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 { /* amd_power */
146 #ifdef CONFIG_HAVE_HW_BREAKPOINT
147 struct { /* breakpoint */
149 * Crufty hack to avoid the chicken and egg
150 * problem hw_breakpoint has with context
151 * creation and event initalization.
153 struct arch_hw_breakpoint info;
154 struct list_head bp_list;
157 struct { /* amd_iommu */
166 * If the event is a per task event, this will point to the task in
167 * question. See the comment in perf_event_alloc().
169 struct task_struct *target;
172 * PMU would store hardware filter configuration
177 /* Last sync'ed generation of filters */
178 unsigned long addr_filters_gen;
181 * hw_perf_event::state flags; used to track the PERF_EF_* state.
183 #define PERF_HES_STOPPED 0x01 /* the counter is stopped */
184 #define PERF_HES_UPTODATE 0x02 /* event->count up-to-date */
185 #define PERF_HES_ARCH 0x04
190 * The last observed hardware counter value, updated with a
191 * local64_cmpxchg() such that pmu::read() can be called nested.
193 local64_t prev_count;
196 * The period to start the next sample with.
201 * The period we started this sample with.
206 * However much is left of the current period; note that this is
207 * a full 64bit value and allows for generation of periods longer
208 * than hardware might allow.
210 local64_t period_left;
213 * State for throttling the event, see __perf_event_overflow() and
214 * perf_adjust_freq_unthr_context().
220 * State for freq target events, see __perf_event_overflow() and
221 * perf_adjust_freq_unthr_context().
224 u64 freq_count_stamp;
231 * Common implementation detail of pmu::{start,commit,cancel}_txn
233 #define PERF_PMU_TXN_ADD 0x1 /* txn to add/schedule event on PMU */
234 #define PERF_PMU_TXN_READ 0x2 /* txn to read event group from PMU */
237 * pmu::capabilities flags
239 #define PERF_PMU_CAP_NO_INTERRUPT 0x01
240 #define PERF_PMU_CAP_NO_NMI 0x02
241 #define PERF_PMU_CAP_AUX_NO_SG 0x04
242 #define PERF_PMU_CAP_AUX_SW_DOUBLEBUF 0x08
243 #define PERF_PMU_CAP_EXCLUSIVE 0x10
244 #define PERF_PMU_CAP_ITRACE 0x20
245 #define PERF_PMU_CAP_HETEROGENEOUS_CPUS 0x40
248 * struct pmu - generic performance monitoring unit
251 struct list_head entry;
253 struct module *module;
255 const struct attribute_group **attr_groups;
260 * various common per-pmu feature flags
264 int * __percpu pmu_disable_count;
265 struct perf_cpu_context * __percpu pmu_cpu_context;
266 atomic_t exclusive_cnt; /* < 0: cpu; > 0: tsk */
268 int hrtimer_interval_ms;
270 /* number of address filters this PMU can do */
271 unsigned int nr_addr_filters;
274 * Fully disable/enable this PMU, can be used to protect from the PMI
275 * as well as for lazy/batch writing of the MSRs.
277 void (*pmu_enable) (struct pmu *pmu); /* optional */
278 void (*pmu_disable) (struct pmu *pmu); /* optional */
281 * Try and initialize the event for this PMU.
284 * -ENOENT -- @event is not for this PMU
286 * -ENODEV -- @event is for this PMU but PMU not present
287 * -EBUSY -- @event is for this PMU but PMU temporarily unavailable
288 * -EINVAL -- @event is for this PMU but @event is not valid
289 * -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
290 * -EACCESS -- @event is for this PMU, @event is valid, but no privilidges
292 * 0 -- @event is for this PMU and valid
294 * Other error return values are allowed.
296 int (*event_init) (struct perf_event *event);
299 * Notification that the event was mapped or unmapped. Called
300 * in the context of the mapping task.
302 void (*event_mapped) (struct perf_event *event, struct mm_struct *mm); /* optional */
303 void (*event_unmapped) (struct perf_event *event, struct mm_struct *mm); /* optional */
306 * Flags for ->add()/->del()/ ->start()/->stop(). There are
307 * matching hw_perf_event::state flags.
309 #define PERF_EF_START 0x01 /* start the counter when adding */
310 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
311 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
314 * Adds/Removes a counter to/from the PMU, can be done inside a
315 * transaction, see the ->*_txn() methods.
