xref: /linux-6.15/include/linux/perf_event.h (revision db181ce0)
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  * pmu::capabilities flags
171  */
172 #define PERF_PMU_CAP_NO_INTERRUPT		0x01
173 
174 /**
175  * struct pmu - generic performance monitoring unit
176  */
177 struct pmu {
178 	struct list_head		entry;
179 
180 	struct module			*module;
181 	struct device			*dev;
182 	const struct attribute_group	**attr_groups;
183 	const char			*name;
184 	int				type;
185 
186 	/*
187 	 * various common per-pmu feature flags
188 	 */
189 	int				capabilities;
190 
191 	int * __percpu			pmu_disable_count;
192 	struct perf_cpu_context * __percpu pmu_cpu_context;
193 	int				task_ctx_nr;
194 	int				hrtimer_interval_ms;
195 
196 	/*
197 	 * Fully disable/enable this PMU, can be used to protect from the PMI
198 	 * as well as for lazy/batch writing of the MSRs.
199 	 */
200 	void (*pmu_enable)		(struct pmu *pmu); /* optional */
201 	void (*pmu_disable)		(struct pmu *pmu); /* optional */
202 
203 	/*
204 	 * Try and initialize the event for this PMU.
205 	 * Should return -ENOENT when the @event doesn't match this PMU.
206 	 */
207 	int (*event_init)		(struct perf_event *event);
208 
209 #define PERF_EF_START	0x01		/* start the counter when adding    */
210 #define PERF_EF_RELOAD	0x02		/* reload the counter when starting */
211 #define PERF_EF_UPDATE	0x04		/* update the counter when stopping */
212 
213 	/*
214 	 * Adds/Removes a counter to/from the PMU, can be done inside
215 	 * a transaction, see the ->*_txn() methods.
216 	 */
217 	int  (*add)			(struct perf_event *event, int flags);
218 	void (*del)			(struct perf_event *event, int flags);
219 
220 	/*
221 	 * Starts/Stops a counter present on the PMU. The PMI handler
222 	 * should stop the counter when perf_event_overflow() returns
223 	 * !0. ->start() will be used to continue.
224 	 */
225 	void (*start)			(struct perf_event *event, int flags);
226 	void (*stop)			(struct perf_event *event, int flags);
227 
228 	/*
229 	 * Updates the counter value of the event.
230 	 */
231 	void (*read)			(struct perf_event *event);
232 
233 	/*
234 	 * Group events scheduling is treated as a transaction, add
235 	 * group events as a whole and perform one schedulability test.
236 	 * If the test fails, roll back the whole group
237 	 *
238 	 * Start the transaction, after this ->add() doesn't need to
239 	 * do schedulability tests.
240 	 */
241 	void (*start_txn)		(struct pmu *pmu); /* optional */
242 	/*
243 	 * If ->start_txn() disabled the ->add() schedulability test
244 	 * then ->commit_txn() is required to perform one. On success
245 	 * the transaction is closed. On error the transaction is kept
246 	 * open until ->cancel_txn() is called.
247 	 */
248 	int  (*commit_txn)		(struct pmu *pmu); /* optional */
249 	/*
250 	 * Will cancel the transaction, assumes ->del() is called
251 	 * for each successful ->add() during the transaction.
252 	 */
253 	void (*cancel_txn)		(struct pmu *pmu); /* optional */
254 
255 	/*
256 	 * Will return the value for perf_event_mmap_page::index for this event,
257 	 * if no implementation is provided it will default to: event->hw.idx + 1.
