4 * Copyright (C) 2008-2009, Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2011, Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2011, Red Hat, Inc., Peter Zijlstra
8 * Data type definitions, declarations, prototypes.
10 * Started by: Thomas Gleixner and Ingo Molnar
12 * For licencing details see kernel-base/COPYING
14 #ifndef _LINUX_PERF_EVENT_H
15 #define _LINUX_PERF_EVENT_H
17 #include <uapi/linux/perf_event.h>
20 * Kernel-internal data types and definitions:
23 #ifdef CONFIG_PERF_EVENTS
24 # include <asm/perf_event.h>
25 # include <asm/local64.h>
28 struct perf_guest_info_callbacks {
29 int (*is_in_guest)(void);
30 int (*is_user_mode)(void);
31 unsigned long (*get_guest_ip)(void);
34 #ifdef CONFIG_HAVE_HW_BREAKPOINT
35 #include <asm/hw_breakpoint.h>
38 #include <linux/list.h>
39 #include <linux/mutex.h>
40 #include <linux/rculist.h>
41 #include <linux/rcupdate.h>
42 #include <linux/spinlock.h>
43 #include <linux/hrtimer.h>
45 #include <linux/pid_namespace.h>
46 #include <linux/workqueue.h>
47 #include <linux/ftrace.h>
48 #include <linux/cpu.h>
49 #include <linux/irq_work.h>
50 #include <linux/static_key.h>
51 #include <linux/jump_label_ratelimit.h>
52 #include <linux/atomic.h>
53 #include <linux/sysfs.h>
54 #include <linux/perf_regs.h>
55 #include <linux/workqueue.h>
56 #include <linux/cgroup.h>
57 #include <asm/local.h>
59 struct perf_callchain_entry {
61 __u64 ip[0]; /* /proc/sys/kernel/perf_event_max_stack */
64 struct perf_callchain_entry_ctx {
65 struct perf_callchain_entry *entry;
72 typedef unsigned long (*perf_copy_f)(void *dst, const void *src,
73 unsigned long off, unsigned long len);
75 struct perf_raw_frag {
77 struct perf_raw_frag *next;
85 struct perf_raw_record {
86 struct perf_raw_frag frag;
91 * branch stack layout:
92 * nr: number of taken branches stored in entries[]
94 * Note that nr can vary from sample to sample
95 * branches (to, from) are stored from most recent
96 * to least recent, i.e., entries[0] contains the most
99 struct perf_branch_stack {
101 struct perf_branch_entry entries[0];
107 * extra PMU register associated with an event
109 struct hw_perf_event_extra {
110 u64 config; /* register value */
111 unsigned int reg; /* register address or index */
112 int alloc; /* extra register already allocated */
113 int idx; /* index in shared_regs->regs[] */
117 * struct hw_perf_event - performance event hardware details:
119 struct hw_perf_event {
120 #ifdef CONFIG_PERF_EVENTS
122 struct { /* hardware */
125 unsigned long config_base;
126 unsigned long event_base;
127 int event_base_rdpmc;
132 struct hw_perf_event_extra extra_reg;
133 struct hw_perf_event_extra branch_reg;
135 struct { /* software */
136 struct hrtimer hrtimer;
138 struct { /* tracepoint */
139 /* for tp_event->class */
140 struct list_head tp_list;
142 struct { /* intel_cqm */
146 struct list_head cqm_events_entry;
147 struct list_head cqm_groups_entry;
148 struct list_head cqm_group_entry;
150 struct { /* itrace */
153 struct { /* amd_power */
157 #ifdef CONFIG_HAVE_HW_BREAKPOINT
158 struct { /* breakpoint */
160 * Crufty hack to avoid the chicken and egg
161 * problem hw_breakpoint has with context
162 * creation and event initalization.
164 struct arch_hw_breakpoint info;
165 struct list_head bp_list;
170 * If the event is a per task event, this will point to the task in
171 * question. See the comment in perf_event_alloc().
173 struct task_struct *target;
176 * PMU would store hardware filter configuration
181 /* Last sync'ed generation of filters */
182 unsigned long addr_filters_gen;
185 * hw_perf_event::state flags; used to track the PERF_EF_* state.
187 #define PERF_HES_STOPPED 0x01 /* the counter is stopped */
188 #define PERF_HES_UPTODATE 0x02 /* event->count up-to-date */
189 #define PERF_HES_ARCH 0x04
194 * The last observed hardware counter value, updated with a
195 * local64_cmpxchg() such that pmu::read() can be called nested.
197 local64_t prev_count;
200 * The period to start the next sample with.
