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 enum perf_addr_filter_action_t {
453 PERF_ADDR_FILTER_ACTION_STOP = 0,
454 PERF_ADDR_FILTER_ACTION_START,
455 PERF_ADDR_FILTER_ACTION_FILTER,
459 * struct perf_addr_filter - address range filter definition
460 * @entry: event's filter list linkage
461 * @inode: object file's inode for file-based filters
462 * @offset: filter range offset
463 * @size: filter range size (size==0 means single address trigger)
464 * @action: filter/start/stop
466 * This is a hardware-agnostic filter configuration as specified by the user.
468 struct perf_addr_filter {
469 struct list_head entry;
471 unsigned long offset;
473 enum perf_addr_filter_action_t action;
477 * struct perf_addr_filters_head - container for address range filters
478 * @list: list of filters for this event
479 * @lock: spinlock that serializes accesses to the @list and event's
480 * (and its children's) filter generations.
481 * @nr_file_filters: number of file-based filters
483 * A child event will use parent's @list (and therefore @lock), so they are
484 * bundled together; see perf_event_addr_filters().
486 struct perf_addr_filters_head {
487 struct list_head list;
489 unsigned int nr_file_filters;
493 * enum perf_event_state - the states of a event
495 enum perf_event_state {
496 PERF_EVENT_STATE_DEAD = -4,
497 PERF_EVENT_STATE_EXIT = -3,
498 PERF_EVENT_STATE_ERROR = -2,
499 PERF_EVENT_STATE_OFF = -1,
500 PERF_EVENT_STATE_INACTIVE = 0,
501 PERF_EVENT_STATE_ACTIVE = 1,
505 struct perf_sample_data;
507 typedef void (*perf_overflow_handler_t)(struct perf_event *,
508 struct perf_sample_data *,
509 struct pt_regs *regs);
512 * Event capabilities. For event_caps and groups caps.
514 * PERF_EV_CAP_SOFTWARE: Is a software event.
515 * PERF_EV_CAP_READ_ACTIVE_PKG: A CPU event (or cgroup event) that can be read
516 * from any CPU in the package where it is active.
518 #define PERF_EV_CAP_SOFTWARE BIT(0)
519 #define PERF_EV_CAP_READ_ACTIVE_PKG BIT(1)
521 #define SWEVENT_HLIST_BITS 8
522 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
524 struct swevent_hlist {
525 struct hlist_head heads[SWEVENT_HLIST_SIZE];
526 struct rcu_head rcu_head;
529 #define PERF_ATTACH_CONTEXT 0x01
530 #define PERF_ATTACH_GROUP 0x02
531 #define PERF_ATTACH_TASK 0x04
532 #define PERF_ATTACH_TASK_DATA 0x08
533 #define PERF_ATTACH_ITRACE 0x10
538 struct pmu_event_list {
540 struct list_head list;
543 #define for_each_sibling_event(sibling, event) \
544 if ((event)->group_leader == (event)) \
545 list_for_each_entry((sibling), &(event)->sibling_list, sibling_list)
548 * struct perf_event - performance event kernel representation:
551 #ifdef CONFIG_PERF_EVENTS
553 * entry onto perf_event_context::event_list;
554 * modifications require ctx->lock
555 * RCU safe iterations.
557 struct list_head event_entry;
560 * Locked for modification by both ctx->mutex and ctx->lock; holding
561 * either sufficies for read.
563 struct list_head sibling_list;
564 struct list_head active_list;
566 * Node on the pinned or flexible tree located at the event context;
568 struct rb_node group_node;
571 * We need storage to track the entries in perf_pmu_migrate_context; we
572 * cannot use the event_entry because of RCU and we want to keep the
573 * group in tact which avoids us using the other two entries.
575 struct list_head migrate_entry;
577 struct hlist_node hlist_entry;
578 struct list_head active_entry;
581 /* Not serialized. Only written during event initialization. */
583 /* The cumulative AND of all event_caps for events in this group. */
586 struct perf_event *group_leader;
590 enum perf_event_state state;
591 unsigned int attach_state;
593 atomic64_t child_count;
596 * These are the total time in nanoseconds that the event
597 * has been enabled (i.e. eligible to run, and the task has
598 * been scheduled in, if this is a per-task event)
599 * and running (scheduled onto the CPU), respectively.
