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_EXTENDED_REGS 0x08
245 #define PERF_PMU_CAP_EXCLUSIVE 0x10
246 #define PERF_PMU_CAP_ITRACE 0x20
247 #define PERF_PMU_CAP_HETEROGENEOUS_CPUS 0x40
248 #define PERF_PMU_CAP_NO_EXCLUDE 0x80
251 * struct pmu - generic performance monitoring unit
254 struct list_head entry;
256 struct module *module;
258 const struct attribute_group **attr_groups;
263 * various common per-pmu feature flags
267 int __percpu *pmu_disable_count;
268 struct perf_cpu_context __percpu *pmu_cpu_context;
269 atomic_t exclusive_cnt; /* < 0: cpu; > 0: tsk */
271 int hrtimer_interval_ms;
273 /* number of address filters this PMU can do */
274 unsigned int nr_addr_filters;
277 * Fully disable/enable this PMU, can be used to protect from the PMI
278 * as well as for lazy/batch writing of the MSRs.
280 void (*pmu_enable) (struct pmu *pmu); /* optional */
281 void (*pmu_disable) (struct pmu *pmu); /* optional */
284 * Try and initialize the event for this PMU.
287 * -ENOENT -- @event is not for this PMU
289 * -ENODEV -- @event is for this PMU but PMU not present
290 * -EBUSY -- @event is for this PMU but PMU temporarily unavailable
291 * -EINVAL -- @event is for this PMU but @event is not valid
292 * -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
293 * -EACCESS -- @event is for this PMU, @event is valid, but no privilidges
295 * 0 -- @event is for this PMU and valid
297 * Other error return values are allowed.
299 int (*event_init) (struct perf_event *event);
302 * Notification that the event was mapped or unmapped. Called
303 * in the context of the mapping task.
305 void (*event_mapped) (struct perf_event *event, struct mm_struct *mm); /* optional */
306 void (*event_unmapped) (struct perf_event *event, struct mm_struct *mm); /* optional */
309 * Flags for ->add()/->del()/ ->start()/->stop(). There are
310 * matching hw_perf_event::state flags.
312 #define PERF_EF_START 0x01 /* start the counter when adding */
313 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
314 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
317 * Adds/Removes a counter to/from the PMU, can be done inside a
318 * transaction, see the ->*_txn() methods.
320 * The add/del callbacks will reserve all hardware resources required
321 * to service the event, this includes any counter constraint
324 * Called with IRQs disabled and the PMU disabled on the CPU the event
327 * ->add() called without PERF_EF_START should result in the same state
328 * as ->add() followed by ->stop().
330 * ->del() must always PERF_EF_UPDATE stop an event. If it calls
331 * ->stop() that must deal with already being stopped without
334 int (*add) (struct perf_event *event, int flags);
335 void (*del) (struct perf_event *event, int flags);
338 * Starts/Stops a counter present on the PMU.
340 * The PMI handler should stop the counter when perf_event_overflow()
341 * returns !0. ->start() will be used to continue.
343 * Also used to change the sample period.
345 * Called with IRQs disabled and the PMU disabled on the CPU the event
346 * is on -- will be called from NMI context with the PMU generates
349 * ->stop() with PERF_EF_UPDATE will read the counter and update
350 * period/count values like ->read() would.
352 * ->start() with PERF_EF_RELOAD will reprogram the the counter
353 * value, must be preceded by a ->stop() with PERF_EF_UPDATE.
355 void (*start) (struct perf_event *event, int flags);
356 void (*stop) (struct perf_event *event, int flags);
359 * Updates the counter value of the event.
361 * For sampling capable PMUs this will also update the software period
362 * hw_perf_event::period_left field.
364 void (*read) (struct perf_event *event);
367 * Group events scheduling is treated as a transaction, add
368 * group events as a whole and perform one schedulability test.
369 * If the test fails, roll back the whole group
371 * Start the transaction, after this ->add() doesn't need to
372 * do schedulability tests.