317 * The add/del callbacks will reserve all hardware resources required
318 * to service the event, this includes any counter constraint
321 * Called with IRQs disabled and the PMU disabled on the CPU the event
324 * ->add() called without PERF_EF_START should result in the same state
325 * as ->add() followed by ->stop().
327 * ->del() must always PERF_EF_UPDATE stop an event. If it calls
328 * ->stop() that must deal with already being stopped without
331 int (*add) (struct perf_event *event, int flags);
332 void (*del) (struct perf_event *event, int flags);
335 * Starts/Stops a counter present on the PMU.
337 * The PMI handler should stop the counter when perf_event_overflow()
338 * returns !0. ->start() will be used to continue.
340 * Also used to change the sample period.
342 * Called with IRQs disabled and the PMU disabled on the CPU the event
343 * is on -- will be called from NMI context with the PMU generates
346 * ->stop() with PERF_EF_UPDATE will read the counter and update
347 * period/count values like ->read() would.
349 * ->start() with PERF_EF_RELOAD will reprogram the the counter
350 * value, must be preceded by a ->stop() with PERF_EF_UPDATE.
352 void (*start) (struct perf_event *event, int flags);
353 void (*stop) (struct perf_event *event, int flags);
356 * Updates the counter value of the event.
358 * For sampling capable PMUs this will also update the software period
359 * hw_perf_event::period_left field.
361 void (*read) (struct perf_event *event);
364 * Group events scheduling is treated as a transaction, add
365 * group events as a whole and perform one schedulability test.
366 * If the test fails, roll back the whole group
368 * Start the transaction, after this ->add() doesn't need to
369 * do schedulability tests.
373 void (*start_txn) (struct pmu *pmu, unsigned int txn_flags);
375 * If ->start_txn() disabled the ->add() schedulability test
376 * then ->commit_txn() is required to perform one. On success
377 * the transaction is closed. On error the transaction is kept
378 * open until ->cancel_txn() is called.
382 int (*commit_txn) (struct pmu *pmu);
384 * Will cancel the transaction, assumes ->del() is called
385 * for each successful ->add() during the transaction.
389 void (*cancel_txn) (struct pmu *pmu);
392 * Will return the value for perf_event_mmap_page::index for this event,
393 * if no implementation is provided it will default to: event->hw.idx + 1.
395 int (*event_idx) (struct perf_event *event); /*optional */
398 * context-switches callback
400 void (*sched_task) (struct perf_event_context *ctx,
403 * PMU specific data size
405 size_t task_ctx_size;
409 * Set up pmu-private data structures for an AUX area
411 void *(*setup_aux) (int cpu, void **pages,
412 int nr_pages, bool overwrite);
416 * Free pmu-private AUX data structures
418 void (*free_aux) (void *aux); /* optional */
421 * Validate address range filters: make sure the HW supports the
422 * requested configuration and number of filters; return 0 if the
423 * supplied filters are valid, -errno otherwise.
425 * Runs in the context of the ioctl()ing process and is not serialized
426 * with the rest of the PMU callbacks.
428 int (*addr_filters_validate) (struct list_head *filters);
432 * Synchronize address range filter configuration:
433 * translate hw-agnostic filters into hardware configuration in
434 * event::hw::addr_filters.
436 * Runs as a part of filter sync sequence that is done in ->start()
437 * callback by calling perf_event_addr_filters_sync().
439 * May (and should) traverse event::addr_filters::list, for which its
440 * caller provides necessary serialization.
442 void (*addr_filters_sync) (struct perf_event *event);
446 * Filter events for PMU-specific reasons.
448 int (*filter_match) (struct perf_event *event); /* optional */
451 * Check period value for PERF_EVENT_IOC_PERIOD ioctl.
453 int (*check_period) (struct perf_event *event, u64 value); /* optional */
457 * struct perf_addr_filter - address range filter definition
458 * @entry: event's filter list linkage
459 * @inode: object file's inode for file-based filters
460 * @offset: filter range offset
461 * @size: filter range size
462 * @range: 1: range, 0: address
463 * @filter: 1: filter/start, 0: stop
465 * This is a hardware-agnostic filter configuration as specified by the user.
