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
249 #define PERF_PMU_CAP_AUX_OUTPUT 0x100
252 * struct pmu - generic performance monitoring unit
255 struct list_head entry;
257 struct module *module;
259 const struct attribute_group **attr_groups;
260 const struct attribute_group **attr_update;
265 * various common per-pmu feature flags
269 int __percpu *pmu_disable_count;
270 struct perf_cpu_context __percpu *pmu_cpu_context;
271 atomic_t exclusive_cnt; /* < 0: cpu; > 0: tsk */
273 int hrtimer_interval_ms;
275 /* number of address filters this PMU can do */
276 unsigned int nr_addr_filters;
279 * Fully disable/enable this PMU, can be used to protect from the PMI
280 * as well as for lazy/batch writing of the MSRs.
282 void (*pmu_enable) (struct pmu *pmu); /* optional */
283 void (*pmu_disable) (struct pmu *pmu); /* optional */
286 * Try and initialize the event for this PMU.
289 * -ENOENT -- @event is not for this PMU
291 * -ENODEV -- @event is for this PMU but PMU not present
292 * -EBUSY -- @event is for this PMU but PMU temporarily unavailable
293 * -EINVAL -- @event is for this PMU but @event is not valid
294 * -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
295 * -EACCES -- @event is for this PMU, @event is valid, but no privileges
297 * 0 -- @event is for this PMU and valid
299 * Other error return values are allowed.
301 int (*event_init) (struct perf_event *event);
304 * Notification that the event was mapped or unmapped. Called
305 * in the context of the mapping task.
307 void (*event_mapped) (struct perf_event *event, struct mm_struct *mm); /* optional */
308 void (*event_unmapped) (struct perf_event *event, struct mm_struct *mm); /* optional */
311 * Flags for ->add()/->del()/ ->start()/->stop(). There are
312 * matching hw_perf_event::state flags.
314 #define PERF_EF_START 0x01 /* start the counter when adding */
315 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
316 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
319 * Adds/Removes a counter to/from the PMU, can be done inside a
320 * transaction, see the ->*_txn() methods.
322 * The add/del callbacks will reserve all hardware resources required
323 * to service the event, this includes any counter constraint
326 * Called with IRQs disabled and the PMU disabled on the CPU the event
329 * ->add() called without PERF_EF_START should result in the same state
330 * as ->add() followed by ->stop().
332 * ->del() must always PERF_EF_UPDATE stop an event. If it calls
333 * ->stop() that must deal with already being stopped without
336 int (*add) (struct perf_event *event, int flags);
337 void (*del) (struct perf_event *event, int flags);
340 * Starts/Stops a counter present on the PMU.
342 * The PMI handler should stop the counter when perf_event_overflow()
343 * returns !0. ->start() will be used to continue.
345 * Also used to change the sample period.
347 * Called with IRQs disabled and the PMU disabled on the CPU the event
348 * is on -- will be called from NMI context with the PMU generates
351 * ->stop() with PERF_EF_UPDATE will read the counter and update
352 * period/count values like ->read() would.
354 * ->start() with PERF_EF_RELOAD will reprogram the the counter
355 * value, must be preceded by a ->stop() with PERF_EF_UPDATE.
357 void (*start) (struct perf_event *event, int flags);
358 void (*stop) (struct perf_event *event, int flags);
361 * Updates the counter value of the event.
363 * For sampling capable PMUs this will also update the software period
364 * hw_perf_event::period_left field.
366 void (*read) (struct perf_event *event);
369 * Group events scheduling is treated as a transaction, add
370 * group events as a whole and perform one schedulability test.
371 * If the test fails, roll back the whole group
373 * Start the transaction, after this ->add() doesn't need to
374 * do schedulability tests.
378 void (*start_txn) (struct pmu *pmu, unsigned int txn_flags);
380 * If ->start_txn() disabled the ->add() schedulability test
381 * then ->commit_txn() is required to perform one. On success
382 * the transaction is closed. On error the transaction is kept
383 * open until ->cancel_txn() is called.
387 int (*commit_txn) (struct pmu *pmu);
389 * Will cancel the transaction, assumes ->del() is called
390 * for each successful ->add() during the transaction.
394 void (*cancel_txn) (struct pmu *pmu);
397 * Will return the value for perf_event_mmap_page::index for this event,
398 * if no implementation is provided it will default to: event->hw.idx + 1.