376 void (*start_txn) (struct pmu *pmu, unsigned int txn_flags);
378 * If ->start_txn() disabled the ->add() schedulability test
379 * then ->commit_txn() is required to perform one. On success
380 * the transaction is closed. On error the transaction is kept
381 * open until ->cancel_txn() is called.
385 int (*commit_txn) (struct pmu *pmu);
387 * Will cancel the transaction, assumes ->del() is called
388 * for each successful ->add() during the transaction.
392 void (*cancel_txn) (struct pmu *pmu);
395 * Will return the value for perf_event_mmap_page::index for this event,
396 * if no implementation is provided it will default to: event->hw.idx + 1.
398 int (*event_idx) (struct perf_event *event); /*optional */
401 * context-switches callback
403 void (*sched_task) (struct perf_event_context *ctx,
406 * PMU specific data size
408 size_t task_ctx_size;
412 * Set up pmu-private data structures for an AUX area
414 void *(*setup_aux) (struct perf_event *event, void **pages,
415 int nr_pages, bool overwrite);
419 * Free pmu-private AUX data structures
421 void (*free_aux) (void *aux); /* optional */
424 * Validate address range filters: make sure the HW supports the
425 * requested configuration and number of filters; return 0 if the
426 * supplied filters are valid, -errno otherwise.
428 * Runs in the context of the ioctl()ing process and is not serialized
429 * with the rest of the PMU callbacks.
431 int (*addr_filters_validate) (struct list_head *filters);
435 * Synchronize address range filter configuration:
436 * translate hw-agnostic filters into hardware configuration in
437 * event::hw::addr_filters.
439 * Runs as a part of filter sync sequence that is done in ->start()
440 * callback by calling perf_event_addr_filters_sync().
442 * May (and should) traverse event::addr_filters::list, for which its
443 * caller provides necessary serialization.
445 void (*addr_filters_sync) (struct perf_event *event);
449 * Filter events for PMU-specific reasons.
451 int (*filter_match) (struct perf_event *event); /* optional */
454 * Check period value for PERF_EVENT_IOC_PERIOD ioctl.
456 int (*check_period) (struct perf_event *event, u64 value); /* optional */
459 enum perf_addr_filter_action_t {
460 PERF_ADDR_FILTER_ACTION_STOP = 0,
461 PERF_ADDR_FILTER_ACTION_START,
462 PERF_ADDR_FILTER_ACTION_FILTER,
466 * struct perf_addr_filter - address range filter definition
467 * @entry: event's filter list linkage
468 * @path: object file's path for file-based filters
469 * @offset: filter range offset
470 * @size: filter range size (size==0 means single address trigger)
471 * @action: filter/start/stop
473 * This is a hardware-agnostic filter configuration as specified by the user.
475 struct perf_addr_filter {
476 struct list_head entry;
478 unsigned long offset;
480 enum perf_addr_filter_action_t action;
484 * struct perf_addr_filters_head - container for address range filters
485 * @list: list of filters for this event
486 * @lock: spinlock that serializes accesses to the @list and event's
487 * (and its children's) filter generations.
488 * @nr_file_filters: number of file-based filters
490 * A child event will use parent's @list (and therefore @lock), so they are
491 * bundled together; see perf_event_addr_filters().
493 struct perf_addr_filters_head {
494 struct list_head list;
496 unsigned int nr_file_filters;
499 struct perf_addr_filter_range {
505 * enum perf_event_state - the states of an event:
507 enum perf_event_state {
508 PERF_EVENT_STATE_DEAD = -4,
509 PERF_EVENT_STATE_EXIT = -3,
510 PERF_EVENT_STATE_ERROR = -2,
511 PERF_EVENT_STATE_OFF = -1,
512 PERF_EVENT_STATE_INACTIVE = 0,
513 PERF_EVENT_STATE_ACTIVE = 1,
517 struct perf_sample_data;
519 typedef void (*perf_overflow_handler_t)(struct perf_event *,
520 struct perf_sample_data *,
521 struct pt_regs *regs);
524 * Event capabilities. For event_caps and groups caps.
526 * PERF_EV_CAP_SOFTWARE: Is a software event.