205 * The period we started this sample with.
210 * However much is left of the current period; note that this is
211 * a full 64bit value and allows for generation of periods longer
212 * than hardware might allow.
214 local64_t period_left;
217 * State for throttling the event, see __perf_event_overflow() and
218 * perf_adjust_freq_unthr_context().
224 * State for freq target events, see __perf_event_overflow() and
225 * perf_adjust_freq_unthr_context().
228 u64 freq_count_stamp;
235 * Common implementation detail of pmu::{start,commit,cancel}_txn
237 #define PERF_PMU_TXN_ADD 0x1 /* txn to add/schedule event on PMU */
238 #define PERF_PMU_TXN_READ 0x2 /* txn to read event group from PMU */
241 * pmu::capabilities flags
243 #define PERF_PMU_CAP_NO_INTERRUPT 0x01
244 #define PERF_PMU_CAP_NO_NMI 0x02
245 #define PERF_PMU_CAP_AUX_NO_SG 0x04
246 #define PERF_PMU_CAP_AUX_SW_DOUBLEBUF 0x08
247 #define PERF_PMU_CAP_EXCLUSIVE 0x10
248 #define PERF_PMU_CAP_ITRACE 0x20
249 #define PERF_PMU_CAP_HETEROGENEOUS_CPUS 0x40
252 * struct pmu - generic performance monitoring unit
255 struct list_head entry;
257 struct module *module;
259 const struct attribute_group **attr_groups;
264 * various common per-pmu feature flags
268 int * __percpu pmu_disable_count;
269 struct perf_cpu_context * __percpu pmu_cpu_context;
270 atomic_t exclusive_cnt; /* < 0: cpu; > 0: tsk */
272 int hrtimer_interval_ms;
274 /* number of address filters this PMU can do */
275 unsigned int nr_addr_filters;
278 * Fully disable/enable this PMU, can be used to protect from the PMI
279 * as well as for lazy/batch writing of the MSRs.
281 void (*pmu_enable) (struct pmu *pmu); /* optional */
282 void (*pmu_disable) (struct pmu *pmu); /* optional */
285 * Try and initialize the event for this PMU.
288 * -ENOENT -- @event is not for this PMU
290 * -ENODEV -- @event is for this PMU but PMU not present
291 * -EBUSY -- @event is for this PMU but PMU temporarily unavailable
292 * -EINVAL -- @event is for this PMU but @event is not valid
293 * -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
294 * -EACCESS -- @event is for this PMU, @event is valid, but no privilidges
296 * 0 -- @event is for this PMU and valid
298 * Other error return values are allowed.
300 int (*event_init) (struct perf_event *event);
303 * Notification that the event was mapped or unmapped. Called
304 * in the context of the mapping task.
306 void (*event_mapped) (struct perf_event *event); /*optional*/
307 void (*event_unmapped) (struct perf_event *event); /*optional*/
310 * Flags for ->add()/->del()/ ->start()/->stop(). There are
311 * matching hw_perf_event::state flags.
313 #define PERF_EF_START 0x01 /* start the counter when adding */
314 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
315 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
318 * Adds/Removes a counter to/from the PMU, can be done inside a
319 * transaction, see the ->*_txn() methods.
321 * The add/del callbacks will reserve all hardware resources required
322 * to service the event, this includes any counter constraint
325 * Called with IRQs disabled and the PMU disabled on the CPU the event
328 * ->add() called without PERF_EF_START should result in the same state
329 * as ->add() followed by ->stop().
331 * ->del() must always PERF_EF_UPDATE stop an event. If it calls
332 * ->stop() that must deal with already being stopped without
335 int (*add) (struct perf_event *event, int flags);
336 void (*del) (struct perf_event *event, int flags);
339 * Starts/Stops a counter present on the PMU.
341 * The PMI handler should stop the counter when perf_event_overflow()
342 * returns !0. ->start() will be used to continue.
344 * Also used to change the sample period.
346 * Called with IRQs disabled and the PMU disabled on the CPU the event
347 * is on -- will be called from NMI context with the PMU generates
350 * ->stop() with PERF_EF_UPDATE will read the counter and update
351 * period/count values like ->read() would.
353 * ->start() with PERF_EF_RELOAD will reprogram the the counter
354 * value, must be preceded by a ->stop() with PERF_EF_UPDATE.
356 void (*start) (struct perf_event *event, int flags);
357 void (*stop) (struct perf_event *event, int flags);
360 * Updates the counter value of the event.
362 * For sampling capable PMUs this will also update the software period
363 * hw_perf_event::period_left field.
