4 * Copyright (C) 2008-2009, Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2011, Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2011, Red Hat, Inc., Peter Zijlstra
8 * Data type definitions, declarations, prototypes.
10 * Started by: Thomas Gleixner and Ingo Molnar
12 * For licencing details see kernel-base/COPYING
14 #ifndef _LINUX_PERF_EVENT_H
15 #define _LINUX_PERF_EVENT_H
17 #include <uapi/linux/perf_event.h>
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);
35 #ifdef CONFIG_HAVE_HW_BREAKPOINT
36 #include <asm/hw_breakpoint.h>
39 #include <linux/list.h>
40 #include <linux/mutex.h>
41 #include <linux/rculist.h>
42 #include <linux/rcupdate.h>
43 #include <linux/spinlock.h>
44 #include <linux/hrtimer.h>
46 #include <linux/pid_namespace.h>
47 #include <linux/workqueue.h>
48 #include <linux/ftrace.h>
49 #include <linux/cpu.h>
50 #include <linux/irq_work.h>
51 #include <linux/static_key.h>
52 #include <linux/jump_label_ratelimit.h>
53 #include <linux/atomic.h>
54 #include <linux/sysfs.h>
55 #include <linux/perf_regs.h>
56 #include <linux/workqueue.h>
57 #include <linux/cgroup.h>
58 #include <asm/local.h>
60 struct perf_callchain_entry {
62 __u64 ip[0]; /* /proc/sys/kernel/perf_event_max_stack */
65 struct perf_callchain_entry_ctx {
66 struct perf_callchain_entry *entry;
73 typedef unsigned long (*perf_copy_f)(void *dst, const void *src,
74 unsigned long off, unsigned long len);
76 struct perf_raw_frag {
78 struct perf_raw_frag *next;
86 struct perf_raw_record {
87 struct perf_raw_frag frag;
92 * branch stack layout:
93 * nr: number of taken branches stored in entries[]
95 * Note that nr can vary from sample to sample
96 * branches (to, from) are stored from most recent
97 * to least recent, i.e., entries[0] contains the most
100 struct perf_branch_stack {
102 struct perf_branch_entry entries[0];
108 * extra PMU register associated with an event
110 struct hw_perf_event_extra {
111 u64 config; /* register value */
112 unsigned int reg; /* register address or index */
113 int alloc; /* extra register already allocated */
114 int idx; /* index in shared_regs->regs[] */
118 * struct hw_perf_event - performance event hardware details:
120 struct hw_perf_event {
121 #ifdef CONFIG_PERF_EVENTS
123 struct { /* hardware */
126 unsigned long config_base;
127 unsigned long event_base;
128 int event_base_rdpmc;
133 struct hw_perf_event_extra extra_reg;
134 struct hw_perf_event_extra branch_reg;
136 struct { /* software */
137 struct hrtimer hrtimer;
139 struct { /* tracepoint */
140 /* for tp_event->class */
141 struct list_head tp_list;
143 struct { /* amd_power */
147 #ifdef CONFIG_HAVE_HW_BREAKPOINT
148 struct { /* breakpoint */
150 * Crufty hack to avoid the chicken and egg
151 * problem hw_breakpoint has with context
152 * creation and event initalization.
154 struct arch_hw_breakpoint info;
155 struct list_head bp_list;
158 struct { /* amd_iommu */
167 * If the event is a per task event, this will point to the task in
168 * question. See the comment in perf_event_alloc().
170 struct task_struct *target;
173 * PMU would store hardware filter configuration
178 /* Last sync'ed generation of filters */
179 unsigned long addr_filters_gen;
182 * hw_perf_event::state flags; used to track the PERF_EF_* state.
184 #define PERF_HES_STOPPED 0x01 /* the counter is stopped */
185 #define PERF_HES_UPTODATE 0x02 /* event->count up-to-date */
186 #define PERF_HES_ARCH 0x04
191 * The last observed hardware counter value, updated with a
192 * local64_cmpxchg() such that pmu::read() can be called nested.
194 local64_t prev_count;
197 * The period to start the next sample with.
202 * The period we started this sample with.
207 * However much is left of the current period; note that this is
208 * a full 64bit value and allows for generation of periods longer
209 * than hardware might allow.
211 local64_t period_left;
214 * State for throttling the event, see __perf_event_overflow() and
215 * perf_adjust_freq_unthr_context().
221 * State for freq target events, see __perf_event_overflow() and
222 * perf_adjust_freq_unthr_context().
