xref: /linux-6.15/include/linux/perf_event.h (revision f60d24d2)
1 /*
2  * Performance events:
3  *
4  *    Copyright (C) 2008-2009, Thomas Gleixner <[email protected]>
5  *    Copyright (C) 2008-2009, Red Hat, Inc., Ingo Molnar
6  *    Copyright (C) 2008-2009, 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 <linux/types.h>
18 #include <linux/ioctl.h>
19 #include <asm/byteorder.h>
20 
21 #ifdef CONFIG_HAVE_HW_BREAKPOINT
22 #include <asm/hw_breakpoint.h>
23 #endif
24 
25 /*
26  * User-space ABI bits:
27  */
28 
29 /*
30  * attr.type
31  */
32 enum perf_type_id {
33 	PERF_TYPE_HARDWARE			= 0,
34 	PERF_TYPE_SOFTWARE			= 1,
35 	PERF_TYPE_TRACEPOINT			= 2,
36 	PERF_TYPE_HW_CACHE			= 3,
37 	PERF_TYPE_RAW				= 4,
38 	PERF_TYPE_BREAKPOINT			= 5,
39 
40 	PERF_TYPE_MAX,				/* non-ABI */
41 };
42 
43 /*
44  * Generalized performance event event_id types, used by the
45  * attr.event_id parameter of the sys_perf_event_open()
46  * syscall:
47  */
48 enum perf_hw_id {
49 	/*
50 	 * Common hardware events, generalized by the kernel:
51 	 */
52 	PERF_COUNT_HW_CPU_CYCLES		= 0,
53 	PERF_COUNT_HW_INSTRUCTIONS		= 1,
54 	PERF_COUNT_HW_CACHE_REFERENCES		= 2,
55 	PERF_COUNT_HW_CACHE_MISSES		= 3,
56 	PERF_COUNT_HW_BRANCH_INSTRUCTIONS	= 4,
57 	PERF_COUNT_HW_BRANCH_MISSES		= 5,
58 	PERF_COUNT_HW_BUS_CYCLES		= 6,
59 
60 	PERF_COUNT_HW_MAX,			/* non-ABI */
61 };
62 
63 /*
64  * Generalized hardware cache events:
65  *
66  *       { L1-D, L1-I, LLC, ITLB, DTLB, BPU } x
67  *       { read, write, prefetch } x
68  *       { accesses, misses }
69  */
70 enum perf_hw_cache_id {
71 	PERF_COUNT_HW_CACHE_L1D			= 0,
72 	PERF_COUNT_HW_CACHE_L1I			= 1,
73 	PERF_COUNT_HW_CACHE_LL			= 2,
74 	PERF_COUNT_HW_CACHE_DTLB		= 3,
75 	PERF_COUNT_HW_CACHE_ITLB		= 4,
76 	PERF_COUNT_HW_CACHE_BPU			= 5,
77 
78 	PERF_COUNT_HW_CACHE_MAX,		/* non-ABI */
79 };
80 
81 enum perf_hw_cache_op_id {
82 	PERF_COUNT_HW_CACHE_OP_READ		= 0,
83 	PERF_COUNT_HW_CACHE_OP_WRITE		= 1,
84 	PERF_COUNT_HW_CACHE_OP_PREFETCH		= 2,
85 
86 	PERF_COUNT_HW_CACHE_OP_MAX,		/* non-ABI */
87 };
88 
89 enum perf_hw_cache_op_result_id {
90 	PERF_COUNT_HW_CACHE_RESULT_ACCESS	= 0,
91 	PERF_COUNT_HW_CACHE_RESULT_MISS		= 1,
92 
93 	PERF_COUNT_HW_CACHE_RESULT_MAX,		/* non-ABI */
94 };
95 
96 /*
97  * Special "software" events provided by the kernel, even if the hardware
98  * does not support performance events. These events measure various
99  * physical and sw events of the kernel (and allow the profiling of them as
100  * well):
101  */
102 enum perf_sw_ids {
103 	PERF_COUNT_SW_CPU_CLOCK			= 0,
104 	PERF_COUNT_SW_TASK_CLOCK		= 1,
105 	PERF_COUNT_SW_PAGE_FAULTS		= 2,
106 	PERF_COUNT_SW_CONTEXT_SWITCHES		= 3,
107 	PERF_COUNT_SW_CPU_MIGRATIONS		= 4,
108 	PERF_COUNT_SW_PAGE_FAULTS_MIN		= 5,
109 	PERF_COUNT_SW_PAGE_FAULTS_MAJ		= 6,
110 
111 	PERF_COUNT_SW_MAX,			/* non-ABI */
112 };
113 
114 /*
115  * Bits that can be set in attr.sample_type to request information
116  * in the overflow packets.