601 u64 total_time_enabled;
602 u64 total_time_running;
606 * timestamp shadows the actual context timing but it can
607 * be safely used in NMI interrupt context. It reflects the
608 * context time as it was when the event was last scheduled in.
610 * ctx_time already accounts for ctx->timestamp. Therefore to
611 * compute ctx_time for a sample, simply add perf_clock().
615 struct perf_event_attr attr;
619 struct hw_perf_event hw;
621 struct perf_event_context *ctx;
622 atomic_long_t refcount;
625 * These accumulate total time (in nanoseconds) that children
626 * events have been enabled and running, respectively.
628 atomic64_t child_total_time_enabled;
629 atomic64_t child_total_time_running;
632 * Protect attach/detach and child_list:
634 struct mutex child_mutex;
635 struct list_head child_list;
636 struct perf_event *parent;
641 struct list_head owner_entry;
642 struct task_struct *owner;
645 struct mutex mmap_mutex;
648 struct ring_buffer *rb;
649 struct list_head rb_entry;
650 unsigned long rcu_batches;
654 wait_queue_head_t waitq;
655 struct fasync_struct *fasync;
657 /* delayed work for NMIs and such */
661 struct irq_work pending;
663 atomic_t event_limit;
665 /* address range filters */
666 struct perf_addr_filters_head addr_filters;
667 /* vma address array for file-based filders */
668 unsigned long *addr_filters_offs;
669 unsigned long addr_filters_gen;
671 void (*destroy)(struct perf_event *);
672 struct rcu_head rcu_head;
674 struct pid_namespace *ns;
678 perf_overflow_handler_t overflow_handler;
679 void *overflow_handler_context;
680 #ifdef CONFIG_BPF_SYSCALL
681 perf_overflow_handler_t orig_overflow_handler;
682 struct bpf_prog *prog;
685 #ifdef CONFIG_EVENT_TRACING
686 struct trace_event_call *tp_event;
687 struct event_filter *filter;
688 #ifdef CONFIG_FUNCTION_TRACER
689 struct ftrace_ops ftrace_ops;
693 #ifdef CONFIG_CGROUP_PERF
694 struct perf_cgroup *cgrp; /* cgroup event is attach to */
697 struct list_head sb_list;
698 #endif /* CONFIG_PERF_EVENTS */
702 struct perf_event_groups {
708 * struct perf_event_context - event context structure
710 * Used as a container for task events and CPU events as well:
712 struct perf_event_context {
715 * Protect the states of the events in the list,
716 * nr_active, and the list:
720 * Protect the list of events. Locking either mutex or lock
721 * is sufficient to ensure the list doesn't change; to change
722 * the list you need to lock both the mutex and the spinlock.
726 struct list_head active_ctx_list;
727 struct perf_event_groups pinned_groups;
728 struct perf_event_groups flexible_groups;
729 struct list_head event_list;
731 struct list_head pinned_active;
732 struct list_head flexible_active;
741 struct task_struct *task;
744 * Context clock, runs when context enabled.
750 * These fields let us detect when two contexts have both
751 * been cloned (inherited) from a common ancestor.
753 struct perf_event_context *parent_ctx;
757 #ifdef CONFIG_CGROUP_PERF
758 int nr_cgroups; /* cgroup evts */
760 void *task_ctx_data; /* pmu specific data */
761 struct rcu_head rcu_head;
765 * Number of contexts where an event can trigger:
766 * task, softirq, hardirq, nmi.
768 #define PERF_NR_CONTEXTS 4
771 * struct perf_event_cpu_context - per cpu event context structure
773 struct perf_cpu_context {
774 struct perf_event_context ctx;
775 struct perf_event_context *task_ctx;
779 raw_spinlock_t hrtimer_lock;
780 struct hrtimer hrtimer;
781 ktime_t hrtimer_interval;
782 unsigned int hrtimer_active;
784 #ifdef CONFIG_CGROUP_PERF
785 struct perf_cgroup *cgrp;
786 struct list_head cgrp_cpuctx_entry;
789 struct list_head sched_cb_entry;
795 struct perf_output_handle {
796 struct perf_event *event;
797 struct ring_buffer *rb;
798 unsigned long wakeup;
808 struct bpf_perf_event_data_kern {
809 bpf_user_pt_regs_t *regs;
810 struct perf_sample_data *data;
811 struct perf_event *event;
814 #ifdef CONFIG_CGROUP_PERF
817 * perf_cgroup_info keeps track of time_enabled for a cgroup.
