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);
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 <linux/security.h>
60 #include <asm/local.h>
62 struct perf_callchain_entry {
64 __u64 ip[0]; /* /proc/sys/kernel/perf_event_max_stack */
67 struct perf_callchain_entry_ctx {
68 struct perf_callchain_entry *entry;
75 typedef unsigned long (*perf_copy_f)(void *dst, const void *src,
76 unsigned long off, unsigned long len);
78 struct perf_raw_frag {
80 struct perf_raw_frag *next;
88 struct perf_raw_record {
89 struct perf_raw_frag frag;
94 * branch stack layout:
95 * nr: number of taken branches stored in entries[]
97 * Note that nr can vary from sample to sample
98 * branches (to, from) are stored from most recent
99 * to least recent, i.e., entries[0] contains the most
102 struct perf_branch_stack {
104 struct perf_branch_entry entries[0];
110 * extra PMU register associated with an event
112 struct hw_perf_event_extra {
113 u64 config; /* register value */
114 unsigned int reg; /* register address or index */
115 int alloc; /* extra register already allocated */
116 int idx; /* index in shared_regs->regs[] */
120 * struct hw_perf_event - performance event hardware details:
122 struct hw_perf_event {
123 #ifdef CONFIG_PERF_EVENTS
125 struct { /* hardware */
128 unsigned long config_base;
129 unsigned long event_base;
130 int event_base_rdpmc;
135 struct hw_perf_event_extra extra_reg;
136 struct hw_perf_event_extra branch_reg;
138 struct { /* software */
139 struct hrtimer hrtimer;
141 struct { /* tracepoint */
142 /* for tp_event->class */
143 struct list_head tp_list;
145 struct { /* amd_power */
149 #ifdef CONFIG_HAVE_HW_BREAKPOINT
150 struct { /* breakpoint */
152 * Crufty hack to avoid the chicken and egg
153 * problem hw_breakpoint has with context
154 * creation and event initalization.
156 struct arch_hw_breakpoint info;
157 struct list_head bp_list;
160 struct { /* amd_iommu */
169 * If the event is a per task event, this will point to the task in
170 * question. See the comment in perf_event_alloc().
172 struct task_struct *target;
175 * PMU would store hardware filter configuration
180 /* Last sync'ed generation of filters */
181 unsigned long addr_filters_gen;
184 * hw_perf_event::state flags; used to track the PERF_EF_* state.
186 #define PERF_HES_STOPPED 0x01 /* the counter is stopped */
187 #define PERF_HES_UPTODATE 0x02 /* event->count up-to-date */
188 #define PERF_HES_ARCH 0x04
193 * The last observed hardware counter value, updated with a
194 * local64_cmpxchg() such that pmu::read() can be called nested.
196 local64_t prev_count;
199 * The period to start the next sample with.
204 * The period we started this sample with.
209 * However much is left of the current period; note that this is
210 * a full 64bit value and allows for generation of periods longer
211 * than hardware might allow.
213 local64_t period_left;
216 * State for throttling the event, see __perf_event_overflow() and
217 * perf_adjust_freq_unthr_context().
223 * State for freq target events, see __perf_event_overflow() and
224 * perf_adjust_freq_unthr_context().
227 u64 freq_count_stamp;
234 * Common implementation detail of pmu::{start,commit,cancel}_txn
236 #define PERF_PMU_TXN_ADD 0x1 /* txn to add/schedule event on PMU */
237 #define PERF_PMU_TXN_READ 0x2 /* txn to read event group from PMU */
240 * pmu::capabilities flags
242 #define PERF_PMU_CAP_NO_INTERRUPT 0x01
243 #define PERF_PMU_CAP_NO_NMI 0x02
244 #define PERF_PMU_CAP_AUX_NO_SG 0x04
245 #define PERF_PMU_CAP_EXTENDED_REGS 0x08
246 #define PERF_PMU_CAP_EXCLUSIVE 0x10
247 #define PERF_PMU_CAP_ITRACE 0x20
248 #define PERF_PMU_CAP_HETEROGENEOUS_CPUS 0x40
249 #define PERF_PMU_CAP_NO_EXCLUDE 0x80
250 #define PERF_PMU_CAP_AUX_OUTPUT 0x100
252 struct perf_output_handle;
255 * struct pmu - generic performance monitoring unit
258 struct list_head entry;
260 struct module *module;
262 const struct attribute_group **attr_groups;
263 const struct attribute_group **attr_update;
268 * various common per-pmu feature flags
272 int __percpu *pmu_disable_count;
273 struct perf_cpu_context __percpu *pmu_cpu_context;
274 atomic_t exclusive_cnt; /* < 0: cpu; > 0: tsk */
276 int hrtimer_interval_ms;
278 /* number of address filters this PMU can do */
279 unsigned int nr_addr_filters;
282 * Fully disable/enable this PMU, can be used to protect from the PMI
283 * as well as for lazy/batch writing of the MSRs.
285 void (*pmu_enable) (struct pmu *pmu); /* optional */
286 void (*pmu_disable) (struct pmu *pmu); /* optional */
289 * Try and initialize the event for this PMU.
