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 * single taken branch record layout:
69 * from: source instruction (may not always be a branch insn)
71 * mispred: branch target was mispredicted
72 * predicted: branch target was predicted
74 * support for mispred, predicted is optional. In case it
75 * is not supported mispred = predicted = 0.
77 struct perf_branch_entry {
80 __u64 mispred:1, /* target mispredicted */
81 predicted:1,/* target predicted */
86 * branch stack layout:
87 * nr: number of taken branches stored in entries[]
89 * Note that nr can vary from sample to sample
90 * branches (to, from) are stored from most recent
91 * to least recent, i.e., entries[0] contains the most
94 struct perf_branch_stack {
96 struct perf_branch_entry entries[0];
99 struct perf_regs_user {
101 struct pt_regs *regs;
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 struct task_struct *tp_target;
140 /* for tp_event->class */
141 struct list_head tp_list;
143 #ifdef CONFIG_HAVE_HW_BREAKPOINT
144 struct { /* breakpoint */
146 * Crufty hack to avoid the chicken and egg
147 * problem hw_breakpoint has with context
148 * creation and event initalization.
150 struct task_struct *bp_target;
151 struct arch_hw_breakpoint info;
152 struct list_head bp_list;
157 local64_t prev_count;
160 local64_t period_left;
165 u64 freq_count_stamp;
170 * hw_perf_event::state flags
172 #define PERF_HES_STOPPED 0x01 /* the counter is stopped */
173 #define PERF_HES_UPTODATE 0x02 /* event->count up-to-date */
174 #define PERF_HES_ARCH 0x04
179 * Common implementation detail of pmu::{start,commit,cancel}_txn
181 #define PERF_EVENT_TXN 0x1
184 * struct pmu - generic performance monitoring unit
187 struct list_head entry;
190 const struct attribute_group **attr_groups;
194 int * __percpu pmu_disable_count;
195 struct perf_cpu_context * __percpu pmu_cpu_context;
199 * Fully disable/enable this PMU, can be used to protect from the PMI
200 * as well as for lazy/batch writing of the MSRs.
202 void (*pmu_enable) (struct pmu *pmu); /* optional */
203 void (*pmu_disable) (struct pmu *pmu); /* optional */
206 * Try and initialize the event for this PMU.
207 * Should return -ENOENT when the @event doesn't match this PMU.
209 int (*event_init) (struct perf_event *event);
211 #define PERF_EF_START 0x01 /* start the counter when adding */
212 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
213 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
216 * Adds/Removes a counter to/from the PMU, can be done inside
217 * a transaction, see the ->*_txn() methods.
219 int (*add) (struct perf_event *event, int flags);
220 void (*del) (struct perf_event *event, int flags);
223 * Starts/Stops a counter present on the PMU. The PMI handler
224 * should stop the counter when perf_event_overflow() returns
225 * !0. ->start() will be used to continue.
227 void (*start) (struct perf_event *event, int flags);
228 void (*stop) (struct perf_event *event, int flags);
231 * Updates the counter value of the event.
233 void (*read) (struct perf_event *event);
236 * Group events scheduling is treated as a transaction, add
237 * group events as a whole and perform one schedulability test.
238 * If the test fails, roll back the whole group
240 * Start the transaction, after this ->add() doesn't need to
241 * do schedulability tests.
243 void (*start_txn) (struct pmu *pmu); /* optional */
245 * If ->start_txn() disabled the ->add() schedulability test
246 * then ->commit_txn() is required to perform one. On success
247 * the transaction is closed. On error the transaction is kept
248 * open until ->cancel_txn() is called.
250 int (*commit_txn) (struct pmu *pmu); /* optional */
252 * Will cancel the transaction, assumes ->del() is called
253 * for each successful ->add() during the transaction.
255 void (*cancel_txn) (struct pmu *pmu); /* optional */
258 * Will return the value for perf_event_mmap_page::index for this event,
259 * if no implementation is provided it will default to: event->hw.idx + 1.