117  */
118 enum perf_event_sample_format {
119 	PERF_SAMPLE_IP				= 1U << 0,
120 	PERF_SAMPLE_TID				= 1U << 1,
121 	PERF_SAMPLE_TIME			= 1U << 2,
122 	PERF_SAMPLE_ADDR			= 1U << 3,
123 	PERF_SAMPLE_READ			= 1U << 4,
124 	PERF_SAMPLE_CALLCHAIN			= 1U << 5,
125 	PERF_SAMPLE_ID				= 1U << 6,
126 	PERF_SAMPLE_CPU				= 1U << 7,
127 	PERF_SAMPLE_PERIOD			= 1U << 8,
128 	PERF_SAMPLE_STREAM_ID			= 1U << 9,
129 	PERF_SAMPLE_RAW				= 1U << 10,
130 
131 	PERF_SAMPLE_MAX = 1U << 11,		/* non-ABI */
132 };
133 
134 /*
135  * The format of the data returned by read() on a perf event fd,
136  * as specified by attr.read_format:
137  *
138  * struct read_format {
139  *	{ u64		value;
140  *	  { u64		time_enabled; } && PERF_FORMAT_ENABLED
141  *	  { u64		time_running; } && PERF_FORMAT_RUNNING
142  *	  { u64		id;           } && PERF_FORMAT_ID
143  *	} && !PERF_FORMAT_GROUP
144  *
145  *	{ u64		nr;
146  *	  { u64		time_enabled; } && PERF_FORMAT_ENABLED
147  *	  { u64		time_running; } && PERF_FORMAT_RUNNING
148  *	  { u64		value;
149  *	    { u64	id;           } && PERF_FORMAT_ID
150  *	  }		cntr[nr];
151  *	} && PERF_FORMAT_GROUP
152  * };
153  */
154 enum perf_event_read_format {
155 	PERF_FORMAT_TOTAL_TIME_ENABLED		= 1U << 0,
156 	PERF_FORMAT_TOTAL_TIME_RUNNING		= 1U << 1,
157 	PERF_FORMAT_ID				= 1U << 2,
158 	PERF_FORMAT_GROUP			= 1U << 3,
159 
160 	PERF_FORMAT_MAX = 1U << 4,		/* non-ABI */
161 };
162 
163 #define PERF_ATTR_SIZE_VER0	64	/* sizeof first published struct */
164 
165 /*
166  * Hardware event_id to monitor via a performance monitoring event:
167  */
168 struct perf_event_attr {
169 
170 	/*
171 	 * Major type: hardware/software/tracepoint/etc.
172 	 */
173 	__u32			type;
174 
175 	/*
176 	 * Size of the attr structure, for fwd/bwd compat.
177 	 */
178 	__u32			size;
179 
180 	/*
181 	 * Type specific configuration information.