365 void (*read) (struct perf_event *event);
368 * Group events scheduling is treated as a transaction, add
369 * group events as a whole and perform one schedulability test.
370 * If the test fails, roll back the whole group
372 * Start the transaction, after this ->add() doesn't need to
373 * do schedulability tests.
377 void (*start_txn) (struct pmu *pmu, unsigned int txn_flags);
379 * If ->start_txn() disabled the ->add() schedulability test
380 * then ->commit_txn() is required to perform one. On success
381 * the transaction is closed. On error the transaction is kept
382 * open until ->cancel_txn() is called.
386 int (*commit_txn) (struct pmu *pmu);
388 * Will cancel the transaction, assumes ->del() is called
389 * for each successful ->add() during the transaction.
393 void (*cancel_txn) (struct pmu *pmu);
396 * Will return the value for perf_event_mmap_page::index for this event,
397 * if no implementation is provided it will default to: event->hw.idx + 1.
399 int (*event_idx) (struct perf_event *event); /*optional */
402 * context-switches callback
404 void (*sched_task) (struct perf_event_context *ctx,
407 * PMU specific data size
409 size_t task_ctx_size;
413 * Return the count value for a counter.
415 u64 (*count) (struct perf_event *event); /*optional*/
418 * Set up pmu-private data structures for an AUX area
420 void *(*setup_aux) (int cpu, void **pages,
421 int nr_pages, bool overwrite);
425 * Free pmu-private AUX data structures
427 void (*free_aux) (void *aux); /* optional */
430 * Validate address range filters: make sure the HW supports the
431 * requested configuration and number of filters; return 0 if the
432 * supplied filters are valid, -errno otherwise.
434 * Runs in the context of the ioctl()ing process and is not serialized
435 * with the rest of the PMU callbacks.
437 int (*addr_filters_validate) (struct list_head *filters);
441 * Synchronize address range filter configuration:
442 * translate hw-agnostic filters into hardware configuration in
443 * event::hw::addr_filters.
445 * Runs as a part of filter sync sequence that is done in ->start()
446 * callback by calling perf_event_addr_filters_sync().
448 * May (and should) traverse event::addr_filters::list, for which its
449 * caller provides necessary serialization.
451 void (*addr_filters_sync) (struct perf_event *event);
455 * Filter events for PMU-specific reasons.
457 int (*filter_match) (struct perf_event *event); /* optional */
461 * struct perf_addr_filter - address range filter definition
462 * @entry: event's filter list linkage
463 * @inode: object file's inode for file-based filters
464 * @offset: filter range offset
465 * @size: filter range size
466 * @range: 1: range, 0: address
467 * @filter: 1: filter/start, 0: stop
469 * This is a hardware-agnostic filter configuration as specified by the user.
471 struct perf_addr_filter {
472 struct list_head entry;
474 unsigned long offset;
476 unsigned int range : 1,
481 * struct perf_addr_filters_head - container for address range filters
482 * @list: list of filters for this event
483 * @lock: spinlock that serializes accesses to the @list and event's
484 * (and its children's) filter generations.
486 * A child event will use parent's @list (and therefore @lock), so they are
487 * bundled together; see perf_event_addr_filters().
489 struct perf_addr_filters_head {
490 struct list_head list;
495 * enum perf_event_active_state - the states of a event
497 enum perf_event_active_state {
498 PERF_EVENT_STATE_DEAD = -4,
499 PERF_EVENT_STATE_EXIT = -3,
500 PERF_EVENT_STATE_ERROR = -2,
501 PERF_EVENT_STATE_OFF = -1,
502 PERF_EVENT_STATE_INACTIVE = 0,
503 PERF_EVENT_STATE_ACTIVE = 1,
507 struct perf_sample_data;
509 typedef void (*perf_overflow_handler_t)(struct perf_event *,
510 struct perf_sample_data *,
511 struct pt_regs *regs);
513 enum perf_group_flag {
514 PERF_GROUP_SOFTWARE = 0x1,
517 #define SWEVENT_HLIST_BITS 8
518 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
520 struct swevent_hlist {
521 struct hlist_head heads[SWEVENT_HLIST_SIZE];
522 struct rcu_head rcu_head;
525 #define PERF_ATTACH_CONTEXT 0x01
526 #define PERF_ATTACH_GROUP 0x02
527 #define PERF_ATTACH_TASK 0x04
528 #define PERF_ATTACH_TASK_DATA 0x08
533 struct pmu_event_list {
535 struct list_head list;
539 * struct perf_event - performance event kernel representation:
542 #ifdef CONFIG_PERF_EVENTS
544 * entry onto perf_event_context::event_list;
545 * modifications require ctx->lock
546 * RCU safe iterations.
