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>
18 #include <uapi/linux/bpf_perf_event.h>
21 * Kernel-internal data types and definitions:
24 #ifdef CONFIG_PERF_EVENTS
25 # include <asm/perf_event.h>
26 # include <asm/local64.h>
29 struct perf_guest_info_callbacks {
30 int (*is_in_guest)(void);
31 int (*is_user_mode)(void);
32 unsigned long (*get_guest_ip)(void);
33 void (*handle_intel_pt_intr)(void);
36 #ifdef CONFIG_HAVE_HW_BREAKPOINT
37 #include <asm/hw_breakpoint.h>
40 #include <linux/list.h>
41 #include <linux/mutex.h>
42 #include <linux/rculist.h>
43 #include <linux/rcupdate.h>
44 #include <linux/spinlock.h>
45 #include <linux/hrtimer.h>
47 #include <linux/pid_namespace.h>
48 #include <linux/workqueue.h>
49 #include <linux/ftrace.h>
50 #include <linux/cpu.h>
51 #include <linux/irq_work.h>
52 #include <linux/static_key.h>
53 #include <linux/jump_label_ratelimit.h>
54 #include <linux/atomic.h>
55 #include <linux/sysfs.h>
56 #include <linux/perf_regs.h>
57 #include <linux/cgroup.h>
58 #include <linux/refcount.h>
59 #include <asm/local.h>
61 struct perf_callchain_entry {
63 __u64 ip[0]; /* /proc/sys/kernel/perf_event_max_stack */
66 struct perf_callchain_entry_ctx {
67 struct perf_callchain_entry *entry;
74 typedef unsigned long (*perf_copy_f)(void *dst, const void *src,
75 unsigned long off, unsigned long len);
77 struct perf_raw_frag {
79 struct perf_raw_frag *next;
87 struct perf_raw_record {
88 struct perf_raw_frag frag;
93 * branch stack layout:
94 * nr: number of taken branches stored in entries[]
96 * Note that nr can vary from sample to sample
97 * branches (to, from) are stored from most recent
98 * to least recent, i.e., entries[0] contains the most
101 struct perf_branch_stack {
103 struct perf_branch_entry entries[0];
109 * extra PMU register associated with an event
111 struct hw_perf_event_extra {
112 u64 config; /* register value */
113 unsigned int reg; /* register address or index */
114 int alloc; /* extra register already allocated */
115 int idx; /* index in shared_regs->regs[] */
119 * struct hw_perf_event - performance event hardware details:
121 struct hw_perf_event {
122 #ifdef CONFIG_PERF_EVENTS
124 struct { /* hardware */
127 unsigned long config_base;
128 unsigned long event_base;
129 int event_base_rdpmc;
134 struct hw_perf_event_extra extra_reg;
135 struct hw_perf_event_extra branch_reg;
137 struct { /* software */
138 struct hrtimer hrtimer;
140 struct { /* tracepoint */
141 /* for tp_event->class */
142 struct list_head tp_list;
144 struct { /* amd_power */
148 #ifdef CONFIG_HAVE_HW_BREAKPOINT
149 struct { /* breakpoint */
151 * Crufty hack to avoid the chicken and egg
152 * problem hw_breakpoint has with context
153 * creation and event initalization.
155 struct arch_hw_breakpoint info;
156 struct list_head bp_list;
159 struct { /* amd_iommu */
168 * If the event is a per task event, this will point to the task in
169 * question. See the comment in perf_event_alloc().
171 struct task_struct *target;
174 * PMU would store hardware filter configuration
179 /* Last sync'ed generation of filters */
180 unsigned long addr_filters_gen;
183 * hw_perf_event::state flags; used to track the PERF_EF_* state.
185 #define PERF_HES_STOPPED 0x01 /* the counter is stopped */
186 #define PERF_HES_UPTODATE 0x02 /* event->count up-to-date */
187 #define PERF_HES_ARCH 0x04
192 * The last observed hardware counter value, updated with a
193 * local64_cmpxchg() such that pmu::read() can be called nested.
195 local64_t prev_count;
198 * The period to start the next sample with.
203 * The period we started this sample with.
208 * However much is left of the current period; note that this is
209 * a full 64bit value and allows for generation of periods longer
210 * than hardware might allow.
212 local64_t period_left;
215 * State for throttling the event, see __perf_event_overflow() and
216 * perf_adjust_freq_unthr_context().
222 * State for freq target events, see __perf_event_overflow() and
223 * perf_adjust_freq_unthr_context().