182 	 */
183 	__u64			config;
184 
185 	union {
186 		__u64		sample_period;
187 		__u64		sample_freq;
188 	};
189 
190 	__u64			sample_type;
191 	__u64			read_format;
192 
193 	__u64			disabled       :  1, /* off by default        */
194 				inherit	       :  1, /* children inherit it   */
195 				pinned	       :  1, /* must always be on PMU */
196 				exclusive      :  1, /* only group on PMU     */
197 				exclude_user   :  1, /* don't count user      */
198 				exclude_kernel :  1, /* ditto kernel          */
199 				exclude_hv     :  1, /* ditto hypervisor      */
200 				exclude_idle   :  1, /* don't count when idle */
201 				mmap           :  1, /* include mmap data     */
202 				comm	       :  1, /* include comm data     */
203 				freq           :  1, /* use freq, not period  */
204 				inherit_stat   :  1, /* per task counts       */
205 				enable_on_exec :  1, /* next exec enables     */
206 				task           :  1, /* trace fork/exit       */
207 				watermark      :  1, /* wakeup_watermark      */
208 
209 				__reserved_1   : 49;
210 
211 	union {
212 		__u32		wakeup_events;	  /* wakeup every n events */
213 		__u32		wakeup_watermark; /* bytes before wakeup   */
214 	};
215 
216 	union {
217 		struct { /* Hardware breakpoint info */
218 			__u64		bp_addr;
219 			__u32		bp_type;
220 			__u32		bp_len;
221 		};
222 	};
223 
224 	__u32			__reserved_2;
225 
226 	__u64			__reserved_3;
227 };
228 
229 /*
230  * Ioctls that can be done on a perf event fd:
231  */
232 #define PERF_EVENT_IOC_ENABLE		_IO ('$', 0)
233 #define PERF_EVENT_IOC_DISABLE		_IO ('$', 1)
234 #define PERF_EVENT_IOC_REFRESH		_IO ('$', 2)
235 #define PERF_EVENT_IOC_RESET		_IO ('$', 3)
236 #define PERF_EVENT_IOC_PERIOD		_IOW('$', 4, u64)
237 #define PERF_EVENT_IOC_SET_OUTPUT	_IO ('$', 5)
238 #define PERF_EVENT_IOC_SET_FILTER	_IOW('$', 6, char *)
239 
240 enum perf_event_ioc_flags {
241 	PERF_IOC_FLAG_GROUP		= 1U << 0,
242 };
243 
244 /*
245  * Structure of the page that can be mapped via mmap
246  */
247 struct perf_event_mmap_page {
248 	__u32	version;		/* version number of this structure */
249 	__u32	compat_version;		/* lowest version this is compat with */
250 
251 	/*
252 	 * Bits needed to read the hw events in user-space.
253 	 *
254 	 *   u32 seq;
255 	 *   s64 count;
256 	 *
257 	 *   do {
258 	 *     seq = pc->lock;
259 	 *
260 	 *     barrier()
261 	 *     if (pc->index) {
262 	 *       count = pmc_read(pc->index - 1);
263 	 *       count += pc->offset;
264 	 *     } else
265 	 *       goto regular_read;
266 	 *
267 	 *     barrier();
268 	 *   } while (pc->lock != seq);
269 	 *
270 	 * NOTE: for obvious reason this only works on self-monitoring
271 	 *       processes.
272 	 */
273 	__u32	lock;			/* seqlock for synchronization */
274 	__u32	index;			/* hardware event identifier */
275 	__s64	offset;			/* add to hardware event value */
276 	__u64	time_enabled;		/* time event active */
277 	__u64	time_running;		/* time event on cpu */
278 
279 		/*
280 		 * Hole for extension of the self monitor capabilities
281 		 */
282 
283 	__u64	__reserved[123];	/* align to 1k */
284 
285 	/*
286 	 * Control data for the mmap() data buffer.
287 	 *
288 	 * User-space reading the @data_head value should issue an rmb(), on
289 	 * SMP capable platforms, after reading this value -- see
290 	 * perf_event_wakeup().
291 	 *
292 	 * When the mapping is PROT_WRITE the @data_tail value should be
293 	 * written by userspace to reflect the last read data. In this case
294 	 * the kernel will not over-write unread data.
