry;
383 wait_queue_head_t waitq;
384 struct fasync_struct *fasync;
386 /* delayed work for NMIs and such */
390 struct irq_work pending;
392 atomic_t event_limit;
394 void (*destroy)(struct perf_event *);
395 struct rcu_head rcu_head;
397 struct pid_namespace *ns;
400 perf_overflow_handler_t overflow_handler;
401 void *overflow_handler_context;
403 #ifdef CONFIG_EVENT_TRACING
404 struct ftrace_event_call *tp_event;
405 struct event_filter *filter;
406 #ifdef CONFIG_FUNCTION_TRACER
407 struct ftrace_ops ftrace_ops;
411 #ifdef CONFIG_CGROUP_PERF
412 struct perf_cgroup *cgrp; /* cgroup event is attach to */
413 int cgrp_defer_enabled;
416 #endif /* CONFIG_PERF_EVENTS */
419 enum perf_event_context_type {
425 * struct perf_event_context - event context structure
427 * Used as a container for task events and CPU events as well:
429 struct perf_event_context {
431 enum perf_event_context_type type;
433 * Protect the states of the events in the list,
434 * nr_active, and the list:
438 * Protect the list of events. Locking either mutex or lock
439 * is sufficient to ensure the list doesn't change; to change
440 * the list you need to lock both the mutex and the spinlock.
444 struct list_head pinned_groups;
445 struct list_head flexible_groups;
446 struct list_head event_list;
454 struct task_struct *task;
457 * Context clock, runs when context enabled.
463 * These fields let us detect when two contexts have both
464 * been cloned (inherited) from a common ancestor.
466 struct perf_event_context *parent_ctx;
470 int nr_cgroups; /* cgroup evts */
471 int nr_branch_stack; /* branch_stack evt */
472 struct rcu_head rcu_head;
476 * Number of contexts where an event can trigger:
477 * task, softirq, hardirq, nmi.
479 #define PERF_NR_CONTEXTS 4
482 * struct perf_event_cpu_context - per cpu event context structure
484 struct perf_cpu_context {
485 struct perf_event_context ctx;
486 struct perf_event_context *task_ctx;
489 struct hrtimer hrtimer;
490 ktime_t hrtimer_interval;
491 struct list_head rotation_list;
492 struct pmu *unique_pmu;
493 struct perf_cgroup *cgrp;
496 struct perf_output_handle {
497 struct perf_event *event;
498 struct ring_buffer *rb;
499 unsigned long wakeup;
505 #ifdef CONFIG_PERF_EVENTS
507 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
508 extern void perf_pmu_unregister(struct pmu *pmu);
510 extern int perf_num_counters(void);
511 extern const char *perf_pmu_name(void);
512 extern void __perf_event_task_sched_in(struct task_struct *prev,
513 struct task_struct *task);
514 extern void __perf_event_task_sched_out(struct task_struct *prev,
515 struct task_struct *next);
516 extern int perf_event_init_task(struct task_struct *child);
517 extern void perf_event_exit_task(struct task_struct *child);
518 extern void perf_event_free_task(struct task_struct *task);
519 extern void perf_event_delayed_put(struct task_struct *task);
520 extern void perf_event_print_debug(void);
521 extern void perf_pmu_disable(struct pmu *pmu);
522 extern void perf_pmu_enable(struct pmu *pmu);
523 extern int perf_event_task_disable(void);
524 extern int perf_event_task_enable(void);
525 extern int perf_event_refresh(struct perf_event *event, int refresh);
526 extern void perf_event_update_userpage(struct perf_event *event);
527 extern int perf_event_release_kernel(struct perf_event *event);
528 extern struct perf_event *
529 perf_event_create_kernel_counter(struct perf_event_attr *attr,
531 struct task_struct *task,
532 perf_overflow_handler_t callback,
534 extern void perf_pmu_migrate_context(struct pmu *pmu,
535 int src_cpu, int dst_cpu);
536 extern u64 perf_event_read_value(struct perf_event *event,
537 u64 *enabled, u64 *running);
540 struct perf_sample_data {
557 union perf_mem_data_src data_src;
558 struct perf_callchain_entry *callchain;
559 struct perf_raw_record *raw;
560 struct perf_branch_stack *br_stack;
561 struct perf_regs_user regs_user;
566 static inline void perf_sample_data_init(struct perf_sample_data *data,
567 u64 addr, u64 period)
569 /* remaining struct members initialized in perf_prepare_sample() */
572 data->br_stack = NULL;
573 data->period = period;
574 data->regs_user.abi = PERF_SAMPLE_REGS_ABI_NONE;
575 data->regs_user.regs = NULL;