527 * PERF_EV_CAP_READ_ACTIVE_PKG: A CPU event (or cgroup event) that can be read
528 * from any CPU in the package where it is active.
530 #define PERF_EV_CAP_SOFTWARE BIT(0)
531 #define PERF_EV_CAP_READ_ACTIVE_PKG BIT(1)
533 #define SWEVENT_HLIST_BITS 8
534 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
536 struct swevent_hlist {
537 struct hlist_head heads[SWEVENT_HLIST_SIZE];
538 struct rcu_head rcu_head;
541 #define PERF_ATTACH_CONTEXT 0x01
542 #define PERF_ATTACH_GROUP 0x02
543 #define PERF_ATTACH_TASK 0x04
544 #define PERF_ATTACH_TASK_DATA 0x08
545 #define PERF_ATTACH_ITRACE 0x10
550 struct pmu_event_list {
552 struct list_head list;
555 #define for_each_sibling_event(sibling, event) \
556 if ((event)->group_leader == (event)) \
557 list_for_each_entry((sibling), &(event)->sibling_list, sibling_list)
560 * struct perf_event - performance event kernel representation:
563 #ifdef CONFIG_PERF_EVENTS
565 * entry onto perf_event_context::event_list;
566 * modifications require ctx->lock
567 * RCU safe iterations.
569 struct list_head event_entry;
572 * Locked for modification by both ctx->mutex and ctx->lock; holding
573 * either sufficies for read.
575 struct list_head sibling_list;
576 struct list_head active_list;
578 * Node on the pinned or flexible tree located at the event context;
580 struct rb_node group_node;
583 * We need storage to track the entries in perf_pmu_migrate_context; we
584 * cannot use the event_entry because of RCU and we want to keep the
585 * group in tact which avoids us using the other two entries.
587 struct list_head migrate_entry;
589 struct hlist_node hlist_entry;
590 struct list_head active_entry;
593 /* Not serialized. Only written during event initialization. */
595 /* The cumulative AND of all event_caps for events in this group. */
598 struct perf_event *group_leader;
602 enum perf_event_state state;
603 unsigned int attach_state;
605 atomic64_t child_count;
608 * These are the total time in nanoseconds that the event
609 * has been enabled (i.e. eligible to run, and the task has
610 * been scheduled in, if this is a per-task event)
611 * and running (scheduled onto the CPU), respectively.
613 u64 total_time_enabled;
614 u64 total_time_running;
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 struct perf_addr_filter_range *addr_filter_ranges;
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 */
709 struct list_head sb_list;
710 #endif /* CONFIG_PERF_EVENTS */
714 struct perf_event_groups {
720 * struct perf_event_context - event context structure
722 * Used as a container for task events and CPU events as well:
724 struct perf_event_context {
727 * Protect the states of the events in the list,
728 * nr_active, and the list:
732 * Protect the list of events. Locking either mutex or lock
733 * is sufficient to ensure the list doesn't change; to change
734 * the list you need to lock both the mutex and the spinlock.
738 struct list_head active_ctx_list;
739 struct perf_event_groups pinned_groups;
740 struct perf_event_groups flexible_groups;
741 struct list_head event_list;
743 struct list_head pinned_active;
744 struct list_head flexible_active;
753 struct task_struct *task;
756 * Context clock, runs when context enabled.
762 * These fields let us detect when two contexts have both
763 * been cloned (inherited) from a common ancestor.
765 struct perf_event_context *parent_ctx;
769 #ifdef CONFIG_CGROUP_PERF
770 int nr_cgroups; /* cgroup evts */
772 void *task_ctx_data; /* pmu specific data */
773 struct rcu_head rcu_head;
777 * Number of contexts where an event can trigger:
778 * task, softirq, hardirq, nmi.