295 	 */
296 	__u64   data_head;		/* head in the data section */
297 	__u64	data_tail;		/* user-space written tail */
298 };
299 
300 #define PERF_RECORD_MISC_CPUMODE_MASK		(3 << 0)
301 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN		(0 << 0)
302 #define PERF_RECORD_MISC_KERNEL			(1 << 0)
303 #define PERF_RECORD_MISC_USER			(2 << 0)
304 #define PERF_RECORD_MISC_HYPERVISOR		(3 << 0)
305 
306 struct perf_event_header {
307 	__u32	type;
308 	__u16	misc;
309 	__u16	size;
310 };
311 
312 enum perf_event_type {
313 
314 	/*
315 	 * The MMAP events record the PROT_EXEC mappings so that we can
316 	 * correlate userspace IPs to code. They have the following structure:
317 	 *
318 	 * struct {
319 	 *	struct perf_event_header	header;
320 	 *
321 	 *	u32				pid, tid;
322 	 *	u64				addr;
323 	 *	u64				len;
324 	 *	u64				pgoff;
325 	 *	char				filename[];
326 	 * };
327 	 */
328 	PERF_RECORD_MMAP			= 1,
329 
330 	/*
331 	 * struct {
332 	 *	struct perf_event_header	header;
333 	 *	u64				id;
334 	 *	u64				lost;
335 	 * };
336 	 */
337 	PERF_RECORD_LOST			= 2,
338 
339 	/*
340 	 * struct {
341 	 *	struct perf_event_header	header;
342 	 *
343 	 *	u32				pid, tid;
344 	 *	char				comm[];
345 	 * };
346 	 */
347 	PERF_RECORD_COMM			= 3,
348 
349 	/*
350 	 * struct {
351 	 *	struct perf_event_header	header;
352 	 *	u32				pid, ppid;
353 	 *	u32				tid, ptid;
354 	 *	u64				time;
355 	 * };
356 	 */
357 	PERF_RECORD_EXIT			= 4,
358 
359 	/*
360 	 * struct {
361 	 *	struct perf_event_header	header;
362 	 *	u64				time;
363 	 *	u64				id;
364 	 *	u64				stream_id;
365 	 * };
366 	 */
367 	PERF_RECORD_THROTTLE		= 5,
368 	PERF_RECORD_UNTHROTTLE		= 6,
369 
370 	/*
371 	 * struct {
372 	 *	struct perf_event_header	header;
373 	 *	u32				pid, ppid;
374 	 *	u32				tid, ptid;
375 	 *	u64				time;
376 	 * };
377 	 */
378 	PERF_RECORD_FORK			= 7,
379 
380 	/*
381 	 * struct {
382 	 * 	struct perf_event_header	header;
383 	 * 	u32				pid, tid;
384 	 *
385 	 * 	struct read_format		values;
386 	 * };
387 	 */
388 	PERF_RECORD_READ			= 8,
389 
390 	/*
391 	 * struct {
392 	 *	struct perf_event_header	header;
393 	 *
394 	 *	{ u64			ip;	  } && PERF_SAMPLE_IP
395 	 *	{ u32			pid, tid; } && PERF_SAMPLE_TID
396 	 *	{ u64			time;     } && PERF_SAMPLE_TIME
397 	 *	{ u64			addr;     } && PERF_SAMPLE_ADDR
398 	 *	{ u64			id;	  } && PERF_SAMPLE_ID
399 	 *	{ u64			stream_id;} && PERF_SAMPLE_STREAM_ID
400 	 *	{ u32			cpu, res; } && PERF_SAMPLE_CPU
401 	 *	{ u64			period;   } && PERF_SAMPLE_PERIOD
402 	 *
403 	 *	{ struct read_format	values;	  } && PERF_SAMPLE_READ
404 	 *
405 	 *	{ u64			nr,
406 	 *	  u64			ips[nr];  } && PERF_SAMPLE_CALLCHAIN
407 	 *
408 	 *	#
409 	 *	# The RAW record below is opaque data wrt the ABI
410 	 *	#
411 	 *	# That is, the ABI doesn't make any promises wrt to
412 	 *	# the stability of its content, it may vary depending
413 	 *	# on event, hardware, kernel version and phase of
414 	 *	# the moon.
415 	 *	#
416 	 *	# In other words, PERF_SAMPLE_RAW contents are not an ABI.
