xref: /linux-6.15/include/linux/perf_event.h (revision 8ab596af)
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 module			*module;
176 	struct device			*dev;
177 	const struct attribute_group	**attr_groups;
178 	const char			*name;
179 	int				type;
180 
181 	int * __percpu			pmu_disable_count;
182 	struct perf_cpu_context * __percpu pmu_cpu_context;
183 	int				task_ctx_nr;
184 	int				hrtimer_interval_ms;
185 
186 	/*
187 	 * Fully disable/enable this PMU, can be used to protect from the PMI
188 	 * as well as for lazy/batch writing of the MSRs.
189 	 */
190 	void (*pmu_enable)		(struct pmu *pmu); /* optional */
191 	void (*pmu_disable)		(struct pmu *pmu); /* optional */
192 
193 	/*
194 	 * Try and initialize the event for this PMU.
195 	 * Should return -ENOENT when the @event doesn't match this PMU.
196 	 */
197 	int (*event_init)		(struct perf_event *event);
198 
199 #define PERF_EF_START	0x01		/* start the counter when adding    */
200 #define PERF_EF_RELOAD	0x02		/* reload the counter when starting */
201 #define PERF_EF_UPDATE	0x04		/* update the counter when stopping */
202 
203 	/*
204 	 * Adds/Removes a counter to/from the PMU, can be done inside
205 	 * a transaction, see the ->*_txn() methods.
206 	 */
207 	int  (*add)			(struct perf_event *event, int flags);
208 	void (*del)			(struct perf_event *event, int flags);
209 
210 	/*
211 	 * Starts/Stops a counter present on the PMU. The PMI handler
212 	 * should stop the counter when perf_event_overflow() returns
213 	 * !0. ->start() will be used to continue.
214 	 */
215 	void (*start)			(struct perf_event *event, int flags);
216 	void (*stop)			(struct perf_event *event, int flags);
217 
218 	/*
219 	 * Updates the counter value of the event.
220 	 */
221 	void (*read)			(struct perf_event *event);
222 
223 	/*
224 	 * Group events scheduling is treated as a transaction, add
225 	 * group events as a whole and perform one schedulability test.
226 	 * If the test fails, roll back the whole group
227 	 *
228 	 * Start the transaction, after this ->add() doesn't need to
229 	 * do schedulability tests.
230 	 */
231 	void (*start_txn)		(struct pmu *pmu); /* optional */
232 	/*
233 	 * If ->start_txn() disabled the ->add() schedulability test
234 	 * then ->commit_txn() is required to perform one. On success
235 	 * the transaction is closed. On error the transaction is kept
236 	 * open until ->cancel_txn() is called.
237 	 */
238 	int  (*commit_txn)		(struct pmu *pmu); /* optional */
239 	/*
240 	 * Will cancel the transaction, assumes ->del() is called
241 	 * for each successful ->add() during the transaction.
242 	 */
243 	void (*cancel_txn)		(struct pmu *pmu); /* optional */
244 
245 	/*
246 	 * Will return the value for perf_event_mmap_page::index for this event,
247 	 * if no implementation is provided it will default to: event->hw.idx + 1.
248 	 */
249 	int (*event_idx)		(struct perf_event *event); /*optional */
250 
251 	/*
252 	 * flush branch stack on context-switches (needed in cpu-wide mode)
253 	 */
254 	void (*flush_branch_stack)	(void);
255 };
256 
257 /**
258  * enum perf_event_active_state - the states of a event
259  */
260 enum perf_event_active_state {
261 	PERF_EVENT_STATE_ERROR		= -2,
262 	PERF_EVENT_STATE_OFF		= -1,
263 	PERF_EVENT_STATE_INACTIVE	=  0,
264 	PERF_EVENT_STATE_ACTIVE		=  1,
265 };
266 
267 struct file;
268 struct perf_sample_data;
269 
270 typedef void (*perf_overflow_handler_t)(struct perf_event *,
271 					struct perf_sample_data *,
272 					struct pt_regs *regs);
273 
274 enum perf_group_flag {
275 	PERF_GROUP_SOFTWARE		= 0x1,
276 };
277 
278 #define SWEVENT_HLIST_BITS		8
279 #define SWEVENT_HLIST_SIZE		(1 << SWEVENT_HLIST_BITS)
280 
281 struct swevent_hlist {
282 	struct hlist_head		heads[SWEVENT_HLIST_SIZE];
283 	struct rcu_head			rcu_head;
284 };
285 
286 #define PERF_ATTACH_CONTEXT	0x01
287 #define PERF_ATTACH_GROUP	0x02
288 #define PERF_ATTACH_TASK	0x04
289 
290 struct perf_cgroup;
291 struct ring_buffer;
292 
293 /**
294  * struct perf_event - performance event kernel representation:
295  */
296 struct perf_event {
297 #ifdef CONFIG_PERF_EVENTS
298 	/*
299 	 * entry onto perf_event_context::event_list;
300 	 *   modifications require ctx->lock
301 	 *   RCU safe iterations.