258 	 */
259 	int (*event_idx)		(struct perf_event *event); /*optional */
260 
261 	/*
262 	 * flush branch stack on context-switches (needed in cpu-wide mode)
263 	 */
264 	void (*flush_branch_stack)	(void);
265 };
266 
267 /**
268  * enum perf_event_active_state - the states of a event
269  */
270 enum perf_event_active_state {
271 	PERF_EVENT_STATE_ERROR		= -2,
272 	PERF_EVENT_STATE_OFF		= -1,
273 	PERF_EVENT_STATE_INACTIVE	=  0,
274 	PERF_EVENT_STATE_ACTIVE		=  1,
275 };
276 
277 struct file;
278 struct perf_sample_data;
279 
280 typedef void (*perf_overflow_handler_t)(struct perf_event *,
281 					struct perf_sample_data *,
282 					struct pt_regs *regs);
283 
284 enum perf_group_flag {
285 	PERF_GROUP_SOFTWARE		= 0x1,
286 };
287 
288 #define SWEVENT_HLIST_BITS		8
289 #define SWEVENT_HLIST_SIZE		(1 << SWEVENT_HLIST_BITS)
290 
291 struct swevent_hlist {
292 	struct hlist_head		heads[SWEVENT_HLIST_SIZE];
293 	struct rcu_head			rcu_head;
294 };
295 
296 #define PERF_ATTACH_CONTEXT	0x01
297 #define PERF_ATTACH_GROUP	0x02
298 #define PERF_ATTACH_TASK	0x04
299 
300 struct perf_cgroup;
301 struct ring_buffer;
302 
303 /**
304  * struct perf_event - performance event kernel representation:
305  */
306 struct perf_event {
307 #ifdef CONFIG_PERF_EVENTS
308 	/*
309 	 * entry onto perf_event_context::event_list;
310 	 *   modifications require ctx->lock
311 	 *   RCU safe iterations.
312 	 */
313 	struct list_head		event_entry;
314 
315 	/*
316 	 * XXX: group_entry and sibling_list should be mutually exclusive;
317 	 * either you're a sibling on a group, or you're the group leader.
318 	 * Rework the code to always use the same list element.
319 	 *
320 	 * Locked for modification by both ctx->mutex and ctx->lock; holding
321 	 * either sufficies for read.
322 	 */
323 	struct list_head		group_entry;
324 	struct list_head		sibling_list;
325 
326 	/*
327 	 * We need storage to track the entries in perf_pmu_migrate_context; we
328 	 * cannot use the event_entry because of RCU and we want to keep the
329 	 * group in tact which avoids us using the other two entries.
330 	 */
331 	struct list_head		migrate_entry;
332 
333 	struct hlist_node		hlist_entry;
334 	struct list_head		active_entry;
335 	int				nr_siblings;
336 	int				group_flags;
337 	struct perf_event		*group_leader;
338 	struct pmu			*pmu;
339 
340 	enum perf_event_active_state	state;
341 	unsigned int			attach_state;
342 	local64_t			count;
343 	atomic64_t			child_count;
344 
345 	/*
346 	 * These are the total time in nanoseconds that the event
347 	 * has been enabled (i.e. eligible to run, and the task has
348 	 * been scheduled in, if this is a per-task event)
349 	 * and running (scheduled onto the CPU), respectively.
350 	 *
351 	 * They are computed from tstamp_enabled, tstamp_running and
352 	 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
353 	 */
354 	u64				total_time_enabled;
355 	u64				total_time_running;
356 
357 	/*
358 	 * These are timestamps used for computing total_time_enabled
359 	 * and total_time_running when the event is in INACTIVE or
360 	 * ACTIVE state, measured in nanoseconds from an arbitrary point
361 	 * in time.
362 	 * tstamp_enabled: the notional time when the event was enabled
363 	 * tstamp_running: the notional time when the event was scheduled on
364 	 * tstamp_stopped: in INACTIVE state, the notional time when the
365 	 *	event was scheduled off.
366 	 */
367 	u64				tstamp_enabled;
368 	u64				tstamp_running;
369 	u64				tstamp_stopped;
370 
371 	/*
372 	 * timestamp shadows the actual context timing but it can
373 	 * be safely used in NMI interrupt context. It reflects the
374 	 * context time as it was when the event was last scheduled in.
375 	 *
376 	 * ctx_time already accounts for ctx->timestamp. Therefore to
377 	 * compute ctx_time for a sample, simply add perf_clock().
378 	 */
379 	u64				shadow_ctx_time;
380 
381 	struct perf_event_attr		attr;
382 	u16				header_size;
383 	u16				id_header_size;
384 	u16				read_size;
385 	struct hw_perf_event		hw;
386 
387 	struct perf_event_context	*ctx;
388 	atomic_long_t			refcount;
389 
390 	/*
391 	 * These accumulate total time (in nanoseconds) that children
392 	 * events have been enabled and running, respectively.