400 int (*event_idx) (struct perf_event *event); /*optional */
403 * context-switches callback
405 void (*sched_task) (struct perf_event_context *ctx,
408 * PMU specific data size
410 size_t task_ctx_size;
414 * Set up pmu-private data structures for an AUX area
416 void *(*setup_aux) (struct perf_event *event, void **pages,
417 int nr_pages, bool overwrite);
421 * Free pmu-private AUX data structures
423 void (*free_aux) (void *aux); /* optional */
426 * Validate address range filters: make sure the HW supports the
427 * requested configuration and number of filters; return 0 if the
428 * supplied filters are valid, -errno otherwise.
430 * Runs in the context of the ioctl()ing process and is not serialized
431 * with the rest of the PMU callbacks.
433 int (*addr_filters_validate) (struct list_head *filters);
437 * Synchronize address range filter configuration:
438 * translate hw-agnostic filters into hardware configuration in
439 * event::hw::addr_filters.
441 * Runs as a part of filter sync sequence that is done in ->start()
442 * callback by calling perf_event_addr_filters_sync().
444 * May (and should) traverse event::addr_filters::list, for which its
445 * caller provides necessary serialization.
447 void (*addr_filters_sync) (struct perf_event *event);
451 * Check if event can be used for aux_output purposes for
452 * events of this PMU.
454 * Runs from perf_event_open(). Should return 0 for "no match"
455 * or non-zero for "match".
457 int (*aux_output_match) (struct perf_event *event);
461 * Filter events for PMU-specific reasons.
463 int (*filter_match) (struct perf_event *event); /* optional */
466 * Check period value for PERF_EVENT_IOC_PERIOD ioctl.
468 int (*check_period) (struct perf_event *event, u64 value); /* optional */
471 enum perf_addr_filter_action_t {
472 PERF_ADDR_FILTER_ACTION_STOP = 0,
473 PERF_ADDR_FILTER_ACTION_START,
474 PERF_ADDR_FILTER_ACTION_FILTER,
478 * struct perf_addr_filter - address range filter definition
479 * @entry: event's filter list linkage
480 * @path: object file's path for file-based filters
481 * @offset: filter range offset
482 * @size: filter range size (size==0 means single address trigger)
483 * @action: filter/start/stop
485 * This is a hardware-agnostic filter configuration as specified by the user.
487 struct perf_addr_filter {
488 struct list_head entry;
490 unsigned long offset;
492 enum perf_addr_filter_action_t action;
496 * struct perf_addr_filters_head - container for address range filters
497 * @list: list of filters for this event
498 * @lock: spinlock that serializes accesses to the @list and event's
499 * (and its children's) filter generations.
500 * @nr_file_filters: number of file-based filters
502 * A child event will use parent's @list (and therefore @lock), so they are
503 * bundled together; see perf_event_addr_filters().
505 struct perf_addr_filters_head {
506 struct list_head list;
508 unsigned int nr_file_filters;
511 struct perf_addr_filter_range {
517 * enum perf_event_state - the states of an event:
519 enum perf_event_state {
520 PERF_EVENT_STATE_DEAD = -4,
521 PERF_EVENT_STATE_EXIT = -3,
522 PERF_EVENT_STATE_ERROR = -2,
523 PERF_EVENT_STATE_OFF = -1,
524 PERF_EVENT_STATE_INACTIVE = 0,
525 PERF_EVENT_STATE_ACTIVE = 1,
529 struct perf_sample_data;
531 typedef void (*perf_overflow_handler_t)(struct perf_event *,
532 struct perf_sample_data *,
533 struct pt_regs *regs);
536 * Event capabilities. For event_caps and groups caps.
538 * PERF_EV_CAP_SOFTWARE: Is a software event.
539 * PERF_EV_CAP_READ_ACTIVE_PKG: A CPU event (or cgroup event) that can be read
540 * from any CPU in the package where it is active.
