xref: /linux-6.15/include/linux/perf_event.h (revision edc7616c)
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 /*
22  * User-space ABI bits:
23  */
24 
25 /*
26  * attr.type
27  */
28 enum perf_type_id {
29 	PERF_TYPE_HARDWARE			= 0,
30 	PERF_TYPE_SOFTWARE			= 1,
31 	PERF_TYPE_TRACEPOINT			= 2,
32 	PERF_TYPE_HW_CACHE			= 3,
33 	PERF_TYPE_RAW				= 4,
34 	PERF_TYPE_BREAKPOINT			= 5,
35 
36 	PERF_TYPE_MAX,				/* non-ABI */
37 };
38 
39 /*
40  * Generalized performance event event_id types, used by the
41  * attr.event_id parameter of the sys_perf_event_open()
42  * syscall:
43  */
44 enum perf_hw_id {
45 	/*
46 	 * Common hardware events, generalized by the kernel:
47 	 */
48 	PERF_COUNT_HW_CPU_CYCLES		= 0,
49 	PERF_COUNT_HW_INSTRUCTIONS		= 1,
50 	PERF_COUNT_HW_CACHE_REFERENCES		= 2,
51 	PERF_COUNT_HW_CACHE_MISSES		= 3,
52 	PERF_COUNT_HW_BRANCH_INSTRUCTIONS	= 4,
53 	PERF_COUNT_HW_BRANCH_MISSES		= 5,
54 	PERF_COUNT_HW_BUS_CYCLES		= 6,
55 
56 	PERF_COUNT_HW_MAX,			/* non-ABI */
57 };
58 
59 /*
60  * Generalized hardware cache events:
61  *
62  *       { L1-D, L1-I, LLC, ITLB, DTLB, BPU } x
63  *       { read, write, prefetch } x
64  *       { accesses, misses }
65  */
66 enum perf_hw_cache_id {
67 	PERF_COUNT_HW_CACHE_L1D			= 0,
68 	PERF_COUNT_HW_CACHE_L1I			= 1,
69 	PERF_COUNT_HW_CACHE_LL			= 2,
70 	PERF_COUNT_HW_CACHE_DTLB		= 3,
71 	PERF_COUNT_HW_CACHE_ITLB		= 4,
72 	PERF_COUNT_HW_CACHE_BPU			= 5,
73 
74 	PERF_COUNT_HW_CACHE_MAX,		/* non-ABI */
75 };
76 
77 enum perf_hw_cache_op_id {
78 	PERF_COUNT_HW_CACHE_OP_READ		= 0,
79 	PERF_COUNT_HW_CACHE_OP_WRITE		= 1,
80 	PERF_COUNT_HW_CACHE_OP_PREFETCH		= 2,
81 
82 	PERF_COUNT_HW_CACHE_OP_MAX,		/* non-ABI */
83 };
84 
85 enum perf_hw_cache_op_result_id {
86 	PERF_COUNT_HW_CACHE_RESULT_ACCESS	= 0,
87 	PERF_COUNT_HW_CACHE_RESULT_MISS		= 1,
88 
89 	PERF_COUNT_HW_CACHE_RESULT_MAX,		/* non-ABI */
90 };
91 
92 /*
93  * Special "software" events provided by the kernel, even if the hardware
94  * does not support performance events. These events measure various
95  * physical and sw events of the kernel (and allow the profiling of them as
96  * well):
97  */
98 enum perf_sw_ids {
99 	PERF_COUNT_SW_CPU_CLOCK			= 0,
100 	PERF_COUNT_SW_TASK_CLOCK		= 1,
101 	PERF_COUNT_SW_PAGE_FAULTS		= 2,
102 	PERF_COUNT_SW_CONTEXT_SWITCHES		= 3,
103 	PERF_COUNT_SW_CPU_MIGRATIONS		= 4,
104 	PERF_COUNT_SW_PAGE_FAULTS_MIN		= 5,
105 	PERF_COUNT_SW_PAGE_FAULTS_MAJ		= 6,
106 	PERF_COUNT_SW_ALIGNMENT_FAULTS		= 7,
107 	PERF_COUNT_SW_EMULATION_FAULTS		= 8,
108 
109 	PERF_COUNT_SW_MAX,			/* non-ABI */
110 };
111 
112 /*
113  * Bits that can be set in attr.sample_type to request information
114  * in the overflow packets.
