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
534 * struct perf_event - performance event kernel representation:
537 #ifdef CONFIG_PERF_EVENTS
539 * entry onto perf_event_context::event_list;
540 * modifications require ctx->lock
541 * RCU safe iterations.
543 struct list_head event_entry;
546 * XXX: group_entry and sibling_list should be mutually exclusive;
547 * either you're a sibling on a group, or you're the group leader.
548 * Rework the code to always use the same list element.
550 * Locked for modification by both ctx->mutex and ctx->lock; holding
551 * either sufficies for read.
553 struct list_head group_entry;
554 struct list_head sibling_list;
557 * We need storage to track the entries in perf_pmu_migrate_context; we
558 * cannot use the event_entry because of RCU and we want to keep the
559 * group in tact which avoids us using the other two entries.
561 struct list_head migrate_entry;
563 struct hlist_node hlist_entry;
564 struct list_head active_entry;
567 struct perf_event *group_leader;
571 enum perf_event_active_state state;
572 unsigned int attach_state;
574 atomic64_t child_count;
577 * These are the total time in nanoseconds that the event
578 * has been enabled (i.e. eligible to run, and the task has
579 * been scheduled in, if this is a per-task event)
580 * and running (scheduled onto the CPU), respectively.
582 * They are computed from tstamp_enabled, tstamp_running and
583 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
585 u64 total_time_enabled;
586 u64 total_time_running;
589 * These are timestamps used for computing total_time_enabled
590 * and total_time_running when the event is in INACTIVE or
591 * ACTIVE state, measured in nanoseconds from an arbitrary point
593 * tstamp_enabled: the notional time when the event was enabled
594 * tstamp_running: the notional time when the event was scheduled on
595 * tstamp_stopped: in INACTIVE state, the notional time when the
596 * event was scheduled off.
603 * timestamp shadows the actual context timing but it can
604 * be safely used in NMI interrupt context. It reflects the
605 * context time as it was when the event was last scheduled in.
607 * ctx_time already accounts for ctx->timestamp. Therefore to
608 * compute ctx_time for a sample, simply add perf_clock().
612 struct perf_event_attr attr;
616 struct hw_perf_event hw;
618 struct perf_event_context *ctx;
619 atomic_long_t refcount;
622 * These accumulate total time (in nanoseconds) that children
623 * events have been enabled and running, respectively.
625 atomic64_t child_total_time_enabled;
626 atomic64_t child_total_time_running;
629 * Protect attach/detach and child_list:
631 struct mutex child_mutex;
632 struct list_head child_list;
633 struct perf_event *parent;
638 struct list_head owner_entry;
639 struct task_struct *owner;
642 struct mutex mmap_mutex;
645 struct ring_buffer *rb;
646 struct list_head rb_entry;
647 unsigned long rcu_batches;
651 wait_queue_head_t waitq;
652 struct fasync_struct *fasync;
654 /* delayed work for NMIs and such */
658 struct irq_work pending;
660 atomic_t event_limit;
662 /* address range filters */
663 struct perf_addr_filters_head addr_filters;
664 /* vma address array for file-based filders */
665 unsigned long *addr_filters_offs;
666 unsigned long addr_filters_gen;
668 void (*destroy)(struct perf_event *);
669 struct rcu_head rcu_head;
671 struct pid_namespace *ns;
675 perf_overflow_handler_t overflow_handler;
676 void *overflow_handler_context;
678 #ifdef CONFIG_EVENT_TRACING
679 struct trace_event_call *tp_event;
680 struct event_filter *filter;
681 #ifdef CONFIG_FUNCTION_TRACER
682 struct ftrace_ops ftrace_ops;
686 #ifdef CONFIG_CGROUP_PERF
687 struct perf_cgroup *cgrp; /* cgroup event is attach to */
688 int cgrp_defer_enabled;
691 #endif /* CONFIG_PERF_EVENTS */
695 * struct perf_event_context - event context structure
697 * Used as a container for task events and CPU events as well:
699 struct perf_event_context {
702 * Protect the states of the events in the list,
703 * nr_active, and the list:
707 * Protect the list of events. Locking either mutex or lock
708 * is sufficient to ensure the list doesn't change; to change
709 * the list you need to lock both the mutex and the spinlock.