261 int (*event_idx) (struct perf_event *event); /*optional */
264 * flush branch stack on context-switches (needed in cpu-wide mode)
266 void (*flush_branch_stack) (void);
270 * enum perf_event_active_state - the states of a event
272 enum perf_event_active_state {
273 PERF_EVENT_STATE_ERROR = -2,
274 PERF_EVENT_STATE_OFF = -1,
275 PERF_EVENT_STATE_INACTIVE = 0,
276 PERF_EVENT_STATE_ACTIVE = 1,
280 struct perf_sample_data;
282 typedef void (*perf_overflow_handler_t)(struct perf_event *,
283 struct perf_sample_data *,
284 struct pt_regs *regs);
286 enum perf_group_flag {
287 PERF_GROUP_SOFTWARE = 0x1,
290 #define SWEVENT_HLIST_BITS 8
291 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
293 struct swevent_hlist {
294 struct hlist_head heads[SWEVENT_HLIST_SIZE];
295 struct rcu_head rcu_head;
298 #define PERF_ATTACH_CONTEXT 0x01
299 #define PERF_ATTACH_GROUP 0x02
300 #define PERF_ATTACH_TASK 0x04
306 * struct perf_event - performance event kernel representation:
309 #ifdef CONFIG_PERF_EVENTS
310 struct list_head group_entry;
311 struct list_head event_entry;
312 struct list_head sibling_list;
313 struct hlist_node hlist_entry;
316 struct perf_event *group_leader;
319 enum perf_event_active_state state;
320 unsigned int attach_state;
322 atomic64_t child_count;
325 * These are the total time in nanoseconds that the event
326 * has been enabled (i.e. eligible to run, and the task has
327 * been scheduled in, if this is a per-task event)
328 * and running (scheduled onto the CPU), respectively.
330 * They are computed from tstamp_enabled, tstamp_running and
331 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
333 u64 total_time_enabled;
334 u64 total_time_running;
337 * These are timestamps used for computing total_time_enabled
338 * and total_time_running when the event is in INACTIVE or
339 * ACTIVE state, measured in nanoseconds from an arbitrary point
341 * tstamp_enabled: the notional time when the event was enabled
342 * tstamp_running: the notional time when the event was scheduled on
343 * tstamp_stopped: in INACTIVE state, the notional time when the
344 * event was scheduled off.
351 * timestamp shadows the actual context timing but it can
352 * be safely used in NMI interrupt context. It reflects the
353 * context time as it was when the event was last scheduled in.
355 * ctx_time already accounts for ctx->timestamp. Therefore to
356 * compute ctx_time for a sample, simply add perf_clock().
360 struct perf_event_attr attr;
364 struct hw_perf_event hw;
366 struct perf_event_context *ctx;
367 atomic_long_t refcount;
370 * These accumulate total time (in nanoseconds) that children
371 * events have been enabled and running, respectively.
373 atomic64_t child_total_time_enabled;
374 atomic64_t child_total_time_running;
377 * Protect attach/detach and child_list:
379 struct mutex child_mutex;
380 struct list_head child_list;
381 struct perf_event *parent;
386 struct list_head owner_entry;
387 struct task_struct *owner;
390 struct mutex mmap_mutex;
393 struct ring_buffer *rb;
394 struct list_head rb_entry;
397 wait_queue_head_t waitq;
398 struct fasync_struct *fasync;
400 /* delayed work for NMIs and such */
404 struct irq_work pending;
406 atomic_t event_limit;
408 void (*destroy)(struct perf_event *);
409 struct rcu_head rcu_head;
411 struct pid_namespace *ns;
414 perf_overflow_handler_t overflow_handler;
415 void *overflow_handler_context;
417 #ifdef CONFIG_EVENT_TRACING
418 struct ftrace_event_call *tp_event;
419 struct event_filter *filter;
420 #ifdef CONFIG_FUNCTION_TRACER
421 struct ftrace_ops ftrace_ops;
425 #ifdef CONFIG_CGROUP_PERF
426 struct perf_cgroup *cgrp; /* cgroup event is attach to */
427 int cgrp_defer_enabled;
430 #endif /* CONFIG_PERF_EVENTS */
433 enum perf_event_context_type {
439 * struct perf_event_context - event context structure
441 * Used as a container for task events and CPU events as well:
443 struct perf_event_context {
445 enum perf_event_context_type type;
447 * Protect the states of the events in the list,
448 * nr_active, and the list:
452 * Protect the list of events. Locking either mutex or lock
453 * is sufficient to ensure the list doesn't change; to change
454 * the list you need to lock both the mutex and the spinlock.