292 * -ENOENT -- @event is not for this PMU
294 * -ENODEV -- @event is for this PMU but PMU not present
295 * -EBUSY -- @event is for this PMU but PMU temporarily unavailable
296 * -EINVAL -- @event is for this PMU but @event is not valid
297 * -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
298 * -EACCES -- @event is for this PMU, @event is valid, but no privileges
300 * 0 -- @event is for this PMU and valid
302 * Other error return values are allowed.
304 int (*event_init) (struct perf_event *event);
307 * Notification that the event was mapped or unmapped. Called
308 * in the context of the mapping task.
310 void (*event_mapped) (struct perf_event *event, struct mm_struct *mm); /* optional */
311 void (*event_unmapped) (struct perf_event *event, struct mm_struct *mm); /* optional */
314 * Flags for ->add()/->del()/ ->start()/->stop(). There are
315 * matching hw_perf_event::state flags.
317 #define PERF_EF_START 0x01 /* start the counter when adding */
318 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
319 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
322 * Adds/Removes a counter to/from the PMU, can be done inside a
323 * transaction, see the ->*_txn() methods.
325 * The add/del callbacks will reserve all hardware resources required
326 * to service the event, this includes any counter constraint
329 * Called with IRQs disabled and the PMU disabled on the CPU the event
332 * ->add() called without PERF_EF_START should result in the same state
333 * as ->add() followed by ->stop().
335 * ->del() must always PERF_EF_UPDATE stop an event. If it calls
336 * ->stop() that must deal with already being stopped without
339 int (*add) (struct perf_event *event, int flags);
340 void (*del) (struct perf_event *event, int flags);
343 * Starts/Stops a counter present on the PMU.
345 * The PMI handler should stop the counter when perf_event_overflow()
346 * returns !0. ->start() will be used to continue.
348 * Also used to change the sample period.
350 * Called with IRQs disabled and the PMU disabled on the CPU the event
351 * is on -- will be called from NMI context with the PMU generates
354 * ->stop() with PERF_EF_UPDATE will read the counter and update
355 * period/count values like ->read() would.
357 * ->start() with PERF_EF_RELOAD will reprogram the the counter
358 * value, must be preceded by a ->stop() with PERF_EF_UPDATE.
360 void (*start) (struct perf_event *event, int flags);
361 void (*stop) (struct perf_event *event, int flags);
364 * Updates the counter value of the event.
366 * For sampling capable PMUs this will also update the software period
367 * hw_perf_event::period_left field.
369 void (*read) (struct perf_event *event);
372 * Group events scheduling is treated as a transaction, add
373 * group events as a whole and perform one schedulability test.
374 * If the test fails, roll back the whole group
376 * Start the transaction, after this ->add() doesn't need to
377 * do schedulability tests.
381 void (*start_txn) (struct pmu *pmu, unsigned int txn_flags);
383 * If ->start_txn() disabled the ->add() schedulability test
384 * then ->commit_txn() is required to perform one. On success
385 * the transaction is closed. On error the transaction is kept
386 * open until ->cancel_txn() is called.
390 int (*commit_txn) (struct pmu *pmu);
392 * Will cancel the transaction, assumes ->del() is called
393 * for each successful ->add() during the transaction.
397 void (*cancel_txn) (struct pmu *pmu);
400 * Will return the value for perf_event_mmap_page::index for this event,
401 * if no implementation is provided it will default to: event->hw.idx + 1.
403 int (*event_idx) (struct perf_event *event); /*optional */
406 * context-switches callback
408 void (*sched_task) (struct perf_event_context *ctx,
411 * PMU specific data size
413 size_t task_ctx_size;
416 * PMU specific parts of task perf event context (i.e. ctx->task_ctx_data)
417 * can be synchronized using this function. See Intel LBR callstack support
418 * implementation and Perf core context switch handling callbacks for usage
421 void (*swap_task_ctx) (struct perf_event_context *prev,
422 struct perf_event_context *next);
426 * Set up pmu-private data structures for an AUX area
428 void *(*setup_aux) (struct perf_event *event, void **pages,
429 int nr_pages, bool overwrite);
433 * Free pmu-private AUX data structures
435 void (*free_aux) (void *aux); /* optional */
438 * Take a snapshot of the AUX buffer without touching the event
439 * state, so that preempting ->start()/->stop() callbacks does
440 * not interfere with their logic. Called in PMI context.
442 * Returns the size of AUX data copied to the output handle.
446 long (*snapshot_aux) (struct perf_event *event,
447 struct perf_output_handle *handle,
451 * Validate address range filters: make sure the HW supports the
452 * requested configuration and number of filters; return 0 if the
453 * supplied filters are valid, -errno otherwise.
455 * Runs in the context of the ioctl()ing process and is not serialized
456 * with the rest of the PMU callbacks.
458 int (*addr_filters_validate) (struct list_head *filters);
462 * Synchronize address range filter configuration:
463 * translate hw-agnostic filters into hardware configuration in
464 * event::hw::addr_filters.
466 * Runs as a part of filter sync sequence that is done in ->start()
467 * callback by calling perf_event_addr_filters_sync().