225 u64 freq_count_stamp;
232 * Common implementation detail of pmu::{start,commit,cancel}_txn
234 #define PERF_PMU_TXN_ADD 0x1 /* txn to add/schedule event on PMU */
235 #define PERF_PMU_TXN_READ 0x2 /* txn to read event group from PMU */
238 * pmu::capabilities flags
240 #define PERF_PMU_CAP_NO_INTERRUPT 0x01
241 #define PERF_PMU_CAP_NO_NMI 0x02
242 #define PERF_PMU_CAP_AUX_NO_SG 0x04
243 #define PERF_PMU_CAP_AUX_SW_DOUBLEBUF 0x08
244 #define PERF_PMU_CAP_EXCLUSIVE 0x10
245 #define PERF_PMU_CAP_ITRACE 0x20
246 #define PERF_PMU_CAP_HETEROGENEOUS_CPUS 0x40
249 * struct pmu - generic performance monitoring unit
252 struct list_head entry;
254 struct module *module;
256 const struct attribute_group **attr_groups;
261 * various common per-pmu feature flags
265 int __percpu *pmu_disable_count;
266 struct perf_cpu_context __percpu *pmu_cpu_context;
267 atomic_t exclusive_cnt; /* < 0: cpu; > 0: tsk */
269 int hrtimer_interval_ms;
271 /* number of address filters this PMU can do */
272 unsigned int nr_addr_filters;
275 * Fully disable/enable this PMU, can be used to protect from the PMI
276 * as well as for lazy/batch writing of the MSRs.
278 void (*pmu_enable) (struct pmu *pmu); /* optional */
279 void (*pmu_disable) (struct pmu *pmu); /* optional */
282 * Try and initialize the event for this PMU.
285 * -ENOENT -- @event is not for this PMU
287 * -ENODEV -- @event is for this PMU but PMU not present
288 * -EBUSY -- @event is for this PMU but PMU temporarily unavailable
289 * -EINVAL -- @event is for this PMU but @event is not valid
290 * -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
291 * -EACCESS -- @event is for this PMU, @event is valid, but no privilidges
293 * 0 -- @event is for this PMU and valid
295 * Other error return values are allowed.
297 int (*event_init) (struct perf_event *event);
300 * Notification that the event was mapped or unmapped. Called
301 * in the context of the mapping task.
303 void (*event_mapped) (struct perf_event *event, struct mm_struct *mm); /* optional */
304 void (*event_unmapped) (struct perf_event *event, struct mm_struct *mm); /* optional */
307 * Flags for ->add()/->del()/ ->start()/->stop(). There are
308 * matching hw_perf_event::state flags.
310 #define PERF_EF_START 0x01 /* start the counter when adding */
311 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
312 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
315 * Adds/Removes a counter to/from the PMU, can be done inside a
316 * transaction, see the ->*_txn() methods.
318 * The add/del callbacks will reserve all hardware resources required
319 * to service the event, this includes any counter constraint
322 * Called with IRQs disabled and the PMU disabled on the CPU the event
325 * ->add() called without PERF_EF_START should result in the same state
326 * as ->add() followed by ->stop().
328 * ->del() must always PERF_EF_UPDATE stop an event. If it calls
329 * ->stop() that must deal with already being stopped without
332 int (*add) (struct perf_event *event, int flags);
333 void (*del) (struct perf_event *event, int flags);
336 * Starts/Stops a counter present on the PMU.
338 * The PMI handler should stop the counter when perf_event_overflow()
339 * returns !0. ->start() will be used to continue.
341 * Also used to change the sample period.
343 * Called with IRQs disabled and the PMU disabled on the CPU the event
344 * is on -- will be called from NMI context with the PMU generates
347 * ->stop() with PERF_EF_UPDATE will read the counter and update
348 * period/count values like ->read() would.
350 * ->start() with PERF_EF_RELOAD will reprogram the the counter
351 * value, must be preceded by a ->stop() with PERF_EF_UPDATE.
353 void (*start) (struct perf_event *event, int flags);
354 void (*stop) (struct perf_event *event, int flags);
357 * Updates the counter value of the event.
359 * For sampling capable PMUs this will also update the software period
360 * hw_perf_event::period_left field.
362 void (*read) (struct perf_event *event);
365 * Group events scheduling is treated as a transaction, add
366 * group events as a whole and perform one schedulability test.
367 * If the test fails, roll back the whole group
369 * Start the transaction, after this ->add() doesn't need to
370 * do schedulability tests.
374 void (*start_txn) (struct pmu *pmu, unsigned int txn_flags);
376 * If ->start_txn() disabled the ->add() schedulability test
377 * then ->commit_txn() is required to perform one. On success
378 * the transaction is closed. On error the transaction is kept
379 * open until ->cancel_txn() is called.
383 int (*commit_txn) (struct pmu *pmu);
385 * Will cancel the transaction, assumes ->del() is called
386 * for each successful ->add() during the transaction.
