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;
259 const struct attribute_group **attr_update;
264 * various common per-pmu feature flags
268 int __percpu *pmu_disable_count;
269 struct perf_cpu_context __percpu *pmu_cpu_context;
270 atomic_t exclusive_cnt; /* < 0: cpu; > 0: tsk */
272 int hrtimer_interval_ms;
274 /* number of address filters this PMU can do */
275 unsigned int nr_addr_filters;
278 * Fully disable/enable this PMU, can be used to protect from the PMI
279 * as well as for lazy/batch writing of the MSRs.
281 void (*pmu_enable) (struct pmu *pmu); /* optional */
282 void (*pmu_disable) (struct pmu *pmu); /* optional */
285 * Try and initialize the event for this PMU.
288 * -ENOENT -- @event is not for this PMU
290 * -ENODEV -- @event is for this PMU but PMU not present
291 * -EBUSY -- @event is for this PMU but PMU temporarily unavailable
292 * -EINVAL -- @event is for this PMU but @event is not valid
293 * -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
294 * -EACCESS -- @event is for this PMU, @event is valid, but no privilidges
296 * 0 -- @event is for this PMU and valid
298 * Other error return values are allowed.
300 int (*event_init) (struct perf_event *event);
303 * Notification that the event was mapped or unmapped. Called
304 * in the context of the mapping task.
306 void (*event_mapped) (struct perf_event *event, struct mm_struct *mm); /* optional */
307 void (*event_unmapped) (struct perf_event *event, struct mm_struct *mm); /* optional */
310 * Flags for ->add()/->del()/ ->start()/->stop(). There are
311 * matching hw_perf_event::state flags.
313 #define PERF_EF_START 0x01 /* start the counter when adding */
314 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
315 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
318 * Adds/Removes a counter to/from the PMU, can be done inside a
319 * transaction, see the ->*_txn() methods.
321 * The add/del callbacks will reserve all hardware resources required
322 * to service the event, this includes any counter constraint
325 * Called with IRQs disabled and the PMU disabled on the CPU the event
328 * ->add() called without PERF_EF_START should result in the same state
329 * as ->add() followed by ->stop().
331 * ->del() must always PERF_EF_UPDATE stop an event. If it calls
332 * ->stop() that must deal with already being stopped without
335 int (*add) (struct perf_event *event, int flags);
336 void (*del) (struct perf_event *event, int flags);
339 * Starts/Stops a counter present on the PMU.
341 * The PMI handler should stop the counter when perf_event_overflow()
342 * returns !0. ->start() will be used to continue.
344 * Also used to change the sample period.
346 * Called with IRQs disabled and the PMU disabled on the CPU the event
347 * is on -- will be called from NMI context with the PMU generates
350 * ->stop() with PERF_EF_UPDATE will read the counter and update
351 * period/count values like ->read() would.
353 * ->start() with PERF_EF_RELOAD will reprogram the the counter
354 * value, must be preceded by a ->stop() with PERF_EF_UPDATE.
356 void (*start) (struct perf_event *event, int flags);
357 void (*stop) (struct perf_event *event, int flags);
360 * Updates the counter value of the event.
362 * For sampling capable PMUs this will also update the software period
363 * hw_perf_event::period_left field.
365 void (*read) (struct perf_event *event);
368 * Group events scheduling is treated as a transaction, add
369 * group events as a whole and perform one schedulability test.
370 * If the test fails, roll back the whole group
372 * Start the transaction, after this ->add() doesn't need to
373 * do schedulability tests.
377 void (*start_txn) (struct pmu *pmu, unsigned int txn_flags);
379 * If ->start_txn() disabled the ->add() schedulability test
380 * then ->commit_txn() is required to perform one. On success
381 * the transaction is closed. On error the transaction is kept
382 * open until ->cancel_txn() is called.
386 int (*commit_txn) (struct pmu *pmu);
388 * Will cancel the transaction, assumes ->del() is called
389 * for each successful ->add() during the transaction.
393 void (*cancel_txn) (struct pmu *pmu);
396 * Will return the value for perf_event_mmap_page::index for this event,
397 * if no implementation is provided it will default to: event->hw.idx + 1.