469 * May (and should) traverse event::addr_filters::list, for which its
470 * caller provides necessary serialization.
472 void (*addr_filters_sync) (struct perf_event *event);
476 * Check if event can be used for aux_output purposes for
477 * events of this PMU.
479 * Runs from perf_event_open(). Should return 0 for "no match"
480 * or non-zero for "match".
482 int (*aux_output_match) (struct perf_event *event);
486 * Filter events for PMU-specific reasons.
488 int (*filter_match) (struct perf_event *event); /* optional */
491 * Check period value for PERF_EVENT_IOC_PERIOD ioctl.
493 int (*check_period) (struct perf_event *event, u64 value); /* optional */
496 enum perf_addr_filter_action_t {
497 PERF_ADDR_FILTER_ACTION_STOP = 0,
498 PERF_ADDR_FILTER_ACTION_START,
499 PERF_ADDR_FILTER_ACTION_FILTER,
503 * struct perf_addr_filter - address range filter definition
504 * @entry: event's filter list linkage
505 * @path: object file's path for file-based filters
506 * @offset: filter range offset
507 * @size: filter range size (size==0 means single address trigger)
508 * @action: filter/start/stop
510 * This is a hardware-agnostic filter configuration as specified by the user.
512 struct perf_addr_filter {
513 struct list_head entry;
515 unsigned long offset;
517 enum perf_addr_filter_action_t action;
521 * struct perf_addr_filters_head - container for address range filters
522 * @list: list of filters for this event
523 * @lock: spinlock that serializes accesses to the @list and event's
524 * (and its children's) filter generations.
525 * @nr_file_filters: number of file-based filters
527 * A child event will use parent's @list (and therefore @lock), so they are
528 * bundled together; see perf_event_addr_filters().
530 struct perf_addr_filters_head {
531 struct list_head list;
533 unsigned int nr_file_filters;
536 struct perf_addr_filter_range {
542 * enum perf_event_state - the states of an event:
544 enum perf_event_state {
545 PERF_EVENT_STATE_DEAD = -4,
546 PERF_EVENT_STATE_EXIT = -3,
547 PERF_EVENT_STATE_ERROR = -2,
548 PERF_EVENT_STATE_OFF = -1,
549 PERF_EVENT_STATE_INACTIVE = 0,
550 PERF_EVENT_STATE_ACTIVE = 1,
554 struct perf_sample_data;
556 typedef void (*perf_overflow_handler_t)(struct perf_event *,
557 struct perf_sample_data *,
558 struct pt_regs *regs);
561 * Event capabilities. For event_caps and groups caps.
563 * PERF_EV_CAP_SOFTWARE: Is a software event.
564 * PERF_EV_CAP_READ_ACTIVE_PKG: A CPU event (or cgroup event) that can be read
565 * from any CPU in the package where it is active.
567 #define PERF_EV_CAP_SOFTWARE BIT(0)
568 #define PERF_EV_CAP_READ_ACTIVE_PKG BIT(1)
570 #define SWEVENT_HLIST_BITS 8
571 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
573 struct swevent_hlist {
574 struct hlist_head heads[SWEVENT_HLIST_SIZE];
575 struct rcu_head rcu_head;
578 #define PERF_ATTACH_CONTEXT 0x01
579 #define PERF_ATTACH_GROUP 0x02
580 #define PERF_ATTACH_TASK 0x04
581 #define PERF_ATTACH_TASK_DATA 0x08
582 #define PERF_ATTACH_ITRACE 0x10
587 struct pmu_event_list {
589 struct list_head list;
592 #define for_each_sibling_event(sibling, event) \
593 if ((event)->group_leader == (event)) \
594 list_for_each_entry((sibling), &(event)->sibling_list, sibling_list)
597 * struct perf_event - performance event kernel representation:
600 #ifdef CONFIG_PERF_EVENTS
602 * entry onto perf_event_context::event_list;
603 * modifications require ctx->lock
604 * RCU safe iterations.
606 struct list_head event_entry;
609 * Locked for modification by both ctx->mutex and ctx->lock; holding
610 * either sufficies for read.
612 struct list_head sibling_list;
613 struct list_head active_list;
615 * Node on the pinned or flexible tree located at the event context;
617 struct rb_node group_node;
620 * We need storage to track the entries in perf_pmu_migrate_context; we
621 * cannot use the event_entry because of RCU and we want to keep the
622 * group in tact which avoids us using the other two entries.
624 struct list_head migrate_entry;
626 struct hlist_node hlist_entry;
627 struct list_head active_entry;
630 /* Not serialized. Only written during event initialization. */
632 /* The cumulative AND of all event_caps for events in this group. */
635 struct perf_event *group_leader;
639 enum perf_event_state state;
640 unsigned int attach_state;
642 atomic64_t child_count;
645 * These are the total time in nanoseconds that the event
646 * has been enabled (i.e. eligible to run, and the task has
647 * been scheduled in, if this is a per-task event)
648 * and running (scheduled onto the CPU), respectively.
650 u64 total_time_enabled;
651 u64 total_time_running;
655 * timestamp shadows the actual context timing but it can
656 * be safely used in NMI interrupt context. It reflects the
657 * context time as it was when the event was last scheduled in.