390 void (*cancel_txn) (struct pmu *pmu);
393 * Will return the value for perf_event_mmap_page::index for this event,
394 * if no implementation is provided it will default to: event->hw.idx + 1.
396 int (*event_idx) (struct perf_event *event); /*optional */
399 * context-switches callback
401 void (*sched_task) (struct perf_event_context *ctx,
404 * PMU specific data size
406 size_t task_ctx_size;
410 * Set up pmu-private data structures for an AUX area
412 void *(*setup_aux) (int cpu, void **pages,
413 int nr_pages, bool overwrite);
417 * Free pmu-private AUX data structures
419 void (*free_aux) (void *aux); /* optional */
422 * Validate address range filters: make sure the HW supports the
423 * requested configuration and number of filters; return 0 if the
424 * supplied filters are valid, -errno otherwise.
426 * Runs in the context of the ioctl()ing process and is not serialized
427 * with the rest of the PMU callbacks.
429 int (*addr_filters_validate) (struct list_head *filters);
433 * Synchronize address range filter configuration:
434 * translate hw-agnostic filters into hardware configuration in
435 * event::hw::addr_filters.
437 * Runs as a part of filter sync sequence that is done in ->start()
438 * callback by calling perf_event_addr_filters_sync().
440 * May (and should) traverse event::addr_filters::list, for which its
441 * caller provides necessary serialization.
443 void (*addr_filters_sync) (struct perf_event *event);
447 * Filter events for PMU-specific reasons.
449 int (*filter_match) (struct perf_event *event); /* optional */
452 * Check period value for PERF_EVENT_IOC_PERIOD ioctl.
454 int (*check_period) (struct perf_event *event, u64 value); /* optional */
457 enum perf_addr_filter_action_t {
458 PERF_ADDR_FILTER_ACTION_STOP = 0,
459 PERF_ADDR_FILTER_ACTION_START,
460 PERF_ADDR_FILTER_ACTION_FILTER,
464 * struct perf_addr_filter - address range filter definition
465 * @entry: event's filter list linkage
466 * @inode: object file's inode for file-based filters
467 * @offset: filter range offset
468 * @size: filter range size (size==0 means single address trigger)
469 * @action: filter/start/stop
471 * This is a hardware-agnostic filter configuration as specified by the user.
473 struct perf_addr_filter {
474 struct list_head entry;
476 unsigned long offset;
478 enum perf_addr_filter_action_t action;
482 * struct perf_addr_filters_head - container for address range filters
483 * @list: list of filters for this event
484 * @lock: spinlock that serializes accesses to the @list and event's
485 * (and its children's) filter generations.
486 * @nr_file_filters: number of file-based filters
488 * A child event will use parent's @list (and therefore @lock), so they are
489 * bundled together; see perf_event_addr_filters().
491 struct perf_addr_filters_head {
492 struct list_head list;
494 unsigned int nr_file_filters;
498 * enum perf_event_state - the states of an event:
500 enum perf_event_state {
501 PERF_EVENT_STATE_DEAD = -4,
502 PERF_EVENT_STATE_EXIT = -3,
503 PERF_EVENT_STATE_ERROR = -2,
504 PERF_EVENT_STATE_OFF = -1,
505 PERF_EVENT_STATE_INACTIVE = 0,
506 PERF_EVENT_STATE_ACTIVE = 1,
510 struct perf_sample_data;
512 typedef void (*perf_overflow_handler_t)(struct perf_event *,
513 struct perf_sample_data *,
514 struct pt_regs *regs);
517 * Event capabilities. For event_caps and groups caps.
519 * PERF_EV_CAP_SOFTWARE: Is a software event.
520 * PERF_EV_CAP_READ_ACTIVE_PKG: A CPU event (or cgroup event) that can be read
521 * from any CPU in the package where it is active.
523 #define PERF_EV_CAP_SOFTWARE BIT(0)
524 #define PERF_EV_CAP_READ_ACTIVE_PKG BIT(1)
526 #define SWEVENT_HLIST_BITS 8
527 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
529 struct swevent_hlist {
530 struct hlist_head heads[SWEVENT_HLIST_SIZE];
531 struct rcu_head rcu_head;
534 #define PERF_ATTACH_CONTEXT 0x01
535 #define PERF_ATTACH_GROUP 0x02
536 #define PERF_ATTACH_TASK 0x04
537 #define PERF_ATTACH_TASK_DATA 0x08
538 #define PERF_ATTACH_ITRACE 0x10
543 struct pmu_event_list {
545 struct list_head list;
548 #define for_each_sibling_event(sibling, event) \
549 if ((event)->group_leader == (event)) \
550 list_for_each_entry((sibling), &(event)->sibling_list, sibling_list)
553 * struct perf_event - performance event kernel representation:
556 #ifdef CONFIG_PERF_EVENTS
558 * entry onto perf_event_context::event_list;
559 * modifications require ctx->lock
560 * RCU safe iterations.
