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
) (struct perf_event
*event
, void **pages
,
413 int nr_pages
, bool overwrite
);
417 * Free pmu-private AUX data structures
419 void (*free_aux
) (void *aux
); /* optional */
422 * Validate address range filters: make sure the HW supports the
423 * requested configuration and number of filters; return 0 if the
424 * supplied filters are valid, -errno otherwise.
426 * Runs in the context of the ioctl()ing process and is not serialized
427 * with the rest of the PMU callbacks.
429 int (*addr_filters_validate
) (struct list_head
*filters
);
433 * Synchronize address range filter configuration:
434 * translate hw-agnostic filters into hardware configuration in
435 * event::hw::addr_filters.
437 * Runs as a part of filter sync sequence that is done in ->start()
438 * callback by calling perf_event_addr_filters_sync().
440 * May (and should) traverse event::addr_filters::list, for which its
441 * caller provides necessary serialization.
443 void (*addr_filters_sync
) (struct perf_event
*event
);
447 * Filter events for PMU-specific reasons.
449 int (*filter_match
) (struct perf_event
*event
); /* optional */
452 * Check period value for PERF_EVENT_IOC_PERIOD ioctl.
454 int (*check_period
) (struct perf_event
*event
, u64 value
); /* optional */
458 * struct perf_addr_filter - address range filter definition
459 * @entry: event's filter list linkage
460 * @inode: object file's inode for file-based filters
461 * @offset: filter range offset
462 * @size: filter range size
463 * @range: 1: range, 0: address
464 * @filter: 1: filter/start, 0: 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 unsigned int range
: 1,
478 * struct perf_addr_filters_head - container for address range filters
479 * @list: list of filters for this event
480 * @lock: spinlock that serializes accesses to the @list and event's
481 * (and its children's) filter generations.
482 * @nr_file_filters: number of file-based filters
484 * A child event will use parent's @list (and therefore @lock), so they are
485 * bundled together; see perf_event_addr_filters().
487 struct perf_addr_filters_head
{
488 struct list_head list
;
490 unsigned int nr_file_filters
;
494 * enum perf_event_state - the states of a event
496 enum perf_event_state
{
497 PERF_EVENT_STATE_DEAD
= -4,
498 PERF_EVENT_STATE_EXIT
= -3,
499 PERF_EVENT_STATE_ERROR
= -2,
500 PERF_EVENT_STATE_OFF
= -1,
501 PERF_EVENT_STATE_INACTIVE
= 0,
502 PERF_EVENT_STATE_ACTIVE
= 1,
506 struct perf_sample_data
;
508 typedef void (*perf_overflow_handler_t
)(struct perf_event
*,
509 struct perf_sample_data
*,
510 struct pt_regs
*regs
);
513 * Event capabilities. For event_caps and groups caps.
515 * PERF_EV_CAP_SOFTWARE: Is a software event.
516 * PERF_EV_CAP_READ_ACTIVE_PKG: A CPU event (or cgroup event) that can be read
517 * from any CPU in the package where it is active.
519 #define PERF_EV_CAP_SOFTWARE BIT(0)
520 #define PERF_EV_CAP_READ_ACTIVE_PKG BIT(1)
522 #define SWEVENT_HLIST_BITS 8
523 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
525 struct swevent_hlist
{
526 struct hlist_head heads
[SWEVENT_HLIST_SIZE
];
527 struct rcu_head rcu_head
;
530 #define PERF_ATTACH_CONTEXT 0x01
531 #define PERF_ATTACH_GROUP 0x02
532 #define PERF_ATTACH_TASK 0x04
533 #define PERF_ATTACH_TASK_DATA 0x08
534 #define PERF_ATTACH_ITRACE 0x10
539 struct pmu_event_list
{
541 struct list_head list
;
545 * struct perf_event - performance event kernel representation:
548 #ifdef CONFIG_PERF_EVENTS
550 * entry onto perf_event_context::event_list;
551 * modifications require ctx->lock
552 * RCU safe iterations.
554 struct list_head event_entry
;
557 * XXX: group_entry and sibling_list should be mutually exclusive;
558 * either you're a sibling on a group, or you're the group leader.
559 * Rework the code to always use the same list element.
561 * Locked for modification by both ctx->mutex and ctx->lock; holding
562 * either sufficies for read.