115  */
116 enum perf_event_sample_format {
117 	PERF_SAMPLE_IP				= 1U << 0,
118 	PERF_SAMPLE_TID				= 1U << 1,
119 	PERF_SAMPLE_TIME			= 1U << 2,
120 	PERF_SAMPLE_ADDR			= 1U << 3,
121 	PERF_SAMPLE_READ			= 1U << 4,
122 	PERF_SAMPLE_CALLCHAIN			= 1U << 5,
123 	PERF_SAMPLE_ID				= 1U << 6,
124 	PERF_SAMPLE_CPU				= 1U << 7,
125 	PERF_SAMPLE_PERIOD			= 1U << 8,
126 	PERF_SAMPLE_STREAM_ID			= 1U << 9,
127 	PERF_SAMPLE_RAW				= 1U << 10,
128 
129 	PERF_SAMPLE_MAX = 1U << 11,		/* non-ABI */
130 };
131 
132 /*
133  * The format of the data returned by read() on a perf event fd,
134  * as specified by attr.read_format:
135  *
136  * struct read_format {
137  *	{ u64		value;
138  *	  { u64		time_enabled; } && PERF_FORMAT_ENABLED
139  *	  { u64		time_running; } && PERF_FORMAT_RUNNING
140  *	  { u64		id;           } && PERF_FORMAT_ID
141  *	} && !PERF_FORMAT_GROUP
142  *
143  *	{ u64		nr;
144  *	  { u64		time_enabled; } && PERF_FORMAT_ENABLED
145  *	  { u64		time_running; } && PERF_FORMAT_RUNNING
146  *	  { u64		value;
147  *	    { u64	id;           } && PERF_FORMAT_ID
148  *	  }		cntr[nr];
149  *	} && PERF_FORMAT_GROUP
150  * };
151  */
152 enum perf_event_read_format {
153 	PERF_FORMAT_TOTAL_TIME_ENABLED		= 1U << 0,
154 	PERF_FORMAT_TOTAL_TIME_RUNNING		= 1U << 1,
155 	PERF_FORMAT_ID				= 1U << 2,
156 	PERF_FORMAT_GROUP			= 1U << 3,
157 
158 	PERF_FORMAT_MAX = 1U << 4,		/* non-ABI */
159 };
160 
161 #define PERF_ATTR_SIZE_VER0	64	/* sizeof first published struct */
162 
163 /*
164  * Hardware event_id to monitor via a performance monitoring event:
165  */
166 struct perf_event_attr {
167 
168 	/*
169 	 * Major type: hardware/software/tracepoint/etc.
170 	 */
171 	__u32			type;
172 
173 	/*
174 	 * Size of the attr structure, for fwd/bwd compat.
175 	 */
176 	__u32			size;
177 
178 	/*
179 	 * Type specific configuration information.
180 	 */
181 	__u64			config;
182 
183 	union {
184 		__u64		sample_period;
185 		__u64		sample_freq;
186 	};
187 
188 	__u64			sample_type;
189 	__u64			read_format;
190 
191 	__u64			disabled       :  1, /* off by default        */
192 				inherit	       :  1, /* children inherit it   */
193 				pinned	       :  1, /* must always be on PMU */
194 				exclusive      :  1, /* only group on PMU     */
195 				exclude_user   :  1, /* don't count user      */
196 				exclude_kernel :  1, /* ditto kernel          */
197 				exclude_hv     :  1, /* ditto hypervisor      */
198 				exclude_idle   :  1, /* don't count when idle */
199 				mmap           :  1, /* include mmap data     */
200 				comm	       :  1, /* include comm data     */
201 				freq           :  1, /* use freq, not period  */
202 				inherit_stat   :  1, /* per task counts       */
203 				enable_on_exec :  1, /* next exec enables     */
204 				task           :  1, /* trace fork/exit       */
205 				watermark      :  1, /* wakeup_watermark      */
206 
207 				__reserved_1   : 49;
208 
209 	union {
210 		__u32		wakeup_events;	  /* wakeup every n events */
211 		__u32		wakeup_watermark; /* bytes before wakeup   */
212 	};
213 
214 	__u32			bp_type;
215 	__u64			bp_addr;
216 	__u64			bp_len;
217 };
218 
219 /*
220  * Ioctls that can be done on a perf event fd:
221  */
222 #define PERF_EVENT_IOC_ENABLE		_IO ('$', 0)
223 #define PERF_EVENT_IOC_DISABLE		_IO ('$', 1)
224 #define PERF_EVENT_IOC_REFRESH		_IO ('$', 2)
225 #define PERF_EVENT_IOC_RESET		_IO ('$', 3)
226 #define PERF_EVENT_IOC_PERIOD		_IOW('$', 4, __u64)
227 #define PERF_EVENT_IOC_SET_OUTPUT	_IO ('$', 5)
228 #define PERF_EVENT_IOC_SET_FILTER	_IOW('$', 6, char *)
229 
230 enum perf_event_ioc_flags {
231 	PERF_IOC_FLAG_GROUP		= 1U << 0,
232 };
233 
234 /*
235  * Structure of the page that can be mapped via mmap
236  */
237 struct perf_event_mmap_page {
238 	__u32	version;		/* version number of this structure */
239 	__u32	compat_version;		/* lowest version this is compat with */
240 
241 	/*
242 	 * Bits needed to read the hw events in user-space.