713 struct list_head active_ctx_list;
714 struct list_head pinned_groups;
715 struct list_head flexible_groups;
716 struct list_head event_list;
724 struct task_struct *task;
727 * Context clock, runs when context enabled.
733 * These fields let us detect when two contexts have both
734 * been cloned (inherited) from a common ancestor.
736 struct perf_event_context *parent_ctx;
740 int nr_cgroups; /* cgroup evts */
741 void *task_ctx_data; /* pmu specific data */
742 struct rcu_head rcu_head;
746 * Number of contexts where an event can trigger:
747 * task, softirq, hardirq, nmi.
749 #define PERF_NR_CONTEXTS 4
752 * struct perf_event_cpu_context - per cpu event context structure
754 struct perf_cpu_context {
755 struct perf_event_context ctx;
756 struct perf_event_context *task_ctx;
760 raw_spinlock_t hrtimer_lock;
761 struct hrtimer hrtimer;
762 ktime_t hrtimer_interval;
763 unsigned int hrtimer_active;
765 struct pmu *unique_pmu;
766 struct perf_cgroup *cgrp;
769 struct perf_output_handle {
770 struct perf_event *event;
771 struct ring_buffer *rb;
772 unsigned long wakeup;
781 #ifdef CONFIG_CGROUP_PERF
784 * perf_cgroup_info keeps track of time_enabled for a cgroup.
785 * This is a per-cpu dynamically allocated data structure.
787 struct perf_cgroup_info {
793 struct cgroup_subsys_state css;
794 struct perf_cgroup_info __percpu *info;
798 * Must ensure cgroup is pinned (css_get) before calling
799 * this function. In other words, we cannot call this function
800 * if there is no cgroup event for the current CPU context.
802 static inline struct perf_cgroup *
803 perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
805 return container_of(task_css_check(task, perf_event_cgrp_id,
806 ctx ? lockdep_is_held(&ctx->lock)
808 struct perf_cgroup, css);
810 #endif /* CONFIG_CGROUP_PERF */
812 #ifdef CONFIG_PERF_EVENTS
814 extern void *perf_aux_output_begin(struct perf_output_handle *handle,
815 struct perf_event *event);
816 extern void perf_aux_output_end(struct perf_output_handle *handle,
817 unsigned long size, bool truncated);
818 extern int perf_aux_output_skip(struct perf_output_handle *handle,
820 extern void *perf_get_aux(struct perf_output_handle *handle);
822 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
823 extern void perf_pmu_unregister(struct pmu *pmu);
825 extern int perf_num_counters(void);
826 extern const char *perf_pmu_name(void);
827 extern void __perf_event_task_sched_in(struct task_struct *prev,
828 struct task_struct *task);
829 extern void __perf_event_task_sched_out(struct task_struct *prev,
830 struct task_struct *next);
831 extern int perf_event_init_task(struct task_struct *child);
832 extern void perf_event_exit_task(struct task_struct *child);
833 extern void perf_event_free_task(struct task_struct *task);
834 extern void perf_event_delayed_put(struct task_struct *task);
835 extern struct file *perf_event_get(unsigned int fd);
836 extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
837 extern void perf_event_print_debug(void);
838 extern void perf_pmu_disable(struct pmu *pmu);
839 extern void perf_pmu_enable(struct pmu *pmu);
840 extern void perf_sched_cb_dec(struct pmu *pmu);
841 extern void perf_sched_cb_inc(struct pmu *pmu);
842 extern int perf_event_task_disable(void);
843 extern int perf_event_task_enable(void);
844 extern int perf_event_refresh(struct perf_event *event, int refresh);
845 extern void perf_event_update_userpage(struct perf_event *event);
846 extern int perf_event_release_kernel(struct perf_event *event);
847 extern struct perf_event *
848 perf_event_create_kernel_counter(struct perf_event_attr *attr,
850 struct task_struct *task,
851 perf_overflow_handler_t callback,
853 extern void perf_pmu_migrate_context(struct pmu *pmu,
854 int src_cpu, int dst_cpu);
855 extern u64 perf_event_read_local(struct perf_event *event);
856 extern u64 perf_event_read_value(struct perf_event *event,
857 u64 *enabled, u64 *running);
860 struct perf_sample_data {
862 * Fields set by perf_sample_data_init(), group so as to
863 * minimize the cachelines touched.