818 * This is a per-cpu dynamically allocated data structure.
820 struct perf_cgroup_info {
826 struct cgroup_subsys_state css;
827 struct perf_cgroup_info __percpu *info;
831 * Must ensure cgroup is pinned (css_get) before calling
832 * this function. In other words, we cannot call this function
833 * if there is no cgroup event for the current CPU context.
835 static inline struct perf_cgroup *
836 perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
838 return container_of(task_css_check(task, perf_event_cgrp_id,
839 ctx ? lockdep_is_held(&ctx->lock)
841 struct perf_cgroup, css);
843 #endif /* CONFIG_CGROUP_PERF */
845 #ifdef CONFIG_PERF_EVENTS
847 extern void *perf_aux_output_begin(struct perf_output_handle *handle,
848 struct perf_event *event);
849 extern void perf_aux_output_end(struct perf_output_handle *handle,
851 extern int perf_aux_output_skip(struct perf_output_handle *handle,
853 extern void *perf_get_aux(struct perf_output_handle *handle);
854 extern void perf_aux_output_flag(struct perf_output_handle *handle, u64 flags);
855 extern void perf_event_itrace_started(struct perf_event *event);
857 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
858 extern void perf_pmu_unregister(struct pmu *pmu);
860 extern int perf_num_counters(void);
861 extern const char *perf_pmu_name(void);
862 extern void __perf_event_task_sched_in(struct task_struct *prev,
863 struct task_struct *task);
864 extern void __perf_event_task_sched_out(struct task_struct *prev,
865 struct task_struct *next);
866 extern int perf_event_init_task(struct task_struct *child);
867 extern void perf_event_exit_task(struct task_struct *child);
868 extern void perf_event_free_task(struct task_struct *task);
869 extern void perf_event_delayed_put(struct task_struct *task);
870 extern struct file *perf_event_get(unsigned int fd);
871 extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
872 extern void perf_event_print_debug(void);
873 extern void perf_pmu_disable(struct pmu *pmu);
874 extern void perf_pmu_enable(struct pmu *pmu);
875 extern void perf_sched_cb_dec(struct pmu *pmu);
876 extern void perf_sched_cb_inc(struct pmu *pmu);
877 extern int perf_event_task_disable(void);
878 extern int perf_event_task_enable(void);
879 extern int perf_event_refresh(struct perf_event *event, int refresh);
880 extern void perf_event_update_userpage(struct perf_event *event);
881 extern int perf_event_release_kernel(struct perf_event *event);
882 extern struct perf_event *
883 perf_event_create_kernel_counter(struct perf_event_attr *attr,
885 struct task_struct *task,
886 perf_overflow_handler_t callback,
888 extern void perf_pmu_migrate_context(struct pmu *pmu,
889 int src_cpu, int dst_cpu);
890 int perf_event_read_local(struct perf_event *event, u64 *value,
891 u64 *enabled, u64 *running);
892 extern u64 perf_event_read_value(struct perf_event *event,
893 u64 *enabled, u64 *running);
896 struct perf_sample_data {
898 * Fields set by perf_sample_data_init(), group so as to
899 * minimize the cachelines touched.
902 struct perf_raw_record *raw;
903 struct perf_branch_stack *br_stack;
907 union perf_mem_data_src data_src;
910 * The other fields, optionally {set,used} by
911 * perf_{prepare,output}_sample().