576 data->stack_user_size = 0;
578 data->data_src.val = 0;
581 extern void perf_output_sample(struct perf_output_handle *handle,
582 struct perf_event_header *header,
583 struct perf_sample_data *data,
584 struct perf_event *event);
585 extern void perf_prepare_sample(struct perf_event_header *header,
586 struct perf_sample_data *data,
587 struct perf_event *event,
588 struct pt_regs *regs);
590 extern int perf_event_overflow(struct perf_event *event,
591 struct perf_sample_data *data,
592 struct pt_regs *regs);
594 static inline bool is_sampling_event(struct perf_event *event)
596 return event->attr.sample_period != 0;
600 * Return 1 for a software event, 0 for a hardware event
602 static inline int is_software_event(struct perf_event *event)
604 return event->pmu->task_ctx_nr == perf_sw_context;
607 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
609 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
611 #ifndef perf_arch_fetch_caller_regs
612 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
616 * Take a snapshot of the regs. Skip ip and frame pointer to
617 * the nth caller. We only need a few of the regs:
618 * - ip for PERF_SAMPLE_IP
619 * - cs for user_mode() tests
620 * - bp for callchains
621 * - eflags, for future purposes, just in case
623 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
625 memset(regs, 0, sizeof(*regs));
627 perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
630 static __always_inline void
631 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
633 struct pt_regs hot_regs;
635 if (static_key_false(&perf_swevent_enabled[event_id])) {
637 perf_fetch_caller_regs(&hot_regs);
640 __perf_sw_event(event_id, nr, regs, addr);
644 extern struct static_key_deferred perf_sched_events;
646 static inline void perf_event_task_sched_in(struct task_struct *prev,
647 struct task_struct *task)
649 if (static_key_false(&perf_sched_events.key))
650 __perf_event_task_sched_in(prev, task);
653 static inline void perf_event_task_sched_out(struct task_struct *prev,
654 struct task_struct *next)
656 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, NULL, 0);
658 if (static_key_false(&perf_sched_events.key))
659 __perf_event_task_sched_out(prev, next);
662 extern void perf_event_mmap(struct vm_area_struct *vma);
663 extern struct perf_guest_info_callbacks *perf_guest_cbs;
664 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
665 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
667 extern void perf_event_comm(struct task_struct *tsk);
668 extern void perf_event_fork(struct task_struct *tsk);
671 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
673 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
674 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
676 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
678 if (entry->nr < PERF_MAX_STACK_DEPTH)
679 entry->ip[entry->nr++] = ip;
682 extern int sysctl_perf_event_paranoid;
683 extern int sysctl_perf_event_mlock;
684 extern int sysctl_perf_event_sample_rate;
685 extern int sysctl_perf_cpu_time_max_percent;
687 extern void perf_sample_event_took(u64 sample_len_ns);
689 extern int perf_proc_update_handler(struct ctl_table *table, int write,
690 void __user *buffer, size_t *lenp,
692 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
693 void __user *buffer, size_t *lenp,
697 static inline bool perf_paranoid_tracepoint_raw(void)
699 return sysctl_perf_event_paranoid > -1;
702 static inline bool perf_paranoid_cpu(void)
704 return sysctl_perf_event_paranoid > 0;
707 static inline bool perf_paranoid_kernel(void)
709 return sysctl_perf_event_paranoid > 1;
712 extern void perf_event_init(void);
713 extern void perf_tp_event(u64 addr, u64 count, void *record,
714 int entry_size, struct pt_regs *regs,
715 struct hlist_head *head, int rctx,
716 struct task_struct *task);
717 extern void perf_bp_event(struct perf_event *event, void *data);
719 #ifndef perf_misc_flags
720 # define perf_misc_flags(regs) \
721 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
722 # define perf_instruction_pointer(regs) instruction_pointer(regs)
725 static inline bool has_branch_stack(struct perf_event *event)
727 return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