243 	 *
244 	 *   u32 seq;
245 	 *   s64 count;
246 	 *
247 	 *   do {
248 	 *     seq = pc->lock;
249 	 *
250 	 *     barrier()
251 	 *     if (pc->index) {
252 	 *       count = pmc_read(pc->index - 1);
253 	 *       count += pc->offset;
254 	 *     } else
255 	 *       goto regular_read;
256 	 *
257 	 *     barrier();
258 	 *   } while (pc->lock != seq);
259 	 *
260 	 * NOTE: for obvious reason this only works on self-monitoring
261 	 *       processes.
262 	 */
263 	__u32	lock;			/* seqlock for synchronization */
264 	__u32	index;			/* hardware event identifier */
265 	__s64	offset;			/* add to hardware event value */
266 	__u64	time_enabled;		/* time event active */
267 	__u64	time_running;		/* time event on cpu */
268 
269 		/*
270 		 * Hole for extension of the self monitor capabilities
271 		 */
272 
273 	__u64	__reserved[123];	/* align to 1k */
274 
275 	/*
276 	 * Control data for the mmap() data buffer.
277 	 *
278 	 * User-space reading the @data_head value should issue an rmb(), on
279 	 * SMP capable platforms, after reading this value -- see
280 	 * perf_event_wakeup().
281 	 *
282 	 * When the mapping is PROT_WRITE the @data_tail value should be
283 	 * written by userspace to reflect the last read data. In this case
284 	 * the kernel will not over-write unread data.
285 	 */
286 	__u64   data_head;		/* head in the data section */
287 	__u64	data_tail;		/* user-space written tail */
288 };
289 
290 #define PERF_RECORD_MISC_CPUMODE_MASK		(3 << 0)
291 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN	(0 << 0)
292 #define PERF_RECORD_MISC_KERNEL			(1 << 0)
293 #define PERF_RECORD_MISC_USER			(2 << 0)
294 #define PERF_RECORD_MISC_HYPERVISOR		(3 << 0)
295 
296 struct perf_event_header {
297 	__u32	type;
298 	__u16	misc;
299 	__u16	size;
300 };
301 
302 enum perf_event_type {
303 
304 	/*
305 	 * The MMAP events record the PROT_EXEC mappings so that we can
306 	 * correlate userspace IPs to code. They have the following structure:
307 	 *
308 	 * struct {
309 	 *	struct perf_event_header	header;
310 	 *
311 	 *	u32				pid, tid;
312 	 *	u64				addr;
313 	 *	u64				len;
314 	 *	u64				pgoff;
315 	 *	char				filename[];
316 	 * };
317 	 */
318 	PERF_RECORD_MMAP			= 1,
319 
320 	/*
321 	 * struct {
322 	 *	struct perf_event_header	header;
323 	 *	u64				id;
324 	 *	u64				lost;
325 	 * };
326 	 */
327 	PERF_RECORD_LOST			= 2,
328 
329 	/*
330 	 * struct {
331 	 *	struct perf_event_header	header;
332 	 *
333 	 *	u32				pid, tid;
334 	 *	char				comm[];
335 	 * };
336 	 */
337 	PERF_RECORD_COMM			= 3,
338 
339 	/*
340 	 * struct {
341 	 *	struct perf_event_header	header;
342 	 *	u32				pid, ppid;
343 	 *	u32				tid, ptid;
344 	 *	u64				time;
345 	 * };
346 	 */
347 	PERF_RECORD_EXIT			= 4,
348 
349 	/*
350 	 * struct {
351 	 *	struct perf_event_header	header;
352 	 *	u64				time;
353 	 *	u64				id;
354 	 *	u64				stream_id;
355 	 * };
356 	 */
357 	PERF_RECORD_THROTTLE			= 5,
358 	PERF_RECORD_UNTHROTTLE			= 6,
359 
360 	/*
361 	 * struct {
362 	 *	struct perf_event_header	header;
363 	 *	u32				pid, ppid;
364 	 *	u32				tid, ptid;
365 	 *	u64				time;
366 	 * };
367 	 */
368 	PERF_RECORD_FORK			= 7,
369 
370 	/*
371 	 * struct {
372 	 *	struct perf_event_header	header;
373 	 *	u32				pid, tid;
374 	 *
375 	 *	struct read_format		values;
376 	 * };
377 	 */
378 	PERF_RECORD_READ			= 8,
379 
380 	/*
381 	 * struct {
382 	 *	struct perf_event_header	header;
383 	 *
384 	 *	{ u64			ip;	  } && PERF_SAMPLE_IP
385 	 *	{ u32			pid, tid; } && PERF_SAMPLE_TID
386 	 *	{ u64			time;     } && PERF_SAMPLE_TIME
387 	 *	{ u64			addr;     } && PERF_SAMPLE_ADDR
388 	 *	{ u64			id;	  } && PERF_SAMPLE_ID
389 	 *	{ u64			stream_id;} && PERF_SAMPLE_STREAM_ID
390 	 *	{ u32			cpu, res; } && PERF_SAMPLE_CPU
391 	 *	{ u64			period;   } && PERF_SAMPLE_PERIOD
392 	 *
393 	 *	{ struct read_format	values;	  } && PERF_SAMPLE_READ
394 	 *
395 	 *	{ u64			nr,
396 	 *	  u64			ips[nr];  } && PERF_SAMPLE_CALLCHAIN
397 	 *
398 	 *	#
399 	 *	# The RAW record below is opaque data wrt the ABI
400 	 *	#
401 	 *	# That is, the ABI doesn't make any promises wrt to
402 	 *	# the stability of its content, it may vary depending
403 	 *	# on event, hardware, kernel version and phase of
404 	 *	# the moon.