542 #define PERF_EV_CAP_SOFTWARE BIT(0)
543 #define PERF_EV_CAP_READ_ACTIVE_PKG BIT(1)
545 #define SWEVENT_HLIST_BITS 8
546 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
548 struct swevent_hlist {
549 struct hlist_head heads[SWEVENT_HLIST_SIZE];
550 struct rcu_head rcu_head;
553 #define PERF_ATTACH_CONTEXT 0x01
554 #define PERF_ATTACH_GROUP 0x02
555 #define PERF_ATTACH_TASK 0x04
556 #define PERF_ATTACH_TASK_DATA 0x08
557 #define PERF_ATTACH_ITRACE 0x10
562 struct pmu_event_list {
564 struct list_head list;
567 #define for_each_sibling_event(sibling, event) \
568 if ((event)->group_leader == (event)) \
569 list_for_each_entry((sibling), &(event)->sibling_list, sibling_list)
572 * struct perf_event - performance event kernel representation:
575 #ifdef CONFIG_PERF_EVENTS
577 * entry onto perf_event_context::event_list;
578 * modifications require ctx->lock
579 * RCU safe iterations.
581 struct list_head event_entry;
584 * Locked for modification by both ctx->mutex and ctx->lock; holding
585 * either sufficies for read.
587 struct list_head sibling_list;
588 struct list_head active_list;
590 * Node on the pinned or flexible tree located at the event context;
592 struct rb_node group_node;
595 * We need storage to track the entries in perf_pmu_migrate_context; we
596 * cannot use the event_entry because of RCU and we want to keep the
597 * group in tact which avoids us using the other two entries.
599 struct list_head migrate_entry;
601 struct hlist_node hlist_entry;
602 struct list_head active_entry;
605 /* Not serialized. Only written during event initialization. */
607 /* The cumulative AND of all event_caps for events in this group. */
610 struct perf_event *group_leader;
614 enum perf_event_state state;
615 unsigned int attach_state;
617 atomic64_t child_count;
620 * These are the total time in nanoseconds that the event
621 * has been enabled (i.e. eligible to run, and the task has
622 * been scheduled in, if this is a per-task event)
623 * and running (scheduled onto the CPU), respectively.
625 u64 total_time_enabled;
626 u64 total_time_running;
630 * timestamp shadows the actual context timing but it can
631 * be safely used in NMI interrupt context. It reflects the
632 * context time as it was when the event was last scheduled in.
634 * ctx_time already accounts for ctx->timestamp. Therefore to
635 * compute ctx_time for a sample, simply add perf_clock().
639 struct perf_event_attr attr;
643 struct hw_perf_event hw;
645 struct perf_event_context *ctx;
646 atomic_long_t refcount;
649 * These accumulate total time (in nanoseconds) that children
650 * events have been enabled and running, respectively.
652 atomic64_t child_total_time_enabled;
653 atomic64_t child_total_time_running;
656 * Protect attach/detach and child_list:
658 struct mutex child_mutex;
659 struct list_head child_list;
660 struct perf_event *parent;
665 struct list_head owner_entry;
666 struct task_struct *owner;
669 struct mutex mmap_mutex;
672 struct ring_buffer *rb;
673 struct list_head rb_entry;
674 unsigned long rcu_batches;
678 wait_queue_head_t waitq;
679 struct fasync_struct *fasync;
681 /* delayed work for NMIs and such */
685 struct irq_work pending;
687 atomic_t event_limit;
689 /* address range filters */
690 struct perf_addr_filters_head addr_filters;
691 /* vma address array for file-based filders */
692 struct perf_addr_filter_range *addr_filter_ranges;
693 unsigned long addr_filters_gen;
695 /* for aux_output events */
696 struct perf_event *aux_event;
698 void (*destroy)(struct perf_event *);
699 struct rcu_head rcu_head;
701 struct pid_namespace *ns;
705 perf_overflow_handler_t overflow_handler;
706 void *overflow_handler_context;
707 #ifdef CONFIG_BPF_SYSCALL
708 perf_overflow_handler_t orig_overflow_handler;
709 struct bpf_prog *prog;
712 #ifdef CONFIG_EVENT_TRACING
713 struct trace_event_call *tp_event;
714 struct event_filter *filter;
715 #ifdef CONFIG_FUNCTION_TRACER
716 struct ftrace_ops ftrace_ops;
720 #ifdef CONFIG_CGROUP_PERF
721 struct perf_cgroup *cgrp; /* cgroup event is attach to */
724 struct list_head sb_list;
725 #endif /* CONFIG_PERF_EVENTS */
729 struct perf_event_groups {
735 * struct perf_event_context - event context structure
737 * Used as a container for task events and CPU events as well:
739 struct perf_event_context {
742 * Protect the states of the events in the list,
743 * nr_active, and the list:
747 * Protect the list of events. Locking either mutex or lock
748 * is sufficient to ensure the list doesn't change; to change
749 * the list you need to lock both the mutex and the spinlock.