417 	 *	#
418 	 *
419 	 *	{ u32			size;
420 	 *	  char                  data[size];}&& PERF_SAMPLE_RAW
421 	 * };
422 	 */
423 	PERF_RECORD_SAMPLE		= 9,
424 
425 	PERF_RECORD_MAX,			/* non-ABI */
426 };
427 
428 enum perf_callchain_context {
429 	PERF_CONTEXT_HV			= (__u64)-32,
430 	PERF_CONTEXT_KERNEL		= (__u64)-128,
431 	PERF_CONTEXT_USER		= (__u64)-512,
432 
433 	PERF_CONTEXT_GUEST		= (__u64)-2048,
434 	PERF_CONTEXT_GUEST_KERNEL	= (__u64)-2176,
435 	PERF_CONTEXT_GUEST_USER		= (__u64)-2560,
436 
437 	PERF_CONTEXT_MAX		= (__u64)-4095,
438 };
439 
440 #define PERF_FLAG_FD_NO_GROUP	(1U << 0)
441 #define PERF_FLAG_FD_OUTPUT	(1U << 1)
442 
443 #ifdef __KERNEL__
444 /*
445  * Kernel-internal data types and definitions:
446  */
447 
448 #ifdef CONFIG_PERF_EVENTS
449 # include <asm/perf_event.h>
450 #endif
451 
452 #include <linux/list.h>
453 #include <linux/mutex.h>
454 #include <linux/rculist.h>
455 #include <linux/rcupdate.h>
456 #include <linux/spinlock.h>
457 #include <linux/hrtimer.h>
458 #include <linux/fs.h>
459 #include <linux/pid_namespace.h>
460 #include <linux/workqueue.h>
461 #include <asm/atomic.h>
462 
463 #define PERF_MAX_STACK_DEPTH		255
464 
465 struct perf_callchain_entry {
466 	__u64				nr;
467 	__u64				ip[PERF_MAX_STACK_DEPTH];
468 };
469 
470 struct perf_raw_record {
471 	u32				size;
472 	void				*data;
473 };
474 
475 struct task_struct;
476 
477 /**
478  * struct hw_perf_event - performance event hardware details:
479  */
480 struct hw_perf_event {
481 #ifdef CONFIG_PERF_EVENTS
482 	union {
483 		struct { /* hardware */
484 			u64		config;
485 			unsigned long	config_base;
486 			unsigned long	event_base;
487 			int		idx;
488 		};
489 		union { /* software */
490 			atomic64_t	count;
491 			struct hrtimer	hrtimer;
492 		};
493 #ifdef CONFIG_HAVE_HW_BREAKPOINT
494 		union { /* breakpoint */
495 			struct arch_hw_breakpoint	info;
496 		};
497 #endif
498 	};
499 	atomic64_t			prev_count;
500 	u64				sample_period;
501 	u64				last_period;
502 	atomic64_t			period_left;
503 	u64				interrupts;
504 
505 	u64				freq_count;
506 	u64				freq_interrupts;
507 	u64				freq_stamp;
508 #endif
509 };
510 
511 struct perf_event;
512 
513 /**
514  * struct pmu - generic performance monitoring unit
515  */
516 struct pmu {
517 	int (*enable)			(struct perf_event *event);
518 	void (*disable)			(struct perf_event *event);
519 	void (*read)			(struct perf_event *event);
520 	void (*unthrottle)		(struct perf_event *event);
521 };
522 
523 /**
524  * enum perf_event_active_state - the states of a event
525  */
526 enum perf_event_active_state {
527 	PERF_EVENT_STATE_ERROR		= -2,
528 	PERF_EVENT_STATE_OFF		= -1,
529 	PERF_EVENT_STATE_INACTIVE	=  0,
530 	PERF_EVENT_STATE_ACTIVE		=  1,
531 };
532 
533 struct file;
534 
535 struct perf_mmap_data {
536 	struct rcu_head			rcu_head;
537 #ifdef CONFIG_PERF_USE_VMALLOC
538 	struct work_struct		work;
539 #endif
540 	int				data_order;
541 	int				nr_pages;	/* nr of data pages  */
542 	int				writable;	/* are we writable   */
543 	int				nr_locked;	/* nr pages mlocked  */
544 
545 	atomic_t			poll;		/* POLL_ for wakeups */
546 	atomic_t			events;		/* event_id limit       */
547 
548 	atomic_long_t			head;		/* write position    */
549 	atomic_long_t			done_head;	/* completed head    */
550 
551 	atomic_t			lock;		/* concurrent writes */
552 	atomic_t			wakeup;		/* needs a wakeup    */
553 	atomic_t			lost;		/* nr records lost   */
554 
555 	long				watermark;	/* wakeup watermark  */
556 
557 	struct perf_event_mmap_page	*user_page;
558 	void				*data_pages[0];
559 };
560 
561 struct perf_pending_entry {
562 	struct perf_pending_entry *next;
563 	void (*func)(struct perf_pending_entry *);
564 };
565 
566 typedef void (*perf_callback_t)(struct perf_event *, void *);
567 
568 /**
569  * struct perf_event - performance event kernel representation:
570  */
571 struct perf_event {
572 #ifdef CONFIG_PERF_EVENTS
573 	struct list_head		group_entry;
574 	struct list_head		event_entry;
575 	struct list_head		sibling_list;
576 	int				nr_siblings;
577 	struct perf_event		*group_leader;
578 	struct perf_event		*output;
579 	const struct pmu		*pmu;
580 
581 	enum perf_event_active_state	state;
582 	atomic64_t			count;
583 
584 	/*
585 	 * These are the total time in nanoseconds that the event
586 	 * has been enabled (i.e. eligible to run, and the task has
587 	 * been scheduled in, if this is a per-task event)
588 	 * and running (scheduled onto the CPU), respectively.