562 struct list_head event_entry;
565 * Locked for modification by both ctx->mutex and ctx->lock; holding
566 * either sufficies for read.
568 struct list_head sibling_list;
569 struct list_head active_list;
571 * Node on the pinned or flexible tree located at the event context;
573 struct rb_node group_node;
576 * We need storage to track the entries in perf_pmu_migrate_context; we
577 * cannot use the event_entry because of RCU and we want to keep the
578 * group in tact which avoids us using the other two entries.
580 struct list_head migrate_entry;
582 struct hlist_node hlist_entry;
583 struct list_head active_entry;
586 /* Not serialized. Only written during event initialization. */
588 /* The cumulative AND of all event_caps for events in this group. */
591 struct perf_event *group_leader;
595 enum perf_event_state state;
596 unsigned int attach_state;
598 atomic64_t child_count;
601 * These are the total time in nanoseconds that the event
602 * has been enabled (i.e. eligible to run, and the task has
603 * been scheduled in, if this is a per-task event)
604 * and running (scheduled onto the CPU), respectively.
606 u64 total_time_enabled;
607 u64 total_time_running;
611 * timestamp shadows the actual context timing but it can
612 * be safely used in NMI interrupt context. It reflects the
613 * context time as it was when the event was last scheduled in.
615 * ctx_time already accounts for ctx->timestamp. Therefore to
616 * compute ctx_time for a sample, simply add perf_clock().
620 struct perf_event_attr attr;
624 struct hw_perf_event hw;
626 struct perf_event_context *ctx;
627 atomic_long_t refcount;
630 * These accumulate total time (in nanoseconds) that children
631 * events have been enabled and running, respectively.
633 atomic64_t child_total_time_enabled;
634 atomic64_t child_total_time_running;
637 * Protect attach/detach and child_list:
639 struct mutex child_mutex;
640 struct list_head child_list;
641 struct perf_event *parent;
646 struct list_head owner_entry;
647 struct task_struct *owner;
650 struct mutex mmap_mutex;
653 struct ring_buffer *rb;
654 struct list_head rb_entry;
655 unsigned long rcu_batches;
659 wait_queue_head_t waitq;
660 struct fasync_struct *fasync;
662 /* delayed work for NMIs and such */
666 struct irq_work pending;
668 atomic_t event_limit;
670 /* address range filters */
671 struct perf_addr_filters_head addr_filters;
672 /* vma address array for file-based filders */
673 unsigned long *addr_filters_offs;
674 unsigned long addr_filters_gen;
676 void (*destroy)(struct perf_event *);
677 struct rcu_head rcu_head;
679 struct pid_namespace *ns;
683 perf_overflow_handler_t overflow_handler;
684 void *overflow_handler_context;
685 #ifdef CONFIG_BPF_SYSCALL
686 perf_overflow_handler_t orig_overflow_handler;
687 struct bpf_prog *prog;
690 #ifdef CONFIG_EVENT_TRACING
691 struct trace_event_call *tp_event;
692 struct event_filter *filter;
693 #ifdef CONFIG_FUNCTION_TRACER
694 struct ftrace_ops ftrace_ops;
698 #ifdef CONFIG_CGROUP_PERF
699 struct perf_cgroup *cgrp; /* cgroup event is attach to */
702 struct list_head sb_list;
703 #endif /* CONFIG_PERF_EVENTS */
707 struct perf_event_groups {
713 * struct perf_event_context - event context structure
715 * Used as a container for task events and CPU events as well:
717 struct perf_event_context {
720 * Protect the states of the events in the list,
721 * nr_active, and the list:
725 * Protect the list of events. Locking either mutex or lock
726 * is sufficient to ensure the list doesn't change; to change
727 * the list you need to lock both the mutex and the spinlock.
731 struct list_head active_ctx_list;
732 struct perf_event_groups pinned_groups;
733 struct perf_event_groups flexible_groups;
734 struct list_head event_list;
736 struct list_head pinned_active;
737 struct list_head flexible_active;
746 struct task_struct *task;
749 * Context clock, runs when context enabled.
755 * These fields let us detect when two contexts have both
756 * been cloned (inherited) from a common ancestor.
758 struct perf_event_context *parent_ctx;
762 #ifdef CONFIG_CGROUP_PERF
763 int nr_cgroups; /* cgroup evts */
765 void *task_ctx_data; /* pmu specific data */
766 struct rcu_head rcu_head;
770 * Number of contexts where an event can trigger:
771 * task, softirq, hardirq, nmi.