780 #define PERF_NR_CONTEXTS 4
783 * struct perf_event_cpu_context - per cpu event context structure
785 struct perf_cpu_context {
786 struct perf_event_context ctx;
787 struct perf_event_context *task_ctx;
791 raw_spinlock_t hrtimer_lock;
792 struct hrtimer hrtimer;
793 ktime_t hrtimer_interval;
794 unsigned int hrtimer_active;
796 #ifdef CONFIG_CGROUP_PERF
797 struct perf_cgroup *cgrp;
798 struct list_head cgrp_cpuctx_entry;
801 struct list_head sched_cb_entry;
807 struct perf_output_handle {
808 struct perf_event *event;
809 struct ring_buffer *rb;
810 unsigned long wakeup;
820 struct bpf_perf_event_data_kern {
821 bpf_user_pt_regs_t *regs;
822 struct perf_sample_data *data;
823 struct perf_event *event;
826 #ifdef CONFIG_CGROUP_PERF
829 * perf_cgroup_info keeps track of time_enabled for a cgroup.
830 * This is a per-cpu dynamically allocated data structure.
832 struct perf_cgroup_info {
838 struct cgroup_subsys_state css;
839 struct perf_cgroup_info __percpu *info;
843 * Must ensure cgroup is pinned (css_get) before calling
844 * this function. In other words, we cannot call this function
845 * if there is no cgroup event for the current CPU context.
847 static inline struct perf_cgroup *
848 perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
850 return container_of(task_css_check(task, perf_event_cgrp_id,
851 ctx ? lockdep_is_held(&ctx->lock)
853 struct perf_cgroup, css);
855 #endif /* CONFIG_CGROUP_PERF */
857 #ifdef CONFIG_PERF_EVENTS
859 extern void *perf_aux_output_begin(struct perf_output_handle *handle,
860 struct perf_event *event);
861 extern void perf_aux_output_end(struct perf_output_handle *handle,
863 extern int perf_aux_output_skip(struct perf_output_handle *handle,
865 extern void *perf_get_aux(struct perf_output_handle *handle);
866 extern void perf_aux_output_flag(struct perf_output_handle *handle, u64 flags);
867 extern void perf_event_itrace_started(struct perf_event *event);
869 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
870 extern void perf_pmu_unregister(struct pmu *pmu);
872 extern int perf_num_counters(void);
873 extern const char *perf_pmu_name(void);
874 extern void __perf_event_task_sched_in(struct task_struct *prev,
875 struct task_struct *task);
876 extern void __perf_event_task_sched_out(struct task_struct *prev,
877 struct task_struct *next);
878 extern int perf_event_init_task(struct task_struct *child);
879 extern void perf_event_exit_task(struct task_struct *child);
880 extern void perf_event_free_task(struct task_struct *task);
881 extern void perf_event_delayed_put(struct task_struct *task);
882 extern struct file *perf_event_get(unsigned int fd);
883 extern const struct perf_event *perf_get_event(struct file *file);
884 extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
885 extern void perf_event_print_debug(void);
886 extern void perf_pmu_disable(struct pmu *pmu);
887 extern void perf_pmu_enable(struct pmu *pmu);
888 extern void perf_sched_cb_dec(struct pmu *pmu);
889 extern void perf_sched_cb_inc(struct pmu *pmu);
890 extern int perf_event_task_disable(void);
891 extern int perf_event_task_enable(void);
893 extern void perf_pmu_resched(struct pmu *pmu);
895 extern int perf_event_refresh(struct perf_event *event, int refresh);
896 extern void perf_event_update_userpage(struct perf_event *event);
897 extern int perf_event_release_kernel(struct perf_event *event);
898 extern struct perf_event *
899 perf_event_create_kernel_counter(struct perf_event_attr *attr,
901 struct task_struct *task,
902 perf_overflow_handler_t callback,
904 extern void perf_pmu_migrate_context(struct pmu *pmu,
905 int src_cpu, int dst_cpu);
906 int perf_event_read_local(struct perf_event *event, u64 *value,
907 u64 *enabled, u64 *running);
908 extern u64 perf_event_read_value(struct perf_event *event,
909 u64 *enabled, u64 *running);
912 struct perf_sample_data {
914 * Fields set by perf_sample_data_init(), group so as to
915 * minimize the cachelines touched.
