xref: /linux-6.15/include/linux/perf_event.h (revision f5e4e7fd)
1 /*
2  * Performance events:
3  *
4  *    Copyright (C) 2008-2009, Thomas Gleixner <[email protected]>
5  *    Copyright (C) 2008-2011, Red Hat, Inc., Ingo Molnar
6  *    Copyright (C) 2008-2011, Red Hat, Inc., Peter Zijlstra
7  *
8  * Data type definitions, declarations, prototypes.
9  *
10  *    Started by: Thomas Gleixner and Ingo Molnar
11  *
12  * For licencing details see kernel-base/COPYING
13  */
14 #ifndef _LINUX_PERF_EVENT_H
15 #define _LINUX_PERF_EVENT_H
16 
17 #include <uapi/linux/perf_event.h>
18 
19 /*
20  * Kernel-internal data types and definitions:
21  */
22 
23 #ifdef CONFIG_PERF_EVENTS
24 # include <asm/perf_event.h>
25 # include <asm/local64.h>
26 #endif
27 
28 struct perf_guest_info_callbacks {
29 	int				(*is_in_guest)(void);
30 	int				(*is_user_mode)(void);
31 	unsigned long			(*get_guest_ip)(void);
32 };
33 
34 #ifdef CONFIG_HAVE_HW_BREAKPOINT
35 #include <asm/hw_breakpoint.h>
36 #endif
37 
38 #include <linux/list.h>
39 #include <linux/mutex.h>
40 #include <linux/rculist.h>
41 #include <linux/rcupdate.h>
42 #include <linux/spinlock.h>
43 #include <linux/hrtimer.h>
44 #include <linux/fs.h>
45 #include <linux/pid_namespace.h>
46 #include <linux/workqueue.h>
47 #include <linux/ftrace.h>
48 #include <linux/cpu.h>
49 #include <linux/irq_work.h>
50 #include <linux/static_key.h>
51 #include <linux/atomic.h>
52 #include <linux/sysfs.h>
53 #include <linux/perf_regs.h>
54 #include <asm/local.h>
55 
56 struct perf_callchain_entry {
57 	__u64				nr;
58 	__u64				ip[PERF_MAX_STACK_DEPTH];
59 };
60 
61 struct perf_raw_record {
62 	u32				size;
63 	void				*data;
64 };
65 
66 /*
67  * single taken branch record layout:
68  *
69  *      from: source instruction (may not always be a branch insn)
70  *        to: branch target
71  *   mispred: branch target was mispredicted
72  * predicted: branch target was predicted
73  *
74  * support for mispred, predicted is optional. In case it
75  * is not supported mispred = predicted = 0.
76  *
77  *     in_tx: running in a hardware transaction
78  *     abort: aborting a hardware transaction
79  */
80 struct perf_branch_entry {
81 	__u64	from;
82 	__u64	to;
83 	__u64	mispred:1,  /* target mispredicted */
84 		predicted:1,/* target predicted */
85 		in_tx:1,    /* in transaction */
86 		abort:1,    /* transaction abort */
87 		reserved:60;
88 };
89 
90 /*
91  * branch stack layout:
92  *  nr: number of taken branches stored in entries[]
93  *
94  * Note that nr can vary from sample to sample
95  * branches (to, from) are stored from most recent
96  * to least recent, i.e., entries[0] contains the most
97  * recent branch.
98  */
99 struct perf_branch_stack {
100 	__u64				nr;
101 	struct perf_branch_entry	entries[0];
102 };
103 
104 struct perf_regs_user {
105 	__u64		abi;
106 	struct pt_regs	*regs;
107 };
108 
109 struct task_struct;
110 
111 /*
112  * extra PMU register associated with an event
113  */
114 struct hw_perf_event_extra {
115 	u64		config;	/* register value */
116 	unsigned int	reg;	/* register address or index */
117 	int		alloc;	/* extra register already allocated */
118 	int		idx;	/* index in shared_regs->regs[] */
119 };
120 
121 struct event_constraint;
122 
123 /**
124  * struct hw_perf_event - performance event hardware details:
125  */
126 struct hw_perf_event {
127 #ifdef CONFIG_PERF_EVENTS
128 	union {
129 		struct { /* hardware */
130 			u64		config;
131 			u64		last_tag;
132 			unsigned long	config_base;
133 			unsigned long	event_base;
134 			int		event_base_rdpmc;
135 			int		idx;
136 			int		last_cpu;
137 			int		flags;
138 
139 			struct hw_perf_event_extra extra_reg;
140 			struct hw_perf_event_extra branch_reg;
141 
142 			struct event_constraint *constraint;
143 		};
144 		struct { /* software */
145 			struct hrtimer	hrtimer;
146 		};
147 		struct { /* tracepoint */
148 			struct task_struct	*tp_target;
149 			/* for tp_event->class */
150 			struct list_head	tp_list;
151 		};
152 #ifdef CONFIG_HAVE_HW_BREAKPOINT
153 		struct { /* breakpoint */
154 			/*
155 			 * Crufty hack to avoid the chicken and egg
156 			 * problem hw_breakpoint has with context
157 			 * creation and event initalization.
