xref: /linux-6.15/include/linux/perf_event.h (revision bb66fc67)
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/jump_label_ratelimit.h>
52 #include <linux/atomic.h>
53 #include <linux/sysfs.h>
54 #include <linux/perf_regs.h>
55 #include <asm/local.h>
56 
57 struct perf_callchain_entry {
58 	__u64				nr;
59 	__u64				ip[PERF_MAX_STACK_DEPTH];
60 };
61 
62 struct perf_raw_record {
63 	u32				size;
64 	void				*data;
65 };
66 
67 /*
68  * branch stack layout:
69  *  nr: number of taken branches stored in entries[]
70  *
71  * Note that nr can vary from sample to sample
72  * branches (to, from) are stored from most recent
73  * to least recent, i.e., entries[0] contains the most
74  * recent branch.
75  */
76 struct perf_branch_stack {
77 	__u64				nr;
78 	struct perf_branch_entry	entries[0];
79 };
80 
81 struct perf_regs_user {
82 	__u64		abi;
83 	struct pt_regs	*regs;
84 };
85 
86 struct task_struct;
87 
88 /*
89  * extra PMU register associated with an event
90  */
91 struct hw_perf_event_extra {
92 	u64		config;	/* register value */
93 	unsigned int	reg;	/* register address or index */
94 	int		alloc;	/* extra register already allocated */
95 	int		idx;	/* index in shared_regs->regs[] */
96 };
97 
98 struct event_constraint;
99 
100 /**
101  * struct hw_perf_event - performance event hardware details:
102  */
103 struct hw_perf_event {
104 #ifdef CONFIG_PERF_EVENTS
105 	union {
106 		struct { /* hardware */
107 			u64		config;
108 			u64		last_tag;
109 			unsigned long	config_base;
110 			unsigned long	event_base;
111 			int		event_base_rdpmc;
112 			int		idx;
113 			int		last_cpu;
114 			int		flags;
115 
116 			struct hw_perf_event_extra extra_reg;
117 			struct hw_perf_event_extra branch_reg;
118 
119 			struct event_constraint *constraint;
120 		};
121 		struct { /* software */
122 			struct hrtimer	hrtimer;
123 		};
124 		struct { /* tracepoint */
125 			struct task_struct	*tp_target;
126 			/* for tp_event->class */
127 			struct list_head	tp_list;
128 		};
129 #ifdef CONFIG_HAVE_HW_BREAKPOINT
130 		struct { /* breakpoint */
131 			/*
132 			 * Crufty hack to avoid the chicken and egg
133 			 * problem hw_breakpoint has with context
134 			 * creation and event initalization.
135 			 */
136 			struct task_struct		*bp_target;
137 			struct arch_hw_breakpoint	info;
138 			struct list_head		bp_list;
139 		};
140 #endif
141 	};
142 	int				state;
143 	local64_t			prev_count;
144 	u64				sample_period;
145 	u64				last_period;
146 	local64_t			period_left;
147 	u64                             interrupts_seq;
148 	u64				interrupts;
149 
150 	u64				freq_time_stamp;
151 	u64				freq_count_stamp;
152 #endif
153 };
154 
155 /*
156  * hw_perf_event::state flags
157  */
158 #define PERF_HES_STOPPED	0x01 /* the counter is stopped */
159 #define PERF_HES_UPTODATE	0x02 /* event->count up-to-date */
160 #define PERF_HES_ARCH		0x04
161 
162 struct perf_event;
163 
164 /*
165  * Common implementation detail of pmu::{start,commit,cancel}_txn
166  */
167 #define PERF_EVENT_TXN 0x1
168 
169 /**
170  * struct pmu - generic performance monitoring unit
171  */
172 struct pmu {
173 	struct list_head		entry;
174 
175 	struct device			*dev;
176 	const struct attribute_group	**attr_groups;
177 	const char			*name;
178 	int				type;
179 
180 	int * __percpu			pmu_disable_count;
181 	struct perf_cpu_context * __percpu pmu_cpu_context;
182 	int				task_ctx_nr;
183 	int				hrtimer_interval_ms;
184 
185 	/*
186 	 * Fully disable/enable this PMU, can be used to protect from the PMI
187 	 * as well as for lazy/batch writing of the MSRs.