918 struct perf_raw_record *raw;
919 struct perf_branch_stack *br_stack;
923 union perf_mem_data_src data_src;
926 * The other fields, optionally {set,used} by
927 * perf_{prepare,output}_sample().
942 struct perf_callchain_entry *callchain;
945 * regs_user may point to task_pt_regs or to regs_user_copy, depending
948 struct perf_regs regs_user;
949 struct pt_regs regs_user_copy;
951 struct perf_regs regs_intr;
955 } ____cacheline_aligned;
957 /* default value for data source */
958 #define PERF_MEM_NA (PERF_MEM_S(OP, NA) |\
959 PERF_MEM_S(LVL, NA) |\
960 PERF_MEM_S(SNOOP, NA) |\
961 PERF_MEM_S(LOCK, NA) |\
964 static inline void perf_sample_data_init(struct perf_sample_data *data,
965 u64 addr, u64 period)
967 /* remaining struct members initialized in perf_prepare_sample() */
970 data->br_stack = NULL;
971 data->period = period;
973 data->data_src.val = PERF_MEM_NA;
977 extern void perf_output_sample(struct perf_output_handle *handle,
978 struct perf_event_header *header,
979 struct perf_sample_data *data,
980 struct perf_event *event);
981 extern void perf_prepare_sample(struct perf_event_header *header,
982 struct perf_sample_data *data,
983 struct perf_event *event,
984 struct pt_regs *regs);
986 extern int perf_event_overflow(struct perf_event *event,
987 struct perf_sample_data *data,
988 struct pt_regs *regs);
990 extern void perf_event_output_forward(struct perf_event *event,
991 struct perf_sample_data *data,
992 struct pt_regs *regs);
993 extern void perf_event_output_backward(struct perf_event *event,
994 struct perf_sample_data *data,
995 struct pt_regs *regs);
996 extern int perf_event_output(struct perf_event *event,
997 struct perf_sample_data *data,
998 struct pt_regs *regs);
1001 is_default_overflow_handler(struct perf_event *event)
1003 if (likely(event->overflow_handler == perf_event_output_forward))
1005 if (unlikely(event->overflow_handler == perf_event_output_backward))
1011 perf_event_header__init_id(struct perf_event_header *header,
1012 struct perf_sample_data *data,
1013 struct perf_event *event);
1015 perf_event__output_id_sample(struct perf_event *event,
1016 struct perf_output_handle *handle,
1017 struct perf_sample_data *sample);
1020 perf_log_lost_samples(struct perf_event *event, u64 lost);
1022 static inline bool event_has_any_exclude_flag(struct perf_event *event)
1024 struct perf_event_attr *attr = &event->attr;
1026 return attr->exclude_idle || attr->exclude_user ||
1027 attr->exclude_kernel || attr->exclude_hv ||
1028 attr->exclude_guest || attr->exclude_host;
1031 static inline bool is_sampling_event(struct perf_event *event)
1033 return event->attr.sample_period != 0;
1037 * Return 1 for a software event, 0 for a hardware event
1039 static inline int is_software_event(struct perf_event *event)
1041 return event->event_caps & PERF_EV_CAP_SOFTWARE;
1045 * Return 1 for event in sw context, 0 for event in hw context
1047 static inline int in_software_context(struct perf_event *event)
1049 return event->ctx->pmu->task_ctx_nr == perf_sw_context;
1052 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
1054 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
1055 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
1057 #ifndef perf_arch_fetch_caller_regs
1058 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
1062 * When generating a perf sample in-line, instead of from an interrupt /
1063 * exception, we lack a pt_regs. This is typically used from software events
1064 * like: SW_CONTEXT_SWITCHES, SW_MIGRATIONS and the tie-in with tracepoints.
1066 * We typically don't need a full set, but (for x86) do require:
1067 * - ip for PERF_SAMPLE_IP
1068 * - cs for user_mode() tests
1069 * - sp for PERF_SAMPLE_CALLCHAIN
1070 * - eflags for MISC bits and CALLCHAIN (see: perf_hw_regs())
1072 * NOTE: assumes @regs is otherwise already 0 filled; this is important for
1073 * things like PERF_SAMPLE_REGS_INTR.