393 	 */
394 	atomic64_t			child_total_time_enabled;
395 	atomic64_t			child_total_time_running;
396 
397 	/*
398 	 * Protect attach/detach and child_list:
399 	 */
400 	struct mutex			child_mutex;
401 	struct list_head		child_list;
402 	struct perf_event		*parent;
403 
404 	int				oncpu;
405 	int				cpu;
406 
407 	struct list_head		owner_entry;
408 	struct task_struct		*owner;
409 
410 	/* mmap bits */
411 	struct mutex			mmap_mutex;
412 	atomic_t			mmap_count;
413 
414 	struct ring_buffer		*rb;
415 	struct list_head		rb_entry;
416 	unsigned long			rcu_batches;
417 	int				rcu_pending;
418 
419 	/* poll related */
420 	wait_queue_head_t		waitq;
421 	struct fasync_struct		*fasync;
422 
423 	/* delayed work for NMIs and such */
424 	int				pending_wakeup;
425 	int				pending_kill;
426 	int				pending_disable;
427 	struct irq_work			pending;
428 
429 	atomic_t			event_limit;
430 
431 	void (*destroy)(struct perf_event *);
432 	struct rcu_head			rcu_head;
433 
434 	struct pid_namespace		*ns;
435 	u64				id;
436 
437 	perf_overflow_handler_t		overflow_handler;
438 	void				*overflow_handler_context;
439 
440 #ifdef CONFIG_EVENT_TRACING
441 	struct ftrace_event_call	*tp_event;
442 	struct event_filter		*filter;
443 #ifdef CONFIG_FUNCTION_TRACER
444 	struct ftrace_ops               ftrace_ops;
445 #endif
446 #endif
447 
448 #ifdef CONFIG_CGROUP_PERF
449 	struct perf_cgroup		*cgrp; /* cgroup event is attach to */
450 	int				cgrp_defer_enabled;
451 #endif
452 
453 #endif /* CONFIG_PERF_EVENTS */
454 };
455 
456 enum perf_event_context_type {
457 	task_context,
458 	cpu_context,
459 };
460 
461 /**
462  * struct perf_event_context - event context structure
463  *
464  * Used as a container for task events and CPU events as well:
465  */
466 struct perf_event_context {
467 	struct pmu			*pmu;
468 	enum perf_event_context_type	type;
469 	/*
470 	 * Protect the states of the events in the list,
471 	 * nr_active, and the list:
472 	 */
473 	raw_spinlock_t			lock;
474 	/*
475 	 * Protect the list of events.  Locking either mutex or lock
476 	 * is sufficient to ensure the list doesn't change; to change
477 	 * the list you need to lock both the mutex and the spinlock.
478 	 */
479 	struct mutex			mutex;
480 
481 	struct list_head		pinned_groups;
482 	struct list_head		flexible_groups;
483 	struct list_head		event_list;
484 	int				nr_events;
485 	int				nr_active;
486 	int				is_active;
487 	int				nr_stat;
488 	int				nr_freq;
489 	int				rotate_disable;
490 	atomic_t			refcount;
491 	struct task_struct		*task;
492 
493 	/*
494 	 * Context clock, runs when context enabled.
495 	 */
496 	u64				time;
497 	u64				timestamp;
498 
499 	/*
500 	 * These fields let us detect when two contexts have both
501 	 * been cloned (inherited) from a common ancestor.
502 	 */
503 	struct perf_event_context	*parent_ctx;
504 	u64				parent_gen;
505 	u64				generation;
506 	int				pin_count;
507 	int				nr_cgroups;	 /* cgroup evts */
508 	int				nr_branch_stack; /* branch_stack evt */
509 	struct rcu_head			rcu_head;
510 };
511 
512 /*
513  * Number of contexts where an event can trigger:
514  *	task, softirq, hardirq, nmi.