773 #define PERF_NR_CONTEXTS 4
776 * struct perf_event_cpu_context - per cpu event context structure
778 struct perf_cpu_context {
779 struct perf_event_context ctx;
780 struct perf_event_context *task_ctx;
784 raw_spinlock_t hrtimer_lock;
785 struct hrtimer hrtimer;
786 ktime_t hrtimer_interval;
787 unsigned int hrtimer_active;
789 #ifdef CONFIG_CGROUP_PERF
790 struct perf_cgroup *cgrp;
791 struct list_head cgrp_cpuctx_entry;
794 struct list_head sched_cb_entry;
800 struct perf_output_handle {
801 struct perf_event *event;
802 struct ring_buffer *rb;
803 unsigned long wakeup;
813 struct bpf_perf_event_data_kern {
814 bpf_user_pt_regs_t *regs;
815 struct perf_sample_data *data;
816 struct perf_event *event;
819 #ifdef CONFIG_CGROUP_PERF
822 * perf_cgroup_info keeps track of time_enabled for a cgroup.
823 * This is a per-cpu dynamically allocated data structure.
825 struct perf_cgroup_info {
831 struct cgroup_subsys_state css;
832 struct perf_cgroup_info __percpu *info;
836 * Must ensure cgroup is pinned (css_get) before calling
837 * this function. In other words, we cannot call this function
838 * if there is no cgroup event for the current CPU context.
840 static inline struct perf_cgroup *
841 perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
843 return container_of(task_css_check(task, perf_event_cgrp_id,
844 ctx ? lockdep_is_held(&ctx->lock)
846 struct perf_cgroup, css);
848 #endif /* CONFIG_CGROUP_PERF */
850 #ifdef CONFIG_PERF_EVENTS
852 extern void *perf_aux_output_begin(struct perf_output_handle *handle,
853 struct perf_event *event);
854 extern void perf_aux_output_end(struct perf_output_handle *handle,
856 extern int perf_aux_output_skip(struct perf_output_handle *handle,
858 extern void *perf_get_aux(struct perf_output_handle *handle);
859 extern void perf_aux_output_flag(struct perf_output_handle *handle, u64 flags);
860 extern void perf_event_itrace_started(struct perf_event *event);
862 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
863 extern void perf_pmu_unregister(struct pmu *pmu);
865 extern int perf_num_counters(void);
866 extern const char *perf_pmu_name(void);
867 extern void __perf_event_task_sched_in(struct task_struct *prev,
868 struct task_struct *task);
869 extern void __perf_event_task_sched_out(struct task_struct *prev,
870 struct task_struct *next);
871 extern int perf_event_init_task(struct task_struct *child);
872 extern void perf_event_exit_task(struct task_struct *child);
873 extern void perf_event_free_task(struct task_struct *task);
874 extern void perf_event_delayed_put(struct task_struct *task);
875 extern struct file *perf_event_get(unsigned int fd);
876 extern const struct perf_event *perf_get_event(struct file *file);
877 extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
878 extern void perf_event_print_debug(void);
879 extern void perf_pmu_disable(struct pmu *pmu);
880 extern void perf_pmu_enable(struct pmu *pmu);
881 extern void perf_sched_cb_dec(struct pmu *pmu);
882 extern void perf_sched_cb_inc(struct pmu *pmu);
883 extern int perf_event_task_disable(void);
884 extern int perf_event_task_enable(void);
885 extern int perf_event_refresh(struct perf_event *event, int refresh);
886 extern void perf_event_update_userpage(struct perf_event *event);
887 extern int perf_event_release_kernel(struct perf_event *event);
888 extern struct perf_event *
889 perf_event_create_kernel_counter(struct perf_event_attr *attr,
891 struct task_struct *task,
892 perf_overflow_handler_t callback,
894 extern void perf_pmu_migrate_context(struct pmu *pmu,
895 int src_cpu, int dst_cpu);
896 int perf_event_read_local(struct perf_event *event, u64 *value,
897 u64 *enabled, u64 *running);
898 extern u64 perf_event_read_value(struct perf_event *event,
899 u64 *enabled, u64 *running);
902 struct perf_sample_data {
904 * Fields set by perf_sample_data_init(), group so as to
905 * minimize the cachelines touched.
908 struct perf_raw_record *raw;
909 struct perf_branch_stack *br_stack;
913 union perf_mem_data_src data_src;
916 * The other fields, optionally {set,used} by
917 * perf_{prepare,output}_sample().