1075 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1077 perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1080 static __always_inline void
1081 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1083 if (static_key_false(&perf_swevent_enabled[event_id]))
1084 __perf_sw_event(event_id, nr, regs, addr);
1087 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
1090 * 'Special' version for the scheduler, it hard assumes no recursion,
1091 * which is guaranteed by us not actually scheduling inside other swevents
1092 * because those disable preemption.
1094 static __always_inline void
1095 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
1097 if (static_key_false(&perf_swevent_enabled[event_id])) {
1098 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1100 perf_fetch_caller_regs(regs);
1101 ___perf_sw_event(event_id, nr, regs, addr);
1105 extern struct static_key_false perf_sched_events;
1107 static __always_inline bool
1108 perf_sw_migrate_enabled(void)
1110 if (static_key_false(&perf_swevent_enabled[PERF_COUNT_SW_CPU_MIGRATIONS]))
1115 static inline void perf_event_task_migrate(struct task_struct *task)
1117 if (perf_sw_migrate_enabled())
1118 task->sched_migrated = 1;
1121 static inline void perf_event_task_sched_in(struct task_struct *prev,
1122 struct task_struct *task)
1124 if (static_branch_unlikely(&perf_sched_events))
1125 __perf_event_task_sched_in(prev, task);
1127 if (perf_sw_migrate_enabled() && task->sched_migrated) {
1128 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1130 perf_fetch_caller_regs(regs);
1131 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, regs, 0);
1132 task->sched_migrated = 0;
1136 static inline void perf_event_task_sched_out(struct task_struct *prev,
1137 struct task_struct *next)
1139 perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
1141 if (static_branch_unlikely(&perf_sched_events))
1142 __perf_event_task_sched_out(prev, next);
1145 extern void perf_event_mmap(struct vm_area_struct *vma);
1147 extern void perf_event_ksymbol(u16 ksym_type, u64 addr, u32 len,
1148 bool unregister, const char *sym);
1149 extern void perf_event_bpf_event(struct bpf_prog *prog,
1150 enum perf_bpf_event_type type,
1153 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1154 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1155 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1157 extern void perf_event_exec(void);
1158 extern void perf_event_comm(struct task_struct *tsk, bool exec);
1159 extern void perf_event_namespaces(struct task_struct *tsk);
1160 extern void perf_event_fork(struct task_struct *tsk);
1163 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1165 extern void perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1166 extern void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1167 extern struct perf_callchain_entry *
1168 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
1169 u32 max_stack, bool crosstask, bool add_mark);
1170 extern struct perf_callchain_entry *perf_callchain(struct perf_event *event, struct pt_regs *regs);
1171 extern int get_callchain_buffers(int max_stack);
1172 extern void put_callchain_buffers(void);
1174 extern int sysctl_perf_event_max_stack;
1175 extern int sysctl_perf_event_max_contexts_per_stack;
1177 static inline int perf_callchain_store_context(struct perf_callchain_entry_ctx *ctx, u64 ip)
1179 if (ctx->contexts < sysctl_perf_event_max_contexts_per_stack) {
1180 struct perf_callchain_entry *entry = ctx->entry;
1181 entry->ip[entry->nr++] = ip;
1185 ctx->contexts_maxed = true;
1186 return -1; /* no more room, stop walking the stack */
1190 static inline int perf_callchain_store(struct perf_callchain_entry_ctx *ctx, u64 ip)
1192 if (ctx->nr < ctx->max_stack && !ctx->contexts_maxed) {
1193 struct perf_callchain_entry *entry = ctx->entry;
1194 entry->ip[entry->nr++] = ip;
1198 return -1; /* no more room, stop walking the stack */
1202 extern int sysctl_perf_event_paranoid;
1203 extern int sysctl_perf_event_mlock;
1204 extern int sysctl_perf_event_sample_rate;
1205 extern int sysctl_perf_cpu_time_max_percent;