158 			 */
159 			struct task_struct		*bp_target;
160 			struct arch_hw_breakpoint	info;
161 			struct list_head		bp_list;
162 		};
163 #endif
164 	};
165 	int				state;
166 	local64_t			prev_count;
167 	u64				sample_period;
168 	u64				last_period;
169 	local64_t			period_left;
170 	u64                             interrupts_seq;
171 	u64				interrupts;
172 
173 	u64				freq_time_stamp;
174 	u64				freq_count_stamp;
175 #endif
176 };
177 
178 /*
179  * hw_perf_event::state flags
180  */
181 #define PERF_HES_STOPPED	0x01 /* the counter is stopped */
182 #define PERF_HES_UPTODATE	0x02 /* event->count up-to-date */
183 #define PERF_HES_ARCH		0x04
184 
185 struct perf_event;
186 
187 /*
188  * Common implementation detail of pmu::{start,commit,cancel}_txn
189  */
190 #define PERF_EVENT_TXN 0x1
191 
192 /**
193  * struct pmu - generic performance monitoring unit
194  */
195 struct pmu {
196 	struct list_head		entry;
197 
198 	struct device			*dev;
199 	const struct attribute_group	**attr_groups;
200 	const char			*name;
201 	int				type;
202 
203 	int * __percpu			pmu_disable_count;
204 	struct perf_cpu_context * __percpu pmu_cpu_context;
205 	int				task_ctx_nr;
206 	int				hrtimer_interval_ms;
207 
208 	/*
209 	 * Fully disable/enable this PMU, can be used to protect from the PMI
210 	 * as well as for lazy/batch writing of the MSRs.
211 	 */
212 	void (*pmu_enable)		(struct pmu *pmu); /* optional */
213 	void (*pmu_disable)		(struct pmu *pmu); /* optional */
214 
215 	/*
216 	 * Try and initialize the event for this PMU.
217 	 * Should return -ENOENT when the @event doesn't match this PMU.
218 	 */
219 	int (*event_init)		(struct perf_event *event);
220 
221 #define PERF_EF_START	0x01		/* start the counter when adding    */
222 #define PERF_EF_RELOAD	0x02		/* reload the counter when starting */
223 #define PERF_EF_UPDATE	0x04		/* update the counter when stopping */
224 
225 	/*
226 	 * Adds/Removes a counter to/from the PMU, can be done inside
227 	 * a transaction, see the ->*_txn() methods.
228 	 */
229 	int  (*add)			(struct perf_event *event, int flags);
230 	void (*del)			(struct perf_event *event, int flags);
231 
232 	/*
233 	 * Starts/Stops a counter present on the PMU. The PMI handler
234 	 * should stop the counter when perf_event_overflow() returns
235 	 * !0. ->start() will be used to continue.
236 	 */
237 	void (*start)			(struct perf_event *event, int flags);
238 	void (*stop)			(struct perf_event *event, int flags);
239 
240 	/*
241 	 * Updates the counter value of the event.
242 	 */
243 	void (*read)			(struct perf_event *event);
244 
245 	/*
246 	 * Group events scheduling is treated as a transaction, add
247 	 * group events as a whole and perform one schedulability test.
248 	 * If the test fails, roll back the whole group
249 	 *
250 	 * Start the transaction, after this ->add() doesn't need to
251 	 * do schedulability tests.
252 	 */
253 	void (*start_txn)		(struct pmu *pmu); /* optional */
254 	/*
255 	 * If ->start_txn() disabled the ->add() schedulability test
256 	 * then ->commit_txn() is required to perform one. On success
257 	 * the transaction is closed. On error the transaction is kept
258 	 * open until ->cancel_txn() is called.