188 	 */
189 	void (*pmu_enable)		(struct pmu *pmu); /* optional */
190 	void (*pmu_disable)		(struct pmu *pmu); /* optional */
191 
192 	/*
193 	 * Try and initialize the event for this PMU.
194 	 * Should return -ENOENT when the @event doesn't match this PMU.
195 	 */
196 	int (*event_init)		(struct perf_event *event);
197 
198 #define PERF_EF_START	0x01		/* start the counter when adding    */
199 #define PERF_EF_RELOAD	0x02		/* reload the counter when starting */
200 #define PERF_EF_UPDATE	0x04		/* update the counter when stopping */
201 
202 	/*
203 	 * Adds/Removes a counter to/from the PMU, can be done inside
204 	 * a transaction, see the ->*_txn() methods.
205 	 */
206 	int  (*add)			(struct perf_event *event, int flags);
207 	void (*del)			(struct perf_event *event, int flags);
208 
209 	/*
210 	 * Starts/Stops a counter present on the PMU. The PMI handler
211 	 * should stop the counter when perf_event_overflow() returns
212 	 * !0. ->start() will be used to continue.
213 	 */
214 	void (*start)			(struct perf_event *event, int flags);
215 	void (*stop)			(struct perf_event *event, int flags);
216 
217 	/*
218 	 * Updates the counter value of the event.
219 	 */
220 	void (*read)			(struct perf_event *event);
221 
222 	/*
223 	 * Group events scheduling is treated as a transaction, add
224 	 * group events as a whole and perform one schedulability test.
225 	 * If the test fails, roll back the whole group
226 	 *
227 	 * Start the transaction, after this ->add() doesn't need to
228 	 * do schedulability tests.
229 	 */
230 	void (*start_txn)		(struct pmu *pmu); /* optional */
231 	/*
232 	 * If ->start_txn() disabled the ->add() schedulability test
233 	 * then ->commit_txn() is required to perform one. On success
234 	 * the transaction is closed. On error the transaction is kept
235 	 * open until ->cancel_txn() is called.
236 	 */
237 	int  (*commit_txn)		(struct pmu *pmu); /* optional */
238 	/*
239 	 * Will cancel the transaction, assumes ->del() is called
240 	 * for each successful ->add() during the transaction.
241 	 */
242 	void (*cancel_txn)		(struct pmu *pmu); /* optional */
243 
244 	/*
245 	 * Will return the value for perf_event_mmap_page::index for this event,
246 	 * if no implementation is provided it will default to: event->hw.idx + 1.
247 	 */
248 	int (*event_idx)		(struct perf_event *event); /*optional */
249 
250 	/*
251 	 * flush branch stack on context-switches (needed in cpu-wide mode)
252 	 */
253 	void (*flush_branch_stack)	(void);
254 };
255 
256 /**
257  * enum perf_event_active_state - the states of a event
258  */
259 enum perf_event_active_state {
260 	PERF_EVENT_STATE_ERROR		= -2,
261 	PERF_EVENT_STATE_OFF		= -1,
262 	PERF_EVENT_STATE_INACTIVE	=  0,
263 	PERF_EVENT_STATE_ACTIVE		=  1,
264 };
265 
266 struct file;
267 struct perf_sample_data;
268 
269 typedef void (*perf_overflow_handler_t)(struct perf_event *,
270 					struct perf_sample_data *,
271 					struct pt_regs *regs);
272 
273 enum perf_group_flag {
274 	PERF_GROUP_SOFTWARE		= 0x1,
275 };
276 
277 #define SWEVENT_HLIST_BITS		8
278 #define SWEVENT_HLIST_SIZE		(1 << SWEVENT_HLIST_BITS)
279 
280 struct swevent_hlist {
281 	struct hlist_head		heads[SWEVENT_HLIST_SIZE];
282 	struct rcu_head			rcu_head;
283 };
284 
285 #define PERF_ATTACH_CONTEXT	0x01
286 #define PERF_ATTACH_GROUP	0x02
287 #define PERF_ATTACH_TASK	0x04
288 
289 struct perf_cgroup;
290 struct ring_buffer;
291 
292 /**
293  * struct perf_event - performance event kernel representation:
294  */
295 struct perf_event {
296 #ifdef CONFIG_PERF_EVENTS
297 	/*
298 	 * entry onto perf_event_context::event_list;
299 	 *   modifications require ctx->lock
300 	 *   RCU safe iterations.