866 struct perf_raw_record *raw;
867 struct perf_branch_stack *br_stack;
871 union perf_mem_data_src data_src;
874 * The other fields, optionally {set,used} by
875 * perf_{prepare,output}_sample().
890 struct perf_callchain_entry *callchain;
893 * regs_user may point to task_pt_regs or to regs_user_copy, depending
896 struct perf_regs regs_user;
897 struct pt_regs regs_user_copy;
899 struct perf_regs regs_intr;
901 } ____cacheline_aligned;
903 /* default value for data source */
904 #define PERF_MEM_NA (PERF_MEM_S(OP, NA) |\
905 PERF_MEM_S(LVL, NA) |\
906 PERF_MEM_S(SNOOP, NA) |\
907 PERF_MEM_S(LOCK, NA) |\
910 static inline void perf_sample_data_init(struct perf_sample_data *data,
911 u64 addr, u64 period)
913 /* remaining struct members initialized in perf_prepare_sample() */
916 data->br_stack = NULL;
917 data->period = period;
919 data->data_src.val = PERF_MEM_NA;
923 extern void perf_output_sample(struct perf_output_handle *handle,
924 struct perf_event_header *header,
925 struct perf_sample_data *data,
926 struct perf_event *event);
927 extern void perf_prepare_sample(struct perf_event_header *header,
928 struct perf_sample_data *data,
929 struct perf_event *event,
930 struct pt_regs *regs);
932 extern int perf_event_overflow(struct perf_event *event,
933 struct perf_sample_data *data,
934 struct pt_regs *regs);
936 extern void perf_event_output_forward(struct perf_event *event,
937 struct perf_sample_data *data,
938 struct pt_regs *regs);
939 extern void perf_event_output_backward(struct perf_event *event,
940 struct perf_sample_data *data,
941 struct pt_regs *regs);
942 extern void perf_event_output(struct perf_event *event,
943 struct perf_sample_data *data,
944 struct pt_regs *regs);
947 is_default_overflow_handler(struct perf_event *event)
949 if (likely(event->overflow_handler == perf_event_output_forward))
951 if (unlikely(event->overflow_handler == perf_event_output_backward))
957 perf_event_header__init_id(struct perf_event_header *header,
958 struct perf_sample_data *data,
959 struct perf_event *event);
961 perf_event__output_id_sample(struct perf_event *event,
962 struct perf_output_handle *handle,
963 struct perf_sample_data *sample);
966 perf_log_lost_samples(struct perf_event *event, u64 lost);
968 static inline bool is_sampling_event(struct perf_event *event)
970 return event->attr.sample_period != 0;
974 * Return 1 for a software event, 0 for a hardware event
976 static inline int is_software_event(struct perf_event *event)
978 return event->pmu->task_ctx_nr == perf_sw_context;
981 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
983 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
984 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
986 #ifndef perf_arch_fetch_caller_regs
987 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
991 * Take a snapshot of the regs. Skip ip and frame pointer to
992 * the nth caller. We only need a few of the regs:
993 * - ip for PERF_SAMPLE_IP
994 * - cs for user_mode() tests
995 * - bp for callchains
996 * - eflags, for future purposes, just in case
998 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1000 perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1003 static __always_inline void
1004 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1006 if (static_key_false(&perf_swevent_enabled[event_id]))
1007 __perf_sw_event(event_id, nr, regs, addr);
1010 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
1013 * 'Special' version for the scheduler, it hard assumes no recursion,
1014 * which is guaranteed by us not actually scheduling inside other swevents
1015 * because those disable preemption.