564 struct list_head group_entry
;
565 struct list_head sibling_list
;
568 * We need storage to track the entries in perf_pmu_migrate_context; we
569 * cannot use the event_entry because of RCU and we want to keep the
570 * group in tact which avoids us using the other two entries.
572 struct list_head migrate_entry
;
574 struct hlist_node hlist_entry
;
575 struct list_head active_entry
;
578 /* Not serialized. Only written during event initialization. */
580 /* The cumulative AND of all event_caps for events in this group. */
583 struct perf_event
*group_leader
;
587 enum perf_event_state state
;
588 unsigned int attach_state
;
590 atomic64_t child_count
;
593 * These are the total time in nanoseconds that the event
594 * has been enabled (i.e. eligible to run, and the task has
595 * been scheduled in, if this is a per-task event)
596 * and running (scheduled onto the CPU), respectively.
598 u64 total_time_enabled
;
599 u64 total_time_running
;
603 * timestamp shadows the actual context timing but it can
604 * be safely used in NMI interrupt context. It reflects the
605 * context time as it was when the event was last scheduled in.
607 * ctx_time already accounts for ctx->timestamp. Therefore to
608 * compute ctx_time for a sample, simply add perf_clock().
612 struct perf_event_attr attr
;
616 struct hw_perf_event hw
;
618 struct perf_event_context
*ctx
;
619 atomic_long_t refcount
;
622 * These accumulate total time (in nanoseconds) that children
623 * events have been enabled and running, respectively.
625 atomic64_t child_total_time_enabled
;
626 atomic64_t child_total_time_running
;
629 * Protect attach/detach and child_list:
631 struct mutex child_mutex
;
632 struct list_head child_list
;
633 struct perf_event
*parent
;
638 struct list_head owner_entry
;
639 struct task_struct
*owner
;
642 struct mutex mmap_mutex
;
645 struct ring_buffer
*rb
;
646 struct list_head rb_entry
;
647 unsigned long rcu_batches
;
651 wait_queue_head_t waitq
;
652 struct fasync_struct
*fasync
;
654 /* delayed work for NMIs and such */
658 struct irq_work pending
;
660 atomic_t event_limit
;
662 /* address range filters */
663 struct perf_addr_filters_head addr_filters
;
664 /* vma address array for file-based filders */
665 unsigned long *addr_filters_offs
;
666 unsigned long addr_filters_gen
;
668 void (*destroy
)(struct perf_event
*);
669 struct rcu_head rcu_head
;
671 struct pid_namespace
*ns
;
675 perf_overflow_handler_t overflow_handler
;
676 void *overflow_handler_context
;
677 #ifdef CONFIG_BPF_SYSCALL
678 perf_overflow_handler_t orig_overflow_handler
;
679 struct bpf_prog
*prog
;
682 #ifdef CONFIG_EVENT_TRACING
683 struct trace_event_call
*tp_event
;
684 struct event_filter
*filter
;
685 #ifdef CONFIG_FUNCTION_TRACER
686 struct ftrace_ops ftrace_ops
;
690 #ifdef CONFIG_CGROUP_PERF
691 struct perf_cgroup
*cgrp
; /* cgroup event is attach to */
694 struct list_head sb_list
;
695 #endif /* CONFIG_PERF_EVENTS */
699 * struct perf_event_context - event context structure
701 * Used as a container for task events and CPU events as well:
703 struct perf_event_context
{
706 * Protect the states of the events in the list,
707 * nr_active, and the list:
711 * Protect the list of events. Locking either mutex or lock
712 * is sufficient to ensure the list doesn't change; to change
713 * the list you need to lock both the mutex and the spinlock.
717 struct list_head active_ctx_list
;
718 struct list_head pinned_groups
;
719 struct list_head flexible_groups
;
720 struct list_head event_list
;
728 struct task_struct
*task
;
731 * Context clock, runs when context enabled.
737 * These fields let us detect when two contexts have both
738 * been cloned (inherited) from a common ancestor.
740 struct perf_event_context
*parent_ctx
;
744 #ifdef CONFIG_CGROUP_PERF
745 int nr_cgroups
; /* cgroup evts */
747 void *task_ctx_data
; /* pmu specific data */
748 struct rcu_head rcu_head
;
752 * Number of contexts where an event can trigger:
753 * task, softirq, hardirq, nmi.