467 struct perf_addr_filter {
468 struct list_head entry;
470 unsigned long offset;
472 unsigned int range : 1,
477 * struct perf_addr_filters_head - container for address range filters
478 * @list: list of filters for this event
479 * @lock: spinlock that serializes accesses to the @list and event's
480 * (and its children's) filter generations.
481 * @nr_file_filters: number of file-based filters
483 * A child event will use parent's @list (and therefore @lock), so they are
484 * bundled together; see perf_event_addr_filters().
486 struct perf_addr_filters_head {
487 struct list_head list;
489 unsigned int nr_file_filters;
493 * enum perf_event_active_state - the states of a event
495 enum perf_event_active_state {
496 PERF_EVENT_STATE_DEAD = -4,
497 PERF_EVENT_STATE_EXIT = -3,
498 PERF_EVENT_STATE_ERROR = -2,
499 PERF_EVENT_STATE_OFF = -1,
500 PERF_EVENT_STATE_INACTIVE = 0,
501 PERF_EVENT_STATE_ACTIVE = 1,
505 struct perf_sample_data;
507 typedef void (*perf_overflow_handler_t)(struct perf_event *,
508 struct perf_sample_data *,
509 struct pt_regs *regs);
512 * Event capabilities. For event_caps and groups caps.
514 * PERF_EV_CAP_SOFTWARE: Is a software event.
515 * PERF_EV_CAP_READ_ACTIVE_PKG: A CPU event (or cgroup event) that can be read
516 * from any CPU in the package where it is active.
518 #define PERF_EV_CAP_SOFTWARE BIT(0)
519 #define PERF_EV_CAP_READ_ACTIVE_PKG BIT(1)
521 #define SWEVENT_HLIST_BITS 8
522 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
524 struct swevent_hlist {
525 struct hlist_head heads[SWEVENT_HLIST_SIZE];
526 struct rcu_head rcu_head;
529 #define PERF_ATTACH_CONTEXT 0x01
530 #define PERF_ATTACH_GROUP 0x02
531 #define PERF_ATTACH_TASK 0x04
532 #define PERF_ATTACH_TASK_DATA 0x08
533 #define PERF_ATTACH_ITRACE 0x10
538 struct pmu_event_list {
540 struct list_head list;
544 * struct perf_event - performance event kernel representation:
547 #ifdef CONFIG_PERF_EVENTS
549 * entry onto perf_event_context::event_list;
550 * modifications require ctx->lock
551 * RCU safe iterations.
553 struct list_head event_entry;
556 * XXX: group_entry and sibling_list should be mutually exclusive;
557 * either you're a sibling on a group, or you're the group leader.
558 * Rework the code to always use the same list element.
560 * Locked for modification by both ctx->mutex and ctx->lock; holding
561 * either sufficies for read.
563 struct list_head group_entry;
564 struct list_head sibling_list;
567 * We need storage to track the entries in perf_pmu_migrate_context; we
568 * cannot use the event_entry because of RCU and we want to keep the
569 * group in tact which avoids us using the other two entries.
571 struct list_head migrate_entry;
573 struct hlist_node hlist_entry;
574 struct list_head active_entry;
577 /* Not serialized. Only written during event initialization. */
579 /* The cumulative AND of all event_caps for events in this group. */
582 struct perf_event *group_leader;
586 enum perf_event_active_state state;
587 unsigned int attach_state;
589 atomic64_t child_count;
592 * These are the total time in nanoseconds that the event
593 * has been enabled (i.e. eligible to run, and the task has
594 * been scheduled in, if this is a per-task event)
595 * and running (scheduled onto the CPU), respectively.
597 * They are computed from tstamp_enabled, tstamp_running and
598 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
600 u64 total_time_enabled;
601 u64 total_time_running;
604 * These are timestamps used for computing total_time_enabled
605 * and total_time_running when the event is in INACTIVE or
606 * ACTIVE state, measured in nanoseconds from an arbitrary point
608 * tstamp_enabled: the notional time when the event was enabled
609 * tstamp_running: the notional time when the event was scheduled on
610 * tstamp_stopped: in INACTIVE state, the notional time when the
611 * event was scheduled off.
618 * timestamp shadows the actual context timing but it can
619 * be safely used in NMI interrupt context. It reflects the
620 * context time as it was when the event was last scheduled in.