659 * ctx_time already accounts for ctx->timestamp. Therefore to
660 * compute ctx_time for a sample, simply add perf_clock().
664 struct perf_event_attr attr;
668 struct hw_perf_event hw;
670 struct perf_event_context *ctx;
671 atomic_long_t refcount;
674 * These accumulate total time (in nanoseconds) that children
675 * events have been enabled and running, respectively.
677 atomic64_t child_total_time_enabled;
678 atomic64_t child_total_time_running;
681 * Protect attach/detach and child_list:
683 struct mutex child_mutex;
684 struct list_head child_list;
685 struct perf_event *parent;
690 struct list_head owner_entry;
691 struct task_struct *owner;
694 struct mutex mmap_mutex;
697 struct ring_buffer *rb;
698 struct list_head rb_entry;
699 unsigned long rcu_batches;
703 wait_queue_head_t waitq;
704 struct fasync_struct *fasync;
706 /* delayed work for NMIs and such */
710 struct irq_work pending;
712 atomic_t event_limit;
714 /* address range filters */
715 struct perf_addr_filters_head addr_filters;
716 /* vma address array for file-based filders */
717 struct perf_addr_filter_range *addr_filter_ranges;
718 unsigned long addr_filters_gen;
720 /* for aux_output events */
721 struct perf_event *aux_event;
723 void (*destroy)(struct perf_event *);
724 struct rcu_head rcu_head;
726 struct pid_namespace *ns;
730 perf_overflow_handler_t overflow_handler;
731 void *overflow_handler_context;
732 #ifdef CONFIG_BPF_SYSCALL
733 perf_overflow_handler_t orig_overflow_handler;
734 struct bpf_prog *prog;
737 #ifdef CONFIG_EVENT_TRACING
738 struct trace_event_call *tp_event;
739 struct event_filter *filter;
740 #ifdef CONFIG_FUNCTION_TRACER
741 struct ftrace_ops ftrace_ops;
745 #ifdef CONFIG_CGROUP_PERF
746 struct perf_cgroup *cgrp; /* cgroup event is attach to */
749 #ifdef CONFIG_SECURITY
752 struct list_head sb_list;
753 #endif /* CONFIG_PERF_EVENTS */
757 struct perf_event_groups {
763 * struct perf_event_context - event context structure
765 * Used as a container for task events and CPU events as well:
767 struct perf_event_context {
770 * Protect the states of the events in the list,
771 * nr_active, and the list:
775 * Protect the list of events. Locking either mutex or lock
776 * is sufficient to ensure the list doesn't change; to change
777 * the list you need to lock both the mutex and the spinlock.
781 struct list_head active_ctx_list;
782 struct perf_event_groups pinned_groups;
783 struct perf_event_groups flexible_groups;
784 struct list_head event_list;
786 struct list_head pinned_active;
787 struct list_head flexible_active;
796 * Set when nr_events != nr_active, except tolerant to events not
797 * necessary to be active due to scheduling constraints, such as cgroups.
799 int rotate_necessary;
801 struct task_struct *task;
804 * Context clock, runs when context enabled.
810 * These fields let us detect when two contexts have both
811 * been cloned (inherited) from a common ancestor.
813 struct perf_event_context *parent_ctx;
817 #ifdef CONFIG_CGROUP_PERF
818 int nr_cgroups; /* cgroup evts */
820 void *task_ctx_data; /* pmu specific data */
821 struct rcu_head rcu_head;
825 * Number of contexts where an event can trigger:
826 * task, softirq, hardirq, nmi.
828 #define PERF_NR_CONTEXTS 4
831 * struct perf_event_cpu_context - per cpu event context structure
833 struct perf_cpu_context {
834 struct perf_event_context ctx;
835 struct perf_event_context *task_ctx;
839 raw_spinlock_t hrtimer_lock;
840 struct hrtimer hrtimer;
841 ktime_t hrtimer_interval;
842 unsigned int hrtimer_active;
844 #ifdef CONFIG_CGROUP_PERF
845 struct perf_cgroup *cgrp;
846 struct list_head cgrp_cpuctx_entry;
849 struct list_head sched_cb_entry;
855 struct perf_output_handle {
856 struct perf_event *event;
857 struct ring_buffer *rb;
858 unsigned long wakeup;
868 struct bpf_perf_event_data_kern {
869 bpf_user_pt_regs_t *regs;
870 struct perf_sample_data *data;
871 struct perf_event *event;
874 #ifdef CONFIG_CGROUP_PERF
877 * perf_cgroup_info keeps track of time_enabled for a cgroup.
878 * This is a per-cpu dynamically allocated data structure.
880 struct perf_cgroup_info {
886 struct cgroup_subsys_state css;
887 struct perf_cgroup_info __percpu *info;
891 * Must ensure cgroup is pinned (css_get) before calling
892 * this function. In other words, we cannot call this function
893 * if there is no cgroup event for the current CPU context.