515  */
516 #define PERF_NR_CONTEXTS	4
517 
518 /**
519  * struct perf_event_cpu_context - per cpu event context structure
520  */
521 struct perf_cpu_context {
522 	struct perf_event_context	ctx;
523 	struct perf_event_context	*task_ctx;
524 	int				active_oncpu;
525 	int				exclusive;
526 	struct hrtimer			hrtimer;
527 	ktime_t				hrtimer_interval;
528 	struct list_head		rotation_list;
529 	struct pmu			*unique_pmu;
530 	struct perf_cgroup		*cgrp;
531 };
532 
533 struct perf_output_handle {
534 	struct perf_event		*event;
535 	struct ring_buffer		*rb;
536 	unsigned long			wakeup;
537 	unsigned long			size;
538 	void				*addr;
539 	int				page;
540 };
541 
542 #ifdef CONFIG_PERF_EVENTS
543 
544 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
545 extern void perf_pmu_unregister(struct pmu *pmu);
546 
547 extern int perf_num_counters(void);
548 extern const char *perf_pmu_name(void);
549 extern void __perf_event_task_sched_in(struct task_struct *prev,
550 				       struct task_struct *task);
551 extern void __perf_event_task_sched_out(struct task_struct *prev,
552 					struct task_struct *next);
553 extern int perf_event_init_task(struct task_struct *child);
554 extern void perf_event_exit_task(struct task_struct *child);
555 extern void perf_event_free_task(struct task_struct *task);
556 extern void perf_event_delayed_put(struct task_struct *task);
557 extern void perf_event_print_debug(void);
558 extern void perf_pmu_disable(struct pmu *pmu);
559 extern void perf_pmu_enable(struct pmu *pmu);
560 extern int perf_event_task_disable(void);
561 extern int perf_event_task_enable(void);
562 extern int perf_event_refresh(struct perf_event *event, int refresh);
563 extern void perf_event_update_userpage(struct perf_event *event);
564 extern int perf_event_release_kernel(struct perf_event *event);
565 extern struct perf_event *
566 perf_event_create_kernel_counter(struct perf_event_attr *attr,
567 				int cpu,
568 				struct task_struct *task,
569 				perf_overflow_handler_t callback,
570 				void *context);
571 extern void perf_pmu_migrate_context(struct pmu *pmu,
572 				int src_cpu, int dst_cpu);
573 extern u64 perf_event_read_value(struct perf_event *event,
574 				 u64 *enabled, u64 *running);
575 
576 
577 struct perf_sample_data {
578 	u64				type;
579 
580 	u64				ip;
581 	struct {
582 		u32	pid;
583 		u32	tid;
584 	}				tid_entry;
585 	u64				time;
586 	u64				addr;
587 	u64				id;
588 	u64				stream_id;
589 	struct {
590 		u32	cpu;
591 		u32	reserved;
592 	}				cpu_entry;
593 	u64				period;
594 	union  perf_mem_data_src	data_src;
595 	struct perf_callchain_entry	*callchain;
596 	struct perf_raw_record		*raw;
597 	struct perf_branch_stack	*br_stack;
598 	struct perf_regs_user		regs_user;
599 	u64				stack_user_size;
600 	u64				weight;
601 	/*
602 	 * Transaction flags for abort events:
603 	 */
604 	u64				txn;
605 };
606 
607 static inline void perf_sample_data_init(struct perf_sample_data *data,
608 					 u64 addr, u64 period)
609 {
610 	/* remaining struct members initialized in perf_prepare_sample() */
611 	data->addr = addr;
612 	data->raw  = NULL;
613 	data->br_stack = NULL;
614 	data->period = period;
615 	data->regs_user.abi = PERF_SAMPLE_REGS_ABI_NONE;
616 	data->regs_user.regs = NULL;
617 	data->stack_user_size = 0;
618 	data->weight = 0;
619 	data->data_src.val = 0;
620 	data->txn = 0;
621 }
622 
623 extern void perf_output_sample(struct perf_output_handle *handle,
624 			       struct perf_event_header *header,
625 			       struct perf_sample_data *data,
626 			       struct perf_event *event);
627 extern void perf_prepare_sample(struct perf_event_header *header,
628 				struct perf_sample_data *data,
629 				struct perf_event *event,
630 				struct pt_regs *regs);
631 
632 extern int perf_event_overflow(struct perf_event *event,
633 				 struct perf_sample_data *data,
634 				 struct pt_regs *regs);
635 
636 static inline bool is_sampling_event(struct perf_event *event)
637 {
638 	return event->attr.sample_period != 0;
639 }
640 
641 /*
642  * Return 1 for a software event, 0 for a hardware event
643  */
644 static inline int is_software_event(struct perf_event *event)
645 {
646 	return event->pmu->task_ctx_nr == perf_sw_context;
647 }
648 
649 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
650 
651 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
652 
653 #ifndef perf_arch_fetch_caller_regs
654 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
655 #endif
656 
657 /*
658  * Take a snapshot of the regs. Skip ip and frame pointer to