259 	 */
260 	int  (*commit_txn)		(struct pmu *pmu); /* optional */
261 	/*
262 	 * Will cancel the transaction, assumes ->del() is called
263 	 * for each successful ->add() during the transaction.
264 	 */
265 	void (*cancel_txn)		(struct pmu *pmu); /* optional */
266 
267 	/*
268 	 * Will return the value for perf_event_mmap_page::index for this event,
269 	 * if no implementation is provided it will default to: event->hw.idx + 1.
270 	 */
271 	int (*event_idx)		(struct perf_event *event); /*optional */
272 
273 	/*
274 	 * flush branch stack on context-switches (needed in cpu-wide mode)
275 	 */
276 	void (*flush_branch_stack)	(void);
277 };
278 
279 /**
280  * enum perf_event_active_state - the states of a event
281  */
282 enum perf_event_active_state {
283 	PERF_EVENT_STATE_ERROR		= -2,
284 	PERF_EVENT_STATE_OFF		= -1,
285 	PERF_EVENT_STATE_INACTIVE	=  0,
286 	PERF_EVENT_STATE_ACTIVE		=  1,
287 };
288 
289 struct file;
290 struct perf_sample_data;
291 
292 typedef void (*perf_overflow_handler_t)(struct perf_event *,
293 					struct perf_sample_data *,
294 					struct pt_regs *regs);
295 
296 enum perf_group_flag {
297 	PERF_GROUP_SOFTWARE		= 0x1,
298 };
299 
300 #define SWEVENT_HLIST_BITS		8
301 #define SWEVENT_HLIST_SIZE		(1 << SWEVENT_HLIST_BITS)
302 
303 struct swevent_hlist {
304 	struct hlist_head		heads[SWEVENT_HLIST_SIZE];
305 	struct rcu_head			rcu_head;
306 };
307 
308 #define PERF_ATTACH_CONTEXT	0x01
309 #define PERF_ATTACH_GROUP	0x02
310 #define PERF_ATTACH_TASK	0x04
311 
312 struct perf_cgroup;
313 struct ring_buffer;
314 
315 /**
316  * struct perf_event - performance event kernel representation:
317  */
318 struct perf_event {
319 #ifdef CONFIG_PERF_EVENTS
320 	struct list_head		group_entry;
321 	struct list_head		event_entry;
322 	struct list_head		sibling_list;
323 	struct hlist_node		hlist_entry;
324 	int				nr_siblings;
325 	int				group_flags;
326 	struct perf_event		*group_leader;
327 	struct pmu			*pmu;
328 
329 	enum perf_event_active_state	state;
330 	unsigned int			attach_state;
331 	local64_t			count;
332 	atomic64_t			child_count;
333 
334 	/*
335 	 * These are the total time in nanoseconds that the event
336 	 * has been enabled (i.e. eligible to run, and the task has
337 	 * been scheduled in, if this is a per-task event)
338 	 * and running (scheduled onto the CPU), respectively.
339 	 *
340 	 * They are computed from tstamp_enabled, tstamp_running and
341 	 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
342 	 */
343 	u64				total_time_enabled;
344 	u64				total_time_running;
345 
346 	/*
347 	 * These are timestamps used for computing total_time_enabled
348 	 * and total_time_running when the event is in INACTIVE or
349 	 * ACTIVE state, measured in nanoseconds from an arbitrary point
350 	 * in time.
351 	 * tstamp_enabled: the notional time when the event was enabled
352 	 * tstamp_running: the notional time when the event was scheduled on
353 	 * tstamp_stopped: in INACTIVE state, the notional time when the
354 	 *	event was scheduled off.
355 	 */
356 	u64				tstamp_enabled;
357 	u64				tstamp_running;
358 	u64				tstamp_stopped;
359 
360 	/*
361 	 * timestamp shadows the actual context timing but it can
362 	 * be safely used in NMI interrupt context. It reflects the
363 	 * context time as it was when the event was last scheduled in.
364 	 *
365 	 * ctx_time already accounts for ctx->timestamp. Therefore to
366 	 * compute ctx_time for a sample, simply add perf_clock().