302 	 */
303 	struct list_head		event_entry;
304 
305 	/*
306 	 * XXX: group_entry and sibling_list should be mutually exclusive;
307 	 * either you're a sibling on a group, or you're the group leader.
308 	 * Rework the code to always use the same list element.
309 	 *
310 	 * Locked for modification by both ctx->mutex and ctx->lock; holding
311 	 * either sufficies for read.
312 	 */
313 	struct list_head		group_entry;
314 	struct list_head		sibling_list;
315 
316 	/*
317 	 * We need storage to track the entries in perf_pmu_migrate_context; we
318 	 * cannot use the event_entry because of RCU and we want to keep the
319 	 * group in tact which avoids us using the other two entries.
320 	 */
321 	struct list_head		migrate_entry;
322 
323 	struct hlist_node		hlist_entry;
324 	struct list_head		active_entry;
325 	int				nr_siblings;
326 	int				group_flags;
327 	struct perf_event		*group_leader;
328 	struct pmu			*pmu;
329 
330 	enum perf_event_active_state	state;
331 	unsigned int			attach_state;
332 	local64_t			count;
333 	atomic64_t			child_count;
334 
335 	/*
336 	 * These are the total time in nanoseconds that the event
337 	 * has been enabled (i.e. eligible to run, and the task has
338 	 * been scheduled in, if this is a per-task event)
339 	 * and running (scheduled onto the CPU), respectively.
340 	 *
341 	 * They are computed from tstamp_enabled, tstamp_running and
342 	 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
343 	 */
344 	u64				total_time_enabled;
345 	u64				total_time_running;
346 
347 	/*
348 	 * These are timestamps used for computing total_time_enabled
349 	 * and total_time_running when the event is in INACTIVE or
350 	 * ACTIVE state, measured in nanoseconds from an arbitrary point
351 	 * in time.
352 	 * tstamp_enabled: the notional time when the event was enabled
353 	 * tstamp_running: the notional time when the event was scheduled on
354 	 * tstamp_stopped: in INACTIVE state, the notional time when the
355 	 *	event was scheduled off.
356 	 */
357 	u64				tstamp_enabled;
358 	u64				tstamp_running;
359 	u64				tstamp_stopped;
360 
361 	/*
362 	 * timestamp shadows the actual context timing but it can
363 	 * be safely used in NMI interrupt context. It reflects the
364 	 * context time as it was when the event was last scheduled in.
365 	 *
366 	 * ctx_time already accounts for ctx->timestamp. Therefore to
367 	 * compute ctx_time for a sample, simply add perf_clock().
368 	 */
369 	u64				shadow_ctx_time;
370 
371 	struct perf_event_attr		attr;
372 	u16				header_size;
373 	u16				id_header_size;
374 	u16				read_size;
375 	struct hw_perf_event		hw;
376 
377 	struct perf_event_context	*ctx;
378 	atomic_long_t			refcount;
379 
380 	/*
381 	 * These accumulate total time (in nanoseconds) that children
382 	 * events have been enabled and running, respectively.