301 	 */
302 	struct list_head		event_entry;
303 
304 	/*
305 	 * XXX: group_entry and sibling_list should be mutually exclusive;
306 	 * either you're a sibling on a group, or you're the group leader.
307 	 * Rework the code to always use the same list element.
308 	 *
309 	 * Locked for modification by both ctx->mutex and ctx->lock; holding
310 	 * either sufficies for read.
311 	 */
312 	struct list_head		group_entry;
313 	struct list_head		sibling_list;
314 
315 	/*
316 	 * We need storage to track the entries in perf_pmu_migrate_context; we
317 	 * cannot use the event_entry because of RCU and we want to keep the
318 	 * group in tact which avoids us using the other two entries.
319 	 */
320 	struct list_head		migrate_entry;
321 
322 	struct hlist_node		hlist_entry;
323 	struct list_head		active_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 	 * Transaction flags for abort events:
590 	 */
591 	u64				txn;
592 };
593 
594 static inline void perf_sample_data_init(struct perf_sample_data *data,
595 					 u64 addr, u64 period)
596 {
597 	/* remaining struct members initialized in perf_prepare_sample() */
598 	data->addr = addr;
599 	data->raw  = NULL;
600 	data->br_stack = NULL;
601 	data->period = period;
602 	data->regs_user.abi = PERF_SAMPLE_REGS_ABI_NONE;
603 	data->regs_user.regs = NULL;
604 	data->stack_user_size = 0;
605 	data->weight = 0;
606 	data->data_src.val = 0;
607 	data->txn = 0;
608 }
609 
610 extern void perf_output_sample(struct perf_output_handle *handle,
611 			       struct perf_event_header *header,
612 			       struct perf_sample_data *data,
613 			       struct perf_event *event);
614 extern void perf_prepare_sample(struct perf_event_header *header,
615 				struct perf_sample_data *data,
616 				struct perf_event *event,
617 				struct pt_regs *regs);
618 
619 extern int perf_event_overflow(struct perf_event *event,
620 				 struct perf_sample_data *data,
621 				 struct pt_regs *regs);
622 
623 static inline bool is_sampling_event(struct perf_event *event)
624 {
625 	return event->attr.sample_period != 0;
626 }
627 
628 /*
629  * Return 1 for a software event, 0 for a hardware event
630  */
631 static inline int is_software_event(struct perf_event *event)
632 {
633 	return event->pmu->task_ctx_nr == perf_sw_context;
634 }
635 
636 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
637 
638 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
639 
640 #ifndef perf_arch_fetch_caller_regs
641 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
642 #endif
643 
644 /*
645  * Take a snapshot of the regs. Skip ip and frame pointer to
646  * the nth caller. We only need a few of the regs:
647  * - ip for PERF_SAMPLE_IP
648  * - cs for user_mode() tests
649  * - bp for callchains
650  * - eflags, for future purposes, just in case
651  */
652 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
653 {
654 	memset(regs, 0, sizeof(*regs));
655 
656 	perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
657 }
658 
659 static __always_inline void
660 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
661 {
662 	struct pt_regs hot_regs;
663 
664 	if (static_key_false(&perf_swevent_enabled[event_id])) {
665 		if (!regs) {
666 			perf_fetch_caller_regs(&hot_regs);
667 			regs = &hot_regs;
668 		}
669 		__perf_sw_event(event_id, nr, regs, addr);
670 	}
671 }
672 
673 extern struct static_key_deferred perf_sched_events;
674 
675 static inline void perf_event_task_sched_in(struct task_struct *prev,
676 					    struct task_struct *task)
677 {
678 	if (static_key_false(&perf_sched_events.key))
679 		__perf_event_task_sched_in(prev, task);
680 }
681 
682 static inline void perf_event_task_sched_out(struct task_struct *prev,
683 					     struct task_struct *next)
684 {
685 	perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, NULL, 0);
686 
687 	if (static_key_false(&perf_sched_events.key))