1017 static __always_inline void
1018 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
1020 if (static_key_false(&perf_swevent_enabled[event_id])) {
1021 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1023 perf_fetch_caller_regs(regs);
1024 ___perf_sw_event(event_id, nr, regs, addr);
1028 extern struct static_key_false perf_sched_events;
1030 static __always_inline bool
1031 perf_sw_migrate_enabled(void)
1033 if (static_key_false(&perf_swevent_enabled[PERF_COUNT_SW_CPU_MIGRATIONS]))
1038 static inline void perf_event_task_migrate(struct task_struct *task)
1040 if (perf_sw_migrate_enabled())
1041 task->sched_migrated = 1;
1044 static inline void perf_event_task_sched_in(struct task_struct *prev,
1045 struct task_struct *task)
1047 if (static_branch_unlikely(&perf_sched_events))
1048 __perf_event_task_sched_in(prev, task);
1050 if (perf_sw_migrate_enabled() && task->sched_migrated) {
1051 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1053 perf_fetch_caller_regs(regs);
1054 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, regs, 0);
1055 task->sched_migrated = 0;
1059 static inline void perf_event_task_sched_out(struct task_struct *prev,
1060 struct task_struct *next)
1062 perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
1064 if (static_branch_unlikely(&perf_sched_events))
1065 __perf_event_task_sched_out(prev, next);
1068 static inline u64 __perf_event_count(struct perf_event *event)
1070 return local64_read(&event->count) + atomic64_read(&event->child_count);
1073 extern void perf_event_mmap(struct vm_area_struct *vma);
1074 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1075 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1076 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1078 extern void perf_event_exec(void);
1079 extern void perf_event_comm(struct task_struct *tsk, bool exec);
1080 extern void perf_event_fork(struct task_struct *tsk);
1083 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1085 extern void perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1086 extern void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1087 extern struct perf_callchain_entry *
1088 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
1089 u32 max_stack, bool crosstask, bool add_mark);
1090 extern int get_callchain_buffers(void);
1091 extern void put_callchain_buffers(void);
1093 extern int sysctl_perf_event_max_stack;
1094 extern int sysctl_perf_event_max_contexts_per_stack;
1096 static inline int perf_callchain_store_context(struct perf_callchain_entry_ctx *ctx, u64 ip)
1098 if (ctx->contexts < sysctl_perf_event_max_contexts_per_stack) {
1099 struct perf_callchain_entry *entry = ctx->entry;
1100 entry->ip[entry->nr++] = ip;
1104 ctx->contexts_maxed = true;
1105 return -1; /* no more room, stop walking the stack */
1109 static inline int perf_callchain_store(struct perf_callchain_entry_ctx *ctx, u64 ip)
1111 if (ctx->nr < ctx->max_stack && !ctx->contexts_maxed) {
1112 struct perf_callchain_entry *entry = ctx->entry;
1113 entry->ip[entry->nr++] = ip;
1117 return -1; /* no more room, stop walking the stack */
1121 extern int sysctl_perf_event_paranoid;
1122 extern int sysctl_perf_event_mlock;
1123 extern int sysctl_perf_event_sample_rate;
1124 extern int sysctl_perf_cpu_time_max_percent;
1126 extern void perf_sample_event_took(u64 sample_len_ns);
1128 extern int perf_proc_update_handler(struct ctl_table *table, int write,
1129 void __user *buffer, size_t *lenp,
1131 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
1132 void __user *buffer, size_t *lenp,
1135 int perf_event_max_stack_handler(struct ctl_table *table, int write,
1136 void __user *buffer, size_t *lenp, loff_t *ppos);
1138 static inline bool perf_paranoid_tracepoint_raw(void)
1140 return sysctl_perf_event_paranoid > -1;