753 struct list_head active_ctx_list;
754 struct perf_event_groups pinned_groups;
755 struct perf_event_groups flexible_groups;
756 struct list_head event_list;
758 struct list_head pinned_active;
759 struct list_head flexible_active;
768 * Set when nr_events != nr_active, except tolerant to events not
769 * necessary to be active due to scheduling constraints, such as cgroups.
771 int rotate_necessary;
773 struct task_struct *task;
776 * Context clock, runs when context enabled.
782 * These fields let us detect when two contexts have both
783 * been cloned (inherited) from a common ancestor.
785 struct perf_event_context *parent_ctx;
789 #ifdef CONFIG_CGROUP_PERF
790 int nr_cgroups; /* cgroup evts */
792 void *task_ctx_data; /* pmu specific data */
793 struct rcu_head rcu_head;
797 * Number of contexts where an event can trigger:
798 * task, softirq, hardirq, nmi.
800 #define PERF_NR_CONTEXTS 4
803 * struct perf_event_cpu_context - per cpu event context structure
805 struct perf_cpu_context {
806 struct perf_event_context ctx;
807 struct perf_event_context *task_ctx;
811 raw_spinlock_t hrtimer_lock;
812 struct hrtimer hrtimer;
813 ktime_t hrtimer_interval;
814 unsigned int hrtimer_active;
816 #ifdef CONFIG_CGROUP_PERF
817 struct perf_cgroup *cgrp;
818 struct list_head cgrp_cpuctx_entry;
821 struct list_head sched_cb_entry;
827 struct perf_output_handle {
828 struct perf_event *event;
829 struct ring_buffer *rb;
830 unsigned long wakeup;
840 struct bpf_perf_event_data_kern {
841 bpf_user_pt_regs_t *regs;
842 struct perf_sample_data *data;
843 struct perf_event *event;
846 #ifdef CONFIG_CGROUP_PERF
849 * perf_cgroup_info keeps track of time_enabled for a cgroup.
850 * This is a per-cpu dynamically allocated data structure.
852 struct perf_cgroup_info {
858 struct cgroup_subsys_state css;
859 struct perf_cgroup_info __percpu *info;
863 * Must ensure cgroup is pinned (css_get) before calling
864 * this function. In other words, we cannot call this function
865 * if there is no cgroup event for the current CPU context.
867 static inline struct perf_cgroup *
868 perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
870 return container_of(task_css_check(task, perf_event_cgrp_id,
871 ctx ? lockdep_is_held(&ctx->lock)
873 struct perf_cgroup, css);
875 #endif /* CONFIG_CGROUP_PERF */
877 #ifdef CONFIG_PERF_EVENTS
879 extern void *perf_aux_output_begin(struct perf_output_handle *handle,
880 struct perf_event *event);
881 extern void perf_aux_output_end(struct perf_output_handle *handle,
883 extern int perf_aux_output_skip(struct perf_output_handle *handle,
885 extern void *perf_get_aux(struct perf_output_handle *handle);
886 extern void perf_aux_output_flag(struct perf_output_handle *handle, u64 flags);
887 extern void perf_event_itrace_started(struct perf_event *event);
889 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
890 extern void perf_pmu_unregister(struct pmu *pmu);
892 extern int perf_num_counters(void);
893 extern const char *perf_pmu_name(void);
894 extern void __perf_event_task_sched_in(struct task_struct *prev,
895 struct task_struct *task);
896 extern void __perf_event_task_sched_out(struct task_struct *prev,
897 struct task_struct *next);
898 extern int perf_event_init_task(struct task_struct *child);
899 extern void perf_event_exit_task(struct task_struct *child);
900 extern void perf_event_free_task(struct task_struct *task);
901 extern void perf_event_delayed_put(struct task_struct *task);
902 extern struct file *perf_event_get(unsigned int fd);
903 extern const struct perf_event *perf_get_event(struct file *file);
904 extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
905 extern void perf_event_print_debug(void);
906 extern void perf_pmu_disable(struct pmu *pmu);
907 extern void perf_pmu_enable(struct pmu *pmu);
908 extern void perf_sched_cb_dec(struct pmu *pmu);
909 extern void perf_sched_cb_inc(struct pmu *pmu);
910 extern int perf_event_task_disable(void);
911 extern int perf_event_task_enable(void);
913 extern void perf_pmu_resched(struct pmu *pmu);
915 extern int perf_event_refresh(struct perf_event *event, int refresh);
916 extern void perf_event_update_userpage(struct perf_event *event);
917 extern int perf_event_release_kernel(struct perf_event *event);
918 extern struct perf_event *
919 perf_event_create_kernel_counter(struct perf_event_attr *attr,
921 struct task_struct *task,
922 perf_overflow_handler_t callback,
924 extern void perf_pmu_migrate_context(struct pmu *pmu,
925 int src_cpu, int dst_cpu);
926 int perf_event_read_local(struct perf_event *event, u64 *value,
927 u64 *enabled, u64 *running);
928 extern u64 perf_event_read_value(struct perf_event *event,
929 u64 *enabled, u64 *running);
932 struct perf_sample_data {
934 * Fields set by perf_sample_data_init(), group so as to
935 * minimize the cachelines touched.