399 int (*event_idx) (struct perf_event *event); /*optional */
402 * context-switches callback
404 void (*sched_task) (struct perf_event_context *ctx,
407 * PMU specific data size
409 size_t task_ctx_size;
413 * Set up pmu-private data structures for an AUX area
415 void *(*setup_aux) (struct perf_event *event, void **pages,
416 int nr_pages, bool overwrite);
420 * Free pmu-private AUX data structures
422 void (*free_aux) (void *aux); /* optional */
425 * Validate address range filters: make sure the HW supports the
426 * requested configuration and number of filters; return 0 if the
427 * supplied filters are valid, -errno otherwise.
429 * Runs in the context of the ioctl()ing process and is not serialized
430 * with the rest of the PMU callbacks.
432 int (*addr_filters_validate) (struct list_head *filters);
436 * Synchronize address range filter configuration:
437 * translate hw-agnostic filters into hardware configuration in
438 * event::hw::addr_filters.
440 * Runs as a part of filter sync sequence that is done in ->start()
441 * callback by calling perf_event_addr_filters_sync().
443 * May (and should) traverse event::addr_filters::list, for which its
444 * caller provides necessary serialization.
446 void (*addr_filters_sync) (struct perf_event *event);
450 * Filter events for PMU-specific reasons.
452 int (*filter_match) (struct perf_event *event); /* optional */
455 * Check period value for PERF_EVENT_IOC_PERIOD ioctl.
457 int (*check_period) (struct perf_event *event, u64 value); /* optional */
460 enum perf_addr_filter_action_t {
461 PERF_ADDR_FILTER_ACTION_STOP = 0,
462 PERF_ADDR_FILTER_ACTION_START,
463 PERF_ADDR_FILTER_ACTION_FILTER,
467 * struct perf_addr_filter - address range filter definition
468 * @entry: event's filter list linkage
469 * @path: object file's path for file-based filters
470 * @offset: filter range offset
471 * @size: filter range size (size==0 means single address trigger)
472 * @action: filter/start/stop
474 * This is a hardware-agnostic filter configuration as specified by the user.
476 struct perf_addr_filter {
477 struct list_head entry;
479 unsigned long offset;
481 enum perf_addr_filter_action_t action;
485 * struct perf_addr_filters_head - container for address range filters
486 * @list: list of filters for this event
487 * @lock: spinlock that serializes accesses to the @list and event's
488 * (and its children's) filter generations.
489 * @nr_file_filters: number of file-based filters
491 * A child event will use parent's @list (and therefore @lock), so they are
492 * bundled together; see perf_event_addr_filters().
494 struct perf_addr_filters_head {
495 struct list_head list;
497 unsigned int nr_file_filters;
500 struct perf_addr_filter_range {
506 * enum perf_event_state - the states of an event:
508 enum perf_event_state {
509 PERF_EVENT_STATE_DEAD = -4,
510 PERF_EVENT_STATE_EXIT = -3,
511 PERF_EVENT_STATE_ERROR = -2,
512 PERF_EVENT_STATE_OFF = -1,
513 PERF_EVENT_STATE_INACTIVE = 0,
514 PERF_EVENT_STATE_ACTIVE = 1,
518 struct perf_sample_data;
520 typedef void (*perf_overflow_handler_t)(struct perf_event *,
521 struct perf_sample_data *,
522 struct pt_regs *regs);
525 * Event capabilities. For event_caps and groups caps.
527 * PERF_EV_CAP_SOFTWARE: Is a software event.
528 * PERF_EV_CAP_READ_ACTIVE_PKG: A CPU event (or cgroup event) that can be read
529 * from any CPU in the package where it is active.
531 #define PERF_EV_CAP_SOFTWARE BIT(0)
532 #define PERF_EV_CAP_READ_ACTIVE_PKG BIT(1)
534 #define SWEVENT_HLIST_BITS 8
535 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
537 struct swevent_hlist {
538 struct hlist_head heads[SWEVENT_HLIST_SIZE];
539 struct rcu_head rcu_head;
542 #define PERF_ATTACH_CONTEXT 0x01
543 #define PERF_ATTACH_GROUP 0x02
544 #define PERF_ATTACH_TASK 0x04
545 #define PERF_ATTACH_TASK_DATA 0x08
546 #define PERF_ATTACH_ITRACE 0x10
551 struct pmu_event_list {
553 struct list_head list;
556 #define for_each_sibling_event(sibling, event) \
557 if ((event)->group_leader == (event)) \
558 list_for_each_entry((sibling), &(event)->sibling_list, sibling_list)
561 * struct perf_event - performance event kernel representation:
564 #ifdef CONFIG_PERF_EVENTS
566 * entry onto perf_event_context::event_list;
567 * modifications require ctx->lock
568 * RCU safe iterations.