622 * ctx_time already accounts for ctx->timestamp. Therefore to
623 * compute ctx_time for a sample, simply add perf_clock().
627 struct perf_event_attr attr;
631 struct hw_perf_event hw;
633 struct perf_event_context *ctx;
634 atomic_long_t refcount;
637 * These accumulate total time (in nanoseconds) that children
638 * events have been enabled and running, respectively.
640 atomic64_t child_total_time_enabled;
641 atomic64_t child_total_time_running;
644 * Protect attach/detach and child_list:
646 struct mutex child_mutex;
647 struct list_head child_list;
648 struct perf_event *parent;
653 struct list_head owner_entry;
654 struct task_struct *owner;
657 struct mutex mmap_mutex;
660 struct ring_buffer *rb;
661 struct list_head rb_entry;
662 unsigned long rcu_batches;
666 wait_queue_head_t waitq;
667 struct fasync_struct *fasync;
669 /* delayed work for NMIs and such */
673 struct irq_work pending;
675 atomic_t event_limit;
677 /* address range filters */
678 struct perf_addr_filters_head addr_filters;
679 /* vma address array for file-based filders */
680 unsigned long *addr_filters_offs;
681 unsigned long addr_filters_gen;
683 void (*destroy)(struct perf_event *);
684 struct rcu_head rcu_head;
686 struct pid_namespace *ns;
690 perf_overflow_handler_t overflow_handler;
691 void *overflow_handler_context;
692 #ifdef CONFIG_BPF_SYSCALL
693 perf_overflow_handler_t orig_overflow_handler;
694 struct bpf_prog *prog;
697 #ifdef CONFIG_EVENT_TRACING
698 struct trace_event_call *tp_event;
699 struct event_filter *filter;
700 #ifdef CONFIG_FUNCTION_TRACER
701 struct ftrace_ops ftrace_ops;
705 #ifdef CONFIG_CGROUP_PERF
706 struct perf_cgroup *cgrp; /* cgroup event is attach to */
707 int cgrp_defer_enabled;
710 struct list_head sb_list;
711 #endif /* CONFIG_PERF_EVENTS */
715 * struct perf_event_context - event context structure
717 * Used as a container for task events and CPU events as well:
719 struct perf_event_context {
722 * Protect the states of the events in the list,
723 * nr_active, and the list:
727 * Protect the list of events. Locking either mutex or lock
728 * is sufficient to ensure the list doesn't change; to change
729 * the list you need to lock both the mutex and the spinlock.
733 struct list_head active_ctx_list;
734 struct list_head pinned_groups;
735 struct list_head flexible_groups;
736 struct list_head event_list;
744 struct task_struct *task;
747 * Context clock, runs when context enabled.
753 * These fields let us detect when two contexts have both
754 * been cloned (inherited) from a common ancestor.
756 struct perf_event_context *parent_ctx;
760 #ifdef CONFIG_CGROUP_PERF
761 int nr_cgroups; /* cgroup evts */
763 void *task_ctx_data; /* pmu specific data */
764 struct rcu_head rcu_head;
768 * Number of contexts where an event can trigger:
769 * task, softirq, hardirq, nmi.
771 #define PERF_NR_CONTEXTS 4
774 * struct perf_event_cpu_context - per cpu event context structure
776 struct perf_cpu_context {
777 struct perf_event_context ctx;
778 struct perf_event_context *task_ctx;
782 raw_spinlock_t hrtimer_lock;
783 struct hrtimer hrtimer;
784 ktime_t hrtimer_interval;
785 unsigned int hrtimer_active;
787 #ifdef CONFIG_CGROUP_PERF
788 struct perf_cgroup *cgrp;
789 struct list_head cgrp_cpuctx_entry;
792 struct list_head sched_cb_entry;
798 struct perf_output_handle {
799 struct perf_event *event;
800 struct ring_buffer *rb;
801 unsigned long wakeup;
811 struct bpf_perf_event_data_kern {
812 struct pt_regs *regs;
813 struct perf_sample_data *data;
816 #ifdef CONFIG_CGROUP_PERF
819 * perf_cgroup_info keeps track of time_enabled for a cgroup.
820 * This is a per-cpu dynamically allocated data structure.