730 extern int perf_output_begin(struct perf_output_handle *handle,
731 struct perf_event *event, unsigned int size);
732 extern void perf_output_end(struct perf_output_handle *handle);
733 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
734 const void *buf, unsigned int len);
735 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
737 extern int perf_swevent_get_recursion_context(void);
738 extern void perf_swevent_put_recursion_context(int rctx);
739 extern u64 perf_swevent_set_period(struct perf_event *event);
740 extern void perf_event_enable(struct perf_event *event);
741 extern void perf_event_disable(struct perf_event *event);
742 extern int __perf_event_disable(void *info);
743 extern void perf_event_task_tick(void);
746 perf_event_task_sched_in(struct task_struct *prev,
747 struct task_struct *task) { }
749 perf_event_task_sched_out(struct task_struct *prev,
750 struct task_struct *next) { }
751 static inline int perf_event_init_task(struct task_struct *child) { return 0; }
752 static inline void perf_event_exit_task(struct task_struct *child) { }
753 static inline void perf_event_free_task(struct task_struct *task) { }
754 static inline void perf_event_delayed_put(struct task_struct *task) { }
755 static inline void perf_event_print_debug(void) { }
756 static inline int perf_event_task_disable(void) { return -EINVAL; }
757 static inline int perf_event_task_enable(void) { return -EINVAL; }
758 static inline int perf_event_refresh(struct perf_event *event, int refresh)
764 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) { }
766 perf_bp_event(struct perf_event *event, void *data) { }
768 static inline int perf_register_guest_info_callbacks
769 (struct perf_guest_info_callbacks *callbacks) { return 0; }
770 static inline int perf_unregister_guest_info_callbacks
771 (struct perf_guest_info_callbacks *callbacks) { return 0; }
773 static inline void perf_event_mmap(struct vm_area_struct *vma) { }
774 static inline void perf_event_comm(struct task_struct *tsk) { }
775 static inline void perf_event_fork(struct task_struct *tsk) { }
776 static inline void perf_event_init(void) { }
777 static inline int perf_swevent_get_recursion_context(void) { return -1; }
778 static inline void perf_swevent_put_recursion_context(int rctx) { }
779 static inline u64 perf_swevent_set_period(struct perf_event *event) { return 0; }
780 static inline void perf_event_enable(struct perf_event *event) { }
781 static inline void perf_event_disable(struct perf_event *event) { }
782 static inline int __perf_event_disable(void *info) { return -1; }
783 static inline void perf_event_task_tick(void) { }
786 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_NO_HZ_FULL)
787 extern bool perf_event_can_stop_tick(void);
789 static inline bool perf_event_can_stop_tick(void) { return true; }
792 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
793 extern void perf_restore_debug_store(void);
795 static inline void perf_restore_debug_store(void) { }
798 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
801 * This has to have a higher priority than migration_notifier in sched/core.c.
803 #define perf_cpu_notifier(fn) \
805 static struct notifier_block fn##_nb = \
806 { .notifier_call = fn, .priority = CPU_PRI_PERF }; \
807 unsigned long cpu = smp_processor_id(); \
808 unsigned long flags; \
809 fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE, \
810 (void *)(unsigned long)cpu); \
811 local_irq_save(flags); \
812 fn(&fn##_nb, (unsigned long)CPU_STARTING, \
813 (void *)(unsigned long)cpu); \
814 local_irq_restore(flags); \
815 fn(&fn##_nb, (unsigned long)CPU_ONLINE, \
816 (void *)(unsigned long)cpu); \
817 register_cpu_notifier(&fn##_nb); \
821 struct perf_pmu_events_attr {
822 struct device_attribute attr;
824 const char *event_str;
827 #define PMU_EVENT_ATTR(_name, _var, _id, _show) \
828 static struct perf_pmu_events_attr _var = { \
829 .attr = __ATTR(_name, 0444, _show, NULL), \
833 #define PMU_FORMAT_ATTR(_name, _format) \
835 _name##_show(struct device *dev, \
836 struct device_attribute *attr, \
839 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
840 return sprintf(page, _format "\n"); \
843 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
845 #endif /* _LINUX_PERF_EVENT_H */