383 	 */
384 	atomic64_t			child_total_time_enabled;
385 	atomic64_t			child_total_time_running;
386 
387 	/*
388 	 * Protect attach/detach and child_list:
389 	 */
390 	struct mutex			child_mutex;
391 	struct list_head		child_list;
392 	struct perf_event		*parent;
393 
394 	int				oncpu;
395 	int				cpu;
396 
397 	struct list_head		owner_entry;
398 	struct task_struct		*owner;
399 
400 	/* mmap bits */
401 	struct mutex			mmap_mutex;
402 	atomic_t			mmap_count;
403 
404 	struct ring_buffer		*rb;
405 	struct list_head		rb_entry;
406 
407 	/* poll related */
408 	wait_queue_head_t		waitq;
409 	struct fasync_struct		*fasync;
410 
411 	/* delayed work for NMIs and such */
412 	int				pending_wakeup;
413 	int				pending_kill;
414 	int				pending_disable;
415 	struct irq_work			pending;
416 
417 	atomic_t			event_limit;
418 
419 	void (*destroy)(struct perf_event *);
420 	struct rcu_head			rcu_head;
421 
422 	struct pid_namespace		*ns;
423 	u64				id;
424 
425 	perf_overflow_handler_t		overflow_handler;
426 	void				*overflow_handler_context;
427 
428 #ifdef CONFIG_EVENT_TRACING
429 	struct ftrace_event_call	*tp_event;
430 	struct event_filter		*filter;
431 #ifdef CONFIG_FUNCTION_TRACER
432 	struct ftrace_ops               ftrace_ops;
433 #endif
434 #endif
435 
436 #ifdef CONFIG_CGROUP_PERF
437 	struct perf_cgroup		*cgrp; /* cgroup event is attach to */
438 	int				cgrp_defer_enabled;
439 #endif
440 
441 #endif /* CONFIG_PERF_EVENTS */
442 };
443 
444 enum perf_event_context_type {
445 	task_context,
446 	cpu_context,
447 };
448 
449 /**
450  * struct perf_event_context - event context structure
451  *
452  * Used as a container for task events and CPU events as well:
453  */
454 struct perf_event_context {
455 	struct pmu			*pmu;
456 	enum perf_event_context_type	type;
457 	/*
458 	 * Protect the states of the events in the list,
459 	 * nr_active, and the list:
460 	 */
461 	raw_spinlock_t			lock;
462 	/*
463 	 * Protect the list of events.  Locking either mutex or lock
464 	 * is sufficient to ensure the list doesn't change; to change
465 	 * the list you need to lock both the mutex and the spinlock.
466 	 */
467 	struct mutex			mutex;
468 
469 	struct list_head		pinned_groups;
470 	struct list_head		flexible_groups;
471 	struct list_head		event_list;
472 	int				nr_events;
473 	int				nr_active;
474 	int				is_active;
475 	int				nr_stat;
476 	int				nr_freq;
477 	int				rotate_disable;
478 	atomic_t			refcount;
479 	struct task_struct		*task;
480 
481 	/*
482 	 * Context clock, runs when context enabled.
483 	 */
484 	u64				time;
485 	u64				timestamp;
486 
487 	/*
488 	 * These fields let us detect when two contexts have both
489 	 * been cloned (inherited) from a common ancestor.
490 	 */
491 	struct perf_event_context	*parent_ctx;
492 	u64				parent_gen;
493 	u64				generation;
494 	int				pin_count;
495 	int				nr_cgroups;	 /* cgroup evts */
496 	int				nr_branch_stack; /* branch_stack evt */
497 	struct rcu_head			rcu_head;
498 };
499 
500 /*
501  * Number of contexts where an event can trigger:
502  *	task, softirq, hardirq, nmi.