570 struct list_head event_entry;
573 * Locked for modification by both ctx->mutex and ctx->lock; holding
574 * either sufficies for read.
576 struct list_head sibling_list;
577 struct list_head active_list;
579 * Node on the pinned or flexible tree located at the event context;
581 struct rb_node group_node;
584 * We need storage to track the entries in perf_pmu_migrate_context; we
585 * cannot use the event_entry because of RCU and we want to keep the
586 * group in tact which avoids us using the other two entries.
588 struct list_head migrate_entry;
590 struct hlist_node hlist_entry;
591 struct list_head active_entry;
594 /* Not serialized. Only written during event initialization. */
596 /* The cumulative AND of all event_caps for events in this group. */
599 struct perf_event *group_leader;
603 enum perf_event_state state;
604 unsigned int attach_state;
606 atomic64_t child_count;
609 * These are the total time in nanoseconds that the event
610 * has been enabled (i.e. eligible to run, and the task has
611 * been scheduled in, if this is a per-task event)
612 * and running (scheduled onto the CPU), respectively.
614 u64 total_time_enabled;
615 u64 total_time_running;
619 * timestamp shadows the actual context timing but it can
620 * be safely used in NMI interrupt context. It reflects the
621 * context time as it was when the event was last scheduled in.
623 * ctx_time already accounts for ctx->timestamp. Therefore to
624 * compute ctx_time for a sample, simply add perf_clock().
628 struct perf_event_attr attr;
632 struct hw_perf_event hw;
634 struct perf_event_context *ctx;
635 atomic_long_t refcount;
638 * These accumulate total time (in nanoseconds) that children
639 * events have been enabled and running, respectively.
641 atomic64_t child_total_time_enabled;
642 atomic64_t child_total_time_running;
645 * Protect attach/detach and child_list:
647 struct mutex child_mutex;
648 struct list_head child_list;
649 struct perf_event *parent;
654 struct list_head owner_entry;
655 struct task_struct *owner;
658 struct mutex mmap_mutex;
661 struct ring_buffer *rb;
662 struct list_head rb_entry;
663 unsigned long rcu_batches;
667 wait_queue_head_t waitq;
668 struct fasync_struct *fasync;
670 /* delayed work for NMIs and such */
674 struct irq_work pending;
676 atomic_t event_limit;
678 /* address range filters */
679 struct perf_addr_filters_head addr_filters;
680 /* vma address array for file-based filders */
681 struct perf_addr_filter_range *addr_filter_ranges;
682 unsigned long addr_filters_gen;
684 void (*destroy)(struct perf_event *);
685 struct rcu_head rcu_head;
687 struct pid_namespace *ns;
691 perf_overflow_handler_t overflow_handler;
692 void *overflow_handler_context;
693 #ifdef CONFIG_BPF_SYSCALL
694 perf_overflow_handler_t orig_overflow_handler;
695 struct bpf_prog *prog;
698 #ifdef CONFIG_EVENT_TRACING
699 struct trace_event_call *tp_event;
700 struct event_filter *filter;
701 #ifdef CONFIG_FUNCTION_TRACER
702 struct ftrace_ops ftrace_ops;
706 #ifdef CONFIG_CGROUP_PERF
707 struct perf_cgroup *cgrp; /* cgroup event is attach to */
710 struct list_head sb_list;
711 #endif /* CONFIG_PERF_EVENTS */
715 struct perf_event_groups {
721 * struct perf_event_context - event context structure
723 * Used as a container for task events and CPU events as well:
725 struct perf_event_context {
728 * Protect the states of the events in the list,
729 * nr_active, and the list:
733 * Protect the list of events. Locking either mutex or lock
734 * is sufficient to ensure the list doesn't change; to change
735 * the list you need to lock both the mutex and the spinlock.