755 #define PERF_NR_CONTEXTS 4
758 * struct perf_event_cpu_context - per cpu event context structure
760 struct perf_cpu_context
{
761 struct perf_event_context ctx
;
762 struct perf_event_context
*task_ctx
;
766 raw_spinlock_t hrtimer_lock
;
767 struct hrtimer hrtimer
;
768 ktime_t hrtimer_interval
;
769 unsigned int hrtimer_active
;
771 #ifdef CONFIG_CGROUP_PERF
772 struct perf_cgroup
*cgrp
;
773 struct list_head cgrp_cpuctx_entry
;
776 struct list_head sched_cb_entry
;
782 struct perf_output_handle
{
783 struct perf_event
*event
;
784 struct ring_buffer
*rb
;
785 unsigned long wakeup
;
795 struct bpf_perf_event_data_kern
{
796 bpf_user_pt_regs_t
*regs
;
797 struct perf_sample_data
*data
;
798 struct perf_event
*event
;
801 #ifdef CONFIG_CGROUP_PERF
804 * perf_cgroup_info keeps track of time_enabled for a cgroup.
805 * This is a per-cpu dynamically allocated data structure.
807 struct perf_cgroup_info
{
813 struct cgroup_subsys_state css
;
814 struct perf_cgroup_info __percpu
*info
;
818 * Must ensure cgroup is pinned (css_get) before calling
819 * this function. In other words, we cannot call this function
820 * if there is no cgroup event for the current CPU context.
822 static inline struct perf_cgroup
*
823 perf_cgroup_from_task(struct task_struct
*task
, struct perf_event_context
*ctx
)
825 return container_of(task_css_check(task
, perf_event_cgrp_id
,
826 ctx
? lockdep_is_held(&ctx
->lock
)
828 struct perf_cgroup
, css
);
830 #endif /* CONFIG_CGROUP_PERF */
832 #ifdef CONFIG_PERF_EVENTS
834 extern void *perf_aux_output_begin(struct perf_output_handle
*handle
,
835 struct perf_event
*event
);
836 extern void perf_aux_output_end(struct perf_output_handle
*handle
,
838 extern int perf_aux_output_skip(struct perf_output_handle
*handle
,
840 extern void *perf_get_aux(struct perf_output_handle
*handle
);
841 extern void perf_aux_output_flag(struct perf_output_handle
*handle
, u64 flags
);
842 extern void perf_event_itrace_started(struct perf_event
*event
);
844 extern int perf_pmu_register(struct pmu
*pmu
, const char *name
, int type
);
845 extern void perf_pmu_unregister(struct pmu
*pmu
);
847 extern int perf_num_counters(void);
848 extern const char *perf_pmu_name(void);
849 extern void __perf_event_task_sched_in(struct task_struct
*prev
,
850 struct task_struct
*task
);
851 extern void __perf_event_task_sched_out(struct task_struct
*prev
,
852 struct task_struct
*next
);
853 extern int perf_event_init_task(struct task_struct
*child
);
854 extern void perf_event_exit_task(struct task_struct
*child
);
855 extern void perf_event_free_task(struct task_struct
*task
);
856 extern void perf_event_delayed_put(struct task_struct
*task
);
857 extern struct file
*perf_event_get(unsigned int fd
);
858 extern const struct perf_event_attr
*perf_event_attrs(struct perf_event
*event
);
859 extern void perf_event_print_debug(void);
860 extern void perf_pmu_disable(struct pmu
*pmu
);
861 extern void perf_pmu_enable(struct pmu
*pmu
);
862 extern void perf_sched_cb_dec(struct pmu
*pmu
);
863 extern void perf_sched_cb_inc(struct pmu
*pmu
);
864 extern int perf_event_task_disable(void);
865 extern int perf_event_task_enable(void);
866 extern int perf_event_refresh(struct perf_event
*event
, int refresh
);
867 extern void perf_event_update_userpage(struct perf_event
*event
);
868 extern int perf_event_release_kernel(struct perf_event
*event
);
869 extern struct perf_event
*
870 perf_event_create_kernel_counter(struct perf_event_attr
*attr
,
872 struct task_struct
*task
,
873 perf_overflow_handler_t callback
,
875 extern void perf_pmu_migrate_context(struct pmu
*pmu
,
876 int src_cpu
, int dst_cpu
);
877 int perf_event_read_local(struct perf_event
*event
, u64
*value
,
878 u64
*enabled
, u64
*running
);
879 extern u64
perf_event_read_value(struct perf_event
*event
,
880 u64
*enabled
, u64
*running
);
883 struct perf_sample_data
{
885 * Fields set by perf_sample_data_init(), group so as to
886 * minimize the cachelines touched.