503  */
504 #define PERF_NR_CONTEXTS	4
505 
506 /**
507  * struct perf_event_cpu_context - per cpu event context structure
508  */
509 struct perf_cpu_context {
510 	struct perf_event_context	ctx;
511 	struct perf_event_context	*task_ctx;
512 	int				active_oncpu;
513 	int				exclusive;
514 	struct hrtimer			hrtimer;
515 	ktime_t				hrtimer_interval;
516 	struct list_head		rotation_list;
517 	struct pmu			*unique_pmu;
518 	struct perf_cgroup		*cgrp;
519 };
520 
521 struct perf_output_handle {
522 	struct perf_event		*event;
523 	struct ring_buffer		*rb;
524 	unsigned long			wakeup;
525 	unsigned long			size;
526 	void				*addr;
527 	int				page;
528 };
529 
530 #ifdef CONFIG_PERF_EVENTS
531 
532 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
533 extern void perf_pmu_unregister(struct pmu *pmu);
534 
535 extern int perf_num_counters(void);
536 extern const char *perf_pmu_name(void);
537 extern void __perf_event_task_sched_in(struct task_struct *prev,
538 				       struct task_struct *task);
539 extern void __perf_event_task_sched_out(struct task_struct *prev,
540 					struct task_struct *next);
541 extern int perf_event_init_task(struct task_struct *child);
542 extern void perf_event_exit_task(struct task_struct *child);
543 extern void perf_event_free_task(struct task_struct *task);
544 extern void perf_event_delayed_put(struct task_struct *task);
545 extern void perf_event_print_debug(void);
546 extern void perf_pmu_disable(struct pmu *pmu);
547 extern void perf_pmu_enable(struct pmu *pmu);
548 extern int perf_event_task_disable(void);
549 extern int perf_event_task_enable(void);
550 extern int perf_event_refresh(struct perf_event *event, int refresh);
551 extern void perf_event_update_userpage(struct perf_event *event);
552 extern int perf_event_release_kernel(struct perf_event *event);
553 extern struct perf_event *
554 perf_event_create_kernel_counter(struct perf_event_attr *attr,
555 				int cpu,
556 				struct task_struct *task,
557 				perf_overflow_handler_t callback,
558 				void *context);
559 extern void perf_pmu_migrate_context(struct pmu *pmu,
560 				int src_cpu, int dst_cpu);
561 extern u64 perf_event_read_value(struct perf_event *event,
562 				 u64 *enabled, u64 *running);
563 
564 
565 struct perf_sample_data {
566 	u64				type;
567 
568 	u64				ip;
569 	struct {
570 		u32	pid;
571 		u32	tid;
572 	}				tid_entry;
573 	u64				time;
574 	u64				addr;
575 	u64				id;
576 	u64				stream_id;
577 	struct {
578 		u32	cpu;
579 		u32	reserved;
580 	}				cpu_entry;
581 	u64				period;
582 	union  perf_mem_data_src	data_src;
583 	struct perf_callchain_entry	*callchain;
584 	struct perf_raw_record		*raw;
585 	struct perf_branch_stack	*br_stack;
586 	struct perf_regs_user		regs_user;
587 	u64				stack_user_size;
588 	u64				weight;
589 	/*
590 	 * Transaction flags for abort events:
591 	 */
592 	u64				txn;
593 };
594 
595 static inline void perf_sample_data_init(struct perf_sample_data *data,
596 					 u64 addr, u64 period)
597 {
598 	/* remaining struct members initialized in perf_prepare_sample() */
599 	data->addr = addr;
600 	data->raw  = NULL;
601 	data->br_stack = NULL;
602 	data->period = period;
603 	data->regs_user.abi = PERF_SAMPLE_REGS_ABI_NONE;
604 	data->regs_user.regs = NULL;
605 	data->stack_user_size = 0;
606 	data->weight = 0;
607 	data->data_src.val = 0;
608 	data->txn = 0;
609 }
610 
611 extern void perf_output_sample(struct perf_output_handle *handle,
612 			       struct perf_event_header *header,
613 			       struct perf_sample_data *data,
614 			       struct perf_event *event);
615 extern void perf_prepare_sample(struct perf_event_header *header,
616 				struct perf_sample_data *data,
617 				struct perf_event *event,
618 				struct pt_regs *regs);
619 
620 extern int perf_event_overflow(struct perf_event *event,
621 				 struct perf_sample_data *data,
622 				 struct pt_regs *regs);
623 
624 static inline bool is_sampling_event(struct perf_event *event)
625 {
626 	return event->attr.sample_period != 0;
627 }
628 
629 /*
630  * Return 1 for a software event, 0 for a hardware event
631  */
632 static inline int is_software_event(struct perf_event *event)
633 {
634 	return event->pmu->task_ctx_nr == perf_sw_context;
635 }
636 
637 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
638 
639 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
640 
641 #ifndef perf_arch_fetch_caller_regs
642 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
643 #endif
644 
645 /*