548 struct list_head event_entry;
551 * XXX: group_entry and sibling_list should be mutually exclusive;
552 * either you're a sibling on a group, or you're the group leader.
553 * Rework the code to always use the same list element.
555 * Locked for modification by both ctx->mutex and ctx->lock; holding
556 * either sufficies for read.
558 struct list_head group_entry;
559 struct list_head sibling_list;
562 * We need storage to track the entries in perf_pmu_migrate_context; we
563 * cannot use the event_entry because of RCU and we want to keep the
564 * group in tact which avoids us using the other two entries.
566 struct list_head migrate_entry;
568 struct hlist_node hlist_entry;
569 struct list_head active_entry;
572 struct perf_event *group_leader;
576 enum perf_event_active_state state;
577 unsigned int attach_state;
579 atomic64_t child_count;
582 * These are the total time in nanoseconds that the event
583 * has been enabled (i.e. eligible to run, and the task has
584 * been scheduled in, if this is a per-task event)
585 * and running (scheduled onto the CPU), respectively.
587 * They are computed from tstamp_enabled, tstamp_running and
588 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
590 u64 total_time_enabled;
591 u64 total_time_running;
594 * These are timestamps used for computing total_time_enabled
595 * and total_time_running when the event is in INACTIVE or
596 * ACTIVE state, measured in nanoseconds from an arbitrary point
598 * tstamp_enabled: the notional time when the event was enabled
599 * tstamp_running: the notional time when the event was scheduled on
600 * tstamp_stopped: in INACTIVE state, the notional time when the
601 * event was scheduled off.
608 * timestamp shadows the actual context timing but it can
609 * be safely used in NMI interrupt context. It reflects the
610 * context time as it was when the event was last scheduled in.
612 * ctx_time already accounts for ctx->timestamp. Therefore to
613 * compute ctx_time for a sample, simply add perf_clock().
617 struct perf_event_attr attr;
621 struct hw_perf_event hw;
623 struct perf_event_context *ctx;
624 atomic_long_t refcount;
627 * These accumulate total time (in nanoseconds) that children
628 * events have been enabled and running, respectively.
630 atomic64_t child_total_time_enabled;
631 atomic64_t child_total_time_running;
634 * Protect attach/detach and child_list:
636 struct mutex child_mutex;
637 struct list_head child_list;
638 struct perf_event *parent;
643 struct list_head owner_entry;
644 struct task_struct *owner;
647 struct mutex mmap_mutex;
650 struct ring_buffer *rb;
651 struct list_head rb_entry;
652 unsigned long rcu_batches;
656 wait_queue_head_t waitq;
657 struct fasync_struct *fasync;
659 /* delayed work for NMIs and such */
663 struct irq_work pending;
665 atomic_t event_limit;
667 /* address range filters */
668 struct perf_addr_filters_head addr_filters;
669 /* vma address array for file-based filders */
670 unsigned long *addr_filters_offs;
671 unsigned long addr_filters_gen;
673 void (*destroy)(struct perf_event *);
674 struct rcu_head rcu_head;
676 struct pid_namespace *ns;
680 perf_overflow_handler_t overflow_handler;
681 void *overflow_handler_context;
683 #ifdef CONFIG_EVENT_TRACING
684 struct trace_event_call *tp_event;
685 struct event_filter *filter;
686 #ifdef CONFIG_FUNCTION_TRACER
687 struct ftrace_ops ftrace_ops;
691 #ifdef CONFIG_CGROUP_PERF
692 struct perf_cgroup *cgrp; /* cgroup event is attach to */
693 int cgrp_defer_enabled;
696 struct list_head sb_list;
697 #endif /* CONFIG_PERF_EVENTS */
701 * struct perf_event_context - event context structure
703 * Used as a container for task events and CPU events as well:
705 struct perf_event_context {
708 * Protect the states of the events in the list,
709 * nr_active, and the list:
713 * Protect the list of events. Locking either mutex or lock
714 * is sufficient to ensure the list doesn't change; to change
715 * the list you need to lock both the mutex and the spinlock.
719 struct list_head active_ctx_list;
720 struct list_head pinned_groups;
721 struct list_head flexible_groups;
722 struct list_head event_list;
730 struct task_struct *task;
733 * Context clock, runs when context enabled.
739 * These fields let us detect when two contexts have both
740 * been cloned (inherited) from a common ancestor.