739 struct list_head active_ctx_list;
740 struct perf_event_groups pinned_groups;
741 struct perf_event_groups flexible_groups;
742 struct list_head event_list;
744 struct list_head pinned_active;
745 struct list_head flexible_active;
754 * Set when nr_events != nr_active, except tolerant to events not
755 * necessary to be active due to scheduling constraints, such as cgroups.
757 int rotate_necessary;
759 struct task_struct *task;
762 * Context clock, runs when context enabled.
768 * These fields let us detect when two contexts have both
769 * been cloned (inherited) from a common ancestor.
771 struct perf_event_context *parent_ctx;
775 #ifdef CONFIG_CGROUP_PERF
776 int nr_cgroups; /* cgroup evts */
778 void *task_ctx_data; /* pmu specific data */
779 struct rcu_head rcu_head;
783 * Number of contexts where an event can trigger:
784 * task, softirq, hardirq, nmi.
786 #define PERF_NR_CONTEXTS 4
789 * struct perf_event_cpu_context - per cpu event context structure
791 struct perf_cpu_context {
792 struct perf_event_context ctx;
793 struct perf_event_context *task_ctx;
797 raw_spinlock_t hrtimer_lock;
798 struct hrtimer hrtimer;
799 ktime_t hrtimer_interval;
800 unsigned int hrtimer_active;
802 #ifdef CONFIG_CGROUP_PERF
803 struct perf_cgroup *cgrp;
804 struct list_head cgrp_cpuctx_entry;
807 struct list_head sched_cb_entry;
813 struct perf_output_handle {
814 struct perf_event *event;
815 struct ring_buffer *rb;
816 unsigned long wakeup;
826 struct bpf_perf_event_data_kern {
827 bpf_user_pt_regs_t *regs;
828 struct perf_sample_data *data;
829 struct perf_event *event;
832 #ifdef CONFIG_CGROUP_PERF
835 * perf_cgroup_info keeps track of time_enabled for a cgroup.
836 * This is a per-cpu dynamically allocated data structure.
838 struct perf_cgroup_info {
844 struct cgroup_subsys_state css;
845 struct perf_cgroup_info __percpu *info;
849 * Must ensure cgroup is pinned (css_get) before calling
850 * this function. In other words, we cannot call this function
851 * if there is no cgroup event for the current CPU context.
853 static inline struct perf_cgroup *
854 perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
856 return container_of(task_css_check(task, perf_event_cgrp_id,
857 ctx ? lockdep_is_held(&ctx->lock)
859 struct perf_cgroup, css);
861 #endif /* CONFIG_CGROUP_PERF */
863 #ifdef CONFIG_PERF_EVENTS
865 extern void *perf_aux_output_begin(struct perf_output_handle *handle,
866 struct perf_event *event);
867 extern void perf_aux_output_end(struct perf_output_handle *handle,
869 extern int perf_aux_output_skip(struct perf_output_handle *handle,
871 extern void *perf_get_aux(struct perf_output_handle *handle);
872 extern void perf_aux_output_flag(struct perf_output_handle *handle, u64 flags);
873 extern void perf_event_itrace_started(struct perf_event *event);
875 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
876 extern void perf_pmu_unregister(struct pmu *pmu);
878 extern int perf_num_counters(void);
879 extern const char *perf_pmu_name(void);
880 extern void __perf_event_task_sched_in(struct task_struct *prev,
881 struct task_struct *task);
882 extern void __perf_event_task_sched_out(struct task_struct *prev,
883 struct task_struct *next);
884 extern int perf_event_init_task(struct task_struct *child);
885 extern void perf_event_exit_task(struct task_struct *child);
886 extern void perf_event_free_task(struct task_struct *task);
887 extern void perf_event_delayed_put(struct task_struct *task);
888 extern struct file *perf_event_get(unsigned int fd);
889 extern const struct perf_event *perf_get_event(struct file *file);
890 extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
891 extern void perf_event_print_debug(void);
892 extern void perf_pmu_disable(struct pmu *pmu);
893 extern void perf_pmu_enable(struct pmu *pmu);
894 extern void perf_sched_cb_dec(struct pmu *pmu);
895 extern void perf_sched_cb_inc(struct pmu *pmu);
896 extern int perf_event_task_disable(void);
897 extern int perf_event_task_enable(void);
899 extern void perf_pmu_resched(struct pmu *pmu);
901 extern int perf_event_refresh(struct perf_event *event, int refresh);
902 extern void perf_event_update_userpage(struct perf_event *event);
903 extern int perf_event_release_kernel(struct perf_event *event);
904 extern struct perf_event *
905 perf_event_create_kernel_counter(struct perf_event_attr *attr,
907 struct task_struct *task,
908 perf_overflow_handler_t callback,
910 extern void perf_pmu_migrate_context(struct pmu *pmu,
911 int src_cpu, int dst_cpu);
912 int perf_event_read_local(struct perf_event *event, u64 *value,
913 u64 *enabled, u64 *running);
914 extern u64 perf_event_read_value(struct perf_event *event,
915 u64 *enabled, u64 *running);
918 struct perf_sample_data {
920 * Fields set by perf_sample_data_init(), group so as to
921 * minimize the cachelines touched.