367 	 */
368 	u64				shadow_ctx_time;
369 
370 	struct perf_event_attr		attr;
371 	u16				header_size;
372 	u16				id_header_size;
373 	u16				read_size;
374 	struct hw_perf_event		hw;
375 
376 	struct perf_event_context	*ctx;
377 	atomic_long_t			refcount;
378 
379 	/*
380 	 * These accumulate total time (in nanoseconds) that children
381 	 * events have been enabled and running, respectively.
382 	 */
383 	atomic64_t			child_total_time_enabled;
384 	atomic64_t			child_total_time_running;
385 
386 	/*
387 	 * Protect attach/detach and child_list:
388 	 */
389 	struct mutex			child_mutex;
390 	struct list_head		child_list;
391 	struct perf_event		*parent;
392 
393 	int				oncpu;
394 	int				cpu;
395 
396 	struct list_head		owner_entry;
397 	struct task_struct		*owner;
398 
399 	/* mmap bits */
400 	struct mutex			mmap_mutex;
401 	atomic_t			mmap_count;
402 
403 	struct ring_buffer		*rb;
404 	struct list_head		rb_entry;
405 
406 	/* poll related */
407 	wait_queue_head_t		waitq;
408 	struct fasync_struct		*fasync;
409 
410 	/* delayed work for NMIs and such */
411 	int				pending_wakeup;
412 	int				pending_kill;
413 	int				pending_disable;
414 	struct irq_work			pending;
415 
416 	atomic_t			event_limit;
417 
418 	void (*destroy)(struct perf_event *);
419 	struct rcu_head			rcu_head;
420 
421 	struct pid_namespace		*ns;
422 	u64				id;
423 
424 	perf_overflow_handler_t		overflow_handler;
425 	void				*overflow_handler_context;
426 
427 #ifdef CONFIG_EVENT_TRACING
428 	struct ftrace_event_call	*tp_event;
429 	struct event_filter		*filter;
430 #ifdef CONFIG_FUNCTION_TRACER
431 	struct ftrace_ops               ftrace_ops;
432 #endif
433 #endif
434 
435 #ifdef CONFIG_CGROUP_PERF
436 	struct perf_cgroup		*cgrp; /* cgroup event is attach to */
437 	int				cgrp_defer_enabled;
438 #endif
439 
440 #endif /* CONFIG_PERF_EVENTS */
441 };
442 
443 enum perf_event_context_type {
444 	task_context,
445 	cpu_context,
446 };
447 
448 /**
449  * struct perf_event_context - event context structure
450  *
451  * Used as a container for task events and CPU events as well:
452  */
453 struct perf_event_context {
454 	struct pmu			*pmu;
455 	enum perf_event_context_type	type;
456 	/*
457 	 * Protect the states of the events in the list,
458 	 * nr_active, and the list:
459 	 */
460 	raw_spinlock_t			lock;
461 	/*
462 	 * Protect the list of events.  Locking either mutex or lock
463 	 * is sufficient to ensure the list doesn't change; to change
464 	 * the list you need to lock both the mutex and the spinlock.
465 	 */
466 	struct mutex			mutex;
467 
468 	struct list_head		pinned_groups;
469 	struct list_head		flexible_groups;
470 	struct list_head		event_list;
471 	int				nr_events;
472 	int				nr_active;
473 	int				is_active;
474 	int				nr_stat;
475 	int				nr_freq;
476 	int				rotate_disable;
477 	atomic_t			refcount;
478 	struct task_struct		*task;
479 
480 	/*
481 	 * Context clock, runs when context enabled.
482 	 */
483 	u64				time;
484 	u64				timestamp;
485 
486 	/*
487 	 * These fields let us detect when two contexts have both
488 	 * been cloned (inherited) from a common ancestor.
489 	 */
490 	struct perf_event_context	*parent_ctx;
491 	u64				parent_gen;
492 	u64				generation;
493 	int				pin_count;
494 	int				nr_cgroups;	 /* cgroup evts */
495 	int				nr_branch_stack; /* branch_stack evt */
496 	struct rcu_head			rcu_head;
497 };
498 
499 /*
500  * Number of contexts where an event can trigger:
501  *	task, softirq, hardirq, nmi.