822 struct perf_cgroup_info {
828 struct cgroup_subsys_state css;
829 struct perf_cgroup_info __percpu *info;
833 * Must ensure cgroup is pinned (css_get) before calling
834 * this function. In other words, we cannot call this function
835 * if there is no cgroup event for the current CPU context.
837 static inline struct perf_cgroup *
838 perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
840 return container_of(task_css_check(task, perf_event_cgrp_id,
841 ctx ? lockdep_is_held(&ctx->lock)
843 struct perf_cgroup, css);
845 #endif /* CONFIG_CGROUP_PERF */
847 #ifdef CONFIG_PERF_EVENTS
849 extern void *perf_aux_output_begin(struct perf_output_handle *handle,
850 struct perf_event *event);
851 extern void perf_aux_output_end(struct perf_output_handle *handle,
853 extern int perf_aux_output_skip(struct perf_output_handle *handle,
855 extern void *perf_get_aux(struct perf_output_handle *handle);
856 extern void perf_aux_output_flag(struct perf_output_handle *handle, u64 flags);
857 extern void perf_event_itrace_started(struct perf_event *event);
859 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
860 extern void perf_pmu_unregister(struct pmu *pmu);
862 extern int perf_num_counters(void);
863 extern const char *perf_pmu_name(void);
864 extern void __perf_event_task_sched_in(struct task_struct *prev,
865 struct task_struct *task);
866 extern void __perf_event_task_sched_out(struct task_struct *prev,
867 struct task_struct *next);
868 extern int perf_event_init_task(struct task_struct *child);
869 extern void perf_event_exit_task(struct task_struct *child);
870 extern void perf_event_free_task(struct task_struct *task);
871 extern void perf_event_delayed_put(struct task_struct *task);
872 extern struct file *perf_event_get(unsigned int fd);
873 extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
874 extern void perf_event_print_debug(void);
875 extern void perf_pmu_disable(struct pmu *pmu);
876 extern void perf_pmu_enable(struct pmu *pmu);
877 extern void perf_sched_cb_dec(struct pmu *pmu);
878 extern void perf_sched_cb_inc(struct pmu *pmu);
879 extern int perf_event_task_disable(void);
880 extern int perf_event_task_enable(void);
881 extern int perf_event_refresh(struct perf_event *event, int refresh);
882 extern void perf_event_update_userpage(struct perf_event *event);
883 extern int perf_event_release_kernel(struct perf_event *event);
884 extern struct perf_event *
885 perf_event_create_kernel_counter(struct perf_event_attr *attr,
887 struct task_struct *task,
888 perf_overflow_handler_t callback,
890 extern void perf_pmu_migrate_context(struct pmu *pmu,
891 int src_cpu, int dst_cpu);
892 int perf_event_read_local(struct perf_event *event, u64 *value);
893 extern u64 perf_event_read_value(struct perf_event *event,
894 u64 *enabled, u64 *running);
897 struct perf_sample_data {
899 * Fields set by perf_sample_data_init(), group so as to
900 * minimize the cachelines touched.
903 struct perf_raw_record *raw;
904 struct perf_branch_stack *br_stack;
908 union perf_mem_data_src data_src;
911 * The other fields, optionally {set,used} by
912 * perf_{prepare,output}_sample().