742 struct perf_event_context *parent_ctx;
746 #ifdef CONFIG_CGROUP_PERF
747 int nr_cgroups; /* cgroup evts */
749 void *task_ctx_data; /* pmu specific data */
750 struct rcu_head rcu_head;
754 * Number of contexts where an event can trigger:
755 * task, softirq, hardirq, nmi.
757 #define PERF_NR_CONTEXTS 4
760 * struct perf_event_cpu_context - per cpu event context structure
762 struct perf_cpu_context {
763 struct perf_event_context ctx;
764 struct perf_event_context *task_ctx;
768 raw_spinlock_t hrtimer_lock;
769 struct hrtimer hrtimer;
770 ktime_t hrtimer_interval;
771 unsigned int hrtimer_active;
773 struct pmu *unique_pmu;
774 #ifdef CONFIG_CGROUP_PERF
775 struct perf_cgroup *cgrp;
779 struct perf_output_handle {
780 struct perf_event *event;
781 struct ring_buffer *rb;
782 unsigned long wakeup;
791 #ifdef CONFIG_CGROUP_PERF
794 * perf_cgroup_info keeps track of time_enabled for a cgroup.
795 * This is a per-cpu dynamically allocated data structure.
797 struct perf_cgroup_info {
803 struct cgroup_subsys_state css;
804 struct perf_cgroup_info __percpu *info;
808 * Must ensure cgroup is pinned (css_get) before calling
809 * this function. In other words, we cannot call this function
810 * if there is no cgroup event for the current CPU context.
812 static inline struct perf_cgroup *
813 perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
815 return container_of(task_css_check(task, perf_event_cgrp_id,
816 ctx ? lockdep_is_held(&ctx->lock)
818 struct perf_cgroup, css);
820 #endif /* CONFIG_CGROUP_PERF */
822 #ifdef CONFIG_PERF_EVENTS
824 extern void *perf_aux_output_begin(struct perf_output_handle *handle,
825 struct perf_event *event);
826 extern void perf_aux_output_end(struct perf_output_handle *handle,
827 unsigned long size, bool truncated);
828 extern int perf_aux_output_skip(struct perf_output_handle *handle,
830 extern void *perf_get_aux(struct perf_output_handle *handle);
832 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
833 extern void perf_pmu_unregister(struct pmu *pmu);
835 extern int perf_num_counters(void);
836 extern const char *perf_pmu_name(void);
837 extern void __perf_event_task_sched_in(struct task_struct *prev,
838 struct task_struct *task);
839 extern void __perf_event_task_sched_out(struct task_struct *prev,
840 struct task_struct *next);
841 extern int perf_event_init_task(struct task_struct *child);
842 extern void perf_event_exit_task(struct task_struct *child);
843 extern void perf_event_free_task(struct task_struct *task);
844 extern void perf_event_delayed_put(struct task_struct *task);
845 extern struct file *perf_event_get(unsigned int fd);
846 extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
847 extern void perf_event_print_debug(void);
848 extern void perf_pmu_disable(struct pmu *pmu);
849 extern void perf_pmu_enable(struct pmu *pmu);
850 extern void perf_sched_cb_dec(struct pmu *pmu);
851 extern void perf_sched_cb_inc(struct pmu *pmu);
852 extern int perf_event_task_disable(void);
853 extern int perf_event_task_enable(void);
854 extern int perf_event_refresh(struct perf_event *event, int refresh);
855 extern void perf_event_update_userpage(struct perf_event *event);
856 extern int perf_event_release_kernel(struct perf_event *event);
857 extern struct perf_event *
858 perf_event_create_kernel_counter(struct perf_event_attr *attr,
860 struct task_struct *task,
861 perf_overflow_handler_t callback,
863 extern void perf_pmu_migrate_context(struct pmu *pmu,
864 int src_cpu, int dst_cpu);
865 extern u64 perf_event_read_local(struct perf_event *event);
866 extern u64 perf_event_read_value(struct perf_event *event,
867 u64 *enabled, u64 *running);
870 struct perf_sample_data {
872 * Fields set by perf_sample_data_init(), group so as to
873 * minimize the cachelines touched.
876 struct perf_raw_record *raw;
877 struct perf_branch_stack *br_stack;
881 union perf_mem_data_src data_src;
884 * The other fields, optionally {set,used} by
885 * perf_{prepare,output}_sample().