659  * the nth caller. We only need a few of the regs:
660  * - ip for PERF_SAMPLE_IP
661  * - cs for user_mode() tests
662  * - bp for callchains
663  * - eflags, for future purposes, just in case
664  */
665 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
666 {
667 	memset(regs, 0, sizeof(*regs));
668 
669 	perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
670 }
671 
672 static __always_inline void
673 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
674 {
675 	struct pt_regs hot_regs;
676 
677 	if (static_key_false(&perf_swevent_enabled[event_id])) {
678 		if (!regs) {
679 			perf_fetch_caller_regs(&hot_regs);
680 			regs = &hot_regs;
681 		}
682 		__perf_sw_event(event_id, nr, regs, addr);
683 	}
684 }
685 
686 extern struct static_key_deferred perf_sched_events;
687 
688 static inline void perf_event_task_sched_in(struct task_struct *prev,
689 					    struct task_struct *task)
690 {
691 	if (static_key_false(&perf_sched_events.key))
692 		__perf_event_task_sched_in(prev, task);
693 }
694 
695 static inline void perf_event_task_sched_out(struct task_struct *prev,
696 					     struct task_struct *next)
697 {
698 	perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, NULL, 0);
699 
700 	if (static_key_false(&perf_sched_events.key))
701 		__perf_event_task_sched_out(prev, next);
702 }
703 
704 extern void perf_event_mmap(struct vm_area_struct *vma);
705 extern struct perf_guest_info_callbacks *perf_guest_cbs;
706 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
707 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
708 
709 extern void perf_event_exec(void);
710 extern void perf_event_comm(struct task_struct *tsk, bool exec);
711 extern void perf_event_fork(struct task_struct *tsk);
712 
713 /* Callchains */
714 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
715 
716 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
717 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
718 
719 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
720 {
721 	if (entry->nr < PERF_MAX_STACK_DEPTH)
722 		entry->ip[entry->nr++] = ip;
723 }
724 
725 extern int sysctl_perf_event_paranoid;
726 extern int sysctl_perf_event_mlock;
727 extern int sysctl_perf_event_sample_rate;
728 extern int sysctl_perf_cpu_time_max_percent;
729 
730 extern void perf_sample_event_took(u64 sample_len_ns);
731 
732 extern int perf_proc_update_handler(struct ctl_table *table, int write,
733 		void __user *buffer, size_t *lenp,
734 		loff_t *ppos);
735 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
736 		void __user *buffer, size_t *lenp,
737 		loff_t *ppos);
738 
739 
740 static inline bool perf_paranoid_tracepoint_raw(void)
741 {
742 	return sysctl_perf_event_paranoid > -1;
743 }
744 
745 static inline bool perf_paranoid_cpu(void)
746 {
747 	return sysctl_perf_event_paranoid > 0;
748 }
749 
750 static inline bool perf_paranoid_kernel(void)
751 {
752 	return sysctl_perf_event_paranoid > 1;
753 }
754 
755 extern void perf_event_init(void);
756 extern void perf_tp_event(u64 addr, u64 count, void *record,
757 			  int entry_size, struct pt_regs *regs,
758 			  struct hlist_head *head, int rctx,
759 			  struct task_struct *task);
760 extern void perf_bp_event(struct perf_event *event, void *data);
761 
762 #ifndef perf_misc_flags
763 # define perf_misc_flags(regs) \
764 		(user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
765 # define perf_instruction_pointer(regs)	instruction_pointer(regs)
766 #endif
767 
768 static inline bool has_branch_stack(struct perf_event *event)
769 {
770 	return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
771 }
772 
773 extern int perf_output_begin(struct perf_output_handle *handle,
774 			     struct perf_event *event, unsigned int size);
775 extern void perf_output_end(struct perf_output_handle *handle);
776 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
777 			     const void *buf, unsigned int len);
778 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
779 				     unsigned int len);
780 extern int perf_swevent_get_recursion_context(void);
781 extern void perf_swevent_put_recursion_context(int rctx);
782 extern u64 perf_swevent_set_period(struct perf_event *event);
783 extern void perf_event_enable(struct perf_event *event);
784 extern void perf_event_disable(struct perf_event *event);
785 extern int __perf_event_disable(void *info);
786 extern void perf_event_task_tick(void);
787 #else /* !CONFIG_PERF_EVENTS: */
788 static inline void
789 perf_event_task_sched_in(struct task_struct *prev,
790 			 struct task_struct *task)			{ }