889 struct perf_raw_record
*raw
;
890 struct perf_branch_stack
*br_stack
;
894 union perf_mem_data_src data_src
;
897 * The other fields, optionally {set,used} by
898 * perf_{prepare,output}_sample().
913 struct perf_callchain_entry
*callchain
;
916 * regs_user may point to task_pt_regs or to regs_user_copy, depending
919 struct perf_regs regs_user
;
920 struct pt_regs regs_user_copy
;
922 struct perf_regs regs_intr
;
926 } ____cacheline_aligned
;
928 /* default value for data source */
929 #define PERF_MEM_NA (PERF_MEM_S(OP, NA) |\
930 PERF_MEM_S(LVL, NA) |\
931 PERF_MEM_S(SNOOP, NA) |\
932 PERF_MEM_S(LOCK, NA) |\
935 static inline void perf_sample_data_init(struct perf_sample_data
*data
,
936 u64 addr
, u64 period
)
938 /* remaining struct members initialized in perf_prepare_sample() */
941 data
->br_stack
= NULL
;
942 data
->period
= period
;
944 data
->data_src
.val
= PERF_MEM_NA
;
948 extern void perf_output_sample(struct perf_output_handle
*handle
,
949 struct perf_event_header
*header
,
950 struct perf_sample_data
*data
,
951 struct perf_event
*event
);
952 extern void perf_prepare_sample(struct perf_event_header
*header
,
953 struct perf_sample_data
*data
,
954 struct perf_event
*event
,
955 struct pt_regs
*regs
);
957 extern int perf_event_overflow(struct perf_event
*event
,
958 struct perf_sample_data
*data
,
959 struct pt_regs
*regs
);
961 extern void perf_event_output_forward(struct perf_event
*event
,
962 struct perf_sample_data
*data
,
963 struct pt_regs
*regs
);
964 extern void perf_event_output_backward(struct perf_event
*event
,
965 struct perf_sample_data
*data
,
966 struct pt_regs
*regs
);
967 extern void perf_event_output(struct perf_event
*event
,
968 struct perf_sample_data
*data
,
969 struct pt_regs
*regs
);
972 is_default_overflow_handler(struct perf_event
*event
)
974 if (likely(event
->overflow_handler
== perf_event_output_forward
))
976 if (unlikely(event
->overflow_handler
== perf_event_output_backward
))
982 perf_event_header__init_id(struct perf_event_header
*header
,
983 struct perf_sample_data
*data
,
984 struct perf_event
*event
);
986 perf_event__output_id_sample(struct perf_event
*event
,
987 struct perf_output_handle
*handle
,
988 struct perf_sample_data
*sample
);
991 perf_log_lost_samples(struct perf_event
*event
, u64 lost
);
993 static inline bool is_sampling_event(struct perf_event
*event
)
995 return event
->attr
.sample_period
!= 0;
999 * Return 1 for a software event, 0 for a hardware event
1001 static inline int is_software_event(struct perf_event
*event
)
1003 return event
->event_caps
& PERF_EV_CAP_SOFTWARE
;
1006 extern struct static_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
1008 extern void ___perf_sw_event(u32
, u64
, struct pt_regs
*, u64
);
1009 extern void __perf_sw_event(u32
, u64
, struct pt_regs
*, u64
);
1011 #ifndef perf_arch_fetch_caller_regs
1012 static inline void perf_arch_fetch_caller_regs(struct pt_regs
*regs
, unsigned long ip
) { }
1016 * Take a snapshot of the regs. Skip ip and frame pointer to
1017 * the nth caller. We only need a few of the regs:
1018 * - ip for PERF_SAMPLE_IP
1019 * - cs for user_mode() tests
1020 * - bp for callchains
1021 * - eflags, for future purposes, just in case
1023 static inline void perf_fetch_caller_regs(struct pt_regs
*regs
)
1025 perf_arch_fetch_caller_regs(regs
, CALLER_ADDR0
);
1028 static __always_inline
void
1029 perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
1031 if (static_key_false(&perf_swevent_enabled
[event_id
]))
1032 __perf_sw_event(event_id
, nr
, regs
, addr
);
1035 DECLARE_PER_CPU(struct pt_regs
, __perf_regs
[4]);
1038 * 'Special' version for the scheduler, it hard assumes no recursion,
1039 * which is guaranteed by us not actually scheduling inside other swevents
1040 * because those disable preemption.