502  */
503 #define PERF_NR_CONTEXTS	4
504 
505 /**
506  * struct perf_event_cpu_context - per cpu event context structure
507  */
508 struct perf_cpu_context {
509 	struct perf_event_context	ctx;
510 	struct perf_event_context	*task_ctx;
511 	int				active_oncpu;
512 	int				exclusive;
513 	struct hrtimer			hrtimer;
514 	ktime_t				hrtimer_interval;
515 	struct list_head		rotation_list;
516 	struct pmu			*unique_pmu;
517 	struct perf_cgroup		*cgrp;
518 };
519 
520 struct perf_output_handle {
521 	struct perf_event		*event;
522 	struct ring_buffer		*rb;
523 	unsigned long			wakeup;
524 	unsigned long			size;
525 	void				*addr;
526 	int				page;
527 };
528 
529 #ifdef CONFIG_PERF_EVENTS
530 
531 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
532 extern void perf_pmu_unregister(struct pmu *pmu);
533 
534 extern int perf_num_counters(void);
535 extern const char *perf_pmu_name(void);
536 extern void __perf_event_task_sched_in(struct task_struct *prev,
537 				       struct task_struct *task);
538 extern void __perf_event_task_sched_out(struct task_struct *prev,
539 					struct task_struct *next);
540 extern int perf_event_init_task(struct task_struct *child);
541 extern void perf_event_exit_task(struct task_struct *child);
542 extern void perf_event_free_task(struct task_struct *task);
543 extern void perf_event_delayed_put(struct task_struct *task);
544 extern void perf_event_print_debug(void);
545 extern void perf_pmu_disable(struct pmu *pmu);
546 extern void perf_pmu_enable(struct pmu *pmu);
547 extern int perf_event_task_disable(void);
548 extern int perf_event_task_enable(void);
549 extern int perf_event_refresh(struct perf_event *event, int refresh);
550 extern void perf_event_update_userpage(struct perf_event *event);
551 extern int perf_event_release_kernel(struct perf_event *event);
552 extern struct perf_event *
553 perf_event_create_kernel_counter(struct perf_event_attr *attr,
554 				int cpu,
555 				struct task_struct *task,
556 				perf_overflow_handler_t callback,
557 				void *context);
558 extern void perf_pmu_migrate_context(struct pmu *pmu,
559 				int src_cpu, int dst_cpu);
560 extern u64 perf_event_read_value(struct perf_event *event,
561 				 u64 *enabled, u64 *running);
562 
563 
564 struct perf_sample_data {
565 	u64				type;
566 
567 	u64				ip;
568 	struct {
569 		u32	pid;
570 		u32	tid;
571 	}				tid_entry;
572 	u64				time;
573 	u64				addr;
574 	u64				id;
575 	u64				stream_id;
576 	struct {
577 		u32	cpu;
578 		u32	reserved;
579 	}				cpu_entry;
580 	u64				period;
581 	union  perf_mem_data_src	data_src;
582 	struct perf_callchain_entry	*callchain;
583 	struct perf_raw_record		*raw;
584 	struct perf_branch_stack	*br_stack;
585 	struct perf_regs_user		regs_user;
586 	u64				stack_user_size;
587 	u64				weight;
588 };
589 
590 static inline void perf_sample_data_init(struct perf_sample_data *data,
591 					 u64 addr, u64 period)
592 {
593 	/* remaining struct members initialized in perf_prepare_sample() */
594 	data->addr = addr;
595 	data->raw  = NULL;
596 	data->br_stack = NULL;
597 	data->period = period;
598 	data->regs_user.abi = PERF_SAMPLE_REGS_ABI_NONE;
599 	data->regs_user.regs = NULL;
600 	data->stack_user_size = 0;
601 	data->weight = 0;
602 	data->data_src.val = 0;
603 }
604 
605 extern void perf_output_sample(struct perf_output_handle *handle,
606 			       struct perf_event_header *header,
607 			       struct perf_sample_data *data,
608 			       struct perf_event *event);
609 extern void perf_prepare_sample(struct perf_event_header *header,
610 				struct perf_sample_data *data,
611 				struct perf_event *event,
612 				struct pt_regs *regs);
613 
614 extern int perf_event_overflow(struct perf_event *event,
615 				 struct perf_sample_data *data,
616 				 struct pt_regs *regs);
617 
618 static inline bool is_sampling_event(struct perf_event *event)
619 {
620 	return event->attr.sample_period != 0;
621 }
622 
623 /*
624  * Return 1 for a software event, 0 for a hardware event
625  */
626 static inline int is_software_event(struct perf_event *event)
627 {
628 	return event->pmu->task_ctx_nr == perf_sw_context;
629 }
630 
631 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
632 
633 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
634 
635 #ifndef perf_arch_fetch_caller_regs