226 u64 freq_count_stamp;
233 * Common implementation detail of pmu::{start,commit,cancel}_txn
235 #define PERF_PMU_TXN_ADD 0x1 /* txn to add/schedule event on PMU */
236 #define PERF_PMU_TXN_READ 0x2 /* txn to read event group from PMU */
239 * pmu::capabilities flags
241 #define PERF_PMU_CAP_NO_INTERRUPT 0x01
242 #define PERF_PMU_CAP_NO_NMI 0x02
243 #define PERF_PMU_CAP_AUX_NO_SG 0x04
244 #define PERF_PMU_CAP_EXCLUSIVE 0x10
245 #define PERF_PMU_CAP_ITRACE 0x20
246 #define PERF_PMU_CAP_HETEROGENEOUS_CPUS 0x40
247 #define PERF_PMU_CAP_NO_EXCLUDE 0x80
250 * struct pmu - generic performance monitoring unit
253 struct list_head entry;
255 struct module *module;
257 const struct attribute_group **attr_groups;
262 * various common per-pmu feature flags
266 int __percpu *pmu_disable_count;
267 struct perf_cpu_context __percpu *pmu_cpu_context;
268 atomic_t exclusive_cnt; /* < 0: cpu; > 0: tsk */
270 int hrtimer_interval_ms;
272 /* number of address filters this PMU can do */
273 unsigned int nr_addr_filters;
276 * Fully disable/enable this PMU, can be used to protect from the PMI
277 * as well as for lazy/batch writing of the MSRs.
279 void (*pmu_enable) (struct pmu *pmu); /* optional */
280 void (*pmu_disable) (struct pmu *pmu); /* optional */
283 * Try and initialize the event for this PMU.
286 * -ENOENT -- @event is not for this PMU
288 * -ENODEV -- @event is for this PMU but PMU not present
289 * -EBUSY -- @event is for this PMU but PMU temporarily unavailable
290 * -EINVAL -- @event is for this PMU but @event is not valid
291 * -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
292 * -EACCESS -- @event is for this PMU, @event is valid, but no privilidges
294 * 0 -- @event is for this PMU and valid
296 * Other error return values are allowed.
298 int (*event_init) (struct perf_event *event);
301 * Notification that the event was mapped or unmapped. Called
302 * in the context of the mapping task.
304 void (*event_mapped) (struct perf_event *event, struct mm_struct *mm); /* optional */
305 void (*event_unmapped) (struct perf_event *event, struct mm_struct *mm); /* optional */
308 * Flags for ->add()/->del()/ ->start()/->stop(). There are
309 * matching hw_perf_event::state flags.
311 #define PERF_EF_START 0x01 /* start the counter when adding */
312 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
313 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
316 * Adds/Removes a counter to/from the PMU, can be done inside a
317 * transaction, see the ->*_txn() methods.
319 * The add/del callbacks will reserve all hardware resources required
320 * to service the event, this includes any counter constraint
323 * Called with IRQs disabled and the PMU disabled on the CPU the event
326 * ->add() called without PERF_EF_START should result in the same state
327 * as ->add() followed by ->stop().
329 * ->del() must always PERF_EF_UPDATE stop an event. If it calls
330 * ->stop() that must deal with already being stopped without
333 int (*add) (struct perf_event *event, int flags);
334 void (*del) (struct perf_event *event, int flags);
337 * Starts/Stops a counter present on the PMU.
339 * The PMI handler should stop the counter when perf_event_overflow()
340 * returns !0. ->start() will be used to continue.
342 * Also used to change the sample period.
344 * Called with IRQs disabled and the PMU disabled on the CPU the event
345 * is on -- will be called from NMI context with the PMU generates
348 * ->stop() with PERF_EF_UPDATE will read the counter and update
349 * period/count values like ->read() would.
351 * ->start() with PERF_EF_RELOAD will reprogram the the counter
352 * value, must be preceded by a ->stop() with PERF_EF_UPDATE.
354 void (*start) (struct perf_event *event, int flags);
355 void (*stop) (struct perf_event *event, int flags);
358 * Updates the counter value of the event.
360 * For sampling capable PMUs this will also update the software period
361 * hw_perf_event::period_left field.
363 void (*read) (struct perf_event *event);
366 * Group events scheduling is treated as a transaction, add
367 * group events as a whole and perform one schedulability test.
368 * If the test fails, roll back the whole group
370 * Start the transaction, after this ->add() doesn't need to
371 * do schedulability tests.
375 void (*start_txn) (struct pmu *pmu, unsigned int txn_flags);
377 * If ->start_txn() disabled the ->add() schedulability test
378 * then ->commit_txn() is required to perform one. On success
379 * the transaction is closed. On error the transaction is kept
380 * open until ->cancel_txn() is called.
384 int (*commit_txn) (struct pmu *pmu);
386 * Will cancel the transaction, assumes ->del() is called
387 * for each successful ->add() during the transaction.
391 void (*cancel_txn) (struct pmu *pmu);
394 * Will return the value for perf_event_mmap_page::index for this event,
395 * if no implementation is provided it will default to: event->hw.idx + 1.