1143 static inline bool perf_paranoid_cpu(void)
1145 return sysctl_perf_event_paranoid > 0;
1148 static inline bool perf_paranoid_kernel(void)
1150 return sysctl_perf_event_paranoid > 1;
1153 extern void perf_event_init(void);
1154 extern void perf_tp_event(u16 event_type, u64 count, void *record,
1155 int entry_size, struct pt_regs *regs,
1156 struct hlist_head *head, int rctx,
1157 struct task_struct *task);
1158 extern void perf_bp_event(struct perf_event *event, void *data);
1160 #ifndef perf_misc_flags
1161 # define perf_misc_flags(regs) \
1162 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1163 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1166 static inline bool has_branch_stack(struct perf_event *event)
1168 return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1171 static inline bool needs_branch_stack(struct perf_event *event)
1173 return event->attr.branch_sample_type != 0;
1176 static inline bool has_aux(struct perf_event *event)
1178 return event->pmu->setup_aux;
1181 static inline bool is_write_backward(struct perf_event *event)
1183 return !!event->attr.write_backward;
1186 static inline bool has_addr_filter(struct perf_event *event)
1188 return event->pmu->nr_addr_filters;
1192 * An inherited event uses parent's filters
1194 static inline struct perf_addr_filters_head *
1195 perf_event_addr_filters(struct perf_event *event)
1197 struct perf_addr_filters_head *ifh = &event->addr_filters;
1200 ifh = &event->parent->addr_filters;
1205 extern void perf_event_addr_filters_sync(struct perf_event *event);
1207 extern int perf_output_begin(struct perf_output_handle *handle,
1208 struct perf_event *event, unsigned int size);
1209 extern int perf_output_begin_forward(struct perf_output_handle *handle,
1210 struct perf_event *event,
1212 extern int perf_output_begin_backward(struct perf_output_handle *handle,
1213 struct perf_event *event,
1216 extern void perf_output_end(struct perf_output_handle *handle);
1217 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
1218 const void *buf, unsigned int len);
1219 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
1221 extern int perf_swevent_get_recursion_context(void);
1222 extern void perf_swevent_put_recursion_context(int rctx);
1223 extern u64 perf_swevent_set_period(struct perf_event *event);
1224 extern void perf_event_enable(struct perf_event *event);
1225 extern void perf_event_disable(struct perf_event *event);
1226 extern void perf_event_disable_local(struct perf_event *event);
1227 extern void perf_event_task_tick(void);
1228 #else /* !CONFIG_PERF_EVENTS: */
1229 static inline void *
1230 perf_aux_output_begin(struct perf_output_handle *handle,
1231 struct perf_event *event) { return NULL; }
1233 perf_aux_output_end(struct perf_output_handle *handle, unsigned long size,
1236 perf_aux_output_skip(struct perf_output_handle *handle,
1237 unsigned long size) { return -EINVAL; }
1238 static inline void *
1239 perf_get_aux(struct perf_output_handle *handle) { return NULL; }
1241 perf_event_task_migrate(struct task_struct *task) { }
1243 perf_event_task_sched_in(struct task_struct *prev,
1244 struct task_struct *task) { }
1246 perf_event_task_sched_out(struct task_struct *prev,
1247 struct task_struct *next) { }
1248 static inline int perf_event_init_task(struct task_struct *child) { return 0; }
1249 static inline void perf_event_exit_task(struct task_struct *child) { }
1250 static inline void perf_event_free_task(struct task_struct *task) { }
1251 static inline void perf_event_delayed_put(struct task_struct *task) { }
1252 static inline struct file *perf_event_get(unsigned int fd) { return ERR_PTR(-EINVAL); }
1253 static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
1255 return ERR_PTR(-EINVAL);
1257 static inline u64 perf_event_read_local(struct perf_event *event) { return -EINVAL; }