589 	 *
590 	 * They are computed from tstamp_enabled, tstamp_running and
591 	 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
592 	 */
593 	u64				total_time_enabled;
594 	u64				total_time_running;
595 
596 	/*
597 	 * These are timestamps used for computing total_time_enabled
598 	 * and total_time_running when the event is in INACTIVE or
599 	 * ACTIVE state, measured in nanoseconds from an arbitrary point
600 	 * in time.
601 	 * tstamp_enabled: the notional time when the event was enabled
602 	 * tstamp_running: the notional time when the event was scheduled on
603 	 * tstamp_stopped: in INACTIVE state, the notional time when the
604 	 *	event was scheduled off.
605 	 */
606 	u64				tstamp_enabled;
607 	u64				tstamp_running;
608 	u64				tstamp_stopped;
609 
610 	struct perf_event_attr		attr;
611 	struct hw_perf_event		hw;
612 
613 	struct perf_event_context	*ctx;
614 	struct file			*filp;
615 
616 	/*
617 	 * These accumulate total time (in nanoseconds) that children
618 	 * events have been enabled and running, respectively.
619 	 */
620 	atomic64_t			child_total_time_enabled;
621 	atomic64_t			child_total_time_running;
622 
623 	/*
624 	 * Protect attach/detach and child_list:
625 	 */
626 	struct mutex			child_mutex;
627 	struct list_head		child_list;
628 	struct perf_event		*parent;
629 
630 	int				oncpu;
631 	int				cpu;
632 
633 	struct list_head		owner_entry;
634 	struct task_struct		*owner;
635 
636 	/* mmap bits */
637 	struct mutex			mmap_mutex;
638 	atomic_t			mmap_count;
639 	struct perf_mmap_data		*data;
640 
641 	/* poll related */
642 	wait_queue_head_t		waitq;
643 	struct fasync_struct		*fasync;
644 
645 	/* delayed work for NMIs and such */
646 	int				pending_wakeup;
647 	int				pending_kill;
648 	int				pending_disable;
649 	struct perf_pending_entry	pending;
650 
651 	atomic_t			event_limit;
652 
653 	void (*destroy)(struct perf_event *);
654 	struct rcu_head			rcu_head;
655 
656 	struct pid_namespace		*ns;
657 	u64				id;
658 
659 #ifdef CONFIG_EVENT_PROFILE
660 	struct event_filter		*filter;
661 #endif
662 
663 	perf_callback_t			callback;
664 
665 	perf_callback_t			event_callback;
666 
667 #endif /* CONFIG_PERF_EVENTS */
668 };
669 
670 /**
671  * struct perf_event_context - event context structure
672  *
673  * Used as a container for task events and CPU events as well:
674  */
675 struct perf_event_context {
676 	/*
677 	 * Protect the states of the events in the list,
678 	 * nr_active, and the list:
679 	 */
680 	spinlock_t			lock;
681 	/*
682 	 * Protect the list of events.  Locking either mutex or lock
683 	 * is sufficient to ensure the list doesn't change; to change
684 	 * the list you need to lock both the mutex and the spinlock.