397 int (*event_idx) (struct perf_event *event); /*optional */
400 * context-switches callback
402 void (*sched_task) (struct perf_event_context *ctx,
405 * PMU specific data size
407 size_t task_ctx_size;
411 * Set up pmu-private data structures for an AUX area
413 void *(*setup_aux) (struct perf_event *event, void **pages,
414 int nr_pages, bool overwrite);
418 * Free pmu-private AUX data structures
420 void (*free_aux) (void *aux); /* optional */
423 * Validate address range filters: make sure the HW supports the
424 * requested configuration and number of filters; return 0 if the
425 * supplied filters are valid, -errno otherwise.
427 * Runs in the context of the ioctl()ing process and is not serialized
428 * with the rest of the PMU callbacks.
430 int (*addr_filters_validate) (struct list_head *filters);
434 * Synchronize address range filter configuration:
435 * translate hw-agnostic filters into hardware configuration in
436 * event::hw::addr_filters.
438 * Runs as a part of filter sync sequence that is done in ->start()
439 * callback by calling perf_event_addr_filters_sync().
441 * May (and should) traverse event::addr_filters::list, for which its
442 * caller provides necessary serialization.
444 void (*addr_filters_sync) (struct perf_event *event);
448 * Filter events for PMU-specific reasons.
450 int (*filter_match) (struct perf_event *event); /* optional */
453 * Check period value for PERF_EVENT_IOC_PERIOD ioctl.
455 int (*check_period) (struct perf_event *event, u64 value); /* optional */
458 enum perf_addr_filter_action_t {
459 PERF_ADDR_FILTER_ACTION_STOP = 0,
460 PERF_ADDR_FILTER_ACTION_START,
461 PERF_ADDR_FILTER_ACTION_FILTER,
465 * struct perf_addr_filter - address range filter definition
466 * @entry: event's filter list linkage
467 * @path: object file's path for file-based filters
468 * @offset: filter range offset
469 * @size: filter range size (size==0 means single address trigger)
470 * @action: filter/start/stop
472 * This is a hardware-agnostic filter configuration as specified by the user.
474 struct perf_addr_filter {
475 struct list_head entry;
477 unsigned long offset;
479 enum perf_addr_filter_action_t action;
483 * struct perf_addr_filters_head - container for address range filters
484 * @list: list of filters for this event
485 * @lock: spinlock that serializes accesses to the @list and event's
486 * (and its children's) filter generations.
487 * @nr_file_filters: number of file-based filters
489 * A child event will use parent's @list (and therefore @lock), so they are
490 * bundled together; see perf_event_addr_filters().
492 struct perf_addr_filters_head {
493 struct list_head list;
495 unsigned int nr_file_filters;
498 struct perf_addr_filter_range {
504 * enum perf_event_state - the states of an event:
506 enum perf_event_state {
507 PERF_EVENT_STATE_DEAD = -4,
508 PERF_EVENT_STATE_EXIT = -3,
509 PERF_EVENT_STATE_ERROR = -2,
510 PERF_EVENT_STATE_OFF = -1,
511 PERF_EVENT_STATE_INACTIVE = 0,
512 PERF_EVENT_STATE_ACTIVE = 1,
516 struct perf_sample_data;
518 typedef void (*perf_overflow_handler_t)(struct perf_event *,
519 struct perf_sample_data *,
520 struct pt_regs *regs);
523 * Event capabilities. For event_caps and groups caps.
525 * PERF_EV_CAP_SOFTWARE: Is a software event.
526 * PERF_EV_CAP_READ_ACTIVE_PKG: A CPU event (or cgroup event) that can be read
527 * from any CPU in the package where it is active.
529 #define PERF_EV_CAP_SOFTWARE BIT(0)
530 #define PERF_EV_CAP_READ_ACTIVE_PKG BIT(1)
532 #define SWEVENT_HLIST_BITS 8
533 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
535 struct swevent_hlist {
536 struct hlist_head heads[SWEVENT_HLIST_SIZE];
537 struct rcu_head rcu_head;
540 #define PERF_ATTACH_CONTEXT 0x01
541 #define PERF_ATTACH_GROUP 0x02
542 #define PERF_ATTACH_TASK 0x04
543 #define PERF_ATTACH_TASK_DATA 0x08
544 #define PERF_ATTACH_ITRACE 0x10
549 struct pmu_event_list {
551 struct list_head list;
554 #define for_each_sibling_event(sibling, event) \
555 if ((event)->group_leader == (event)) \
556 list_for_each_entry((sibling), &(event)->sibling_list, sibling_list)
559 * struct perf_event - performance event kernel representation:
562 #ifdef CONFIG_PERF_EVENTS
564 * entry onto perf_event_context::event_list;
565 * modifications require ctx->lock
566 * RCU safe iterations.
568 struct list_head event_entry;
571 * Locked for modification by both ctx->mutex and ctx->lock; holding
572 * either sufficies for read.
574 struct list_head sibling_list;
575 struct list_head active_list;
577 * Node on the pinned or flexible tree located at the event context;
579 struct rb_node group_node;
582 * We need storage to track the entries in perf_pmu_migrate_context; we
583 * cannot use the event_entry because of RCU and we want to keep the
584 * group in tact which avoids us using the other two entries.