938 struct perf_raw_record *raw;
939 struct perf_branch_stack *br_stack;
943 union perf_mem_data_src data_src;
946 * The other fields, optionally {set,used} by
947 * perf_{prepare,output}_sample().
962 struct perf_callchain_entry *callchain;
965 * regs_user may point to task_pt_regs or to regs_user_copy, depending
968 struct perf_regs regs_user;
969 struct pt_regs regs_user_copy;
971 struct perf_regs regs_intr;
975 } ____cacheline_aligned;
977 /* default value for data source */
978 #define PERF_MEM_NA (PERF_MEM_S(OP, NA) |\
979 PERF_MEM_S(LVL, NA) |\
980 PERF_MEM_S(SNOOP, NA) |\
981 PERF_MEM_S(LOCK, NA) |\
984 static inline void perf_sample_data_init(struct perf_sample_data *data,
985 u64 addr, u64 period)
987 /* remaining struct members initialized in perf_prepare_sample() */
990 data->br_stack = NULL;
991 data->period = period;
993 data->data_src.val = PERF_MEM_NA;
997 extern void perf_output_sample(struct perf_output_handle *handle,
998 struct perf_event_header *header,
999 struct perf_sample_data *data,
1000 struct perf_event *event);
1001 extern void perf_prepare_sample(struct perf_event_header *header,
1002 struct perf_sample_data *data,
1003 struct perf_event *event,
1004 struct pt_regs *regs);
1006 extern int perf_event_overflow(struct perf_event *event,
1007 struct perf_sample_data *data,
1008 struct pt_regs *regs);
1010 extern void perf_event_output_forward(struct perf_event *event,
1011 struct perf_sample_data *data,
1012 struct pt_regs *regs);
1013 extern void perf_event_output_backward(struct perf_event *event,
1014 struct perf_sample_data *data,
1015 struct pt_regs *regs);
1016 extern int perf_event_output(struct perf_event *event,
1017 struct perf_sample_data *data,
1018 struct pt_regs *regs);
1021 is_default_overflow_handler(struct perf_event *event)
1023 if (likely(event->overflow_handler == perf_event_output_forward))
1025 if (unlikely(event->overflow_handler == perf_event_output_backward))
1031 perf_event_header__init_id(struct perf_event_header *header,
1032 struct perf_sample_data *data,
1033 struct perf_event *event);
1035 perf_event__output_id_sample(struct perf_event *event,
1036 struct perf_output_handle *handle,
1037 struct perf_sample_data *sample);
1040 perf_log_lost_samples(struct perf_event *event, u64 lost);
1042 static inline bool event_has_any_exclude_flag(struct perf_event *event)
1044 struct perf_event_attr *attr = &event->attr;
1046 return attr->exclude_idle || attr->exclude_user ||
1047 attr->exclude_kernel || attr->exclude_hv ||
1048 attr->exclude_guest || attr->exclude_host;
1051 static inline bool is_sampling_event(struct perf_event *event)
1053 return event->attr.sample_period != 0;
1057 * Return 1 for a software event, 0 for a hardware event
1059 static inline int is_software_event(struct perf_event *event)
1061 return event->event_caps & PERF_EV_CAP_SOFTWARE;
1065 * Return 1 for event in sw context, 0 for event in hw context
1067 static inline int in_software_context(struct perf_event *event)
1069 return event->ctx->pmu->task_ctx_nr == perf_sw_context;
1072 static inline int is_exclusive_pmu(struct pmu *pmu)
1074 return pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE;
1077 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
1079 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
1080 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
1082 #ifndef perf_arch_fetch_caller_regs
1083 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
1087 * When generating a perf sample in-line, instead of from an interrupt /
1088 * exception, we lack a pt_regs. This is typically used from software events
1089 * like: SW_CONTEXT_SWITCHES, SW_MIGRATIONS and the tie-in with tracepoints.