636 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
637 #endif
638 
639 /*
640  * Take a snapshot of the regs. Skip ip and frame pointer to
641  * the nth caller. We only need a few of the regs:
642  * - ip for PERF_SAMPLE_IP
643  * - cs for user_mode() tests
644  * - bp for callchains
645  * - eflags, for future purposes, just in case
646  */
647 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
648 {
649 	memset(regs, 0, sizeof(*regs));
650 
651 	perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
652 }
653 
654 static __always_inline void
655 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
656 {
657 	struct pt_regs hot_regs;
658 
659 	if (static_key_false(&perf_swevent_enabled[event_id])) {
660 		if (!regs) {
661 			perf_fetch_caller_regs(&hot_regs);
662 			regs = &hot_regs;
663 		}
664 		__perf_sw_event(event_id, nr, regs, addr);
665 	}
666 }
667 
668 extern struct static_key_deferred perf_sched_events;
669 
670 static inline void perf_event_task_sched_in(struct task_struct *prev,
671 					    struct task_struct *task)
672 {
673 	if (static_key_false(&perf_sched_events.key))
674 		__perf_event_task_sched_in(prev, task);
675 }
676 
677 static inline void perf_event_task_sched_out(struct task_struct *prev,
678 					     struct task_struct *next)
679 {
680 	perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, NULL, 0);
681 
682 	if (static_key_false(&perf_sched_events.key))
683 		__perf_event_task_sched_out(prev, next);
684 }
685 
686 extern void perf_event_mmap(struct vm_area_struct *vma);
687 extern struct perf_guest_info_callbacks *perf_guest_cbs;
688 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
689 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
690 
691 extern void perf_event_comm(struct task_struct *tsk);
692 extern void perf_event_fork(struct task_struct *tsk);
693 
694 /* Callchains */
695 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
696 
697 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
698 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
699 
700 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
701 {
702 	if (entry->nr < PERF_MAX_STACK_DEPTH)
703 		entry->ip[entry->nr++] = ip;
704 }
705 
706 extern int sysctl_perf_event_paranoid;
707 extern int sysctl_perf_event_mlock;
708 extern int sysctl_perf_event_sample_rate;
709 extern int sysctl_perf_cpu_time_max_percent;
710 
711 extern void perf_sample_event_took(u64 sample_len_ns);
712 
713 extern int perf_proc_update_handler(struct ctl_table *table, int write,
714 		void __user *buffer, size_t *lenp,
715 		loff_t *ppos);
716 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
717 		void __user *buffer, size_t *lenp,
718 		loff_t *ppos);
719 
720 
721 static inline bool perf_paranoid_tracepoint_raw(void)
722 {
723 	return sysctl_perf_event_paranoid > -1;
724 }
725 
726 static inline bool perf_paranoid_cpu(void)
727 {
728 	return sysctl_perf_event_paranoid > 0;
729 }
730 
731 static inline bool perf_paranoid_kernel(void)
732 {
733 	return sysctl_perf_event_paranoid > 1;
734 }
735 
736 extern void perf_event_init(void);
737 extern void perf_tp_event(u64 addr, u64 count, void *record,
738 			  int entry_size, struct pt_regs *regs,
739 			  struct hlist_head *head, int rctx,
740 			  struct task_struct *task);
741 extern void perf_bp_event(struct perf_event *event, void *data);
742 
743 #ifndef perf_misc_flags
744 # define perf_misc_flags(regs) \
745 		(user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
746 # define perf_instruction_pointer(regs)	instruction_pointer(regs)
747 #endif
748 
749 static inline bool has_branch_stack(struct perf_event *event)
750 {
751 	return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
752 }
753 
754 extern int perf_output_begin(struct perf_output_handle *handle,
755 			     struct perf_event *event, unsigned int size);
756 extern void perf_output_end(struct perf_output_handle *handle);
757 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
758 			     const void *buf, unsigned int len);
759 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
760 				     unsigned int len);
761 extern int perf_swevent_get_recursion_context(void);
762 extern void perf_swevent_put_recursion_context(int rctx);
763 extern u64 perf_swevent_set_period(struct perf_event *event);
764 extern void perf_event_enable(struct perf_event *event);