895 static inline struct perf_cgroup *
896 perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
898 return container_of(task_css_check(task, perf_event_cgrp_id,
899 ctx ? lockdep_is_held(&ctx->lock)
901 struct perf_cgroup, css);
903 #endif /* CONFIG_CGROUP_PERF */
905 #ifdef CONFIG_PERF_EVENTS
907 extern void *perf_aux_output_begin(struct perf_output_handle *handle,
908 struct perf_event *event);
909 extern void perf_aux_output_end(struct perf_output_handle *handle,
911 extern int perf_aux_output_skip(struct perf_output_handle *handle,
913 extern void *perf_get_aux(struct perf_output_handle *handle);
914 extern void perf_aux_output_flag(struct perf_output_handle *handle, u64 flags);
915 extern void perf_event_itrace_started(struct perf_event *event);
917 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
918 extern void perf_pmu_unregister(struct pmu *pmu);
920 extern int perf_num_counters(void);
921 extern const char *perf_pmu_name(void);
922 extern void __perf_event_task_sched_in(struct task_struct *prev,
923 struct task_struct *task);
924 extern void __perf_event_task_sched_out(struct task_struct *prev,
925 struct task_struct *next);
926 extern int perf_event_init_task(struct task_struct *child);
927 extern void perf_event_exit_task(struct task_struct *child);
928 extern void perf_event_free_task(struct task_struct *task);
929 extern void perf_event_delayed_put(struct task_struct *task);
930 extern struct file *perf_event_get(unsigned int fd);
931 extern const struct perf_event *perf_get_event(struct file *file);
932 extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
933 extern void perf_event_print_debug(void);
934 extern void perf_pmu_disable(struct pmu *pmu);
935 extern void perf_pmu_enable(struct pmu *pmu);
936 extern void perf_sched_cb_dec(struct pmu *pmu);
937 extern void perf_sched_cb_inc(struct pmu *pmu);
938 extern int perf_event_task_disable(void);
939 extern int perf_event_task_enable(void);
941 extern void perf_pmu_resched(struct pmu *pmu);
943 extern int perf_event_refresh(struct perf_event *event, int refresh);
944 extern void perf_event_update_userpage(struct perf_event *event);
945 extern int perf_event_release_kernel(struct perf_event *event);
946 extern struct perf_event *
947 perf_event_create_kernel_counter(struct perf_event_attr *attr,
949 struct task_struct *task,
950 perf_overflow_handler_t callback,
952 extern void perf_pmu_migrate_context(struct pmu *pmu,
953 int src_cpu, int dst_cpu);
954 int perf_event_read_local(struct perf_event *event, u64 *value,
955 u64 *enabled, u64 *running);
956 extern u64 perf_event_read_value(struct perf_event *event,
957 u64 *enabled, u64 *running);
960 struct perf_sample_data {
962 * Fields set by perf_sample_data_init(), group so as to
963 * minimize the cachelines touched.
966 struct perf_raw_record *raw;
967 struct perf_branch_stack *br_stack;
971 union perf_mem_data_src data_src;
974 * The other fields, optionally {set,used} by
975 * perf_{prepare,output}_sample().
990 struct perf_callchain_entry *callchain;
994 * regs_user may point to task_pt_regs or to regs_user_copy, depending
997 struct perf_regs regs_user;
998 struct pt_regs regs_user_copy;
1000 struct perf_regs regs_intr;
1001 u64 stack_user_size;
1004 } ____cacheline_aligned;
1006 /* default value for data source */
1007 #define PERF_MEM_NA (PERF_MEM_S(OP, NA) |\
1008 PERF_MEM_S(LVL, NA) |\
1009 PERF_MEM_S(SNOOP, NA) |\
1010 PERF_MEM_S(LOCK, NA) |\
1011 PERF_MEM_S(TLB, NA))
1013 static inline void perf_sample_data_init(struct perf_sample_data *data,
1014 u64 addr, u64 period)
1016 /* remaining struct members initialized in perf_prepare_sample() */
1019 data->br_stack = NULL;
1020 data->period = period;
1022 data->data_src.val = PERF_MEM_NA;
1026 extern void perf_output_sample(struct perf_output_handle *handle,
1027 struct perf_event_header *header,
1028 struct perf_sample_data *data,
1029 struct perf_event *event);
1030 extern void perf_prepare_sample(struct perf_event_header *header,
1031 struct perf_sample_data *data,
1032 struct perf_event *event,
1033 struct pt_regs *regs);
1035 extern int perf_event_overflow(struct perf_event *event,
1036 struct perf_sample_data *data,
1037 struct pt_regs *regs);
1039 extern void perf_event_output_forward(struct perf_event *event,
1040 struct perf_sample_data *data,
1041 struct pt_regs *regs);
1042 extern void perf_event_output_backward(struct perf_event *event,
1043 struct perf_sample_data *data,
1044 struct pt_regs *regs);
1045 extern int perf_event_output(struct perf_event *event,
1046 struct perf_sample_data *data,
1047 struct pt_regs *regs);
1050 is_default_overflow_handler(struct perf_event *event)
1052 if (likely(event->overflow_handler == perf_event_output_forward))
1054 if (unlikely(event->overflow_handler == perf_event_output_backward))
1060 perf_event_header__init_id(struct perf_event_header *header,
1061 struct perf_sample_data *data,
1062 struct perf_event *event);
1064 perf_event__output_id_sample(struct perf_event *event,
1065 struct perf_output_handle *handle,
1066 struct perf_sample_data *sample);
1069 perf_log_lost_samples(struct perf_event *event, u64 lost);
1071 static inline bool event_has_any_exclude_flag(struct perf_event *event)
1073 struct perf_event_attr *attr = &event->attr;
1075 return attr->exclude_idle || attr->exclude_user ||
1076 attr->exclude_kernel || attr->exclude_hv ||
1077 attr->exclude_guest || attr->exclude_host;
1080 static inline bool is_sampling_event(struct perf_event *event)
1082 return event->attr.sample_period != 0;
1086 * Return 1 for a software event, 0 for a hardware event
1088 static inline int is_software_event(struct perf_event *event)
1090 return event->event_caps & PERF_EV_CAP_SOFTWARE;
1094 * Return 1 for event in sw context, 0 for event in hw context
1096 static inline int in_software_context(struct perf_event *event)
1098 return event->ctx->pmu->task_ctx_nr == perf_sw_context;
1101 static inline int is_exclusive_pmu(struct pmu *pmu)
1103 return pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE;
1106 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
1108 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
1109 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
1111 #ifndef perf_arch_fetch_caller_regs
1112 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
1116 * When generating a perf sample in-line, instead of from an interrupt /
1117 * exception, we lack a pt_regs. This is typically used from software events
1118 * like: SW_CONTEXT_SWITCHES, SW_MIGRATIONS and the tie-in with tracepoints.