932 struct perf_callchain_entry *callchain;
935 * regs_user may point to task_pt_regs or to regs_user_copy, depending
938 struct perf_regs regs_user;
939 struct pt_regs regs_user_copy;
941 struct perf_regs regs_intr;
945 } ____cacheline_aligned;
947 /* default value for data source */
948 #define PERF_MEM_NA (PERF_MEM_S(OP, NA) |\
949 PERF_MEM_S(LVL, NA) |\
950 PERF_MEM_S(SNOOP, NA) |\
951 PERF_MEM_S(LOCK, NA) |\
954 static inline void perf_sample_data_init(struct perf_sample_data *data,
955 u64 addr, u64 period)
957 /* remaining struct members initialized in perf_prepare_sample() */
960 data->br_stack = NULL;
961 data->period = period;
963 data->data_src.val = PERF_MEM_NA;
967 extern void perf_output_sample(struct perf_output_handle *handle,
968 struct perf_event_header *header,
969 struct perf_sample_data *data,
970 struct perf_event *event);
971 extern void perf_prepare_sample(struct perf_event_header *header,
972 struct perf_sample_data *data,
973 struct perf_event *event,
974 struct pt_regs *regs);
976 extern int perf_event_overflow(struct perf_event *event,
977 struct perf_sample_data *data,
978 struct pt_regs *regs);
980 extern void perf_event_output_forward(struct perf_event *event,
981 struct perf_sample_data *data,
982 struct pt_regs *regs);
983 extern void perf_event_output_backward(struct perf_event *event,
984 struct perf_sample_data *data,
985 struct pt_regs *regs);
986 extern void perf_event_output(struct perf_event *event,
987 struct perf_sample_data *data,
988 struct pt_regs *regs);
991 is_default_overflow_handler(struct perf_event *event)
993 if (likely(event->overflow_handler == perf_event_output_forward))
995 if (unlikely(event->overflow_handler == perf_event_output_backward))
1001 perf_event_header__init_id(struct perf_event_header *header,
1002 struct perf_sample_data *data,
1003 struct perf_event *event);
1005 perf_event__output_id_sample(struct perf_event *event,
1006 struct perf_output_handle *handle,
1007 struct perf_sample_data *sample);
1010 perf_log_lost_samples(struct perf_event *event, u64 lost);
1012 static inline bool is_sampling_event(struct perf_event *event)
1014 return event->attr.sample_period != 0;
1018 * Return 1 for a software event, 0 for a hardware event
1020 static inline int is_software_event(struct perf_event *event)
1022 return event->event_caps & PERF_EV_CAP_SOFTWARE;
1026 * Return 1 for event in sw context, 0 for event in hw context
1028 static inline int in_software_context(struct perf_event *event)
1030 return event->ctx->pmu->task_ctx_nr == perf_sw_context;
1033 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
1035 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
1036 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
1038 #ifndef perf_arch_fetch_caller_regs
1039 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
1043 * Take a snapshot of the regs. Skip ip and frame pointer to
1044 * the nth caller. We only need a few of the regs:
1045 * - ip for PERF_SAMPLE_IP
1046 * - cs for user_mode() tests
1047 * - bp for callchains
1048 * - eflags, for future purposes, just in case
1050 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1052 perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1055 static __always_inline void
1056 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1058 if (static_key_false(&perf_swevent_enabled[event_id]))
1059 __perf_sw_event(event_id, nr, regs, addr);
1062 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
1065 * 'Special' version for the scheduler, it hard assumes no recursion,
1066 * which is guaranteed by us not actually scheduling inside other swevents
1067 * because those disable preemption.
1069 static __always_inline void
1070 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
1072 if (static_key_false(&perf_swevent_enabled[event_id])) {
1073 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1075 perf_fetch_caller_regs(regs);
1076 ___perf_sw_event(event_id, nr, regs, addr);
1080 extern struct static_key_false perf_sched_events;
1082 static __always_inline bool
1083 perf_sw_migrate_enabled(void)
1085 if (static_key_false(&perf_swevent_enabled[PERF_COUNT_SW_CPU_MIGRATIONS]))
1090 static inline void perf_event_task_migrate(struct task_struct *task)
1092 if (perf_sw_migrate_enabled())
1093 task->sched_migrated = 1;
1096 static inline void perf_event_task_sched_in(struct task_struct *prev,
1097 struct task_struct *task)
1099 if (static_branch_unlikely(&perf_sched_events))
1100 __perf_event_task_sched_in(prev, task);
1102 if (perf_sw_migrate_enabled() && task->sched_migrated) {
1103 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1105 perf_fetch_caller_regs(regs);
1106 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, regs, 0);
1107 task->sched_migrated = 0;
1111 static inline void perf_event_task_sched_out(struct task_struct *prev,
1112 struct task_struct *next)
1114 perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
1116 if (static_branch_unlikely(&perf_sched_events))