900 struct perf_callchain_entry *callchain;
903 * regs_user may point to task_pt_regs or to regs_user_copy, depending
906 struct perf_regs regs_user;
907 struct pt_regs regs_user_copy;
909 struct perf_regs regs_intr;
911 } ____cacheline_aligned;
913 /* default value for data source */
914 #define PERF_MEM_NA (PERF_MEM_S(OP, NA) |\
915 PERF_MEM_S(LVL, NA) |\
916 PERF_MEM_S(SNOOP, NA) |\
917 PERF_MEM_S(LOCK, NA) |\
920 static inline void perf_sample_data_init(struct perf_sample_data *data,
921 u64 addr, u64 period)
923 /* remaining struct members initialized in perf_prepare_sample() */
926 data->br_stack = NULL;
927 data->period = period;
929 data->data_src.val = PERF_MEM_NA;
933 extern void perf_output_sample(struct perf_output_handle *handle,
934 struct perf_event_header *header,
935 struct perf_sample_data *data,
936 struct perf_event *event);
937 extern void perf_prepare_sample(struct perf_event_header *header,
938 struct perf_sample_data *data,
939 struct perf_event *event,
940 struct pt_regs *regs);
942 extern int perf_event_overflow(struct perf_event *event,
943 struct perf_sample_data *data,
944 struct pt_regs *regs);
946 extern void perf_event_output_forward(struct perf_event *event,
947 struct perf_sample_data *data,
948 struct pt_regs *regs);
949 extern void perf_event_output_backward(struct perf_event *event,
950 struct perf_sample_data *data,
951 struct pt_regs *regs);
952 extern void perf_event_output(struct perf_event *event,
953 struct perf_sample_data *data,
954 struct pt_regs *regs);
957 is_default_overflow_handler(struct perf_event *event)
959 if (likely(event->overflow_handler == perf_event_output_forward))
961 if (unlikely(event->overflow_handler == perf_event_output_backward))
967 perf_event_header__init_id(struct perf_event_header *header,
968 struct perf_sample_data *data,
969 struct perf_event *event);
971 perf_event__output_id_sample(struct perf_event *event,
972 struct perf_output_handle *handle,
973 struct perf_sample_data *sample);
976 perf_log_lost_samples(struct perf_event *event, u64 lost);
978 static inline bool is_sampling_event(struct perf_event *event)
980 return event->attr.sample_period != 0;
984 * Return 1 for a software event, 0 for a hardware event
986 static inline int is_software_event(struct perf_event *event)
988 return event->pmu->task_ctx_nr == perf_sw_context;
991 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
993 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
994 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
996 #ifndef perf_arch_fetch_caller_regs
997 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
1001 * Take a snapshot of the regs. Skip ip and frame pointer to
1002 * the nth caller. We only need a few of the regs:
1003 * - ip for PERF_SAMPLE_IP
1004 * - cs for user_mode() tests
1005 * - bp for callchains
1006 * - eflags, for future purposes, just in case
1008 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1010 perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1013 static __always_inline void
1014 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1016 if (static_key_false(&perf_swevent_enabled[event_id]))
1017 __perf_sw_event(event_id, nr, regs, addr);
1020 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
1023 * 'Special' version for the scheduler, it hard assumes no recursion,
1024 * which is guaranteed by us not actually scheduling inside other swevents
1025 * because those disable preemption.
1027 static __always_inline void
1028 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
1030 if (static_key_false(&perf_swevent_enabled[event_id])) {
1031 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1033 perf_fetch_caller_regs(regs);
1034 ___perf_sw_event(event_id, nr, regs, addr);
1038 extern struct static_key_false perf_sched_events;
1040 static __always_inline bool
1041 perf_sw_migrate_enabled(void)
1043 if (static_key_false(&perf_swevent_enabled[PERF_COUNT_SW_CPU_MIGRATIONS]))
1048 static inline void perf_event_task_migrate(struct task_struct *task)
1050 if (perf_sw_migrate_enabled())
1051 task->sched_migrated = 1;
1054 static inline void perf_event_task_sched_in(struct task_struct *prev,
1055 struct task_struct *task)
1057 if (static_branch_unlikely(&perf_sched_events))
1058 __perf_event_task_sched_in(prev, task);
1060 if (perf_sw_migrate_enabled() && task->sched_migrated) {
1061 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1063 perf_fetch_caller_regs(regs);
1064 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, regs, 0);
1065 task->sched_migrated = 0;
1069 static inline void perf_event_task_sched_out(struct task_struct *prev,
1070 struct task_struct *next)
1072 perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
1074 if (static_branch_unlikely(&perf_sched_events))