458 struct list_head pinned_groups;
459 struct list_head flexible_groups;
460 struct list_head event_list;
468 struct task_struct *task;
471 * Context clock, runs when context enabled.
477 * These fields let us detect when two contexts have both
478 * been cloned (inherited) from a common ancestor.
480 struct perf_event_context *parent_ctx;
484 int nr_cgroups; /* cgroup evts */
485 int nr_branch_stack; /* branch_stack evt */
486 struct rcu_head rcu_head;
490 * Number of contexts where an event can trigger:
491 * task, softirq, hardirq, nmi.
493 #define PERF_NR_CONTEXTS 4
496 * struct perf_event_cpu_context - per cpu event context structure
498 struct perf_cpu_context {
499 struct perf_event_context ctx;
500 struct perf_event_context *task_ctx;
503 struct list_head rotation_list;
504 int jiffies_interval;
505 struct pmu *unique_pmu;
506 struct perf_cgroup *cgrp;
509 struct perf_output_handle {
510 struct perf_event *event;
511 struct ring_buffer *rb;
512 unsigned long wakeup;
518 #ifdef CONFIG_PERF_EVENTS
520 extern int perf_pmu_register(struct pmu *pmu, char *name, int type);
521 extern void perf_pmu_unregister(struct pmu *pmu);
523 extern int perf_num_counters(void);
524 extern const char *perf_pmu_name(void);
525 extern void __perf_event_task_sched_in(struct task_struct *prev,
526 struct task_struct *task);
527 extern void __perf_event_task_sched_out(struct task_struct *prev,
528 struct task_struct *next);
529 extern int perf_event_init_task(struct task_struct *child);
530 extern void perf_event_exit_task(struct task_struct *child);
531 extern void perf_event_free_task(struct task_struct *task);
532 extern void perf_event_delayed_put(struct task_struct *task);
533 extern void perf_event_print_debug(void);
534 extern void perf_pmu_disable(struct pmu *pmu);
535 extern void perf_pmu_enable(struct pmu *pmu);
536 extern int perf_event_task_disable(void);
537 extern int perf_event_task_enable(void);
538 extern int perf_event_refresh(struct perf_event *event, int refresh);
539 extern void perf_event_update_userpage(struct perf_event *event);
540 extern int perf_event_release_kernel(struct perf_event *event);
541 extern struct perf_event *
542 perf_event_create_kernel_counter(struct perf_event_attr *attr,
544 struct task_struct *task,
545 perf_overflow_handler_t callback,
547 extern void perf_pmu_migrate_context(struct pmu *pmu,
548 int src_cpu, int dst_cpu);
549 extern u64 perf_event_read_value(struct perf_event *event,
550 u64 *enabled, u64 *running);
553 struct perf_sample_data {
570 union perf_mem_data_src data_src;
571 struct perf_callchain_entry *callchain;
572 struct perf_raw_record *raw;
573 struct perf_branch_stack *br_stack;
574 struct perf_regs_user regs_user;
579 static inline void perf_sample_data_init(struct perf_sample_data *data,
580 u64 addr, u64 period)
582 /* remaining struct members initialized in perf_prepare_sample() */
585 data->br_stack = NULL;
586 data->period = period;
587 data->regs_user.abi = PERF_SAMPLE_REGS_ABI_NONE;
588 data->regs_user.regs = NULL;
589 data->stack_user_size = 0;
591 data->data_src.val = 0;
594 extern void perf_output_sample(struct perf_output_handle *handle,
595 struct perf_event_header *header,
596 struct perf_sample_data *data,
597 struct perf_event *event);
598 extern void perf_prepare_sample(struct perf_event_header *header,
599 struct perf_sample_data *data,
600 struct perf_event *event,
601 struct pt_regs *regs);
603 extern int perf_event_overflow(struct perf_event *event,
604 struct perf_sample_data *data,
605 struct pt_regs *regs);
607 static inline bool is_sampling_event(struct perf_event *event)
609 return event->attr.sample_period != 0;
613 * Return 1 for a software event, 0 for a hardware event
615 static inline int is_software_event(struct perf_event *event)