1120 * We typically don't need a full set, but (for x86) do require:
1121 * - ip for PERF_SAMPLE_IP
1122 * - cs for user_mode() tests
1123 * - sp for PERF_SAMPLE_CALLCHAIN
1124 * - eflags for MISC bits and CALLCHAIN (see: perf_hw_regs())
1126 * NOTE: assumes @regs is otherwise already 0 filled; this is important for
1127 * things like PERF_SAMPLE_REGS_INTR.
1129 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1131 perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1134 static __always_inline void
1135 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1137 if (static_key_false(&perf_swevent_enabled[event_id]))
1138 __perf_sw_event(event_id, nr, regs, addr);
1141 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
1144 * 'Special' version for the scheduler, it hard assumes no recursion,
1145 * which is guaranteed by us not actually scheduling inside other swevents
1146 * because those disable preemption.
1148 static __always_inline void
1149 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
1151 if (static_key_false(&perf_swevent_enabled[event_id])) {
1152 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1154 perf_fetch_caller_regs(regs);
1155 ___perf_sw_event(event_id, nr, regs, addr);
1159 extern struct static_key_false perf_sched_events;
1161 static __always_inline bool
1162 perf_sw_migrate_enabled(void)
1164 if (static_key_false(&perf_swevent_enabled[PERF_COUNT_SW_CPU_MIGRATIONS]))
1169 static inline void perf_event_task_migrate(struct task_struct *task)
1171 if (perf_sw_migrate_enabled())
1172 task->sched_migrated = 1;
1175 static inline void perf_event_task_sched_in(struct task_struct *prev,
1176 struct task_struct *task)
1178 if (static_branch_unlikely(&perf_sched_events))
1179 __perf_event_task_sched_in(prev, task);
1181 if (perf_sw_migrate_enabled() && task->sched_migrated) {
1182 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1184 perf_fetch_caller_regs(regs);
1185 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, regs, 0);
1186 task->sched_migrated = 0;
1190 static inline void perf_event_task_sched_out(struct task_struct *prev,
1191 struct task_struct *next)
1193 perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
1195 if (static_branch_unlikely(&perf_sched_events))
1196 __perf_event_task_sched_out(prev, next);
1199 extern void perf_event_mmap(struct vm_area_struct *vma);
1201 extern void perf_event_ksymbol(u16 ksym_type, u64 addr, u32 len,
1202 bool unregister, const char *sym);
1203 extern void perf_event_bpf_event(struct bpf_prog *prog,
1204 enum perf_bpf_event_type type,
1207 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1208 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1209 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1211 extern void perf_event_exec(void);
1212 extern void perf_event_comm(struct task_struct *tsk, bool exec);
1213 extern void perf_event_namespaces(struct task_struct *tsk);
1214 extern void perf_event_fork(struct task_struct *tsk);
1217 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1219 extern void perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1220 extern void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1221 extern struct perf_callchain_entry *
1222 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
1223 u32 max_stack, bool crosstask, bool add_mark);
1224 extern struct perf_callchain_entry *perf_callchain(struct perf_event *event, struct pt_regs *regs);
1225 extern int get_callchain_buffers(int max_stack);
1226 extern void put_callchain_buffers(void);
1228 extern int sysctl_perf_event_max_stack;
1229 extern int sysctl_perf_event_max_contexts_per_stack;
1231 static inline int perf_callchain_store_context(struct perf_callchain_entry_ctx *ctx, u64 ip)
1233 if (ctx->contexts < sysctl_perf_event_max_contexts_per_stack) {
1234 struct perf_callchain_entry *entry = ctx->entry;
1235 entry->ip[entry->nr++] = ip;
1239 ctx->contexts_maxed = true;
1240 return -1; /* no more room, stop walking the stack */
1244 static inline int perf_callchain_store(struct perf_callchain_entry_ctx *ctx, u64 ip)
1246 if (ctx->nr < ctx->max_stack && !ctx->contexts_maxed) {
1247 struct perf_callchain_entry *entry = ctx->entry;
1248 entry->ip[entry->nr++] = ip;
1252 return -1; /* no more room, stop walking the stack */
1256 extern int sysctl_perf_event_paranoid;
1257 extern int sysctl_perf_event_mlock;
1258 extern int sysctl_perf_event_sample_rate;
1259 extern int sysctl_perf_cpu_time_max_percent;
1261 extern void perf_sample_event_took(u64 sample_len_ns);
1263 extern int perf_proc_update_handler(struct ctl_table *table, int write,