688 		__perf_event_task_sched_out(prev, next);
689 }
690 
691 extern void perf_event_mmap(struct vm_area_struct *vma);
692 extern struct perf_guest_info_callbacks *perf_guest_cbs;
693 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
694 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
695 
696 extern void perf_event_comm(struct task_struct *tsk);
697 extern void perf_event_fork(struct task_struct *tsk);
698 
699 /* Callchains */
700 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
701 
702 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
703 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
704 
705 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
706 {
707 	if (entry->nr < PERF_MAX_STACK_DEPTH)
708 		entry->ip[entry->nr++] = ip;
709 }
710 
711 extern int sysctl_perf_event_paranoid;
712 extern int sysctl_perf_event_mlock;
713 extern int sysctl_perf_event_sample_rate;
714 extern int sysctl_perf_cpu_time_max_percent;
715 
716 extern void perf_sample_event_took(u64 sample_len_ns);
717 
718 extern int perf_proc_update_handler(struct ctl_table *table, int write,
719 		void __user *buffer, size_t *lenp,
720 		loff_t *ppos);
721 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
722 		void __user *buffer, size_t *lenp,
723 		loff_t *ppos);
724 
725 
726 static inline bool perf_paranoid_tracepoint_raw(void)
727 {
728 	return sysctl_perf_event_paranoid > -1;
729 }
730 
731 static inline bool perf_paranoid_cpu(void)
732 {
733 	return sysctl_perf_event_paranoid > 0;
734 }
735 
736 static inline bool perf_paranoid_kernel(void)
737 {
738 	return sysctl_perf_event_paranoid > 1;
739 }
740 
741 extern void perf_event_init(void);
742 extern void perf_tp_event(u64 addr, u64 count, void *record,
743 			  int entry_size, struct pt_regs *regs,
744 			  struct hlist_head *head, int rctx,
745 			  struct task_struct *task);
746 extern void perf_bp_event(struct perf_event *event, void *data);
747 
748 #ifndef perf_misc_flags
749 # define perf_misc_flags(regs) \
750 		(user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
751 # define perf_instruction_pointer(regs)	instruction_pointer(regs)
752 #endif
753 
754 static inline bool has_branch_stack(struct perf_event *event)
755 {
756 	return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
757 }
758 
759 extern int perf_output_begin(struct perf_output_handle *handle,
760 			     struct perf_event *event, unsigned int size);
761 extern void perf_output_end(struct perf_output_handle *handle);
762 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
763 			     const void *buf, unsigned int len);
764 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
765 				     unsigned int len);
766 extern int perf_swevent_get_recursion_context(void);
767 extern void perf_swevent_put_recursion_context(int rctx);
768 extern u64 perf_swevent_set_period(struct perf_event *event);
769 extern void perf_event_enable(struct perf_event *event);
770 extern void perf_event_disable(struct perf_event *event);
771 extern int __perf_event_disable(void *info);
772 extern void perf_event_task_tick(void);
773 #else
774 static inline void
775 perf_event_task_sched_in(struct task_struct *prev,
776 			 struct task_struct *task)			{ }
777 static inline void
778 perf_event_task_sched_out(struct task_struct *prev,
779 			  struct task_struct *next)			{ }
780 static inline int perf_event_init_task(struct task_struct *child)	{ return 0; }
781 static inline void perf_event_exit_task(struct task_struct *child)	{ }
782 static inline void perf_event_free_task(struct task_struct *task)	{ }
783 static inline void perf_event_delayed_put(struct task_struct *task)	{ }
784 static inline void perf_event_print_debug(void)				{ }
785 static inline int perf_event_task_disable(void)				{ return -EINVAL; }
786 static inline int perf_event_task_enable(void)				{ return -EINVAL; }