926 struct perf_callchain_entry *callchain;
929 * regs_user may point to task_pt_regs or to regs_user_copy, depending
932 struct perf_regs regs_user;
933 struct pt_regs regs_user_copy;
935 struct perf_regs regs_intr;
939 } ____cacheline_aligned;
941 /* default value for data source */
942 #define PERF_MEM_NA (PERF_MEM_S(OP, NA) |\
943 PERF_MEM_S(LVL, NA) |\
944 PERF_MEM_S(SNOOP, NA) |\
945 PERF_MEM_S(LOCK, NA) |\
948 static inline void perf_sample_data_init(struct perf_sample_data *data,
949 u64 addr, u64 period)
951 /* remaining struct members initialized in perf_prepare_sample() */
954 data->br_stack = NULL;
955 data->period = period;
957 data->data_src.val = PERF_MEM_NA;
961 extern void perf_output_sample(struct perf_output_handle *handle,
962 struct perf_event_header *header,
963 struct perf_sample_data *data,
964 struct perf_event *event);
965 extern void perf_prepare_sample(struct perf_event_header *header,
966 struct perf_sample_data *data,
967 struct perf_event *event,
968 struct pt_regs *regs);
970 extern int perf_event_overflow(struct perf_event *event,
971 struct perf_sample_data *data,
972 struct pt_regs *regs);
974 extern void perf_event_output_forward(struct perf_event *event,
975 struct perf_sample_data *data,
976 struct pt_regs *regs);
977 extern void perf_event_output_backward(struct perf_event *event,
978 struct perf_sample_data *data,
979 struct pt_regs *regs);
980 extern void perf_event_output(struct perf_event *event,
981 struct perf_sample_data *data,
982 struct pt_regs *regs);
985 is_default_overflow_handler(struct perf_event *event)
987 if (likely(event->overflow_handler == perf_event_output_forward))
989 if (unlikely(event->overflow_handler == perf_event_output_backward))
995 perf_event_header__init_id(struct perf_event_header *header,
996 struct perf_sample_data *data,
997 struct perf_event *event);
999 perf_event__output_id_sample(struct perf_event *event,
1000 struct perf_output_handle *handle,
1001 struct perf_sample_data *sample);
1004 perf_log_lost_samples(struct perf_event *event, u64 lost);
1006 static inline bool is_sampling_event(struct perf_event *event)
1008 return event->attr.sample_period != 0;
1012 * Return 1 for a software event, 0 for a hardware event
1014 static inline int is_software_event(struct perf_event *event)
1016 return event->event_caps & PERF_EV_CAP_SOFTWARE;
1019 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
1021 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
1022 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
1024 #ifndef perf_arch_fetch_caller_regs
1025 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
1029 * Take a snapshot of the regs. Skip ip and frame pointer to
1030 * the nth caller. We only need a few of the regs:
1031 * - ip for PERF_SAMPLE_IP
1032 * - cs for user_mode() tests
1033 * - bp for callchains
1034 * - eflags, for future purposes, just in case
1036 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1038 perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1041 static __always_inline void
1042 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1044 if (static_key_false(&perf_swevent_enabled[event_id]))
1045 __perf_sw_event(event_id, nr, regs, addr);
1048 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
1051 * 'Special' version for the scheduler, it hard assumes no recursion,
1052 * which is guaranteed by us not actually scheduling inside other swevents
1053 * because those disable preemption.
1055 static __always_inline void
1056 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
1058 if (static_key_false(&perf_swevent_enabled[event_id])) {
1059 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1061 perf_fetch_caller_regs(regs);
1062 ___perf_sw_event(event_id, nr, regs, addr);
1066 extern struct static_key_false perf_sched_events;
1068 static __always_inline bool
1069 perf_sw_migrate_enabled(void)
1071 if (static_key_false(&perf_swevent_enabled[PERF_COUNT_SW_CPU_MIGRATIONS]))
1076 static inline void perf_event_task_migrate(struct task_struct *task)
1078 if (perf_sw_migrate_enabled())
1079 task->sched_migrated = 1;
1082 static inline void perf_event_task_sched_in(struct task_struct *prev,
1083 struct task_struct *task)
1085 if (static_branch_unlikely(&perf_sched_events))
1086 __perf_event_task_sched_in(prev, task);
1088 if (perf_sw_migrate_enabled() && task->sched_migrated) {
1089 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1091 perf_fetch_caller_regs(regs);
1092 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, regs, 0);
1093 task->sched_migrated = 0;
1097 static inline void perf_event_task_sched_out(struct task_struct *prev,
1098 struct task_struct *next)
1100 perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
1102 if (static_branch_unlikely(&perf_sched_events))
1103 __perf_event_task_sched_out(prev, next);