927 struct perf_callchain_entry *callchain;
930 * regs_user may point to task_pt_regs or to regs_user_copy, depending
933 struct perf_regs regs_user;
934 struct pt_regs regs_user_copy;
936 struct perf_regs regs_intr;
940 } ____cacheline_aligned;
942 /* default value for data source */
943 #define PERF_MEM_NA (PERF_MEM_S(OP, NA) |\
944 PERF_MEM_S(LVL, NA) |\
945 PERF_MEM_S(SNOOP, NA) |\
946 PERF_MEM_S(LOCK, NA) |\
949 static inline void perf_sample_data_init(struct perf_sample_data *data,
950 u64 addr, u64 period)
952 /* remaining struct members initialized in perf_prepare_sample() */
955 data->br_stack = NULL;
956 data->period = period;
958 data->data_src.val = PERF_MEM_NA;
962 extern void perf_output_sample(struct perf_output_handle *handle,
963 struct perf_event_header *header,
964 struct perf_sample_data *data,
965 struct perf_event *event);
966 extern void perf_prepare_sample(struct perf_event_header *header,
967 struct perf_sample_data *data,
968 struct perf_event *event,
969 struct pt_regs *regs);
971 extern int perf_event_overflow(struct perf_event *event,
972 struct perf_sample_data *data,
973 struct pt_regs *regs);
975 extern void perf_event_output_forward(struct perf_event *event,
976 struct perf_sample_data *data,
977 struct pt_regs *regs);
978 extern void perf_event_output_backward(struct perf_event *event,
979 struct perf_sample_data *data,
980 struct pt_regs *regs);
981 extern void perf_event_output(struct perf_event *event,
982 struct perf_sample_data *data,
983 struct pt_regs *regs);
986 is_default_overflow_handler(struct perf_event *event)
988 if (likely(event->overflow_handler == perf_event_output_forward))
990 if (unlikely(event->overflow_handler == perf_event_output_backward))
996 perf_event_header__init_id(struct perf_event_header *header,
997 struct perf_sample_data *data,
998 struct perf_event *event);
1000 perf_event__output_id_sample(struct perf_event *event,
1001 struct perf_output_handle *handle,
1002 struct perf_sample_data *sample);
1005 perf_log_lost_samples(struct perf_event *event, u64 lost);
1007 static inline bool is_sampling_event(struct perf_event *event)
1009 return event->attr.sample_period != 0;
1013 * Return 1 for a software event, 0 for a hardware event
1015 static inline int is_software_event(struct perf_event *event)
1017 return event->event_caps & PERF_EV_CAP_SOFTWARE;
1020 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
1022 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
1023 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
1025 #ifndef perf_arch_fetch_caller_regs
1026 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
1030 * Take a snapshot of the regs. Skip ip and frame pointer to
1031 * the nth caller. We only need a few of the regs:
1032 * - ip for PERF_SAMPLE_IP
1033 * - cs for user_mode() tests
1034 * - bp for callchains
1035 * - eflags, for future purposes, just in case
1037 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1039 perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1042 static __always_inline void
1043 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1045 if (static_key_false(&perf_swevent_enabled[event_id]))
1046 __perf_sw_event(event_id, nr, regs, addr);
1049 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
1052 * 'Special' version for the scheduler, it hard assumes no recursion,
1053 * which is guaranteed by us not actually scheduling inside other swevents
1054 * because those disable preemption.
1056 static __always_inline void
1057 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
1059 if (static_key_false(&perf_swevent_enabled[event_id])) {
1060 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1062 perf_fetch_caller_regs(regs);
1063 ___perf_sw_event(event_id, nr, regs, addr);
1067 extern struct static_key_false perf_sched_events;
1069 static __always_inline bool
1070 perf_sw_migrate_enabled(void)
1072 if (static_key_false(&perf_swevent_enabled[PERF_COUNT_SW_CPU_MIGRATIONS]))
1077 static inline void perf_event_task_migrate(struct task_struct *task)
1079 if (perf_sw_migrate_enabled())
1080 task->sched_migrated = 1;
1083 static inline void perf_event_task_sched_in(struct task_struct *prev,
1084 struct task_struct *task)
1086 if (static_branch_unlikely(&perf_sched_events))
1087 __perf_event_task_sched_in(prev, task);
1089 if (perf_sw_migrate_enabled() && task->sched_migrated) {
1090 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1092 perf_fetch_caller_regs(regs);
1093 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, regs, 0);
1094 task->sched_migrated = 0;
1098 static inline void perf_event_task_sched_out(struct task_struct *prev,
1099 struct task_struct *next)
1101 perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
1103 if (static_branch_unlikely(&perf_sched_events))
1104 __perf_event_task_sched_out(prev, next);