1091 * We typically don't need a full set, but (for x86) do require:
1092 * - ip for PERF_SAMPLE_IP
1093 * - cs for user_mode() tests
1094 * - sp for PERF_SAMPLE_CALLCHAIN
1095 * - eflags for MISC bits and CALLCHAIN (see: perf_hw_regs())
1097 * NOTE: assumes @regs is otherwise already 0 filled; this is important for
1098 * things like PERF_SAMPLE_REGS_INTR.
1100 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1102 perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1105 static __always_inline void
1106 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1108 if (static_key_false(&perf_swevent_enabled[event_id]))
1109 __perf_sw_event(event_id, nr, regs, addr);
1112 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
1115 * 'Special' version for the scheduler, it hard assumes no recursion,
1116 * which is guaranteed by us not actually scheduling inside other swevents
1117 * because those disable preemption.
1119 static __always_inline void
1120 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
1122 if (static_key_false(&perf_swevent_enabled[event_id])) {
1123 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1125 perf_fetch_caller_regs(regs);
1126 ___perf_sw_event(event_id, nr, regs, addr);
1130 extern struct static_key_false perf_sched_events;
1132 static __always_inline bool
1133 perf_sw_migrate_enabled(void)
1135 if (static_key_false(&perf_swevent_enabled[PERF_COUNT_SW_CPU_MIGRATIONS]))
1140 static inline void perf_event_task_migrate(struct task_struct *task)
1142 if (perf_sw_migrate_enabled())
1143 task->sched_migrated = 1;
1146 static inline void perf_event_task_sched_in(struct task_struct *prev,
1147 struct task_struct *task)
1149 if (static_branch_unlikely(&perf_sched_events))
1150 __perf_event_task_sched_in(prev, task);
1152 if (perf_sw_migrate_enabled() && task->sched_migrated) {
1153 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1155 perf_fetch_caller_regs(regs);
1156 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, regs, 0);
1157 task->sched_migrated = 0;
1161 static inline void perf_event_task_sched_out(struct task_struct *prev,
1162 struct task_struct *next)
1164 perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
1166 if (static_branch_unlikely(&perf_sched_events))
1167 __perf_event_task_sched_out(prev, next);
1170 extern void perf_event_mmap(struct vm_area_struct *vma);
1172 extern void perf_event_ksymbol(u16 ksym_type, u64 addr, u32 len,
1173 bool unregister, const char *sym);
1174 extern void perf_event_bpf_event(struct bpf_prog *prog,
1175 enum perf_bpf_event_type type,
1178 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1179 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1180 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1182 extern void perf_event_exec(void);
1183 extern void perf_event_comm(struct task_struct *tsk, bool exec);
1184 extern void perf_event_namespaces(struct task_struct *tsk);
1185 extern void perf_event_fork(struct task_struct *tsk);
1188 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1190 extern void perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1191 extern void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1192 extern struct perf_callchain_entry *
1193 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
1194 u32 max_stack, bool crosstask, bool add_mark);
1195 extern struct perf_callchain_entry *perf_callchain(struct perf_event *event, struct pt_regs *regs);
1196 extern int get_callchain_buffers(int max_stack);
1197 extern void put_callchain_buffers(void);
1199 extern int sysctl_perf_event_max_stack;
1200 extern int sysctl_perf_event_max_contexts_per_stack;
1202 static inline int perf_callchain_store_context(struct perf_callchain_entry_ctx *ctx, u64 ip)
1204 if (ctx->contexts < sysctl_perf_event_max_contexts_per_stack) {
1205 struct perf_callchain_entry *entry = ctx->entry;
1206 entry->ip[entry->nr++] = ip;
1210 ctx->contexts_maxed = true;
1211 return -1; /* no more room, stop walking the stack */
1215 static inline int perf_callchain_store(struct perf_callchain_entry_ctx *ctx, u64 ip)
1217 if (ctx->nr < ctx->max_stack && !ctx->contexts_maxed) {
1218 struct perf_callchain_entry *entry = ctx->entry;
1219 entry->ip[entry->nr++] = ip;
1223 return -1; /* no more room, stop walking the stack */
1227 extern int sysctl_perf_event_paranoid;