1258 static inline void perf_event_print_debug(void) { }
1259 static inline int perf_event_task_disable(void) { return -EINVAL; }
1260 static inline int perf_event_task_enable(void) { return -EINVAL; }
1261 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1267 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) { }
1269 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr) { }
1271 perf_bp_event(struct perf_event *event, void *data) { }
1273 static inline int perf_register_guest_info_callbacks
1274 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1275 static inline int perf_unregister_guest_info_callbacks
1276 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1278 static inline void perf_event_mmap(struct vm_area_struct *vma) { }
1279 static inline void perf_event_exec(void) { }
1280 static inline void perf_event_comm(struct task_struct *tsk, bool exec) { }
1281 static inline void perf_event_fork(struct task_struct *tsk) { }
1282 static inline void perf_event_init(void) { }
1283 static inline int perf_swevent_get_recursion_context(void) { return -1; }
1284 static inline void perf_swevent_put_recursion_context(int rctx) { }
1285 static inline u64 perf_swevent_set_period(struct perf_event *event) { return 0; }
1286 static inline void perf_event_enable(struct perf_event *event) { }
1287 static inline void perf_event_disable(struct perf_event *event) { }
1288 static inline int __perf_event_disable(void *info) { return -1; }
1289 static inline void perf_event_task_tick(void) { }
1290 static inline int perf_event_release_kernel(struct perf_event *event) { return 0; }
1293 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1294 extern void perf_restore_debug_store(void);
1296 static inline void perf_restore_debug_store(void) { }
1299 static __always_inline bool perf_raw_frag_last(const struct perf_raw_frag *frag)
1301 return frag->pad < sizeof(u64);
1304 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1307 * This has to have a higher priority than migration_notifier in sched/core.c.
1309 #define perf_cpu_notifier(fn) \
1311 static struct notifier_block fn##_nb = \
1312 { .notifier_call = fn, .priority = CPU_PRI_PERF }; \
1313 unsigned long cpu = smp_processor_id(); \
1314 unsigned long flags; \
1316 cpu_notifier_register_begin(); \
1317 fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE, \
1318 (void *)(unsigned long)cpu); \
1319 local_irq_save(flags); \
1320 fn(&fn##_nb, (unsigned long)CPU_STARTING, \
1321 (void *)(unsigned long)cpu); \
1322 local_irq_restore(flags); \
1323 fn(&fn##_nb, (unsigned long)CPU_ONLINE, \
1324 (void *)(unsigned long)cpu); \
1325 __register_cpu_notifier(&fn##_nb); \
1326 cpu_notifier_register_done(); \
1330 * Bare-bones version of perf_cpu_notifier(), which doesn't invoke the
1331 * callback for already online CPUs.
1333 #define __perf_cpu_notifier(fn) \
1335 static struct notifier_block fn##_nb = \
1336 { .notifier_call = fn, .priority = CPU_PRI_PERF }; \
1338 __register_cpu_notifier(&fn##_nb); \
1341 struct perf_pmu_events_attr {
1342 struct device_attribute attr;
1344 const char *event_str;
1347 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1350 #define PMU_EVENT_ATTR(_name, _var, _id, _show) \
1351 static struct perf_pmu_events_attr _var = { \
1352 .attr = __ATTR(_name, 0444, _show, NULL), \
1356 #define PMU_EVENT_ATTR_STRING(_name, _var, _str) \
1357 static struct perf_pmu_events_attr _var = { \
1358 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1360 .event_str = _str, \
1363 #define PMU_FORMAT_ATTR(_name, _format) \
1365 _name##_show(struct device *dev, \
1366 struct device_attribute *attr, \
1369 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
1370 return sprintf(page, _format "\n"); \
1373 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1375 #endif /* _LINUX_PERF_EVENT_H */