1107 extern void perf_event_mmap(struct vm_area_struct *vma);
1108 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1109 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1110 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1112 extern void perf_event_exec(void);
1113 extern void perf_event_comm(struct task_struct *tsk, bool exec);
1114 extern void perf_event_namespaces(struct task_struct *tsk);
1115 extern void perf_event_fork(struct task_struct *tsk);
1118 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1120 extern void perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1121 extern void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1122 extern struct perf_callchain_entry *
1123 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
1124 u32 max_stack, bool crosstask, bool add_mark);
1125 extern int get_callchain_buffers(int max_stack);
1126 extern void put_callchain_buffers(void);
1128 extern int sysctl_perf_event_max_stack;
1129 extern int sysctl_perf_event_max_contexts_per_stack;
1131 static inline int perf_callchain_store_context(struct perf_callchain_entry_ctx *ctx, u64 ip)
1133 if (ctx->contexts < sysctl_perf_event_max_contexts_per_stack) {
1134 struct perf_callchain_entry *entry = ctx->entry;
1135 entry->ip[entry->nr++] = ip;
1139 ctx->contexts_maxed = true;
1140 return -1; /* no more room, stop walking the stack */
1144 static inline int perf_callchain_store(struct perf_callchain_entry_ctx *ctx, u64 ip)
1146 if (ctx->nr < ctx->max_stack && !ctx->contexts_maxed) {
1147 struct perf_callchain_entry *entry = ctx->entry;
1148 entry->ip[entry->nr++] = ip;
1152 return -1; /* no more room, stop walking the stack */
1156 extern int sysctl_perf_event_paranoid;
1157 extern int sysctl_perf_event_mlock;
1158 extern int sysctl_perf_event_sample_rate;
1159 extern int sysctl_perf_cpu_time_max_percent;
1161 extern void perf_sample_event_took(u64 sample_len_ns);
1163 extern int perf_proc_update_handler(struct ctl_table *table, int write,
1164 void __user *buffer, size_t *lenp,
1166 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
1167 void __user *buffer, size_t *lenp,
1170 int perf_event_max_stack_handler(struct ctl_table *table, int write,
1171 void __user *buffer, size_t *lenp, loff_t *ppos);
1173 static inline bool perf_paranoid_tracepoint_raw(void)
1175 return sysctl_perf_event_paranoid > -1;
1178 static inline bool perf_paranoid_cpu(void)
1180 return sysctl_perf_event_paranoid > 0;
1183 static inline bool perf_paranoid_kernel(void)
1185 return sysctl_perf_event_paranoid > 1;
1188 extern void perf_event_init(void);
1189 extern void perf_tp_event(u16 event_type, u64 count, void *record,
1190 int entry_size, struct pt_regs *regs,
1191 struct hlist_head *head, int rctx,
1192 struct task_struct *task, struct perf_event *event);
1193 extern void perf_bp_event(struct perf_event *event, void *data);
1195 #ifndef perf_misc_flags
1196 # define perf_misc_flags(regs) \
1197 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1198 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1201 static inline bool has_branch_stack(struct perf_event *event)
1203 return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1206 static inline bool needs_branch_stack(struct perf_event *event)
1208 return event->attr.branch_sample_type != 0;
1211 static inline bool has_aux(struct perf_event *event)
1213 return event->pmu->setup_aux;
1216 static inline bool is_write_backward(struct perf_event *event)
1218 return !!event->attr.write_backward;
1221 static inline bool has_addr_filter(struct perf_event *event)
1223 return event->pmu->nr_addr_filters;
1227 * An inherited event uses parent's filters
1229 static inline struct perf_addr_filters_head *
1230 perf_event_addr_filters(struct perf_event *event)
1232 struct perf_addr_filters_head *ifh = &event->addr_filters;
1235 ifh = &event->parent->addr_filters;
1240 extern void perf_event_addr_filters_sync(struct perf_event *event);
1242 extern int perf_output_begin(struct perf_output_handle *handle,
1243 struct perf_event *event, unsigned int size);
1244 extern int perf_output_begin_forward(struct perf_output_handle *handle,
1245 struct perf_event *event,
1247 extern int perf_output_begin_backward(struct perf_output_handle *handle,
1248 struct perf_event *event,
1251 extern void perf_output_end(struct perf_output_handle *handle);
1252 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
1253 const void *buf, unsigned int len);
1254 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
1256 extern int perf_swevent_get_recursion_context(void);
1257 extern void perf_swevent_put_recursion_context(int rctx);
1258 extern u64 perf_swevent_set_period(struct perf_event *event);
1259 extern void perf_event_enable(struct perf_event *event);