924 struct perf_raw_record *raw;
925 struct perf_branch_stack *br_stack;
929 union perf_mem_data_src data_src;
932 * The other fields, optionally {set,used} by
933 * perf_{prepare,output}_sample().
948 struct perf_callchain_entry *callchain;
951 * regs_user may point to task_pt_regs or to regs_user_copy, depending
954 struct perf_regs regs_user;
955 struct pt_regs regs_user_copy;
957 struct perf_regs regs_intr;
961 } ____cacheline_aligned;
963 /* default value for data source */
964 #define PERF_MEM_NA (PERF_MEM_S(OP, NA) |\
965 PERF_MEM_S(LVL, NA) |\
966 PERF_MEM_S(SNOOP, NA) |\
967 PERF_MEM_S(LOCK, NA) |\
970 static inline void perf_sample_data_init(struct perf_sample_data *data,
971 u64 addr, u64 period)
973 /* remaining struct members initialized in perf_prepare_sample() */
976 data->br_stack = NULL;
977 data->period = period;
979 data->data_src.val = PERF_MEM_NA;
983 extern void perf_output_sample(struct perf_output_handle *handle,
984 struct perf_event_header *header,
985 struct perf_sample_data *data,
986 struct perf_event *event);
987 extern void perf_prepare_sample(struct perf_event_header *header,
988 struct perf_sample_data *data,
989 struct perf_event *event,
990 struct pt_regs *regs);
992 extern int perf_event_overflow(struct perf_event *event,
993 struct perf_sample_data *data,
994 struct pt_regs *regs);
996 extern void perf_event_output_forward(struct perf_event *event,
997 struct perf_sample_data *data,
998 struct pt_regs *regs);
999 extern void perf_event_output_backward(struct perf_event *event,
1000 struct perf_sample_data *data,
1001 struct pt_regs *regs);
1002 extern int perf_event_output(struct perf_event *event,
1003 struct perf_sample_data *data,
1004 struct pt_regs *regs);
1007 is_default_overflow_handler(struct perf_event *event)
1009 if (likely(event->overflow_handler == perf_event_output_forward))
1011 if (unlikely(event->overflow_handler == perf_event_output_backward))
1017 perf_event_header__init_id(struct perf_event_header *header,
1018 struct perf_sample_data *data,
1019 struct perf_event *event);
1021 perf_event__output_id_sample(struct perf_event *event,
1022 struct perf_output_handle *handle,
1023 struct perf_sample_data *sample);
1026 perf_log_lost_samples(struct perf_event *event, u64 lost);
1028 static inline bool event_has_any_exclude_flag(struct perf_event *event)
1030 struct perf_event_attr *attr = &event->attr;
1032 return attr->exclude_idle || attr->exclude_user ||
1033 attr->exclude_kernel || attr->exclude_hv ||
1034 attr->exclude_guest || attr->exclude_host;
1037 static inline bool is_sampling_event(struct perf_event *event)
1039 return event->attr.sample_period != 0;
1043 * Return 1 for a software event, 0 for a hardware event
1045 static inline int is_software_event(struct perf_event *event)
1047 return event->event_caps & PERF_EV_CAP_SOFTWARE;
1051 * Return 1 for event in sw context, 0 for event in hw context
1053 static inline int in_software_context(struct perf_event *event)
1055 return event->ctx->pmu->task_ctx_nr == perf_sw_context;
1058 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
1060 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
1061 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
1063 #ifndef perf_arch_fetch_caller_regs
1064 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
1068 * When generating a perf sample in-line, instead of from an interrupt /
1069 * exception, we lack a pt_regs. This is typically used from software events
1070 * like: SW_CONTEXT_SWITCHES, SW_MIGRATIONS and the tie-in with tracepoints.