685 	 */
686 	struct mutex			mutex;
687 
688 	struct list_head		group_list;
689 	struct list_head		event_list;
690 	int				nr_events;
691 	int				nr_active;
692 	int				is_active;
693 	int				nr_stat;
694 	atomic_t			refcount;
695 	struct task_struct		*task;
696 
697 	/*
698 	 * Context clock, runs when context enabled.
699 	 */
700 	u64				time;
701 	u64				timestamp;
702 
703 	/*
704 	 * These fields let us detect when two contexts have both
705 	 * been cloned (inherited) from a common ancestor.
706 	 */
707 	struct perf_event_context	*parent_ctx;
708 	u64				parent_gen;
709 	u64				generation;
710 	int				pin_count;
711 	struct rcu_head			rcu_head;
712 };
713 
714 /**
715  * struct perf_event_cpu_context - per cpu event context structure
716  */
717 struct perf_cpu_context {
718 	struct perf_event_context	ctx;
719 	struct perf_event_context	*task_ctx;
720 	int				active_oncpu;
721 	int				max_pertask;
722 	int				exclusive;
723 
724 	/*
725 	 * Recursion avoidance:
726 	 *
727 	 * task, softirq, irq, nmi context
728 	 */
729 	int				recursion[4];
730 };
731 
732 struct perf_output_handle {
733 	struct perf_event		*event;
734 	struct perf_mmap_data		*data;
735 	unsigned long			head;
736 	unsigned long			offset;
737 	int				nmi;
738 	int				sample;
739 	int				locked;
740 	unsigned long			flags;
741 };
742 
743 #ifdef CONFIG_PERF_EVENTS
744 
745 /*
746  * Set by architecture code:
747  */
748 extern int perf_max_events;
749 
750 extern const struct pmu *hw_perf_event_init(struct perf_event *event);
751 
752 extern void perf_event_task_sched_in(struct task_struct *task, int cpu);
753 extern void perf_event_task_sched_out(struct task_struct *task,
754 					struct task_struct *next, int cpu);
755 extern void perf_event_task_tick(struct task_struct *task, int cpu);
756 extern int perf_event_init_task(struct task_struct *child);
757 extern void perf_event_exit_task(struct task_struct *child);
758 extern void perf_event_free_task(struct task_struct *task);
759 extern void set_perf_event_pending(void);
760 extern void perf_event_do_pending(void);
761 extern void perf_event_print_debug(void);
762 extern void __perf_disable(void);
763 extern bool __perf_enable(void);
764 extern void perf_disable(void);
765 extern void perf_enable(void);
766 extern int perf_event_task_disable(void);
767 extern int perf_event_task_enable(void);
768 extern int hw_perf_group_sched_in(struct perf_event *group_leader,
769 	       struct perf_cpu_context *cpuctx,
770 	       struct perf_event_context *ctx, int cpu);
771 extern void perf_event_update_userpage(struct perf_event *event);
772 extern int perf_event_release_kernel(struct perf_event *event);
773 extern struct perf_event *
774 perf_event_create_kernel_counter(struct perf_event_attr *attr,
775 				int cpu,
776 				pid_t pid,
777 				perf_callback_t callback);
778 extern u64 perf_event_read_value(struct perf_event *event);
779 
780 struct perf_sample_data {
781 	u64				type;
782 
783 	u64				ip;
784 	struct {
785 		u32	pid;
786 		u32	tid;
787 	}				tid_entry;
788 	u64				time;
789 	u64				addr;
790 	u64				id;
791 	u64				stream_id;
792 	struct {
793 		u32	cpu;
794 		u32	reserved;
795 	}				cpu_entry;
796 	u64				period;
797 	struct perf_callchain_entry	*callchain;
798 	struct perf_raw_record		*raw;
799 };
800 
801 extern void perf_output_sample(struct perf_output_handle *handle,
802 			       struct perf_event_header *header,
803 			       struct perf_sample_data *data,
804 			       struct perf_event *event);
805 extern void perf_prepare_sample(struct perf_event_header *header,
806 				struct perf_sample_data *data,
807 				struct perf_event *event,
808 				struct pt_regs *regs);
809 