646  * Take a snapshot of the regs. Skip ip and frame pointer to
647  * the nth caller. We only need a few of the regs:
648  * - ip for PERF_SAMPLE_IP
649  * - cs for user_mode() tests
650  * - bp for callchains
651  * - eflags, for future purposes, just in case
652  */
653 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
654 {
655 	memset(regs, 0, sizeof(*regs));
656 
657 	perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
658 }
659 
660 static __always_inline void
661 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
662 {
663 	struct pt_regs hot_regs;
664 
665 	if (static_key_false(&perf_swevent_enabled[event_id])) {
666 		if (!regs) {
667 			perf_fetch_caller_regs(&hot_regs);
668 			regs = &hot_regs;
669 		}
670 		__perf_sw_event(event_id, nr, regs, addr);
671 	}
672 }
673 
674 extern struct static_key_deferred perf_sched_events;
675 
676 static inline void perf_event_task_sched_in(struct task_struct *prev,
677 					    struct task_struct *task)
678 {
679 	if (static_key_false(&perf_sched_events.key))
680 		__perf_event_task_sched_in(prev, task);
681 }
682 
683 static inline void perf_event_task_sched_out(struct task_struct *prev,
684 					     struct task_struct *next)
685 {
686 	perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, NULL, 0);
687 
688 	if (static_key_false(&perf_sched_events.key))
689 		__perf_event_task_sched_out(prev, next);
690 }
691 
692 extern void perf_event_mmap(struct vm_area_struct *vma);
693 extern struct perf_guest_info_callbacks *perf_guest_cbs;
694 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
695 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
696 
697 extern void perf_event_comm(struct task_struct *tsk);
698 extern void perf_event_fork(struct task_struct *tsk);
699 
700 /* Callchains */
701 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
702 
703 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
704 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
705 
706 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
707 {
708 	if (entry->nr < PERF_MAX_STACK_DEPTH)
709 		entry->ip[entry->nr++] = ip;
710 }
711 
712 extern int sysctl_perf_event_paranoid;
713 extern int sysctl_perf_event_mlock;
714 extern int sysctl_perf_event_sample_rate;
715 extern int sysctl_perf_cpu_time_max_percent;
716 
717 extern void perf_sample_event_took(u64 sample_len_ns);
718 
719 extern int perf_proc_update_handler(struct ctl_table *table, int write,
720 		void __user *buffer, size_t *lenp,
721 		loff_t *ppos);
722 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
723 		void __user *buffer, size_t *lenp,
724 		loff_t *ppos);
725 
726 
727 static inline bool perf_paranoid_tracepoint_raw(void)
728 {
729 	return sysctl_perf_event_paranoid > -1;
730 }
731 
732 static inline bool perf_paranoid_cpu(void)
733 {
734 	return sysctl_perf_event_paranoid > 0;
735 }
736 
737 static inline bool perf_paranoid_kernel(void)
738 {
739 	return sysctl_perf_event_paranoid > 1;
740 }
741 
742 extern void perf_event_init(void);
743 extern void perf_tp_event(u64 addr, u64 count, void *record,
744 			  int entry_size, struct pt_regs *regs,
745 			  struct hlist_head *head, int rctx,
746 			  struct task_struct *task);
747 extern void perf_bp_event(struct perf_event *event, void *data);
748 
749 #ifndef perf_misc_flags
750 # define perf_misc_flags(regs) \
751 		(user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
752 # define perf_instruction_pointer(regs)	instruction_pointer(regs)
753 #endif
754 
755 static inline bool has_branch_stack(struct perf_event *event)
756 {
757 	return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
758 }
759 
760 extern int perf_output_begin(struct perf_output_handle *handle,
761 			     struct perf_event *event, unsigned int size);
762 extern void perf_output_end(struct perf_output_handle *handle);
763 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
764 			     const void *buf, unsigned int len);
765 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
766 				     unsigned int len);
767 extern int perf_swevent_get_recursion_context(void);
768 extern void perf_swevent_put_recursion_context(int rctx);
769 extern u64 perf_swevent_set_period(struct perf_event *event);
770 extern void perf_event_enable(struct perf_event *event);
771 extern void perf_event_disable(struct perf_event *event);
772 extern int __perf_event_disable(void *info);
773 extern void perf_event_task_tick(void);
774 #else
775 static inline void
776 perf_event_task_sched_in(struct task_struct *prev,