1260 extern void perf_event_disable(struct perf_event *event);
1261 extern void perf_event_disable_local(struct perf_event *event);
1262 extern void perf_event_disable_inatomic(struct perf_event *event);
1263 extern void perf_event_task_tick(void);
1264 extern int perf_event_account_interrupt(struct perf_event *event);
1265 #else /* !CONFIG_PERF_EVENTS: */
1266 static inline void *
1267 perf_aux_output_begin(struct perf_output_handle *handle,
1268 struct perf_event *event) { return NULL; }
1270 perf_aux_output_end(struct perf_output_handle *handle, unsigned long size)
1273 perf_aux_output_skip(struct perf_output_handle *handle,
1274 unsigned long size) { return -EINVAL; }
1275 static inline void *
1276 perf_get_aux(struct perf_output_handle *handle) { return NULL; }
1278 perf_event_task_migrate(struct task_struct *task) { }
1280 perf_event_task_sched_in(struct task_struct *prev,
1281 struct task_struct *task) { }
1283 perf_event_task_sched_out(struct task_struct *prev,
1284 struct task_struct *next) { }
1285 static inline int perf_event_init_task(struct task_struct *child) { return 0; }
1286 static inline void perf_event_exit_task(struct task_struct *child) { }
1287 static inline void perf_event_free_task(struct task_struct *task) { }
1288 static inline void perf_event_delayed_put(struct task_struct *task) { }
1289 static inline struct file *perf_event_get(unsigned int fd) { return ERR_PTR(-EINVAL); }
1290 static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
1292 return ERR_PTR(-EINVAL);
1294 static inline int perf_event_read_local(struct perf_event *event, u64 *value)
1298 static inline void perf_event_print_debug(void) { }
1299 static inline int perf_event_task_disable(void) { return -EINVAL; }
1300 static inline int perf_event_task_enable(void) { return -EINVAL; }
1301 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1307 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) { }
1309 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr) { }
1311 perf_bp_event(struct perf_event *event, void *data) { }
1313 static inline int perf_register_guest_info_callbacks
1314 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1315 static inline int perf_unregister_guest_info_callbacks
1316 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1318 static inline void perf_event_mmap(struct vm_area_struct *vma) { }
1319 static inline void perf_event_exec(void) { }
1320 static inline void perf_event_comm(struct task_struct *tsk, bool exec) { }
1321 static inline void perf_event_namespaces(struct task_struct *tsk) { }
1322 static inline void perf_event_fork(struct task_struct *tsk) { }
1323 static inline void perf_event_init(void) { }
1324 static inline int perf_swevent_get_recursion_context(void) { return -1; }
1325 static inline void perf_swevent_put_recursion_context(int rctx) { }
1326 static inline u64 perf_swevent_set_period(struct perf_event *event) { return 0; }
1327 static inline void perf_event_enable(struct perf_event *event) { }
1328 static inline void perf_event_disable(struct perf_event *event) { }
1329 static inline int __perf_event_disable(void *info) { return -1; }
1330 static inline void perf_event_task_tick(void) { }
1331 static inline int perf_event_release_kernel(struct perf_event *event) { return 0; }
1334 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1335 extern void perf_restore_debug_store(void);
1337 static inline void perf_restore_debug_store(void) { }
1340 static __always_inline bool perf_raw_frag_last(const struct perf_raw_frag *frag)
1342 return frag->pad < sizeof(u64);
1345 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1347 struct perf_pmu_events_attr {
1348 struct device_attribute attr;
1350 const char *event_str;
1353 struct perf_pmu_events_ht_attr {
1354 struct device_attribute attr;
1356 const char *event_str_ht;
1357 const char *event_str_noht;
1360 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1363 #define PMU_EVENT_ATTR(_name, _var, _id, _show) \
1364 static struct perf_pmu_events_attr _var = { \
1365 .attr = __ATTR(_name, 0444, _show, NULL), \
1369 #define PMU_EVENT_ATTR_STRING(_name, _var, _str) \
1370 static struct perf_pmu_events_attr _var = { \
1371 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1373 .event_str = _str, \
1376 #define PMU_FORMAT_ATTR(_name, _format) \
1378 _name##_show(struct device *dev, \
1379 struct device_attribute *attr, \
1382 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
1383 return sprintf(page, _format "\n"); \
1386 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1388 /* Performance counter hotplug functions */
1389 #ifdef CONFIG_PERF_EVENTS
1390 int perf_event_init_cpu(unsigned int cpu);
1391 int perf_event_exit_cpu(unsigned int cpu);
1393 #define perf_event_init_cpu NULL
1394 #define perf_event_exit_cpu NULL
1397 #endif /* _LINUX_PERF_EVENT_H */