1042 static __always_inline
void
1043 perf_sw_event_sched(u32 event_id
, u64 nr
, u64 addr
)
1045 if (static_key_false(&perf_swevent_enabled
[event_id
])) {
1046 struct pt_regs
*regs
= this_cpu_ptr(&__perf_regs
[0]);
1048 perf_fetch_caller_regs(regs
);
1049 ___perf_sw_event(event_id
, nr
, regs
, addr
);
1053 extern struct static_key_false perf_sched_events
;
1055 static __always_inline
bool
1056 perf_sw_migrate_enabled(void)
1058 if (static_key_false(&perf_swevent_enabled
[PERF_COUNT_SW_CPU_MIGRATIONS
]))
1063 static inline void perf_event_task_migrate(struct task_struct
*task
)
1065 if (perf_sw_migrate_enabled())
1066 task
->sched_migrated
= 1;
1069 static inline void perf_event_task_sched_in(struct task_struct
*prev
,
1070 struct task_struct
*task
)
1072 if (static_branch_unlikely(&perf_sched_events
))
1073 __perf_event_task_sched_in(prev
, task
);
1075 if (perf_sw_migrate_enabled() && task
->sched_migrated
) {
1076 struct pt_regs
*regs
= this_cpu_ptr(&__perf_regs
[0]);
1078 perf_fetch_caller_regs(regs
);
1079 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, regs
, 0);
1080 task
->sched_migrated
= 0;
1084 static inline void perf_event_task_sched_out(struct task_struct
*prev
,
1085 struct task_struct
*next
)
1087 perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES
, 1, 0);
1089 if (static_branch_unlikely(&perf_sched_events
))
1090 __perf_event_task_sched_out(prev
, next
);
1093 extern void perf_event_mmap(struct vm_area_struct
*vma
);
1094 extern struct perf_guest_info_callbacks
*perf_guest_cbs
;
1095 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*callbacks
);
1096 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*callbacks
);
1098 extern void perf_event_exec(void);
1099 extern void perf_event_comm(struct task_struct
*tsk
, bool exec
);
1100 extern void perf_event_namespaces(struct task_struct
*tsk
);
1101 extern void perf_event_fork(struct task_struct
*tsk
);
1104 DECLARE_PER_CPU(struct perf_callchain_entry
, perf_callchain_entry
);
1106 extern void perf_callchain_user(struct perf_callchain_entry_ctx
*entry
, struct pt_regs
*regs
);
1107 extern void perf_callchain_kernel(struct perf_callchain_entry_ctx
*entry
, struct pt_regs
*regs
);
1108 extern struct perf_callchain_entry
*
1109 get_perf_callchain(struct pt_regs
*regs
, u32 init_nr
, bool kernel
, bool user
,
1110 u32 max_stack
, bool crosstask
, bool add_mark
);
1111 extern int get_callchain_buffers(int max_stack
);
1112 extern void put_callchain_buffers(void);
1114 extern int sysctl_perf_event_max_stack
;
1115 extern int sysctl_perf_event_max_contexts_per_stack
;
1117 static inline int perf_callchain_store_context(struct perf_callchain_entry_ctx
*ctx
, u64 ip
)
1119 if (ctx
->contexts
< sysctl_perf_event_max_contexts_per_stack
) {
1120 struct perf_callchain_entry
*entry
= ctx
->entry
;
1121 entry
->ip
[entry
->nr
++] = ip
;
1125 ctx
->contexts_maxed
= true;
1126 return -1; /* no more room, stop walking the stack */
1130 static inline int perf_callchain_store(struct perf_callchain_entry_ctx
*ctx
, u64 ip
)
1132 if (ctx
->nr
< ctx
->max_stack
&& !ctx
->contexts_maxed
) {
1133 struct perf_callchain_entry
*entry
= ctx
->entry
;
1134 entry
->ip
[entry
->nr
++] = ip
;
1138 return -1; /* no more room, stop walking the stack */
1142 extern int sysctl_perf_event_paranoid
;
1143 extern int sysctl_perf_event_mlock
;
1144 extern int sysctl_perf_event_sample_rate
;
1145 extern int sysctl_perf_cpu_time_max_percent
;
1147 extern void perf_sample_event_took(u64 sample_len_ns
);
1149 extern int perf_proc_update_handler(struct ctl_table
*table
, int write
,
1150 void __user
*buffer
, size_t *lenp
,
1152 extern int perf_cpu_time_max_percent_handler(struct ctl_table
*table
, int write
,
1153 void __user
*buffer
, size_t *lenp
,
1156 int perf_event_max_stack_handler(struct ctl_table
*table
, int write
,
1157 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