1207 extern void perf_sample_event_took(u64 sample_len_ns);
1209 extern int perf_proc_update_handler(struct ctl_table *table, int write,
1210 void __user *buffer, size_t *lenp,
1212 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
1213 void __user *buffer, size_t *lenp,
1216 int perf_event_max_stack_handler(struct ctl_table *table, int write,
1217 void __user *buffer, size_t *lenp, loff_t *ppos);
1219 static inline bool perf_paranoid_tracepoint_raw(void)
1221 return sysctl_perf_event_paranoid > -1;
1224 static inline bool perf_paranoid_cpu(void)
1226 return sysctl_perf_event_paranoid > 0;
1229 static inline bool perf_paranoid_kernel(void)
1231 return sysctl_perf_event_paranoid > 1;
1234 extern void perf_event_init(void);
1235 extern void perf_tp_event(u16 event_type, u64 count, void *record,
1236 int entry_size, struct pt_regs *regs,
1237 struct hlist_head *head, int rctx,
1238 struct task_struct *task);
1239 extern void perf_bp_event(struct perf_event *event, void *data);
1241 #ifndef perf_misc_flags
1242 # define perf_misc_flags(regs) \
1243 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1244 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1246 #ifndef perf_arch_bpf_user_pt_regs
1247 # define perf_arch_bpf_user_pt_regs(regs) regs
1250 static inline bool has_branch_stack(struct perf_event *event)
1252 return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1255 static inline bool needs_branch_stack(struct perf_event *event)
1257 return event->attr.branch_sample_type != 0;
1260 static inline bool has_aux(struct perf_event *event)
1262 return event->pmu->setup_aux;
1265 static inline bool is_write_backward(struct perf_event *event)
1267 return !!event->attr.write_backward;
1270 static inline bool has_addr_filter(struct perf_event *event)
1272 return event->pmu->nr_addr_filters;
1276 * An inherited event uses parent's filters
1278 static inline struct perf_addr_filters_head *
1279 perf_event_addr_filters(struct perf_event *event)
1281 struct perf_addr_filters_head *ifh = &event->addr_filters;
1284 ifh = &event->parent->addr_filters;
1289 extern void perf_event_addr_filters_sync(struct perf_event *event);
1291 extern int perf_output_begin(struct perf_output_handle *handle,
1292 struct perf_event *event, unsigned int size);
1293 extern int perf_output_begin_forward(struct perf_output_handle *handle,
1294 struct perf_event *event,
1296 extern int perf_output_begin_backward(struct perf_output_handle *handle,
1297 struct perf_event *event,
1300 extern void perf_output_end(struct perf_output_handle *handle);
1301 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
1302 const void *buf, unsigned int len);
1303 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
1305 extern int perf_swevent_get_recursion_context(void);
1306 extern void perf_swevent_put_recursion_context(int rctx);
1307 extern u64 perf_swevent_set_period(struct perf_event *event);
1308 extern void perf_event_enable(struct perf_event *event);
1309 extern void perf_event_disable(struct perf_event *event);
1310 extern void perf_event_disable_local(struct perf_event *event);
1311 extern void perf_event_disable_inatomic(struct perf_event *event);
1312 extern void perf_event_task_tick(void);
1313 extern int perf_event_account_interrupt(struct perf_event *event);
1314 #else /* !CONFIG_PERF_EVENTS: */
1315 static inline void *
1316 perf_aux_output_begin(struct perf_output_handle *handle,
1317 struct perf_event *event) { return NULL; }
1319 perf_aux_output_end(struct perf_output_handle *handle, unsigned long size)
1322 perf_aux_output_skip(struct perf_output_handle *handle,
1323 unsigned long size) { return -EINVAL; }
1324 static inline void *
1325 perf_get_aux(struct perf_output_handle *handle) { return NULL; }
1327 perf_event_task_migrate(struct task_struct *task) { }
1329 perf_event_task_sched_in(struct task_struct *prev,
1330 struct task_struct *task) { }
1332 perf_event_task_sched_out(struct task_struct *prev,
1333 struct task_struct *next) { }
1334 static inline int perf_event_init_task(struct task_struct *child) { return 0; }
1335 static inline void perf_event_exit_task(struct task_struct *child) { }