1072 * We typically don't need a full set, but (for x86) do require:
1073 * - ip for PERF_SAMPLE_IP
1074 * - cs for user_mode() tests
1075 * - sp for PERF_SAMPLE_CALLCHAIN
1076 * - eflags for MISC bits and CALLCHAIN (see: perf_hw_regs())
1078 * NOTE: assumes @regs is otherwise already 0 filled; this is important for
1079 * things like PERF_SAMPLE_REGS_INTR.
1081 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1083 perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1086 static __always_inline void
1087 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1089 if (static_key_false(&perf_swevent_enabled[event_id]))
1090 __perf_sw_event(event_id, nr, regs, addr);
1093 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
1096 * 'Special' version for the scheduler, it hard assumes no recursion,
1097 * which is guaranteed by us not actually scheduling inside other swevents
1098 * because those disable preemption.
1100 static __always_inline void
1101 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
1103 if (static_key_false(&perf_swevent_enabled[event_id])) {
1104 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1106 perf_fetch_caller_regs(regs);
1107 ___perf_sw_event(event_id, nr, regs, addr);
1111 extern struct static_key_false perf_sched_events;
1113 static __always_inline bool
1114 perf_sw_migrate_enabled(void)
1116 if (static_key_false(&perf_swevent_enabled[PERF_COUNT_SW_CPU_MIGRATIONS]))
1121 static inline void perf_event_task_migrate(struct task_struct *task)
1123 if (perf_sw_migrate_enabled())
1124 task->sched_migrated = 1;
1127 static inline void perf_event_task_sched_in(struct task_struct *prev,
1128 struct task_struct *task)
1130 if (static_branch_unlikely(&perf_sched_events))
1131 __perf_event_task_sched_in(prev, task);
1133 if (perf_sw_migrate_enabled() && task->sched_migrated) {
1134 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1136 perf_fetch_caller_regs(regs);
1137 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, regs, 0);
1138 task->sched_migrated = 0;
1142 static inline void perf_event_task_sched_out(struct task_struct *prev,
1143 struct task_struct *next)
1145 perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
1147 if (static_branch_unlikely(&perf_sched_events))
1148 __perf_event_task_sched_out(prev, next);
1151 extern void perf_event_mmap(struct vm_area_struct *vma);
1153 extern void perf_event_ksymbol(u16 ksym_type, u64 addr, u32 len,
1154 bool unregister, const char *sym);
1155 extern void perf_event_bpf_event(struct bpf_prog *prog,
1156 enum perf_bpf_event_type type,
1159 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1160 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1161 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1163 extern void perf_event_exec(void);
1164 extern void perf_event_comm(struct task_struct *tsk, bool exec);
1165 extern void perf_event_namespaces(struct task_struct *tsk);
1166 extern void perf_event_fork(struct task_struct *tsk);
1169 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1171 extern void perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1172 extern void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1173 extern struct perf_callchain_entry *
1174 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
1175 u32 max_stack, bool crosstask, bool add_mark);
1176 extern struct perf_callchain_entry *perf_callchain(struct perf_event *event, struct pt_regs *regs);
1177 extern int get_callchain_buffers(int max_stack);
1178 extern void put_callchain_buffers(void);
1180 extern int sysctl_perf_event_max_stack;
1181 extern int sysctl_perf_event_max_contexts_per_stack;
1183 static inline int perf_callchain_store_context(struct perf_callchain_entry_ctx *ctx, u64 ip)
1185 if (ctx->contexts < sysctl_perf_event_max_contexts_per_stack) {
1186 struct perf_callchain_entry *entry = ctx->entry;
1187 entry->ip[entry->nr++] = ip;
1191 ctx->contexts_maxed = true;
1192 return -1; /* no more room, stop walking the stack */
1196 static inline int perf_callchain_store(struct perf_callchain_entry_ctx *ctx, u64 ip)
1198 if (ctx->nr < ctx->max_stack && !ctx->contexts_maxed) {
1199 struct perf_callchain_entry *entry = ctx->entry;
1200 entry->ip[entry->nr++] = ip;