1228 extern int sysctl_perf_event_mlock;
1229 extern int sysctl_perf_event_sample_rate;
1230 extern int sysctl_perf_cpu_time_max_percent;
1232 extern void perf_sample_event_took(u64 sample_len_ns);
1234 extern int perf_proc_update_handler(struct ctl_table *table, int write,
1235 void __user *buffer, size_t *lenp,
1237 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
1238 void __user *buffer, size_t *lenp,
1241 int perf_event_max_stack_handler(struct ctl_table *table, int write,
1242 void __user *buffer, size_t *lenp, loff_t *ppos);
1244 static inline bool perf_paranoid_tracepoint_raw(void)
1246 return sysctl_perf_event_paranoid > -1;
1249 static inline bool perf_paranoid_cpu(void)
1251 return sysctl_perf_event_paranoid > 0;
1254 static inline bool perf_paranoid_kernel(void)
1256 return sysctl_perf_event_paranoid > 1;
1259 extern void perf_event_init(void);
1260 extern void perf_tp_event(u16 event_type, u64 count, void *record,
1261 int entry_size, struct pt_regs *regs,
1262 struct hlist_head *head, int rctx,
1263 struct task_struct *task);
1264 extern void perf_bp_event(struct perf_event *event, void *data);
1266 #ifndef perf_misc_flags
1267 # define perf_misc_flags(regs) \
1268 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1269 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1271 #ifndef perf_arch_bpf_user_pt_regs
1272 # define perf_arch_bpf_user_pt_regs(regs) regs
1275 static inline bool has_branch_stack(struct perf_event *event)
1277 return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1280 static inline bool needs_branch_stack(struct perf_event *event)
1282 return event->attr.branch_sample_type != 0;
1285 static inline bool has_aux(struct perf_event *event)
1287 return event->pmu->setup_aux;
1290 static inline bool is_write_backward(struct perf_event *event)
1292 return !!event->attr.write_backward;
1295 static inline bool has_addr_filter(struct perf_event *event)
1297 return event->pmu->nr_addr_filters;
1301 * An inherited event uses parent's filters
1303 static inline struct perf_addr_filters_head *
1304 perf_event_addr_filters(struct perf_event *event)
1306 struct perf_addr_filters_head *ifh = &event->addr_filters;
1309 ifh = &event->parent->addr_filters;
1314 extern void perf_event_addr_filters_sync(struct perf_event *event);
1316 extern int perf_output_begin(struct perf_output_handle *handle,
1317 struct perf_event *event, unsigned int size);
1318 extern int perf_output_begin_forward(struct perf_output_handle *handle,
1319 struct perf_event *event,
1321 extern int perf_output_begin_backward(struct perf_output_handle *handle,
1322 struct perf_event *event,
1325 extern void perf_output_end(struct perf_output_handle *handle);
1326 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
1327 const void *buf, unsigned int len);
1328 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
1330 extern int perf_swevent_get_recursion_context(void);
1331 extern void perf_swevent_put_recursion_context(int rctx);
1332 extern u64 perf_swevent_set_period(struct perf_event *event);
1333 extern void perf_event_enable(struct perf_event *event);
1334 extern void perf_event_disable(struct perf_event *event);
1335 extern void perf_event_disable_local(struct perf_event *event);
1336 extern void perf_event_disable_inatomic(struct perf_event *event);
1337 extern void perf_event_task_tick(void);
1338 extern int perf_event_account_interrupt(struct perf_event *event);
1339 extern int perf_event_period(struct perf_event *event, u64 value);
1340 extern u64 perf_event_pause(struct perf_event *event, bool reset);
1341 #else /* !CONFIG_PERF_EVENTS: */
1342 static inline void *
1343 perf_aux_output_begin(struct perf_output_handle *handle,
1344 struct perf_event *event) { return NULL; }
1346 perf_aux_output_end(struct perf_output_handle *handle, unsigned long size)
1349 perf_aux_output_skip(struct perf_output_handle *handle,
1350 unsigned long size) { return -EINVAL; }
1351 static inline void *
1352 perf_get_aux(struct perf_output_handle *handle) { return NULL; }
1354 perf_event_task_migrate(struct task_struct *task) { }
1356 perf_event_task_sched_in(struct task_struct *prev,
1357 struct task_struct *task) { }
1359 perf_event_task_sched_out(struct task_struct *prev,
1360 struct task_struct *next) { }
1361 static inline int perf_event_init_task(struct task_struct *child) { return 0; }