586 struct list_head migrate_entry;
588 struct hlist_node hlist_entry;
589 struct list_head active_entry;
592 /* Not serialized. Only written during event initialization. */
594 /* The cumulative AND of all event_caps for events in this group. */
597 struct perf_event *group_leader;
601 enum perf_event_state state;
602 unsigned int attach_state;
604 atomic64_t child_count;
607 * These are the total time in nanoseconds that the event
608 * has been enabled (i.e. eligible to run, and the task has
609 * been scheduled in, if this is a per-task event)
610 * and running (scheduled onto the CPU), respectively.
612 u64 total_time_enabled;
613 u64 total_time_running;
617 * timestamp shadows the actual context timing but it can
618 * be safely used in NMI interrupt context. It reflects the
619 * context time as it was when the event was last scheduled in.
621 * ctx_time already accounts for ctx->timestamp. Therefore to
622 * compute ctx_time for a sample, simply add perf_clock().
626 struct perf_event_attr attr;
630 struct hw_perf_event hw;
632 struct perf_event_context *ctx;
633 atomic_long_t refcount;
636 * These accumulate total time (in nanoseconds) that children
637 * events have been enabled and running, respectively.
639 atomic64_t child_total_time_enabled;
640 atomic64_t child_total_time_running;
643 * Protect attach/detach and child_list:
645 struct mutex child_mutex;
646 struct list_head child_list;
647 struct perf_event *parent;
652 struct list_head owner_entry;
653 struct task_struct *owner;
656 struct mutex mmap_mutex;
659 struct ring_buffer *rb;
660 struct list_head rb_entry;
661 unsigned long rcu_batches;
665 wait_queue_head_t waitq;
666 struct fasync_struct *fasync;
668 /* delayed work for NMIs and such */
672 struct irq_work pending;
674 atomic_t event_limit;
676 /* address range filters */
677 struct perf_addr_filters_head addr_filters;
678 /* vma address array for file-based filders */
679 struct perf_addr_filter_range *addr_filter_ranges;
680 unsigned long addr_filters_gen;
682 void (*destroy)(struct perf_event *);
683 struct rcu_head rcu_head;
685 struct pid_namespace *ns;
689 perf_overflow_handler_t overflow_handler;
690 void *overflow_handler_context;
691 #ifdef CONFIG_BPF_SYSCALL
692 perf_overflow_handler_t orig_overflow_handler;
693 struct bpf_prog *prog;
696 #ifdef CONFIG_EVENT_TRACING
697 struct trace_event_call *tp_event;
698 struct event_filter *filter;
699 #ifdef CONFIG_FUNCTION_TRACER
700 struct ftrace_ops ftrace_ops;
704 #ifdef CONFIG_CGROUP_PERF
705 struct perf_cgroup *cgrp; /* cgroup event is attach to */
708 struct list_head sb_list;
709 #endif /* CONFIG_PERF_EVENTS */
713 struct perf_event_groups {
719 * struct perf_event_context - event context structure
721 * Used as a container for task events and CPU events as well:
723 struct perf_event_context {
726 * Protect the states of the events in the list,
727 * nr_active, and the list:
731 * Protect the list of events. Locking either mutex or lock
732 * is sufficient to ensure the list doesn't change; to change
733 * the list you need to lock both the mutex and the spinlock.
737 struct list_head active_ctx_list;
738 struct perf_event_groups pinned_groups;
739 struct perf_event_groups flexible_groups;
740 struct list_head event_list;
742 struct list_head pinned_active;
743 struct list_head flexible_active;
752 struct task_struct *task;
755 * Context clock, runs when context enabled.
761 * These fields let us detect when two contexts have both
762 * been cloned (inherited) from a common ancestor.
764 struct perf_event_context *parent_ctx;
768 #ifdef CONFIG_CGROUP_PERF
769 int nr_cgroups; /* cgroup evts */
771 void *task_ctx_data; /* pmu specific data */
772 struct rcu_head rcu_head;
776 * Number of contexts where an event can trigger:
777 * task, softirq, hardirq, nmi.
779 #define PERF_NR_CONTEXTS 4
782 * struct perf_event_cpu_context - per cpu event context structure
784 struct perf_cpu_context {
785 struct perf_event_context ctx;
786 struct perf_event_context *task_ctx;
790 raw_spinlock_t hrtimer_lock;
791 struct hrtimer hrtimer;
792 ktime_t hrtimer_interval;
793 unsigned int hrtimer_active;
795 #ifdef CONFIG_CGROUP_PERF
796 struct perf_cgroup *cgrp;
797 struct list_head cgrp_cpuctx_entry;
800 struct list_head sched_cb_entry;
806 struct perf_output_handle {
807 struct perf_event *event;
808 struct ring_buffer *rb;
809 unsigned long wakeup;
819 struct bpf_perf_event_data_kern {
820 bpf_user_pt_regs_t *regs;
821 struct perf_sample_data *data;
822 struct perf_event *event;
825 #ifdef CONFIG_CGROUP_PERF
828 * perf_cgroup_info keeps track of time_enabled for a cgroup.