1264 void __user *buffer, size_t *lenp,
1266 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
1267 void __user *buffer, size_t *lenp,
1270 int perf_event_max_stack_handler(struct ctl_table *table, int write,
1271 void __user *buffer, size_t *lenp, loff_t *ppos);
1273 /* Access to perf_event_open(2) syscall. */
1274 #define PERF_SECURITY_OPEN 0
1276 /* Finer grained perf_event_open(2) access control. */
1277 #define PERF_SECURITY_CPU 1
1278 #define PERF_SECURITY_KERNEL 2
1279 #define PERF_SECURITY_TRACEPOINT 3
1281 static inline int perf_is_paranoid(void)
1283 return sysctl_perf_event_paranoid > -1;
1286 static inline int perf_allow_kernel(struct perf_event_attr *attr)
1288 if (sysctl_perf_event_paranoid > 1 && !capable(CAP_SYS_ADMIN))
1291 return security_perf_event_open(attr, PERF_SECURITY_KERNEL);
1294 static inline int perf_allow_cpu(struct perf_event_attr *attr)
1296 if (sysctl_perf_event_paranoid > 0 && !capable(CAP_SYS_ADMIN))
1299 return security_perf_event_open(attr, PERF_SECURITY_CPU);
1302 static inline int perf_allow_tracepoint(struct perf_event_attr *attr)
1304 if (sysctl_perf_event_paranoid > -1 && !capable(CAP_SYS_ADMIN))
1307 return security_perf_event_open(attr, PERF_SECURITY_TRACEPOINT);
1310 extern void perf_event_init(void);
1311 extern void perf_tp_event(u16 event_type, u64 count, void *record,
1312 int entry_size, struct pt_regs *regs,
1313 struct hlist_head *head, int rctx,
1314 struct task_struct *task);
1315 extern void perf_bp_event(struct perf_event *event, void *data);
1317 #ifndef perf_misc_flags
1318 # define perf_misc_flags(regs) \
1319 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1320 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1322 #ifndef perf_arch_bpf_user_pt_regs
1323 # define perf_arch_bpf_user_pt_regs(regs) regs
1326 static inline bool has_branch_stack(struct perf_event *event)
1328 return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1331 static inline bool needs_branch_stack(struct perf_event *event)
1333 return event->attr.branch_sample_type != 0;
1336 static inline bool has_aux(struct perf_event *event)
1338 return event->pmu->setup_aux;
1341 static inline bool is_write_backward(struct perf_event *event)
1343 return !!event->attr.write_backward;
1346 static inline bool has_addr_filter(struct perf_event *event)
1348 return event->pmu->nr_addr_filters;
1352 * An inherited event uses parent's filters
1354 static inline struct perf_addr_filters_head *
1355 perf_event_addr_filters(struct perf_event *event)
1357 struct perf_addr_filters_head *ifh = &event->addr_filters;
1360 ifh = &event->parent->addr_filters;
1365 extern void perf_event_addr_filters_sync(struct perf_event *event);
1367 extern int perf_output_begin(struct perf_output_handle *handle,
1368 struct perf_event *event, unsigned int size);
1369 extern int perf_output_begin_forward(struct perf_output_handle *handle,
1370 struct perf_event *event,
1372 extern int perf_output_begin_backward(struct perf_output_handle *handle,
1373 struct perf_event *event,
1376 extern void perf_output_end(struct perf_output_handle *handle);
1377 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
1378 const void *buf, unsigned int len);
1379 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
1381 extern long perf_output_copy_aux(struct perf_output_handle *aux_handle,
1382 struct perf_output_handle *handle,
1383 unsigned long from, unsigned long to);
1384 extern int perf_swevent_get_recursion_context(void);
1385 extern void perf_swevent_put_recursion_context(int rctx);
1386 extern u64 perf_swevent_set_period(struct perf_event *event);
1387 extern void perf_event_enable(struct perf_event *event);
1388 extern void perf_event_disable(struct perf_event *event);
1389 extern void perf_event_disable_local(struct perf_event *event);
1390 extern void perf_event_disable_inatomic(struct perf_event *event);
1391 extern void perf_event_task_tick(void);
1392 extern int perf_event_account_interrupt(struct perf_event *event);
1393 extern int perf_event_period(struct perf_event *event, u64 value);
1394 extern u64 perf_event_pause(struct perf_event *event, bool reset);
1395 #else /* !CONFIG_PERF_EVENTS: */
1396 static inline void *
1397 perf_aux_output_begin(struct perf_output_handle *handle,
1398 struct perf_event *event) { return NULL; }
1400 perf_aux_output_end(struct perf_output_handle *handle, unsigned long size)
1403 perf_aux_output_skip(struct perf_output_handle *handle,
1404 unsigned long size) { return -EINVAL; }
1405 static inline void *
1406 perf_get_aux(struct perf_output_handle *handle) { return NULL; }
1408 perf_event_task_migrate(struct task_struct *task) { }