1336 static inline void perf_event_free_task(struct task_struct *task) { }
1337 static inline void perf_event_delayed_put(struct task_struct *task) { }
1338 static inline struct file *perf_event_get(unsigned int fd) { return ERR_PTR(-EINVAL); }
1339 static inline const struct perf_event *perf_get_event(struct file *file)
1341 return ERR_PTR(-EINVAL);
1343 static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
1345 return ERR_PTR(-EINVAL);
1347 static inline int perf_event_read_local(struct perf_event *event, u64 *value,
1348 u64 *enabled, u64 *running)
1352 static inline void perf_event_print_debug(void) { }
1353 static inline int perf_event_task_disable(void) { return -EINVAL; }
1354 static inline int perf_event_task_enable(void) { return -EINVAL; }
1355 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1361 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) { }
1363 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr) { }
1365 perf_bp_event(struct perf_event *event, void *data) { }
1367 static inline int perf_register_guest_info_callbacks
1368 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1369 static inline int perf_unregister_guest_info_callbacks
1370 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1372 static inline void perf_event_mmap(struct vm_area_struct *vma) { }
1374 typedef int (perf_ksymbol_get_name_f)(char *name, int name_len, void *data);
1375 static inline void perf_event_ksymbol(u16 ksym_type, u64 addr, u32 len,
1376 bool unregister, const char *sym) { }
1377 static inline void perf_event_bpf_event(struct bpf_prog *prog,
1378 enum perf_bpf_event_type type,
1380 static inline void perf_event_exec(void) { }
1381 static inline void perf_event_comm(struct task_struct *tsk, bool exec) { }
1382 static inline void perf_event_namespaces(struct task_struct *tsk) { }
1383 static inline void perf_event_fork(struct task_struct *tsk) { }
1384 static inline void perf_event_init(void) { }
1385 static inline int perf_swevent_get_recursion_context(void) { return -1; }
1386 static inline void perf_swevent_put_recursion_context(int rctx) { }
1387 static inline u64 perf_swevent_set_period(struct perf_event *event) { return 0; }
1388 static inline void perf_event_enable(struct perf_event *event) { }
1389 static inline void perf_event_disable(struct perf_event *event) { }
1390 static inline int __perf_event_disable(void *info) { return -1; }
1391 static inline void perf_event_task_tick(void) { }
1392 static inline int perf_event_release_kernel(struct perf_event *event) { return 0; }
1395 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1396 extern void perf_restore_debug_store(void);
1398 static inline void perf_restore_debug_store(void) { }
1401 static __always_inline bool perf_raw_frag_last(const struct perf_raw_frag *frag)
1403 return frag->pad < sizeof(u64);
1406 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1408 struct perf_pmu_events_attr {
1409 struct device_attribute attr;
1411 const char *event_str;
1414 struct perf_pmu_events_ht_attr {
1415 struct device_attribute attr;
1417 const char *event_str_ht;
1418 const char *event_str_noht;
1421 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1424 #define PMU_EVENT_ATTR(_name, _var, _id, _show) \
1425 static struct perf_pmu_events_attr _var = { \
1426 .attr = __ATTR(_name, 0444, _show, NULL), \
1430 #define PMU_EVENT_ATTR_STRING(_name, _var, _str) \
1431 static struct perf_pmu_events_attr _var = { \
1432 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1434 .event_str = _str, \
1437 #define PMU_FORMAT_ATTR(_name, _format) \
1439 _name##_show(struct device *dev, \
1440 struct device_attribute *attr, \
1443 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
1444 return sprintf(page, _format "\n"); \
1447 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1449 /* Performance counter hotplug functions */
1450 #ifdef CONFIG_PERF_EVENTS
1451 int perf_event_init_cpu(unsigned int cpu);
1452 int perf_event_exit_cpu(unsigned int cpu);
1454 #define perf_event_init_cpu NULL
1455 #define perf_event_exit_cpu NULL
1458 #endif /* _LINUX_PERF_EVENT_H */