829 * This is a per-cpu dynamically allocated data structure.
831 struct perf_cgroup_info {
837 struct cgroup_subsys_state css;
838 struct perf_cgroup_info __percpu *info;
842 * Must ensure cgroup is pinned (css_get) before calling
843 * this function. In other words, we cannot call this function
844 * if there is no cgroup event for the current CPU context.
846 static inline struct perf_cgroup *
847 perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
849 return container_of(task_css_check(task, perf_event_cgrp_id,
850 ctx ? lockdep_is_held(&ctx->lock)
852 struct perf_cgroup, css);
854 #endif /* CONFIG_CGROUP_PERF */
856 #ifdef CONFIG_PERF_EVENTS
858 extern void *perf_aux_output_begin(struct perf_output_handle *handle,
859 struct perf_event *event);
860 extern void perf_aux_output_end(struct perf_output_handle *handle,
862 extern int perf_aux_output_skip(struct perf_output_handle *handle,
864 extern void *perf_get_aux(struct perf_output_handle *handle);
865 extern void perf_aux_output_flag(struct perf_output_handle *handle, u64 flags);
866 extern void perf_event_itrace_started(struct perf_event *event);
868 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
869 extern void perf_pmu_unregister(struct pmu *pmu);
871 extern int perf_num_counters(void);
872 extern const char *perf_pmu_name(void);
873 extern void __perf_event_task_sched_in(struct task_struct *prev,
874 struct task_struct *task);
875 extern void __perf_event_task_sched_out(struct task_struct *prev,
876 struct task_struct *next);
877 extern int perf_event_init_task(struct task_struct *child);
878 extern void perf_event_exit_task(struct task_struct *child);
879 extern void perf_event_free_task(struct task_struct *task);
880 extern void perf_event_delayed_put(struct task_struct *task);
881 extern struct file *perf_event_get(unsigned int fd);
882 extern const struct perf_event *perf_get_event(struct file *file);
883 extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
884 extern void perf_event_print_debug(void);
885 extern void perf_pmu_disable(struct pmu *pmu);
886 extern void perf_pmu_enable(struct pmu *pmu);
887 extern void perf_sched_cb_dec(struct pmu *pmu);
888 extern void perf_sched_cb_inc(struct pmu *pmu);
889 extern int perf_event_task_disable(void);
890 extern int perf_event_task_enable(void);
892 extern void perf_pmu_resched(struct pmu *pmu);
894 extern int perf_event_refresh(struct perf_event *event, int refresh);
895 extern void perf_event_update_userpage(struct perf_event *event);
896 extern int perf_event_release_kernel(struct perf_event *event);
897 extern struct perf_event *
898 perf_event_create_kernel_counter(struct perf_event_attr *attr,
900 struct task_struct *task,
901 perf_overflow_handler_t callback,
903 extern void perf_pmu_migrate_context(struct pmu *pmu,
904 int src_cpu, int dst_cpu);
905 int perf_event_read_local(struct perf_event *event, u64 *value,
906 u64 *enabled, u64 *running);
907 extern u64 perf_event_read_value(struct perf_event *event,
908 u64 *enabled, u64 *running);
911 struct perf_sample_data {
913 * Fields set by perf_sample_data_init(), group so as to
914 * minimize the cachelines touched.
917 struct perf_raw_record *raw;
918 struct perf_branch_stack *br_stack;
922 union perf_mem_data_src data_src;
925 * The other fields, optionally {set,used} by
926 * perf_{prepare,output}_sample().