1362 static inline void perf_event_exit_task(struct task_struct *child) { }
1363 static inline void perf_event_free_task(struct task_struct *task) { }
1364 static inline void perf_event_delayed_put(struct task_struct *task) { }
1365 static inline struct file *perf_event_get(unsigned int fd) { return ERR_PTR(-EINVAL); }
1366 static inline const struct perf_event *perf_get_event(struct file *file)
1368 return ERR_PTR(-EINVAL);
1370 static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
1372 return ERR_PTR(-EINVAL);
1374 static inline int perf_event_read_local(struct perf_event *event, u64 *value,
1375 u64 *enabled, u64 *running)
1379 static inline void perf_event_print_debug(void) { }
1380 static inline int perf_event_task_disable(void) { return -EINVAL; }
1381 static inline int perf_event_task_enable(void) { return -EINVAL; }
1382 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1388 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) { }
1390 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr) { }
1392 perf_bp_event(struct perf_event *event, void *data) { }
1394 static inline int perf_register_guest_info_callbacks
1395 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1396 static inline int perf_unregister_guest_info_callbacks
1397 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1399 static inline void perf_event_mmap(struct vm_area_struct *vma) { }
1401 typedef int (perf_ksymbol_get_name_f)(char *name, int name_len, void *data);
1402 static inline void perf_event_ksymbol(u16 ksym_type, u64 addr, u32 len,
1403 bool unregister, const char *sym) { }
1404 static inline void perf_event_bpf_event(struct bpf_prog *prog,
1405 enum perf_bpf_event_type type,
1407 static inline void perf_event_exec(void) { }
1408 static inline void perf_event_comm(struct task_struct *tsk, bool exec) { }
1409 static inline void perf_event_namespaces(struct task_struct *tsk) { }
1410 static inline void perf_event_fork(struct task_struct *tsk) { }
1411 static inline void perf_event_init(void) { }
1412 static inline int perf_swevent_get_recursion_context(void) { return -1; }
1413 static inline void perf_swevent_put_recursion_context(int rctx) { }
1414 static inline u64 perf_swevent_set_period(struct perf_event *event) { return 0; }
1415 static inline void perf_event_enable(struct perf_event *event) { }
1416 static inline void perf_event_disable(struct perf_event *event) { }
1417 static inline int __perf_event_disable(void *info) { return -1; }
1418 static inline void perf_event_task_tick(void) { }
1419 static inline int perf_event_release_kernel(struct perf_event *event) { return 0; }
1420 static inline int perf_event_period(struct perf_event *event, u64 value)
1424 static inline u64 perf_event_pause(struct perf_event *event, bool reset)
1430 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1431 extern void perf_restore_debug_store(void);
1433 static inline void perf_restore_debug_store(void) { }
1436 static __always_inline bool perf_raw_frag_last(const struct perf_raw_frag *frag)
1438 return frag->pad < sizeof(u64);
1441 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1443 struct perf_pmu_events_attr {
1444 struct device_attribute attr;
1446 const char *event_str;
1449 struct perf_pmu_events_ht_attr {
1450 struct device_attribute attr;
1452 const char *event_str_ht;
1453 const char *event_str_noht;
1456 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1459 #define PMU_EVENT_ATTR(_name, _var, _id, _show) \
1460 static struct perf_pmu_events_attr _var = { \
1461 .attr = __ATTR(_name, 0444, _show, NULL), \
1465 #define PMU_EVENT_ATTR_STRING(_name, _var, _str) \
1466 static struct perf_pmu_events_attr _var = { \
1467 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1469 .event_str = _str, \
1472 #define PMU_FORMAT_ATTR(_name, _format) \
1474 _name##_show(struct device *dev, \
1475 struct device_attribute *attr, \
1478 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
1479 return sprintf(page, _format "\n"); \
1482 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1484 /* Performance counter hotplug functions */
1485 #ifdef CONFIG_PERF_EVENTS
1486 int perf_event_init_cpu(unsigned int cpu);
1487 int perf_event_exit_cpu(unsigned int cpu);
1489 #define perf_event_init_cpu NULL
1490 #define perf_event_exit_cpu NULL
1493 #endif /* _LINUX_PERF_EVENT_H */