1204 return -1; /* no more room, stop walking the stack */
1208 extern int sysctl_perf_event_paranoid;
1209 extern int sysctl_perf_event_mlock;
1210 extern int sysctl_perf_event_sample_rate;
1211 extern int sysctl_perf_cpu_time_max_percent;
1213 extern void perf_sample_event_took(u64 sample_len_ns);
1215 extern int perf_proc_update_handler(struct ctl_table *table, int write,
1216 void __user *buffer, size_t *lenp,
1218 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
1219 void __user *buffer, size_t *lenp,
1222 int perf_event_max_stack_handler(struct ctl_table *table, int write,
1223 void __user *buffer, size_t *lenp, loff_t *ppos);
1225 static inline bool perf_paranoid_tracepoint_raw(void)
1227 return sysctl_perf_event_paranoid > -1;
1230 static inline bool perf_paranoid_cpu(void)
1232 return sysctl_perf_event_paranoid > 0;
1235 static inline bool perf_paranoid_kernel(void)
1237 return sysctl_perf_event_paranoid > 1;
1240 extern void perf_event_init(void);
1241 extern void perf_tp_event(u16 event_type, u64 count, void *record,
1242 int entry_size, struct pt_regs *regs,
1243 struct hlist_head *head, int rctx,
1244 struct task_struct *task);
1245 extern void perf_bp_event(struct perf_event *event, void *data);
1247 #ifndef perf_misc_flags
1248 # define perf_misc_flags(regs) \
1249 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1250 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1252 #ifndef perf_arch_bpf_user_pt_regs
1253 # define perf_arch_bpf_user_pt_regs(regs) regs
1256 static inline bool has_branch_stack(struct perf_event *event)
1258 return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1261 static inline bool needs_branch_stack(struct perf_event *event)
1263 return event->attr.branch_sample_type != 0;
1266 static inline bool has_aux(struct perf_event *event)
1268 return event->pmu->setup_aux;
1271 static inline bool is_write_backward(struct perf_event *event)
1273 return !!event->attr.write_backward;
1276 static inline bool has_addr_filter(struct perf_event *event)
1278 return event->pmu->nr_addr_filters;
1282 * An inherited event uses parent's filters
1284 static inline struct perf_addr_filters_head *
1285 perf_event_addr_filters(struct perf_event *event)
1287 struct perf_addr_filters_head *ifh = &event->addr_filters;
1290 ifh = &event->parent->addr_filters;
1295 extern void perf_event_addr_filters_sync(struct perf_event *event);
1297 extern int perf_output_begin(struct perf_output_handle *handle,
1298 struct perf_event *event, unsigned int size);
1299 extern int perf_output_begin_forward(struct perf_output_handle *handle,
1300 struct perf_event *event,
1302 extern int perf_output_begin_backward(struct perf_output_handle *handle,
1303 struct perf_event *event,
1306 extern void perf_output_end(struct perf_output_handle *handle);
1307 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
1308 const void *buf, unsigned int len);
1309 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
1311 extern int perf_swevent_get_recursion_context(void);
1312 extern void perf_swevent_put_recursion_context(int rctx);
1313 extern u64 perf_swevent_set_period(struct perf_event *event);
1314 extern void perf_event_enable(struct perf_event *event);
1315 extern void perf_event_disable(struct perf_event *event);
1316 extern void perf_event_disable_local(struct perf_event *event);
1317 extern void perf_event_disable_inatomic(struct perf_event *event);
1318 extern void perf_event_task_tick(void);
1319 extern int perf_event_account_interrupt(struct perf_event *event);
1320 #else /* !CONFIG_PERF_EVENTS: */
1321 static inline void *
1322 perf_aux_output_begin(struct perf_output_handle *handle,
1323 struct perf_event *event) { return NULL; }
1325 perf_aux_output_end(struct perf_output_handle *handle, unsigned long size)
1328 perf_aux_output_skip(struct perf_output_handle *handle,
1329 unsigned long size) { return -EINVAL; }
1330 static inline void *
1331 perf_get_aux(struct perf_output_handle *handle) { return NULL; }
1333 perf_event_task_migrate(struct task_struct *task) { }
1335 perf_event_task_sched_in(struct task_struct *prev,
1336 struct task_struct *task) { }
1338 perf_event_task_sched_out(struct task_struct *prev,
1339 struct task_struct *next) { }