);
1159 static inline bool perf_paranoid_any(void)
1161 return sysctl_perf_event_paranoid
> 2;
1164 static inline bool perf_paranoid_tracepoint_raw(void)
1166 return sysctl_perf_event_paranoid
> -1;
1169 static inline bool perf_paranoid_cpu(void)
1171 return sysctl_perf_event_paranoid
> 0;
1174 static inline bool perf_paranoid_kernel(void)
1176 return sysctl_perf_event_paranoid
> 1;
1179 extern void perf_event_init(void);
1180 extern void perf_tp_event(u16 event_type
, u64 count
, void *record
,
1181 int entry_size
, struct pt_regs
*regs
,
1182 struct hlist_head
*head
, int rctx
,
1183 struct task_struct
*task
);
1184 extern void perf_bp_event(struct perf_event
*event
, void *data
);
1186 #ifndef perf_misc_flags
1187 # define perf_misc_flags(regs) \
1188 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1189 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1191 #ifndef perf_arch_bpf_user_pt_regs
1192 # define perf_arch_bpf_user_pt_regs(regs) regs
1195 static inline bool has_branch_stack(struct perf_event
*event
)
1197 return event
->attr
.sample_type
& PERF_SAMPLE_BRANCH_STACK
;
1200 static inline bool needs_branch_stack(struct perf_event
*event
)
1202 return event
->attr
.branch_sample_type
!= 0;
1205 static inline bool has_aux(struct perf_event
*event
)
1207 return event
->pmu
->setup_aux
;
1210 static inline bool is_write_backward(struct perf_event
*event
)
1212 return !!event
->attr
.write_backward
;
1215 static inline bool has_addr_filter(struct perf_event
*event
)
1217 return event
->pmu
->nr_addr_filters
;
1221 * An inherited event uses parent's filters
1223 static inline struct perf_addr_filters_head
*
1224 perf_event_addr_filters(struct perf_event
*event
)
1226 struct perf_addr_filters_head
*ifh
= &event
->addr_filters
;
1229 ifh
= &event
->parent
->addr_filters
;
1234 extern void perf_event_addr_filters_sync(struct perf_event
*event
);
1236 extern int perf_output_begin(struct perf_output_handle
*handle
,
1237 struct perf_event
*event
, unsigned int size
);
1238 extern int perf_output_begin_forward(struct perf_output_handle
*handle
,
1239 struct perf_event
*event
,
1241 extern int perf_output_begin_backward(struct perf_output_handle
*handle
,
1242 struct perf_event
*event
,
1245 extern void perf_output_end(struct perf_output_handle
*handle
);
1246 extern unsigned int perf_output_copy(struct perf_output_handle
*handle
,
1247 const void *buf
, unsigned int len
);
1248 extern unsigned int perf_output_skip(struct perf_output_handle
*handle
,
1250 extern int perf_swevent_get_recursion_context(void);
1251 extern void perf_swevent_put_recursion_context(int rctx
);
1252 extern u64
perf_swevent_set_period(struct perf_event
*event
);
1253 extern void perf_event_enable(struct perf_event
*event
);
1254 extern void perf_event_disable(struct perf_event
*event
);
1255 extern void perf_event_disable_local(struct perf_event
*event
);
1256 extern void perf_event_disable_inatomic(struct perf_event
*event
);
1257 extern void perf_event_task_tick(void);
1258 extern int perf_event_account_interrupt(struct perf_event
*event
);
1259 #else /* !CONFIG_PERF_EVENTS: */
1260 static inline void *
1261 perf_aux_output_begin(struct perf_output_handle
*handle
,
1262 struct perf_event
*event
) { return NULL
; }
1264 perf_aux_output_end(struct perf_output_handle
*handle
, unsigned long size
)
1267 perf_aux_output_skip(struct perf_output_handle
*handle
,
1268 unsigned long size
) { return -EINVAL
; }
1269 static inline void *
1270 perf_get_aux(struct perf_output_handle
*handle
) { return NULL
; }
1272 perf_event_task_migrate(struct task_struct
*task
) { }
1274 perf_event_task_sched_in(struct task_struct
*prev
,
1275 struct task_struct
*task
) { }
1277 perf_event_task_sched_out(struct task_struct
*prev
,
1278 struct task_struct
*next
) { }
1279 static inline int perf_event_init_task(struct task_struct
*child
) { return 0; }