777 			 struct task_struct *task)			{ }
778 static inline void
779 perf_event_task_sched_out(struct task_struct *prev,
780 			  struct task_struct *next)			{ }
781 static inline int perf_event_init_task(struct task_struct *child)	{ return 0; }
782 static inline void perf_event_exit_task(struct task_struct *child)	{ }
783 static inline void perf_event_free_task(struct task_struct *task)	{ }
784 static inline void perf_event_delayed_put(struct task_struct *task)	{ }
785 static inline void perf_event_print_debug(void)				{ }
786 static inline int perf_event_task_disable(void)				{ return -EINVAL; }
787 static inline int perf_event_task_enable(void)				{ return -EINVAL; }
788 static inline int perf_event_refresh(struct perf_event *event, int refresh)
789 {
790 	return -EINVAL;
791 }
792 
793 static inline void
794 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)	{ }
795 static inline void
796 perf_bp_event(struct perf_event *event, void *data)			{ }
797 
798 static inline int perf_register_guest_info_callbacks
799 (struct perf_guest_info_callbacks *callbacks)				{ return 0; }
800 static inline int perf_unregister_guest_info_callbacks
801 (struct perf_guest_info_callbacks *callbacks)				{ return 0; }
802 
803 static inline void perf_event_mmap(struct vm_area_struct *vma)		{ }
804 static inline void perf_event_comm(struct task_struct *tsk)		{ }
805 static inline void perf_event_fork(struct task_struct *tsk)		{ }
806 static inline void perf_event_init(void)				{ }
807 static inline int  perf_swevent_get_recursion_context(void)		{ return -1; }
808 static inline void perf_swevent_put_recursion_context(int rctx)		{ }
809 static inline u64 perf_swevent_set_period(struct perf_event *event)	{ return 0; }
810 static inline void perf_event_enable(struct perf_event *event)		{ }
811 static inline void perf_event_disable(struct perf_event *event)		{ }
812 static inline int __perf_event_disable(void *info)			{ return -1; }
813 static inline void perf_event_task_tick(void)				{ }
814 #endif
815 
816 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_NO_HZ_FULL)
817 extern bool perf_event_can_stop_tick(void);
818 #else
819 static inline bool perf_event_can_stop_tick(void)			{ return true; }
820 #endif
821 
822 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
823 extern void perf_restore_debug_store(void);
824 #else
825 static inline void perf_restore_debug_store(void)			{ }
826 #endif
827 
828 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
829 
830 /*
831  * This has to have a higher priority than migration_notifier in sched/core.c.
832  */
833 #define perf_cpu_notifier(fn)						\
834 do {									\
835 	static struct notifier_block fn##_nb =				\
836 		{ .notifier_call = fn, .priority = CPU_PRI_PERF };	\
837 	unsigned long cpu = smp_processor_id();				\
838 	unsigned long flags;						\
839 									\
840 	cpu_notifier_register_begin();					\
841 	fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,			\
842 		(void *)(unsigned long)cpu);				\
843 	local_irq_save(flags);						\
844 	fn(&fn##_nb, (unsigned long)CPU_STARTING,			\
845 		(void *)(unsigned long)cpu);				\
846 	local_irq_restore(flags);					\
847 	fn(&fn##_nb, (unsigned long)CPU_ONLINE,				\
848 		(void *)(unsigned long)cpu);				\
849 	__register_cpu_notifier(&fn##_nb);				\
850 	cpu_notifier_register_done();					\
851 } while (0)
852 
853 /*
854  * Bare-bones version of perf_cpu_notifier(), which doesn't invoke the
855  * callback for already online CPUs.
856  */
857 #define __perf_cpu_notifier(fn)						\
858 do {									\
859 	static struct notifier_block fn##_nb =				\
860 		{ .notifier_call = fn, .priority = CPU_PRI_PERF };	\
861 									\
862 	__register_cpu_notifier(&fn##_nb);				\
863 } while (0)
864 
865 struct perf_pmu_events_attr {
866 	struct device_attribute attr;
867 	u64 id;
868 	const char *event_str;
869 };
870 
871 #define PMU_EVENT_ATTR(_name, _var, _id, _show)				\
872 static struct perf_pmu_events_attr _var = {				\
873 	.attr = __ATTR(_name, 0444, _show, NULL),			\
874 	.id   =  _id,							\
875 };
876 
877 #define PMU_FORMAT_ATTR(_name, _format)					\
878 static ssize_t								\
879 _name##_show(struct device *dev,					\
880 			       struct device_attribute *attr,		\
881 			       char *page)				\
882 {									\
883 	BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);			\
884 	return sprintf(page, _format "\n");				\
885 }									\
886 									\
887 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
888 
889 #endif /* _LINUX_PERF_EVENT_H */
890