1075 __perf_event_task_sched_out(prev, next);
1078 static inline u64 __perf_event_count(struct perf_event *event)
1080 return local64_read(&event->count) + atomic64_read(&event->child_count);
1083 extern void perf_event_mmap(struct vm_area_struct *vma);
1084 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1085 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1086 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1088 extern void perf_event_exec(void);
1089 extern void perf_event_comm(struct task_struct *tsk, bool exec);
1090 extern void perf_event_fork(struct task_struct *tsk);
1093 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1095 extern void perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1096 extern void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1097 extern struct perf_callchain_entry *
1098 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
1099 u32 max_stack, bool crosstask, bool add_mark);
1100 extern int get_callchain_buffers(int max_stack);
1101 extern void put_callchain_buffers(void);
1103 extern int sysctl_perf_event_max_stack;
1104 extern int sysctl_perf_event_max_contexts_per_stack;
1106 static inline int perf_callchain_store_context(struct perf_callchain_entry_ctx *ctx, u64 ip)
1108 if (ctx->contexts < sysctl_perf_event_max_contexts_per_stack) {
1109 struct perf_callchain_entry *entry = ctx->entry;
1110 entry->ip[entry->nr++] = ip;
1114 ctx->contexts_maxed = true;
1115 return -1; /* no more room, stop walking the stack */
1119 static inline int perf_callchain_store(struct perf_callchain_entry_ctx *ctx, u64 ip)
1121 if (ctx->nr < ctx->max_stack && !ctx->contexts_maxed) {
1122 struct perf_callchain_entry *entry = ctx->entry;
1123 entry->ip[entry->nr++] = ip;
1127 return -1; /* no more room, stop walking the stack */
1131 extern int sysctl_perf_event_paranoid;
1132 extern int sysctl_perf_event_mlock;
1133 extern int sysctl_perf_event_sample_rate;
1134 extern int sysctl_perf_cpu_time_max_percent;
1136 extern void perf_sample_event_took(u64 sample_len_ns);
1138 extern int perf_proc_update_handler(struct ctl_table *table, int write,
1139 void __user *buffer, size_t *lenp,
1141 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
1142 void __user *buffer, size_t *lenp,
1145 int perf_event_max_stack_handler(struct ctl_table *table, int write,
1146 void __user *buffer, size_t *lenp, loff_t *ppos);
1148 static inline bool perf_paranoid_tracepoint_raw(void)
1150 return sysctl_perf_event_paranoid > -1;
1153 static inline bool perf_paranoid_cpu(void)
1155 return sysctl_perf_event_paranoid > 0;
1158 static inline bool perf_paranoid_kernel(void)
1160 return sysctl_perf_event_paranoid > 1;
1163 extern void perf_event_init(void);
1164 extern void perf_tp_event(u16 event_type, u64 count, void *record,
1165 int entry_size, struct pt_regs *regs,
1166 struct hlist_head *head, int rctx,
1167 struct task_struct *task);
1168 extern void perf_bp_event(struct perf_event *event, void *data);
1170 #ifndef perf_misc_flags
1171 # define perf_misc_flags(regs) \
1172 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1173 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1176 static inline bool has_branch_stack(struct perf_event *event)
1178 return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1181 static inline bool needs_branch_stack(struct perf_event *event)
1183 return event->attr.branch_sample_type != 0;
1186 static inline bool has_aux(struct perf_event *event)
1188 return event->pmu->setup_aux;
1191 static inline bool is_write_backward(struct perf_event *event)
1193 return !!event->attr.write_backward;
1196 static inline bool has_addr_filter(struct perf_event *event)
1198 return event->pmu->nr_addr_filters;
1202 * An inherited event uses parent's filters
1204 static inline struct perf_addr_filters_head *
1205 perf_event_addr_filters(struct perf_event *event)
1207 struct perf_addr_filters_head *ifh = &event->addr_filters;
1210 ifh = &event->parent->addr_filters;
1215 extern void perf_event_addr_filters_sync(struct perf_event *event);
1217 extern int perf_output_begin(struct perf_output_handle *handle,
1218 struct perf_event *event, unsigned int size);
1219 extern int perf_output_begin_forward(struct perf_output_handle *handle,
1220 struct perf_event *event,
1222 extern int perf_output_begin_backward(struct perf_output_handle *handle,
1223 struct perf_event *event,
1226 extern void perf_output_end(struct perf_output_handle *handle);
1227 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
1228 const void *buf, unsigned int len);
1229 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
1231 extern int perf_swevent_get_recursion_context(void);
1232 extern void perf_swevent_put_recursion_context(int rctx);