941 struct perf_callchain_entry *callchain;
944 * regs_user may point to task_pt_regs or to regs_user_copy, depending
947 struct perf_regs regs_user;
948 struct pt_regs regs_user_copy;
950 struct perf_regs regs_intr;
954 } ____cacheline_aligned;
956 /* default value for data source */
957 #define PERF_MEM_NA (PERF_MEM_S(OP, NA) |\
958 PERF_MEM_S(LVL, NA) |\
959 PERF_MEM_S(SNOOP, NA) |\
960 PERF_MEM_S(LOCK, NA) |\
963 static inline void perf_sample_data_init(struct perf_sample_data *data,
964 u64 addr, u64 period)
966 /* remaining struct members initialized in perf_prepare_sample() */
969 data->br_stack = NULL;
970 data->period = period;
972 data->data_src.val = PERF_MEM_NA;
976 extern void perf_output_sample(struct perf_output_handle *handle,
977 struct perf_event_header *header,
978 struct perf_sample_data *data,
979 struct perf_event *event);
980 extern void perf_prepare_sample(struct perf_event_header *header,
981 struct perf_sample_data *data,
982 struct perf_event *event,
983 struct pt_regs *regs);
985 extern int perf_event_overflow(struct perf_event *event,
986 struct perf_sample_data *data,
987 struct pt_regs *regs);
989 extern void perf_event_output_forward(struct perf_event *event,
990 struct perf_sample_data *data,
991 struct pt_regs *regs);
992 extern void perf_event_output_backward(struct perf_event *event,
993 struct perf_sample_data *data,
994 struct pt_regs *regs);
995 extern int perf_event_output(struct perf_event *event,
996 struct perf_sample_data *data,
997 struct pt_regs *regs);
1000 is_default_overflow_handler(struct perf_event *event)
1002 if (likely(event->overflow_handler == perf_event_output_forward))
1004 if (unlikely(event->overflow_handler == perf_event_output_backward))
1010 perf_event_header__init_id(struct perf_event_header *header,
1011 struct perf_sample_data *data,
1012 struct perf_event *event);
1014 perf_event__output_id_sample(struct perf_event *event,
1015 struct perf_output_handle *handle,
1016 struct perf_sample_data *sample);
1019 perf_log_lost_samples(struct perf_event *event, u64 lost);
1021 static inline bool event_has_any_exclude_flag(struct perf_event *event)
1023 struct perf_event_attr *attr = &event->attr;
1025 return attr->exclude_idle || attr->exclude_user ||
1026 attr->exclude_kernel || attr->exclude_hv ||
1027 attr->exclude_guest || attr->exclude_host;
1030 static inline bool is_sampling_event(struct perf_event *event)
1032 return event->attr.sample_period != 0;
1036 * Return 1 for a software event, 0 for a hardware event
1038 static inline int is_software_event(struct perf_event *event)
1040 return event->event_caps & PERF_EV_CAP_SOFTWARE;
1044 * Return 1 for event in sw context, 0 for event in hw context
1046 static inline int in_software_context(struct perf_event *event)
1048 return event->ctx->pmu->task_ctx_nr == perf_sw_context;
1051 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
1053 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
1054 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
1056 #ifndef perf_arch_fetch_caller_regs
1057 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
1061 * When generating a perf sample in-line, instead of from an interrupt /
1062 * exception, we lack a pt_regs. This is typically used from software events
1063 * like: SW_CONTEXT_SWITCHES, SW_MIGRATIONS and the tie-in with tracepoints.
1065 * We typically don't need a full set, but (for x86) do require:
1066 * - ip for PERF_SAMPLE_IP
1067 * - cs for user_mode() tests
1068 * - sp for PERF_SAMPLE_CALLCHAIN
1069 * - eflags for MISC bits and CALLCHAIN (see: perf_hw_regs())
1071 * NOTE: assumes @regs is otherwise already 0 filled; this is important for
1072 * things like PERF_SAMPLE_REGS_INTR.
1074 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1076 perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1079 static __always_inline void
1080 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1082 if (static_key_false(&perf_swevent_enabled[event_id]))
1083 __perf_sw_event(event_id, nr, regs, addr);
1086 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
1089 * 'Special' version for the scheduler, it hard assumes no recursion,
1090 * which is guaranteed by us not actually scheduling inside other swevents
1091 * because those disable preemption.
1093 static __always_inline void
1094 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
1096 if (static_key_false(&perf_swevent_enabled[event_id])) {
1097 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1099 perf_fetch_caller_regs(regs);
1100 ___perf_sw_event(event_id, nr, regs, addr);
1104 extern struct static_key_false perf_sched_events;
1106 static __always_inline bool
1107 perf_sw_migrate_enabled(void)
1109 if (static_key_false(&perf_swevent_enabled[PERF_COUNT_SW_CPU_MIGRATIONS]))
1114 static inline void perf_event_task_migrate(struct task_struct *task)
1116 if (perf_sw_migrate_enabled())
1117 task->sched_migrated = 1;
1120 static inline void perf_event_task_sched_in(struct task_struct *prev,
1121 struct task_struct *task)
1123 if (static_branch_unlikely(&perf_sched_events))
1124 __perf_event_task_sched_in(prev, task);
1126 if (perf_sw_migrate_enabled() && task->sched_migrated) {
1127 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1129 perf_fetch_caller_regs(regs);
1130 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, regs, 0);