1106 extern void perf_event_mmap(struct vm_area_struct *vma);
1107 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1108 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1109 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1111 extern void perf_event_exec(void);
1112 extern void perf_event_comm(struct task_struct *tsk, bool exec);
1113 extern void perf_event_namespaces(struct task_struct *tsk);
1114 extern void perf_event_fork(struct task_struct *tsk);
1117 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1119 extern void perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1120 extern void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1121 extern struct perf_callchain_entry *
1122 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
1123 u32 max_stack, bool crosstask, bool add_mark);
1124 extern int get_callchain_buffers(int max_stack);
1125 extern void put_callchain_buffers(void);
1127 extern int sysctl_perf_event_max_stack;
1128 extern int sysctl_perf_event_max_contexts_per_stack;
1130 static inline int perf_callchain_store_context(struct perf_callchain_entry_ctx *ctx, u64 ip)
1132 if (ctx->contexts < sysctl_perf_event_max_contexts_per_stack) {
1133 struct perf_callchain_entry *entry = ctx->entry;
1134 entry->ip[entry->nr++] = ip;
1138 ctx->contexts_maxed = true;
1139 return -1; /* no more room, stop walking the stack */
1143 static inline int perf_callchain_store(struct perf_callchain_entry_ctx *ctx, u64 ip)
1145 if (ctx->nr < ctx->max_stack && !ctx->contexts_maxed) {
1146 struct perf_callchain_entry *entry = ctx->entry;
1147 entry->ip[entry->nr++] = ip;
1151 return -1; /* no more room, stop walking the stack */
1155 extern int sysctl_perf_event_paranoid;
1156 extern int sysctl_perf_event_mlock;
1157 extern int sysctl_perf_event_sample_rate;
1158 extern int sysctl_perf_cpu_time_max_percent;
1160 extern void perf_sample_event_took(u64 sample_len_ns);
1162 extern int perf_proc_update_handler(struct ctl_table *table, int write,
1163 void __user *buffer, size_t *lenp,
1165 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
1166 void __user *buffer, size_t *lenp,
1169 int perf_event_max_stack_handler(struct ctl_table *table, int write,
1170 void __user *buffer, size_t *lenp, loff_t *ppos);
1172 static inline bool perf_paranoid_tracepoint_raw(void)
1174 return sysctl_perf_event_paranoid > -1;
1177 static inline bool perf_paranoid_cpu(void)
1179 return sysctl_perf_event_paranoid > 0;
1182 static inline bool perf_paranoid_kernel(void)
1184 return sysctl_perf_event_paranoid > 1;
1187 extern void perf_event_init(void);
1188 extern void perf_tp_event(u16 event_type, u64 count, void *record,
1189 int entry_size, struct pt_regs *regs,
1190 struct hlist_head *head, int rctx,
1191 struct task_struct *task);
1192 extern void perf_bp_event(struct perf_event *event, void *data);
1194 #ifndef perf_misc_flags
1195 # define perf_misc_flags(regs) \
1196 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1197 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1199 #ifndef perf_arch_bpf_user_pt_regs
1200 # define perf_arch_bpf_user_pt_regs(regs) regs
1203 static inline bool has_branch_stack(struct perf_event *event)
1205 return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1208 static inline bool needs_branch_stack(struct perf_event *event)
1210 return event->attr.branch_sample_type != 0;
1213 static inline bool has_aux(struct perf_event *event)
1215 return event->pmu->setup_aux;
1218 static inline bool is_write_backward(struct perf_event *event)
1220 return !!event->attr.write_backward;
1223 static inline bool has_addr_filter(struct perf_event *event)
1225 return event->pmu->nr_addr_filters;
1229 * An inherited event uses parent's filters
1231 static inline struct perf_addr_filters_head *
1232 perf_event_addr_filters(struct perf_event *event)
1234 struct perf_addr_filters_head *ifh = &event->addr_filters;
1237 ifh = &event->parent->addr_filters;
1242 extern void perf_event_addr_filters_sync(struct perf_event *event);
1244 extern int perf_output_begin(struct perf_output_handle *handle,
1245 struct perf_event *event, unsigned int size);
1246 extern int perf_output_begin_forward(struct perf_output_handle *handle,
1247 struct perf_event *event,
1249 extern int perf_output_begin_backward(struct perf_output_handle *handle,
1250 struct perf_event *event,
1253 extern void perf_output_end(struct perf_output_handle *handle);
1254 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
1255 const void *buf, unsigned int len);
1256 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
1258 extern int perf_swevent_get_recursion_context(void);
1259 extern void perf_swevent_put_recursion_context(int rctx);
1260 extern u64 perf_swevent_set_period(struct perf_event *event);
1261 extern void perf_event_enable(struct perf_event *event);