791 static inline void
792 perf_event_task_sched_out(struct task_struct *prev,
793 			  struct task_struct *next)			{ }
794 static inline int perf_event_init_task(struct task_struct *child)	{ return 0; }
795 static inline void perf_event_exit_task(struct task_struct *child)	{ }
796 static inline void perf_event_free_task(struct task_struct *task)	{ }
797 static inline void perf_event_delayed_put(struct task_struct *task)	{ }
798 static inline void perf_event_print_debug(void)				{ }
799 static inline int perf_event_task_disable(void)				{ return -EINVAL; }
800 static inline int perf_event_task_enable(void)				{ return -EINVAL; }
801 static inline int perf_event_refresh(struct perf_event *event, int refresh)
802 {
803 	return -EINVAL;
804 }
805 
806 static inline void
807 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)	{ }
808 static inline void
809 perf_bp_event(struct perf_event *event, void *data)			{ }
810 
811 static inline int perf_register_guest_info_callbacks
812 (struct perf_guest_info_callbacks *callbacks)				{ return 0; }
813 static inline int perf_unregister_guest_info_callbacks
814 (struct perf_guest_info_callbacks *callbacks)				{ return 0; }
815 
816 static inline void perf_event_mmap(struct vm_area_struct *vma)		{ }
817 static inline void perf_event_exec(void)				{ }
818 static inline void perf_event_comm(struct task_struct *tsk, bool exec)	{ }
819 static inline void perf_event_fork(struct task_struct *tsk)		{ }
820 static inline void perf_event_init(void)				{ }
821 static inline int  perf_swevent_get_recursion_context(void)		{ return -1; }
822 static inline void perf_swevent_put_recursion_context(int rctx)		{ }
823 static inline u64 perf_swevent_set_period(struct perf_event *event)	{ return 0; }
824 static inline void perf_event_enable(struct perf_event *event)		{ }
825 static inline void perf_event_disable(struct perf_event *event)		{ }
826 static inline int __perf_event_disable(void *info)			{ return -1; }
827 static inline void perf_event_task_tick(void)				{ }
828 #endif
829 
830 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_NO_HZ_FULL)
831 extern bool perf_event_can_stop_tick(void);
832 #else
833 static inline bool perf_event_can_stop_tick(void)			{ return true; }
834 #endif
835 
836 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
837 extern void perf_restore_debug_store(void);
838 #else
839 static inline void perf_restore_debug_store(void)			{ }
840 #endif
841 
842 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
843 
844 /*
845  * This has to have a higher priority than migration_notifier in sched/core.c.
846  */
847 #define perf_cpu_notifier(fn)						\
848 do {									\
849 	static struct notifier_block fn##_nb =				\
850 		{ .notifier_call = fn, .priority = CPU_PRI_PERF };	\
851 	unsigned long cpu = smp_processor_id();				\
852 	unsigned long flags;						\
853 									\
854 	cpu_notifier_register_begin();					\
855 	fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,			\
856 		(void *)(unsigned long)cpu);				\
857 	local_irq_save(flags);						\
858 	fn(&fn##_nb, (unsigned long)CPU_STARTING,			\
859 		(void *)(unsigned long)cpu);				\
860 	local_irq_restore(flags);					\
861 	fn(&fn##_nb, (unsigned long)CPU_ONLINE,				\
862 		(void *)(unsigned long)cpu);				\
863 	__register_cpu_notifier(&fn##_nb);				\
864 	cpu_notifier_register_done();					\
865 } while (0)
866 
867 /*
868  * Bare-bones version of perf_cpu_notifier(), which doesn't invoke the
869  * callback for already online CPUs.
870  */
871 #define __perf_cpu_notifier(fn)						\
872 do {									\
873 	static struct notifier_block fn##_nb =				\
874 		{ .notifier_call = fn, .priority = CPU_PRI_PERF };	\
875 									\
876 	__register_cpu_notifier(&fn##_nb);				\
877 } while (0)
878 
879 struct perf_pmu_events_attr {
880 	struct device_attribute attr;
881 	u64 id;
882 	const char *event_str;
883 };
884 
885 #define PMU_EVENT_ATTR(_name, _var, _id, _show)				\
886 static struct perf_pmu_events_attr _var = {				\
887 	.attr = __ATTR(_name, 0444, _show, NULL),			\
888 	.id   =  _id,							\
889 };
890 
891 #define PMU_FORMAT_ATTR(_name, _format)					\
892 static ssize_t								\
893 _name##_show(struct device *dev,					\
894 			       struct device_attribute *attr,		\
895 			       char *page)				\
896 {									\
897 	BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);			\
898 	return sprintf(page, _format "\n");				\
899 }									\
900 									\
901 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
902 
903 #endif /* _LINUX_PERF_EVENT_H */
904