1131 task->sched_migrated = 0;
1135 static inline void perf_event_task_sched_out(struct task_struct *prev,
1136 struct task_struct *next)
1138 perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
1140 if (static_branch_unlikely(&perf_sched_events))
1141 __perf_event_task_sched_out(prev, next);
1144 extern void perf_event_mmap(struct vm_area_struct *vma);
1146 extern void perf_event_ksymbol(u16 ksym_type, u64 addr, u32 len,
1147 bool unregister, const char *sym);
1148 extern void perf_event_bpf_event(struct bpf_prog *prog,
1149 enum perf_bpf_event_type type,
1152 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1153 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1154 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1156 extern void perf_event_exec(void);
1157 extern void perf_event_comm(struct task_struct *tsk, bool exec);
1158 extern void perf_event_namespaces(struct task_struct *tsk);
1159 extern void perf_event_fork(struct task_struct *tsk);
1162 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1164 extern void perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1165 extern void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1166 extern struct perf_callchain_entry *
1167 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
1168 u32 max_stack, bool crosstask, bool add_mark);
1169 extern struct perf_callchain_entry *perf_callchain(struct perf_event *event, struct pt_regs *regs);
1170 extern int get_callchain_buffers(int max_stack);
1171 extern void put_callchain_buffers(void);
1173 extern int sysctl_perf_event_max_stack;
1174 extern int sysctl_perf_event_max_contexts_per_stack;
1176 static inline int perf_callchain_store_context(struct perf_callchain_entry_ctx *ctx, u64 ip)
1178 if (ctx->contexts < sysctl_perf_event_max_contexts_per_stack) {
1179 struct perf_callchain_entry *entry = ctx->entry;
1180 entry->ip[entry->nr++] = ip;
1184 ctx->contexts_maxed = true;
1185 return -1; /* no more room, stop walking the stack */
1189 static inline int perf_callchain_store(struct perf_callchain_entry_ctx *ctx, u64 ip)
1191 if (ctx->nr < ctx->max_stack && !ctx->contexts_maxed) {
1192 struct perf_callchain_entry *entry = ctx->entry;
1193 entry->ip[entry->nr++] = ip;
1197 return -1; /* no more room, stop walking the stack */
1201 extern int sysctl_perf_event_paranoid;
1202 extern int sysctl_perf_event_mlock;
1203 extern int sysctl_perf_event_sample_rate;
1204 extern int sysctl_perf_cpu_time_max_percent;
1206 extern void perf_sample_event_took(u64 sample_len_ns);
1208 extern int perf_proc_update_handler(struct ctl_table *table, int write,
1209 void __user *buffer, size_t *lenp,
1211 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
1212 void __user *buffer, size_t *lenp,
1215 int perf_event_max_stack_handler(struct ctl_table *table, int write,
1216 void __user *buffer, size_t *lenp, loff_t *ppos);
1218 static inline bool perf_paranoid_tracepoint_raw(void)
1220 return sysctl_perf_event_paranoid > -1;
1223 static inline bool perf_paranoid_cpu(void)
1225 return sysctl_perf_event_paranoid > 0;
1228 static inline bool perf_paranoid_kernel(void)
1230 return sysctl_perf_event_paranoid > 1;
1233 extern void perf_event_init(void);
1234 extern void perf_tp_event(u16 event_type, u64 count, void *record,
1235 int entry_size, struct pt_regs *regs,
1236 struct hlist_head *head, int rctx,
1237 struct task_struct *task);
1238 extern void perf_bp_event(struct perf_event *event, void *data);
1240 #ifndef perf_misc_flags
1241 # define perf_misc_flags(regs) \
1242 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1243 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1245 #ifndef perf_arch_bpf_user_pt_regs
1246 # define perf_arch_bpf_user_pt_regs(regs) regs
1249 static inline bool has_branch_stack(struct perf_event *event)
1251 return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1254 static inline bool needs_branch_stack(struct perf_event *event)
1256 return event->attr.branch_sample_type != 0;
1259 static inline bool has_aux(struct perf_event *event)
1261 return event->pmu->setup_aux;
1264 static inline bool is_write_backward(struct perf_event *event)
1266 return !!event->attr.write_backward;
1269 static inline bool has_addr_filter(struct perf_event *event)
1271 return event->pmu->nr_addr_filters;
1275 * An inherited event uses parent's filters
1277 static inline struct perf_addr_filters_head *
1278 perf_event_addr_filters(struct perf_event *event)
1280 struct perf_addr_filters_head *ifh = &event->addr_filters;
1283 ifh = &event->parent->addr_filters;
1288 extern void perf_event_addr_filters_sync(struct perf_event *event);
1290 extern int perf_output_begin(struct perf_output_handle *handle,
1291 struct perf_event *event, unsigned int size);
1292 extern int perf_output_begin_forward(struct perf_output_handle *handle,
1293 struct perf_event *event,
1295 extern int perf_output_begin_backward(struct perf_output_handle *handle,
1296 struct perf_event *event,
1299 extern void perf_output_end(struct perf_output_handle *handle);
1300 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
1301 const void *buf, unsigned int len);
1302 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
1304 extern int perf_swevent_get_recursion_context(void);
1305 extern void perf_swevent_put_recursion_context(int rctx);
1306 extern u64 perf_swevent_set_period(struct perf_event *event);