1117 __perf_event_task_sched_out(prev, next);
1120 extern void perf_event_mmap(struct vm_area_struct *vma);
1121 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1122 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1123 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1125 extern void perf_event_exec(void);
1126 extern void perf_event_comm(struct task_struct *tsk, bool exec);
1127 extern void perf_event_namespaces(struct task_struct *tsk);
1128 extern void perf_event_fork(struct task_struct *tsk);
1131 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1133 extern void perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1134 extern void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1135 extern struct perf_callchain_entry *
1136 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
1137 u32 max_stack, bool crosstask, bool add_mark);
1138 extern struct perf_callchain_entry *perf_callchain(struct perf_event *event, struct pt_regs *regs);
1139 extern int get_callchain_buffers(int max_stack);
1140 extern void put_callchain_buffers(void);
1142 extern int sysctl_perf_event_max_stack;
1143 extern int sysctl_perf_event_max_contexts_per_stack;
1145 static inline int perf_callchain_store_context(struct perf_callchain_entry_ctx *ctx, u64 ip)
1147 if (ctx->contexts < sysctl_perf_event_max_contexts_per_stack) {
1148 struct perf_callchain_entry *entry = ctx->entry;
1149 entry->ip[entry->nr++] = ip;
1153 ctx->contexts_maxed = true;
1154 return -1; /* no more room, stop walking the stack */
1158 static inline int perf_callchain_store(struct perf_callchain_entry_ctx *ctx, u64 ip)
1160 if (ctx->nr < ctx->max_stack && !ctx->contexts_maxed) {
1161 struct perf_callchain_entry *entry = ctx->entry;
1162 entry->ip[entry->nr++] = ip;
1166 return -1; /* no more room, stop walking the stack */
1170 extern int sysctl_perf_event_paranoid;
1171 extern int sysctl_perf_event_mlock;
1172 extern int sysctl_perf_event_sample_rate;
1173 extern int sysctl_perf_cpu_time_max_percent;
1175 extern void perf_sample_event_took(u64 sample_len_ns);
1177 extern int perf_proc_update_handler(struct ctl_table *table, int write,
1178 void __user *buffer, size_t *lenp,
1180 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
1181 void __user *buffer, size_t *lenp,
1184 int perf_event_max_stack_handler(struct ctl_table *table, int write,
1185 void __user *buffer, size_t *lenp, loff_t *ppos);
1187 static inline bool perf_paranoid_tracepoint_raw(void)
1189 return sysctl_perf_event_paranoid > -1;
1192 static inline bool perf_paranoid_cpu(void)
1194 return sysctl_perf_event_paranoid > 0;
1197 static inline bool perf_paranoid_kernel(void)
1199 return sysctl_perf_event_paranoid > 1;
1202 extern void perf_event_init(void);
1203 extern void perf_tp_event(u16 event_type, u64 count, void *record,
1204 int entry_size, struct pt_regs *regs,
1205 struct hlist_head *head, int rctx,
1206 struct task_struct *task);
1207 extern void perf_bp_event(struct perf_event *event, void *data);
1209 #ifndef perf_misc_flags
1210 # define perf_misc_flags(regs) \
1211 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1212 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1214 #ifndef perf_arch_bpf_user_pt_regs
1215 # define perf_arch_bpf_user_pt_regs(regs) regs
1218 static inline bool has_branch_stack(struct perf_event *event)
1220 return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1223 static inline bool needs_branch_stack(struct perf_event *event)
1225 return event->attr.branch_sample_type != 0;
1228 static inline bool has_aux(struct perf_event *event)
1230 return event->pmu->setup_aux;
1233 static inline bool is_write_backward(struct perf_event *event)
1235 return !!event->attr.write_backward;
1238 static inline bool has_addr_filter(struct perf_event *event)
1240 return event->pmu->nr_addr_filters;
1244 * An inherited event uses parent's filters
1246 static inline struct perf_addr_filters_head *
1247 perf_event_addr_filters(struct perf_event *event)
1249 struct perf_addr_filters_head *ifh = &event->addr_filters;
1252 ifh = &event->parent->addr_filters;
1257 extern void perf_event_addr_filters_sync(struct perf_event *event);
1259 extern int perf_output_begin(struct perf_output_handle *handle,
1260 struct perf_event *event, unsigned int size);
1261 extern int perf_output_begin_forward(struct perf_output_handle *handle,
1262 struct perf_event *event,
1264 extern int perf_output_begin_backward(struct perf_output_handle *handle,
1265 struct perf_event *event,
1268 extern void perf_output_end(struct perf_output_handle *handle);
1269 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
1270 const void *buf, unsigned int len);
1271 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
1273 extern int perf_swevent_get_recursion_context(void);
1274 extern void perf_swevent_put_recursion_context(int rctx);
1275 extern u64 perf_swevent_set_period(struct perf_event *event);
1276 extern void perf_event_enable(struct perf_event *event);