1280 static inline void perf_event_exit_task(struct task_struct
*child
) { }
1281 static inline void perf_event_free_task(struct task_struct
*task
) { }
1282 static inline void perf_event_delayed_put(struct task_struct
*task
) { }
1283 static inline struct file
*perf_event_get(unsigned int fd
) { return ERR_PTR(-EINVAL
); }
1284 static inline const struct perf_event_attr
*perf_event_attrs(struct perf_event
*event
)
1286 return ERR_PTR(-EINVAL
);
1288 static inline int perf_event_read_local(struct perf_event
*event
, u64
*value
,
1289 u64
*enabled
, u64
*running
)
1293 static inline void perf_event_print_debug(void) { }
1294 static inline int perf_event_task_disable(void) { return -EINVAL
; }
1295 static inline int perf_event_task_enable(void) { return -EINVAL
; }
1296 static inline int perf_event_refresh(struct perf_event
*event
, int refresh
)
1302 perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
) { }
1304 perf_sw_event_sched(u32 event_id
, u64 nr
, u64 addr
) { }
1306 perf_bp_event(struct perf_event
*event
, void *data
) { }
1308 static inline int perf_register_guest_info_callbacks
1309 (struct perf_guest_info_callbacks
*callbacks
) { return 0; }
1310 static inline int perf_unregister_guest_info_callbacks
1311 (struct perf_guest_info_callbacks
*callbacks
) { return 0; }
1313 static inline void perf_event_mmap(struct vm_area_struct
*vma
) { }
1314 static inline void perf_event_exec(void) { }
1315 static inline void perf_event_comm(struct task_struct
*tsk
, bool exec
) { }
1316 static inline void perf_event_namespaces(struct task_struct
*tsk
) { }
1317 static inline void perf_event_fork(struct task_struct
*tsk
) { }
1318 static inline void perf_event_init(void) { }
1319 static inline int perf_swevent_get_recursion_context(void) { return -1; }
1320 static inline void perf_swevent_put_recursion_context(int rctx
) { }
1321 static inline u64
perf_swevent_set_period(struct perf_event
*event
) { return 0; }
1322 static inline void perf_event_enable(struct perf_event
*event
) { }
1323 static inline void perf_event_disable(struct perf_event
*event
) { }
1324 static inline int __perf_event_disable(void *info
) { return -1; }
1325 static inline void perf_event_task_tick(void) { }
1326 static inline int perf_event_release_kernel(struct perf_event
*event
) { return 0; }
1329 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1330 extern void perf_restore_debug_store(void);
1332 static inline void perf_restore_debug_store(void) { }
1335 static __always_inline
bool perf_raw_frag_last(const struct perf_raw_frag
*frag
)
1337 return frag
->pad
< sizeof(u64
);
1340 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1342 struct perf_pmu_events_attr
{
1343 struct device_attribute attr
;
1345 const char *event_str
;
1348 struct perf_pmu_events_ht_attr
{
1349 struct device_attribute attr
;
1351 const char *event_str_ht
;
1352 const char *event_str_noht
;
1355 ssize_t
perf_event_sysfs_show(struct device
*dev
, struct device_attribute
*attr
,
1358 #define PMU_EVENT_ATTR(_name, _var, _id, _show) \
1359 static struct perf_pmu_events_attr _var = { \
1360 .attr = __ATTR(_name, 0444, _show, NULL), \
1364 #define PMU_EVENT_ATTR_STRING(_name, _var, _str) \
1365 static struct perf_pmu_events_attr _var = { \
1366 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1368 .event_str = _str, \
1371 #define PMU_FORMAT_ATTR(_name, _format) \
1373 _name##_show(struct device *dev, \
1374 struct device_attribute *attr, \
1377 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
1378 return sprintf(page, _format "\n"); \
1381 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1383 /* Performance counter hotplug functions */
1384 #ifdef CONFIG_PERF_EVENTS
1385 int perf_event_init_cpu(unsigned int cpu
);
1386 int perf_event_exit_cpu(unsigned int cpu
);
1388 #define perf_event_init_cpu NULL
1389 #define perf_event_exit_cpu NULL
1392 #endif /* _LINUX_PERF_EVENT_H */