765 extern void perf_event_disable(struct perf_event *event);
766 extern int __perf_event_disable(void *info);
767 extern void perf_event_task_tick(void);
768 #else
769 static inline void
770 perf_event_task_sched_in(struct task_struct *prev,
771 			 struct task_struct *task)			{ }
772 static inline void
773 perf_event_task_sched_out(struct task_struct *prev,
774 			  struct task_struct *next)			{ }
775 static inline int perf_event_init_task(struct task_struct *child)	{ return 0; }
776 static inline void perf_event_exit_task(struct task_struct *child)	{ }
777 static inline void perf_event_free_task(struct task_struct *task)	{ }
778 static inline void perf_event_delayed_put(struct task_struct *task)	{ }
779 static inline void perf_event_print_debug(void)				{ }
780 static inline int perf_event_task_disable(void)				{ return -EINVAL; }
781 static inline int perf_event_task_enable(void)				{ return -EINVAL; }
782 static inline int perf_event_refresh(struct perf_event *event, int refresh)
783 {
784 	return -EINVAL;
785 }
786 
787 static inline void
788 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)	{ }
789 static inline void
790 perf_bp_event(struct perf_event *event, void *data)			{ }
791 
792 static inline int perf_register_guest_info_callbacks
793 (struct perf_guest_info_callbacks *callbacks)				{ return 0; }
794 static inline int perf_unregister_guest_info_callbacks
795 (struct perf_guest_info_callbacks *callbacks)				{ return 0; }
796 
797 static inline void perf_event_mmap(struct vm_area_struct *vma)		{ }
798 static inline void perf_event_comm(struct task_struct *tsk)		{ }
799 static inline void perf_event_fork(struct task_struct *tsk)		{ }
800 static inline void perf_event_init(void)				{ }
801 static inline int  perf_swevent_get_recursion_context(void)		{ return -1; }
802 static inline void perf_swevent_put_recursion_context(int rctx)		{ }
803 static inline u64 perf_swevent_set_period(struct perf_event *event)	{ return 0; }
804 static inline void perf_event_enable(struct perf_event *event)		{ }
805 static inline void perf_event_disable(struct perf_event *event)		{ }
806 static inline int __perf_event_disable(void *info)			{ return -1; }
807 static inline void perf_event_task_tick(void)				{ }
808 #endif
809 
810 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_NO_HZ_FULL)
811 extern bool perf_event_can_stop_tick(void);
812 #else
813 static inline bool perf_event_can_stop_tick(void)			{ return true; }
814 #endif
815 
816 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
817 extern void perf_restore_debug_store(void);
818 #else
819 static inline void perf_restore_debug_store(void)			{ }
820 #endif
821 
822 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
823 
824 /*
825  * This has to have a higher priority than migration_notifier in sched/core.c.
826  */
827 #define perf_cpu_notifier(fn)						\
828 do {									\
829 	static struct notifier_block fn##_nb =				\
830 		{ .notifier_call = fn, .priority = CPU_PRI_PERF };	\
831 	unsigned long cpu = smp_processor_id();				\
832 	unsigned long flags;						\
833 	fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,			\
834 		(void *)(unsigned long)cpu);				\
835 	local_irq_save(flags);						\
836 	fn(&fn##_nb, (unsigned long)CPU_STARTING,			\
837 		(void *)(unsigned long)cpu);				\
838 	local_irq_restore(flags);					\
839 	fn(&fn##_nb, (unsigned long)CPU_ONLINE,				\
840 		(void *)(unsigned long)cpu);				\
841 	register_cpu_notifier(&fn##_nb);				\
842 } while (0)
843 
844 
845 struct perf_pmu_events_attr {
846 	struct device_attribute attr;
847 	u64 id;
848 	const char *event_str;
849 };
850 
851 #define PMU_EVENT_ATTR(_name, _var, _id, _show)				\
852 static struct perf_pmu_events_attr _var = {				\
853 	.attr = __ATTR(_name, 0444, _show, NULL),			\
854 	.id   =  _id,							\
855 };
856 
857 #define PMU_FORMAT_ATTR(_name, _format)					\
858 static ssize_t								\
859 _name##_show(struct device *dev,					\
860 			       struct device_attribute *attr,		\
861 			       char *page)				\
862 {									\
863 	BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);			\
864 	return sprintf(page, _format "\n");				\
865 }									\
866 									\
867 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
868 
869 #endif /* _LINUX_PERF_EVENT_H */
870