1277 extern void perf_event_disable(struct perf_event *event);
1278 extern void perf_event_disable_local(struct perf_event *event);
1279 extern void perf_event_disable_inatomic(struct perf_event *event);
1280 extern void perf_event_task_tick(void);
1281 extern int perf_event_account_interrupt(struct perf_event *event);
1282 #else /* !CONFIG_PERF_EVENTS: */
1283 static inline void *
1284 perf_aux_output_begin(struct perf_output_handle *handle,
1285 struct perf_event *event) { return NULL; }
1287 perf_aux_output_end(struct perf_output_handle *handle, unsigned long size)
1290 perf_aux_output_skip(struct perf_output_handle *handle,
1291 unsigned long size) { return -EINVAL; }
1292 static inline void *
1293 perf_get_aux(struct perf_output_handle *handle) { return NULL; }
1295 perf_event_task_migrate(struct task_struct *task) { }
1297 perf_event_task_sched_in(struct task_struct *prev,
1298 struct task_struct *task) { }
1300 perf_event_task_sched_out(struct task_struct *prev,
1301 struct task_struct *next) { }
1302 static inline int perf_event_init_task(struct task_struct *child) { return 0; }
1303 static inline void perf_event_exit_task(struct task_struct *child) { }
1304 static inline void perf_event_free_task(struct task_struct *task) { }
1305 static inline void perf_event_delayed_put(struct task_struct *task) { }
1306 static inline struct file *perf_event_get(unsigned int fd) { return ERR_PTR(-EINVAL); }
1307 static inline const struct perf_event *perf_get_event(struct file *file)
1309 return ERR_PTR(-EINVAL);
1311 static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
1313 return ERR_PTR(-EINVAL);
1315 static inline int perf_event_read_local(struct perf_event *event, u64 *value,
1316 u64 *enabled, u64 *running)
1320 static inline void perf_event_print_debug(void) { }
1321 static inline int perf_event_task_disable(void) { return -EINVAL; }
1322 static inline int perf_event_task_enable(void) { return -EINVAL; }
1323 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1329 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) { }
1331 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr) { }
1333 perf_bp_event(struct perf_event *event, void *data) { }
1335 static inline int perf_register_guest_info_callbacks
1336 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1337 static inline int perf_unregister_guest_info_callbacks
1338 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1340 static inline void perf_event_mmap(struct vm_area_struct *vma) { }
1341 static inline void perf_event_exec(void) { }
1342 static inline void perf_event_comm(struct task_struct *tsk, bool exec) { }
1343 static inline void perf_event_namespaces(struct task_struct *tsk) { }
1344 static inline void perf_event_fork(struct task_struct *tsk) { }
1345 static inline void perf_event_init(void) { }
1346 static inline int perf_swevent_get_recursion_context(void) { return -1; }
1347 static inline void perf_swevent_put_recursion_context(int rctx) { }
1348 static inline u64 perf_swevent_set_period(struct perf_event *event) { return 0; }
1349 static inline void perf_event_enable(struct perf_event *event) { }
1350 static inline void perf_event_disable(struct perf_event *event) { }
1351 static inline int __perf_event_disable(void *info) { return -1; }
1352 static inline void perf_event_task_tick(void) { }
1353 static inline int perf_event_release_kernel(struct perf_event *event) { return 0; }
1356 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1357 extern void perf_restore_debug_store(void);
1359 static inline void perf_restore_debug_store(void) { }
1362 static __always_inline bool perf_raw_frag_last(const struct perf_raw_frag *frag)
1364 return frag->pad < sizeof(u64);
1367 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1369 struct perf_pmu_events_attr {
1370 struct device_attribute attr;
1372 const char *event_str;
1375 struct perf_pmu_events_ht_attr {
1376 struct device_attribute attr;
1378 const char *event_str_ht;
1379 const char *event_str_noht;
1382 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1385 #define PMU_EVENT_ATTR(_name, _var, _id, _show) \
1386 static struct perf_pmu_events_attr _var = { \
1387 .attr = __ATTR(_name, 0444, _show, NULL), \
1391 #define PMU_EVENT_ATTR_STRING(_name, _var, _str) \
1392 static struct perf_pmu_events_attr _var = { \
1393 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1395 .event_str = _str, \
1398 #define PMU_FORMAT_ATTR(_name, _format) \
1400 _name##_show(struct device *dev, \
1401 struct device_attribute *attr, \
1404 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
1405 return sprintf(page, _format "\n"); \
1408 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1410 /* Performance counter hotplug functions */
1411 #ifdef CONFIG_PERF_EVENTS
1412 int perf_event_init_cpu(unsigned int cpu);
1413 int perf_event_exit_cpu(unsigned int cpu);
1415 #define perf_event_init_cpu NULL
1416 #define perf_event_exit_cpu NULL
1419 #endif /* _LINUX_PERF_EVENT_H */