1307 extern void perf_event_enable(struct perf_event *event);
1308 extern void perf_event_disable(struct perf_event *event);
1309 extern void perf_event_disable_local(struct perf_event *event);
1310 extern void perf_event_disable_inatomic(struct perf_event *event);
1311 extern void perf_event_task_tick(void);
1312 extern int perf_event_account_interrupt(struct perf_event *event);
1313 #else /* !CONFIG_PERF_EVENTS: */
1314 static inline void *
1315 perf_aux_output_begin(struct perf_output_handle *handle,
1316 struct perf_event *event) { return NULL; }
1318 perf_aux_output_end(struct perf_output_handle *handle, unsigned long size)
1321 perf_aux_output_skip(struct perf_output_handle *handle,
1322 unsigned long size) { return -EINVAL; }
1323 static inline void *
1324 perf_get_aux(struct perf_output_handle *handle) { return NULL; }
1326 perf_event_task_migrate(struct task_struct *task) { }
1328 perf_event_task_sched_in(struct task_struct *prev,
1329 struct task_struct *task) { }
1331 perf_event_task_sched_out(struct task_struct *prev,
1332 struct task_struct *next) { }
1333 static inline int perf_event_init_task(struct task_struct *child) { return 0; }
1334 static inline void perf_event_exit_task(struct task_struct *child) { }
1335 static inline void perf_event_free_task(struct task_struct *task) { }
1336 static inline void perf_event_delayed_put(struct task_struct *task) { }
1337 static inline struct file *perf_event_get(unsigned int fd) { return ERR_PTR(-EINVAL); }
1338 static inline const struct perf_event *perf_get_event(struct file *file)
1340 return ERR_PTR(-EINVAL);
1342 static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
1344 return ERR_PTR(-EINVAL);
1346 static inline int perf_event_read_local(struct perf_event *event, u64 *value,
1347 u64 *enabled, u64 *running)
1351 static inline void perf_event_print_debug(void) { }
1352 static inline int perf_event_task_disable(void) { return -EINVAL; }
1353 static inline int perf_event_task_enable(void) { return -EINVAL; }
1354 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1360 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) { }
1362 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr) { }
1364 perf_bp_event(struct perf_event *event, void *data) { }
1366 static inline int perf_register_guest_info_callbacks
1367 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1368 static inline int perf_unregister_guest_info_callbacks
1369 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1371 static inline void perf_event_mmap(struct vm_area_struct *vma) { }
1373 typedef int (perf_ksymbol_get_name_f)(char *name, int name_len, void *data);
1374 static inline void perf_event_ksymbol(u16 ksym_type, u64 addr, u32 len,
1375 bool unregister, const char *sym) { }
1376 static inline void perf_event_bpf_event(struct bpf_prog *prog,
1377 enum perf_bpf_event_type type,
1379 static inline void perf_event_exec(void) { }
1380 static inline void perf_event_comm(struct task_struct *tsk, bool exec) { }
1381 static inline void perf_event_namespaces(struct task_struct *tsk) { }
1382 static inline void perf_event_fork(struct task_struct *tsk) { }
1383 static inline void perf_event_init(void) { }
1384 static inline int perf_swevent_get_recursion_context(void) { return -1; }
1385 static inline void perf_swevent_put_recursion_context(int rctx) { }
1386 static inline u64 perf_swevent_set_period(struct perf_event *event) { return 0; }
1387 static inline void perf_event_enable(struct perf_event *event) { }
1388 static inline void perf_event_disable(struct perf_event *event) { }
1389 static inline int __perf_event_disable(void *info) { return -1; }
1390 static inline void perf_event_task_tick(void) { }
1391 static inline int perf_event_release_kernel(struct perf_event *event) { return 0; }
1394 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1395 extern void perf_restore_debug_store(void);
1397 static inline void perf_restore_debug_store(void) { }
1400 static __always_inline bool perf_raw_frag_last(const struct perf_raw_frag *frag)
1402 return frag->pad < sizeof(u64);
1405 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1407 struct perf_pmu_events_attr {
1408 struct device_attribute attr;
1410 const char *event_str;
1413 struct perf_pmu_events_ht_attr {
1414 struct device_attribute attr;
1416 const char *event_str_ht;
1417 const char *event_str_noht;
1420 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1423 #define PMU_EVENT_ATTR(_name, _var, _id, _show) \
1424 static struct perf_pmu_events_attr _var = { \
1425 .attr = __ATTR(_name, 0444, _show, NULL), \
1429 #define PMU_EVENT_ATTR_STRING(_name, _var, _str) \
1430 static struct perf_pmu_events_attr _var = { \
1431 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1433 .event_str = _str, \
1436 #define PMU_FORMAT_ATTR(_name, _format) \
1438 _name##_show(struct device *dev, \
1439 struct device_attribute *attr, \
1442 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
1443 return sprintf(page, _format "\n"); \
1446 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1448 /* Performance counter hotplug functions */
1449 #ifdef CONFIG_PERF_EVENTS
1450 int perf_event_init_cpu(unsigned int cpu);
1451 int perf_event_exit_cpu(unsigned int cpu);
1453 #define perf_event_init_cpu NULL
1454 #define perf_event_exit_cpu NULL
1457 #endif /* _LINUX_PERF_EVENT_H */