810 extern int perf_event_overflow(struct perf_event *event, int nmi,
811 				 struct perf_sample_data *data,
812 				 struct pt_regs *regs);
813 
814 /*
815  * Return 1 for a software event, 0 for a hardware event
816  */
817 static inline int is_software_event(struct perf_event *event)
818 {
819 	return (event->attr.type != PERF_TYPE_RAW) &&
820 		(event->attr.type != PERF_TYPE_HARDWARE) &&
821 		(event->attr.type != PERF_TYPE_HW_CACHE);
822 }
823 
824 extern atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
825 
826 extern void __perf_sw_event(u32, u64, int, struct pt_regs *, u64);
827 
828 static inline void
829 perf_sw_event(u32 event_id, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
830 {
831 	if (atomic_read(&perf_swevent_enabled[event_id]))
832 		__perf_sw_event(event_id, nr, nmi, regs, addr);
833 }
834 
835 extern void __perf_event_mmap(struct vm_area_struct *vma);
836 
837 static inline void perf_event_mmap(struct vm_area_struct *vma)
838 {
839 	if (vma->vm_flags & VM_EXEC)
840 		__perf_event_mmap(vma);
841 }
842 
843 extern void perf_event_comm(struct task_struct *tsk);
844 extern void perf_event_fork(struct task_struct *tsk);
845 
846 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
847 
848 extern int sysctl_perf_event_paranoid;
849 extern int sysctl_perf_event_mlock;
850 extern int sysctl_perf_event_sample_rate;
851 
852 extern void perf_event_init(void);
853 extern void perf_tp_event(int event_id, u64 addr, u64 count,
854 				 void *record, int entry_size);
855 extern void perf_bp_event(struct perf_event *event, void *data);
856 
857 #ifndef perf_misc_flags
858 #define perf_misc_flags(regs)	(user_mode(regs) ? PERF_RECORD_MISC_USER : \
859 				 PERF_RECORD_MISC_KERNEL)
860 #define perf_instruction_pointer(regs)	instruction_pointer(regs)
861 #endif
862 
863 extern int perf_output_begin(struct perf_output_handle *handle,
864 			     struct perf_event *event, unsigned int size,
865 			     int nmi, int sample);
866 extern void perf_output_end(struct perf_output_handle *handle);
867 extern void perf_output_copy(struct perf_output_handle *handle,
868 			     const void *buf, unsigned int len);
869 #else
870 static inline void
871 perf_event_task_sched_in(struct task_struct *task, int cpu)		{ }
872 static inline void
873 perf_event_task_sched_out(struct task_struct *task,
874 			    struct task_struct *next, int cpu)		{ }
875 static inline void
876 perf_event_task_tick(struct task_struct *task, int cpu)			{ }
877 static inline int perf_event_init_task(struct task_struct *child)	{ return 0; }
878 static inline void perf_event_exit_task(struct task_struct *child)	{ }
879 static inline void perf_event_free_task(struct task_struct *task)	{ }
880 static inline void perf_event_do_pending(void)				{ }
881 static inline void perf_event_print_debug(void)				{ }
882 static inline void perf_disable(void)					{ }
883 static inline void perf_enable(void)					{ }
884 static inline int perf_event_task_disable(void)				{ return -EINVAL; }
885 static inline int perf_event_task_enable(void)				{ return -EINVAL; }
886 
887 static inline void
888 perf_sw_event(u32 event_id, u64 nr, int nmi,
889 		     struct pt_regs *regs, u64 addr)			{ }
890 static inline void
891 perf_bp_event(struct perf_event *event, void *data)		{ }
892 
893 static inline void perf_event_mmap(struct vm_area_struct *vma)		{ }
894 static inline void perf_event_comm(struct task_struct *tsk)		{ }
895 static inline void perf_event_fork(struct task_struct *tsk)		{ }
896 static inline void perf_event_init(void)				{ }
897 
898 #endif
899 
900 #define perf_output_put(handle, x) \
901 	perf_output_copy((handle), &(x), sizeof(x))
902 
903 #endif /* __KERNEL__ */
904 #endif /* _LINUX_PERF_EVENT_H */
905