1410 perf_event_task_sched_in(struct task_struct *prev,
1411 struct task_struct *task) { }
1413 perf_event_task_sched_out(struct task_struct *prev,
1414 struct task_struct *next) { }
1415 static inline int perf_event_init_task(struct task_struct *child) { return 0; }
1416 static inline void perf_event_exit_task(struct task_struct *child) { }
1417 static inline void perf_event_free_task(struct task_struct *task) { }
1418 static inline void perf_event_delayed_put(struct task_struct *task) { }
1419 static inline struct file *perf_event_get(unsigned int fd) { return ERR_PTR(-EINVAL); }
1420 static inline const struct perf_event *perf_get_event(struct file *file)
1422 return ERR_PTR(-EINVAL);
1424 static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
1426 return ERR_PTR(-EINVAL);
1428 static inline int perf_event_read_local(struct perf_event *event, u64 *value,
1429 u64 *enabled, u64 *running)
1433 static inline void perf_event_print_debug(void) { }
1434 static inline int perf_event_task_disable(void) { return -EINVAL; }
1435 static inline int perf_event_task_enable(void) { return -EINVAL; }
1436 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1442 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) { }
1444 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr) { }
1446 perf_bp_event(struct perf_event *event, void *data) { }
1448 static inline int perf_register_guest_info_callbacks
1449 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1450 static inline int perf_unregister_guest_info_callbacks
1451 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1453 static inline void perf_event_mmap(struct vm_area_struct *vma) { }
1455 typedef int (perf_ksymbol_get_name_f)(char *name, int name_len, void *data);
1456 static inline void perf_event_ksymbol(u16 ksym_type, u64 addr, u32 len,
1457 bool unregister, const char *sym) { }
1458 static inline void perf_event_bpf_event(struct bpf_prog *prog,
1459 enum perf_bpf_event_type type,
1461 static inline void perf_event_exec(void) { }
1462 static inline void perf_event_comm(struct task_struct *tsk, bool exec) { }
1463 static inline void perf_event_namespaces(struct task_struct *tsk) { }
1464 static inline void perf_event_fork(struct task_struct *tsk) { }
1465 static inline void perf_event_init(void) { }
1466 static inline int perf_swevent_get_recursion_context(void) { return -1; }
1467 static inline void perf_swevent_put_recursion_context(int rctx) { }
1468 static inline u64 perf_swevent_set_period(struct perf_event *event) { return 0; }
1469 static inline void perf_event_enable(struct perf_event *event) { }
1470 static inline void perf_event_disable(struct perf_event *event) { }
1471 static inline int __perf_event_disable(void *info) { return -1; }
1472 static inline void perf_event_task_tick(void) { }
1473 static inline int perf_event_release_kernel(struct perf_event *event) { return 0; }
1474 static inline int perf_event_period(struct perf_event *event, u64 value)
1478 static inline u64 perf_event_pause(struct perf_event *event, bool reset)
1484 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1485 extern void perf_restore_debug_store(void);
1487 static inline void perf_restore_debug_store(void) { }
1490 static __always_inline bool perf_raw_frag_last(const struct perf_raw_frag *frag)
1492 return frag->pad < sizeof(u64);
1495 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1497 struct perf_pmu_events_attr {
1498 struct device_attribute attr;
1500 const char *event_str;
1503 struct perf_pmu_events_ht_attr {
1504 struct device_attribute attr;
1506 const char *event_str_ht;
1507 const char *event_str_noht;
1510 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1513 #define PMU_EVENT_ATTR(_name, _var, _id, _show) \
1514 static struct perf_pmu_events_attr _var = { \
1515 .attr = __ATTR(_name, 0444, _show, NULL), \
1519 #define PMU_EVENT_ATTR_STRING(_name, _var, _str) \
1520 static struct perf_pmu_events_attr _var = { \
1521 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1523 .event_str = _str, \
1526 #define PMU_FORMAT_ATTR(_name, _format) \
1528 _name##_show(struct device *dev, \
1529 struct device_attribute *attr, \
1532 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
1533 return sprintf(page, _format "\n"); \
1536 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1538 /* Performance counter hotplug functions */
1539 #ifdef CONFIG_PERF_EVENTS
1540 int perf_event_init_cpu(unsigned int cpu);
1541 int perf_event_exit_cpu(unsigned int cpu);
1543 #define perf_event_init_cpu NULL
1544 #define perf_event_exit_cpu NULL
1547 #endif /* _LINUX_PERF_EVENT_H */