787 static inline int perf_event_refresh(struct perf_event *event, int refresh)
788 {
789 	return -EINVAL;
790 }
791 
792 static inline void
793 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)	{ }
794 static inline void
795 perf_bp_event(struct perf_event *event, void *data)			{ }
796 
797 static inline int perf_register_guest_info_callbacks
798 (struct perf_guest_info_callbacks *callbacks)				{ return 0; }
799 static inline int perf_unregister_guest_info_callbacks
800 (struct perf_guest_info_callbacks *callbacks)				{ return 0; }
801 
802 static inline void perf_event_mmap(struct vm_area_struct *vma)		{ }
803 static inline void perf_event_comm(struct task_struct *tsk)		{ }
804 static inline void perf_event_fork(struct task_struct *tsk)		{ }
805 static inline void perf_event_init(void)				{ }
806 static inline int  perf_swevent_get_recursion_context(void)		{ return -1; }
807 static inline void perf_swevent_put_recursion_context(int rctx)		{ }
808 static inline u64 perf_swevent_set_period(struct perf_event *event)	{ return 0; }
809 static inline void perf_event_enable(struct perf_event *event)		{ }
810 static inline void perf_event_disable(struct perf_event *event)		{ }
811 static inline int __perf_event_disable(void *info)			{ return -1; }
812 static inline void perf_event_task_tick(void)				{ }
813 #endif
814 
815 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_NO_HZ_FULL)
816 extern bool perf_event_can_stop_tick(void);
817 #else
818 static inline bool perf_event_can_stop_tick(void)			{ return true; }
819 #endif
820 
821 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
822 extern void perf_restore_debug_store(void);
823 #else
824 static inline void perf_restore_debug_store(void)			{ }
825 #endif
826 
827 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
828 
829 /*
830  * This has to have a higher priority than migration_notifier in sched/core.c.
831  */
832 #define perf_cpu_notifier(fn)						\
833 do {									\
834 	static struct notifier_block fn##_nb =				\
835 		{ .notifier_call = fn, .priority = CPU_PRI_PERF };	\
836 	unsigned long cpu = smp_processor_id();				\
837 	unsigned long flags;						\
838 									\
839 	cpu_notifier_register_begin();					\
840 	fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,			\
841 		(void *)(unsigned long)cpu);				\
842 	local_irq_save(flags);						\
843 	fn(&fn##_nb, (unsigned long)CPU_STARTING,			\
844 		(void *)(unsigned long)cpu);				\
845 	local_irq_restore(flags);					\
846 	fn(&fn##_nb, (unsigned long)CPU_ONLINE,				\
847 		(void *)(unsigned long)cpu);				\
848 	__register_cpu_notifier(&fn##_nb);				\
849 	cpu_notifier_register_done();					\
850 } while (0)
851 
852 /*
853  * Bare-bones version of perf_cpu_notifier(), which doesn't invoke the
854  * callback for already online CPUs.
855  */
856 #define __perf_cpu_notifier(fn)						\
857 do {									\
858 	static struct notifier_block fn##_nb =				\
859 		{ .notifier_call = fn, .priority = CPU_PRI_PERF };	\
860 									\
861 	__register_cpu_notifier(&fn##_nb);				\
862 } while (0)
863 
864 struct perf_pmu_events_attr {
865 	struct device_attribute attr;
866 	u64 id;
867 	const char *event_str;
868 };
869 
870 #define PMU_EVENT_ATTR(_name, _var, _id, _show)				\
871 static struct perf_pmu_events_attr _var = {				\
872 	.attr = __ATTR(_name, 0444, _show, NULL),			\
873 	.id   =  _id,							\
874 };
875 
876 #define PMU_FORMAT_ATTR(_name, _format)					\
877 static ssize_t								\
878 _name##_show(struct device *dev,					\
879 			       struct device_attribute *attr,		\
880 			       char *page)				\
881 {									\
882 	BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);			\
883 	return sprintf(page, _format "\n");				\
884 }									\
885 									\
886 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
887 
888 #endif /* _LINUX_PERF_EVENT_H */
889