1340 static inline int perf_event_init_task(struct task_struct *child) { return 0; }
1341 static inline void perf_event_exit_task(struct task_struct *child) { }
1342 static inline void perf_event_free_task(struct task_struct *task) { }
1343 static inline void perf_event_delayed_put(struct task_struct *task) { }
1344 static inline struct file *perf_event_get(unsigned int fd) { return ERR_PTR(-EINVAL); }
1345 static inline const struct perf_event *perf_get_event(struct file *file)
1347 return ERR_PTR(-EINVAL);
1349 static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
1351 return ERR_PTR(-EINVAL);
1353 static inline int perf_event_read_local(struct perf_event *event, u64 *value,
1354 u64 *enabled, u64 *running)
1358 static inline void perf_event_print_debug(void) { }
1359 static inline int perf_event_task_disable(void) { return -EINVAL; }
1360 static inline int perf_event_task_enable(void) { return -EINVAL; }
1361 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1367 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) { }
1369 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr) { }
1371 perf_bp_event(struct perf_event *event, void *data) { }
1373 static inline int perf_register_guest_info_callbacks
1374 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1375 static inline int perf_unregister_guest_info_callbacks
1376 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1378 static inline void perf_event_mmap(struct vm_area_struct *vma) { }
1380 typedef int (perf_ksymbol_get_name_f)(char *name, int name_len, void *data);
1381 static inline void perf_event_ksymbol(u16 ksym_type, u64 addr, u32 len,
1382 bool unregister, const char *sym) { }
1383 static inline void perf_event_bpf_event(struct bpf_prog *prog,
1384 enum perf_bpf_event_type type,
1386 static inline void perf_event_exec(void) { }
1387 static inline void perf_event_comm(struct task_struct *tsk, bool exec) { }
1388 static inline void perf_event_namespaces(struct task_struct *tsk) { }
1389 static inline void perf_event_fork(struct task_struct *tsk) { }
1390 static inline void perf_event_init(void) { }
1391 static inline int perf_swevent_get_recursion_context(void) { return -1; }
1392 static inline void perf_swevent_put_recursion_context(int rctx) { }
1393 static inline u64 perf_swevent_set_period(struct perf_event *event) { return 0; }
1394 static inline void perf_event_enable(struct perf_event *event) { }
1395 static inline void perf_event_disable(struct perf_event *event) { }
1396 static inline int __perf_event_disable(void *info) { return -1; }
1397 static inline void perf_event_task_tick(void) { }
1398 static inline int perf_event_release_kernel(struct perf_event *event) { return 0; }
1401 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1402 extern void perf_restore_debug_store(void);
1404 static inline void perf_restore_debug_store(void) { }
1407 static __always_inline bool perf_raw_frag_last(const struct perf_raw_frag *frag)
1409 return frag->pad < sizeof(u64);
1412 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1414 struct perf_pmu_events_attr {
1415 struct device_attribute attr;
1417 const char *event_str;
1420 struct perf_pmu_events_ht_attr {
1421 struct device_attribute attr;
1423 const char *event_str_ht;
1424 const char *event_str_noht;
1427 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1430 #define PMU_EVENT_ATTR(_name, _var, _id, _show) \
1431 static struct perf_pmu_events_attr _var = { \
1432 .attr = __ATTR(_name, 0444, _show, NULL), \
1436 #define PMU_EVENT_ATTR_STRING(_name, _var, _str) \
1437 static struct perf_pmu_events_attr _var = { \
1438 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1440 .event_str = _str, \
1443 #define PMU_FORMAT_ATTR(_name, _format) \
1445 _name##_show(struct device *dev, \
1446 struct device_attribute *attr, \
1449 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
1450 return sprintf(page, _format "\n"); \
1453 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1455 /* Performance counter hotplug functions */
1456 #ifdef CONFIG_PERF_EVENTS
1457 int perf_event_init_cpu(unsigned int cpu);
1458 int perf_event_exit_cpu(unsigned int cpu);
1460 #define perf_event_init_cpu NULL
1461 #define perf_event_exit_cpu NULL
1464 #endif /* _LINUX_PERF_EVENT_H */