617 return event->pmu->task_ctx_nr == perf_sw_context;
620 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
622 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
624 #ifndef perf_arch_fetch_caller_regs
625 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
629 * Take a snapshot of the regs. Skip ip and frame pointer to
630 * the nth caller. We only need a few of the regs:
631 * - ip for PERF_SAMPLE_IP
632 * - cs for user_mode() tests
633 * - bp for callchains
634 * - eflags, for future purposes, just in case
636 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
638 memset(regs, 0, sizeof(*regs));
640 perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
643 static __always_inline void
644 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
646 struct pt_regs hot_regs;
648 if (static_key_false(&perf_swevent_enabled[event_id])) {
650 perf_fetch_caller_regs(&hot_regs);
653 __perf_sw_event(event_id, nr, regs, addr);
657 extern struct static_key_deferred perf_sched_events;
659 static inline void perf_event_task_sched_in(struct task_struct *prev,
660 struct task_struct *task)
662 if (static_key_false(&perf_sched_events.key))
663 __perf_event_task_sched_in(prev, task);
666 static inline void perf_event_task_sched_out(struct task_struct *prev,
667 struct task_struct *next)
669 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, NULL, 0);
671 if (static_key_false(&perf_sched_events.key))
672 __perf_event_task_sched_out(prev, next);
675 extern void perf_event_mmap(struct vm_area_struct *vma);
676 extern struct perf_guest_info_callbacks *perf_guest_cbs;
677 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
678 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
680 extern void perf_event_comm(struct task_struct *tsk);
681 extern void perf_event_fork(struct task_struct *tsk);
684 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
686 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
687 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
689 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
691 if (entry->nr < PERF_MAX_STACK_DEPTH)
692 entry->ip[entry->nr++] = ip;
695 extern int sysctl_perf_event_paranoid;
696 extern int sysctl_perf_event_mlock;
697 extern int sysctl_perf_event_sample_rate;
698 extern int sysctl_perf_cpu_time_max_percent;
700 extern void perf_sample_event_took(u64 sample_len_ns);
702 extern int perf_proc_update_handler(struct ctl_table *table, int write,
703 void __user *buffer, size_t *lenp,
705 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
706 void __user *buffer, size_t *lenp,
710 static inline bool perf_paranoid_tracepoint_raw(void)
712 return sysctl_perf_event_paranoid > -1;
715 static inline bool perf_paranoid_cpu(void)
717 return sysctl_perf_event_paranoid > 0;
720 static inline bool perf_paranoid_kernel(void)
722 return sysctl_perf_event_paranoid > 1;
725 extern void perf_event_init(void);
726 extern void perf_tp_event(u64 addr, u64 count, void *record,
727 int entry_size, struct pt_regs *regs,
728 struct hlist_head *head, int rctx,
729 struct task_struct *task);
730 extern void perf_bp_event(struct perf_event *event, void *data);
732 #ifndef perf_misc_flags
733 # define perf_misc_flags(regs) \
734 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
735 # define perf_instruction_pointer(regs) instruction_pointer(regs)
738 static inline bool has_branch_stack(struct perf_event *event)
740 return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
743 extern int perf_output_begin(struct perf_output_handle *handle,
744 struct perf_event *event, unsigned int size);
745 extern void perf_output_end(struct perf_output_handle *handle);
746 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
747 const void *buf, unsigned int len);
748 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