1262 extern void perf_event_disable(struct perf_event *event);
1263 extern void perf_event_disable_local(struct perf_event *event);
1264 extern void perf_event_disable_inatomic(struct perf_event *event);
1265 extern void perf_event_task_tick(void);
1266 extern int perf_event_account_interrupt(struct perf_event *event);
1267 #else /* !CONFIG_PERF_EVENTS: */
1268 static inline void *
1269 perf_aux_output_begin(struct perf_output_handle *handle,
1270 struct perf_event *event) { return NULL; }
1272 perf_aux_output_end(struct perf_output_handle *handle, unsigned long size)
1275 perf_aux_output_skip(struct perf_output_handle *handle,
1276 unsigned long size) { return -EINVAL; }
1277 static inline void *
1278 perf_get_aux(struct perf_output_handle *handle) { return NULL; }
1280 perf_event_task_migrate(struct task_struct *task) { }
1282 perf_event_task_sched_in(struct task_struct *prev,
1283 struct task_struct *task) { }
1285 perf_event_task_sched_out(struct task_struct *prev,
1286 struct task_struct *next) { }
1287 static inline int perf_event_init_task(struct task_struct *child) { return 0; }
1288 static inline void perf_event_exit_task(struct task_struct *child) { }
1289 static inline void perf_event_free_task(struct task_struct *task) { }
1290 static inline void perf_event_delayed_put(struct task_struct *task) { }
1291 static inline struct file *perf_event_get(unsigned int fd) { return ERR_PTR(-EINVAL); }
1292 static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
1294 return ERR_PTR(-EINVAL);
1296 static inline int perf_event_read_local(struct perf_event *event, u64 *value,
1297 u64 *enabled, u64 *running)
1301 static inline void perf_event_print_debug(void) { }
1302 static inline int perf_event_task_disable(void) { return -EINVAL; }
1303 static inline int perf_event_task_enable(void) { return -EINVAL; }
1304 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1310 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) { }
1312 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr) { }
1314 perf_bp_event(struct perf_event *event, void *data) { }
1316 static inline int perf_register_guest_info_callbacks
1317 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1318 static inline int perf_unregister_guest_info_callbacks
1319 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1321 static inline void perf_event_mmap(struct vm_area_struct *vma) { }
1322 static inline void perf_event_exec(void) { }
1323 static inline void perf_event_comm(struct task_struct *tsk, bool exec) { }
1324 static inline void perf_event_namespaces(struct task_struct *tsk) { }
1325 static inline void perf_event_fork(struct task_struct *tsk) { }
1326 static inline void perf_event_init(void) { }
1327 static inline int perf_swevent_get_recursion_context(void) { return -1; }
1328 static inline void perf_swevent_put_recursion_context(int rctx) { }
1329 static inline u64 perf_swevent_set_period(struct perf_event *event) { return 0; }
1330 static inline void perf_event_enable(struct perf_event *event) { }
1331 static inline void perf_event_disable(struct perf_event *event) { }
1332 static inline int __perf_event_disable(void *info) { return -1; }
1333 static inline void perf_event_task_tick(void) { }
1334 static inline int perf_event_release_kernel(struct perf_event *event) { return 0; }
1337 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1338 extern void perf_restore_debug_store(void);
1340 static inline void perf_restore_debug_store(void) { }
1343 static __always_inline bool perf_raw_frag_last(const struct perf_raw_frag *frag)
1345 return frag->pad < sizeof(u64);
1348 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1350 struct perf_pmu_events_attr {
1351 struct device_attribute attr;
1353 const char *event_str;
1356 struct perf_pmu_events_ht_attr {
1357 struct device_attribute attr;
1359 const char *event_str_ht;
1360 const char *event_str_noht;
1363 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1366 #define PMU_EVENT_ATTR(_name, _var, _id, _show) \
1367 static struct perf_pmu_events_attr _var = { \
1368 .attr = __ATTR(_name, 0444, _show, NULL), \
1372 #define PMU_EVENT_ATTR_STRING(_name, _var, _str) \
1373 static struct perf_pmu_events_attr _var = { \
1374 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1376 .event_str = _str, \
1379 #define PMU_FORMAT_ATTR(_name, _format) \
1381 _name##_show(struct device *dev, \
1382 struct device_attribute *attr, \
1385 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
1386 return sprintf(page, _format "\n"); \
1389 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1391 /* Performance counter hotplug functions */
1392 #ifdef CONFIG_PERF_EVENTS
1393 int perf_event_init_cpu(unsigned int cpu);
1394 int perf_event_exit_cpu(unsigned int cpu);
1396 #define perf_event_init_cpu NULL
1397 #define perf_event_exit_cpu NULL
1400 #endif /* _LINUX_PERF_EVENT_H */