1233 extern u64 perf_swevent_set_period(struct perf_event *event);
1234 extern void perf_event_enable(struct perf_event *event);
1235 extern void perf_event_disable(struct perf_event *event);
1236 extern void perf_event_disable_local(struct perf_event *event);
1237 extern void perf_event_task_tick(void);
1238 #else /* !CONFIG_PERF_EVENTS: */
1239 static inline void *
1240 perf_aux_output_begin(struct perf_output_handle *handle,
1241 struct perf_event *event) { return NULL; }
1243 perf_aux_output_end(struct perf_output_handle *handle, unsigned long size,
1246 perf_aux_output_skip(struct perf_output_handle *handle,
1247 unsigned long size) { return -EINVAL; }
1248 static inline void *
1249 perf_get_aux(struct perf_output_handle *handle) { return NULL; }
1251 perf_event_task_migrate(struct task_struct *task) { }
1253 perf_event_task_sched_in(struct task_struct *prev,
1254 struct task_struct *task) { }
1256 perf_event_task_sched_out(struct task_struct *prev,
1257 struct task_struct *next) { }
1258 static inline int perf_event_init_task(struct task_struct *child) { return 0; }
1259 static inline void perf_event_exit_task(struct task_struct *child) { }
1260 static inline void perf_event_free_task(struct task_struct *task) { }
1261 static inline void perf_event_delayed_put(struct task_struct *task) { }
1262 static inline struct file *perf_event_get(unsigned int fd) { return ERR_PTR(-EINVAL); }
1263 static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
1265 return ERR_PTR(-EINVAL);
1267 static inline u64 perf_event_read_local(struct perf_event *event) { return -EINVAL; }
1268 static inline void perf_event_print_debug(void) { }
1269 static inline int perf_event_task_disable(void) { return -EINVAL; }
1270 static inline int perf_event_task_enable(void) { return -EINVAL; }
1271 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1277 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) { }
1279 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr) { }
1281 perf_bp_event(struct perf_event *event, void *data) { }
1283 static inline int perf_register_guest_info_callbacks
1284 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1285 static inline int perf_unregister_guest_info_callbacks
1286 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1288 static inline void perf_event_mmap(struct vm_area_struct *vma) { }
1289 static inline void perf_event_exec(void) { }
1290 static inline void perf_event_comm(struct task_struct *tsk, bool exec) { }
1291 static inline void perf_event_fork(struct task_struct *tsk) { }
1292 static inline void perf_event_init(void) { }
1293 static inline int perf_swevent_get_recursion_context(void) { return -1; }
1294 static inline void perf_swevent_put_recursion_context(int rctx) { }
1295 static inline u64 perf_swevent_set_period(struct perf_event *event) { return 0; }
1296 static inline void perf_event_enable(struct perf_event *event) { }
1297 static inline void perf_event_disable(struct perf_event *event) { }
1298 static inline int __perf_event_disable(void *info) { return -1; }
1299 static inline void perf_event_task_tick(void) { }
1300 static inline int perf_event_release_kernel(struct perf_event *event) { return 0; }
1303 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1304 extern void perf_restore_debug_store(void);
1306 static inline void perf_restore_debug_store(void) { }
1309 static __always_inline bool perf_raw_frag_last(const struct perf_raw_frag *frag)
1311 return frag->pad < sizeof(u64);
1314 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1316 struct perf_pmu_events_attr {
1317 struct device_attribute attr;
1319 const char *event_str;
1322 struct perf_pmu_events_ht_attr {
1323 struct device_attribute attr;
1325 const char *event_str_ht;
1326 const char *event_str_noht;
1329 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1332 #define PMU_EVENT_ATTR(_name, _var, _id, _show) \
1333 static struct perf_pmu_events_attr _var = { \
1334 .attr = __ATTR(_name, 0444, _show, NULL), \
1338 #define PMU_EVENT_ATTR_STRING(_name, _var, _str) \
1339 static struct perf_pmu_events_attr _var = { \
1340 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1342 .event_str = _str, \
1345 #define PMU_FORMAT_ATTR(_name, _format) \
1347 _name##_show(struct device *dev, \
1348 struct device_attribute *attr, \
1351 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
1352 return sprintf(page, _format "\n"); \
1355 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1357 /* Performance counter hotplug functions */
1358 #ifdef CONFIG_PERF_EVENTS
1359 int perf_event_init_cpu(unsigned int cpu);
1360 int perf_event_exit_cpu(unsigned int cpu);
1362 #define perf_event_init_cpu NULL
1363 #define perf_event_exit_cpu NULL
1366 #endif /* _LINUX_PERF_EVENT_H */