750 extern int perf_swevent_get_recursion_context(void);
751 extern void perf_swevent_put_recursion_context(int rctx);
752 extern void perf_event_enable(struct perf_event *event);
753 extern void perf_event_disable(struct perf_event *event);
754 extern int __perf_event_disable(void *info);
755 extern void perf_event_task_tick(void);
758 perf_event_task_sched_in(struct task_struct *prev,
759 struct task_struct *task) { }
761 perf_event_task_sched_out(struct task_struct *prev,
762 struct task_struct *next) { }
763 static inline int perf_event_init_task(struct task_struct *child) { return 0; }
764 static inline void perf_event_exit_task(struct task_struct *child) { }
765 static inline void perf_event_free_task(struct task_struct *task) { }
766 static inline void perf_event_delayed_put(struct task_struct *task) { }
767 static inline void perf_event_print_debug(void) { }
768 static inline int perf_event_task_disable(void) { return -EINVAL; }
769 static inline int perf_event_task_enable(void) { return -EINVAL; }
770 static inline int perf_event_refresh(struct perf_event *event, int refresh)
776 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) { }
778 perf_bp_event(struct perf_event *event, void *data) { }
780 static inline int perf_register_guest_info_callbacks
781 (struct perf_guest_info_callbacks *callbacks) { return 0; }
782 static inline int perf_unregister_guest_info_callbacks
783 (struct perf_guest_info_callbacks *callbacks) { return 0; }
785 static inline void perf_event_mmap(struct vm_area_struct *vma) { }
786 static inline void perf_event_comm(struct task_struct *tsk) { }
787 static inline void perf_event_fork(struct task_struct *tsk) { }
788 static inline void perf_event_init(void) { }
789 static inline int perf_swevent_get_recursion_context(void) { return -1; }
790 static inline void perf_swevent_put_recursion_context(int rctx) { }
791 static inline void perf_event_enable(struct perf_event *event) { }
792 static inline void perf_event_disable(struct perf_event *event) { }
793 static inline int __perf_event_disable(void *info) { return -1; }
794 static inline void perf_event_task_tick(void) { }
797 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_NO_HZ_FULL)
798 extern bool perf_event_can_stop_tick(void);
800 static inline bool perf_event_can_stop_tick(void) { return true; }
803 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
804 extern void perf_restore_debug_store(void);
806 static inline void perf_restore_debug_store(void) { }
809 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
812 * This has to have a higher priority than migration_notifier in sched.c.
814 #define perf_cpu_notifier(fn) \
816 static struct notifier_block fn##_nb __cpuinitdata = \
817 { .notifier_call = fn, .priority = CPU_PRI_PERF }; \
818 unsigned long cpu = smp_processor_id(); \
819 unsigned long flags; \
820 fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE, \
821 (void *)(unsigned long)cpu); \
822 local_irq_save(flags); \
823 fn(&fn##_nb, (unsigned long)CPU_STARTING, \
824 (void *)(unsigned long)cpu); \
825 local_irq_restore(flags); \
826 fn(&fn##_nb, (unsigned long)CPU_ONLINE, \
827 (void *)(unsigned long)cpu); \
828 register_cpu_notifier(&fn##_nb); \
832 struct perf_pmu_events_attr {
833 struct device_attribute attr;
835 const char *event_str;
838 #define PMU_EVENT_ATTR(_name, _var, _id, _show) \
839 static struct perf_pmu_events_attr _var = { \
840 .attr = __ATTR(_name, 0444, _show, NULL), \
844 #define PMU_FORMAT_ATTR(_name, _format) \
846 _name##_show(struct device *dev, \
847 struct device_attribute *attr, \
850 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
851 return sprintf(page, _format "\n"); \
854 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
856 #endif /* _LINUX_PERF_EVENT_H */