405 	 *	#
406 	 *	# In other words, PERF_SAMPLE_RAW contents are not an ABI.
407 	 *	#
408 	 *
409 	 *	{ u32			size;
410 	 *	  char                  data[size];}&& PERF_SAMPLE_RAW
411 	 * };
412 	 */
413 	PERF_RECORD_SAMPLE			= 9,
414 
415 	PERF_RECORD_MAX,			/* non-ABI */
416 };
417 
418 enum perf_callchain_context {
419 	PERF_CONTEXT_HV			= (__u64)-32,
420 	PERF_CONTEXT_KERNEL		= (__u64)-128,
421 	PERF_CONTEXT_USER		= (__u64)-512,
422 
423 	PERF_CONTEXT_GUEST		= (__u64)-2048,
424 	PERF_CONTEXT_GUEST_KERNEL	= (__u64)-2176,
425 	PERF_CONTEXT_GUEST_USER		= (__u64)-2560,
426 
427 	PERF_CONTEXT_MAX		= (__u64)-4095,
428 };
429 
430 #define PERF_FLAG_FD_NO_GROUP	(1U << 0)
431 #define PERF_FLAG_FD_OUTPUT	(1U << 1)
432 
433 #ifdef __KERNEL__
434 /*
435  * Kernel-internal data types and definitions:
436  */
437 
438 #ifdef CONFIG_PERF_EVENTS
439 # include <asm/perf_event.h>
440 #endif
441 
442 #ifdef CONFIG_HAVE_HW_BREAKPOINT
443 #include <asm/hw_breakpoint.h>
444 #endif
445 
446 #include <linux/list.h>
447 #include <linux/mutex.h>
448 #include <linux/rculist.h>
449 #include <linux/rcupdate.h>
450 #include <linux/spinlock.h>
451 #include <linux/hrtimer.h>
452 #include <linux/fs.h>
453 #include <linux/pid_namespace.h>
454 #include <linux/workqueue.h>
455 #include <asm/atomic.h>
456 
457 #define PERF_MAX_STACK_DEPTH		255
458 
459 struct perf_callchain_entry {
460 	__u64				nr;
461 	__u64				ip[PERF_MAX_STACK_DEPTH];
462 };
463 
464 struct perf_raw_record {
465 	u32				size;
466 	void				*data;
467 };
468 
469 struct task_struct;
470 
471 /**
472  * struct hw_perf_event - performance event hardware details:
473  */
474 struct hw_perf_event {
475 #ifdef CONFIG_PERF_EVENTS
476 	union {
477 		struct { /* hardware */
478 			u64		config;
479 			u64		last_tag;
480 			unsigned long	config_base;
481 			unsigned long	event_base;
482 			int		idx;
483 			int		last_cpu;
484 		};
485 		struct { /* software */
486 			s64		remaining;
487 			struct hrtimer	hrtimer;
488 		};
489 #ifdef CONFIG_HAVE_HW_BREAKPOINT
490 		/* breakpoint */
491 		struct arch_hw_breakpoint	info;
492 #endif
493 	};
494 	atomic64_t			prev_count;
495 	u64				sample_period;
496 	u64				last_period;
497 	atomic64_t			period_left;
498 	u64				interrupts;
499 
500 	u64				freq_time_stamp;
501 	u64				freq_count_stamp;
502 #endif
503 };
504 
505 struct perf_event;
506 
507 /**
508  * struct pmu - generic performance monitoring unit
509  */
510 struct pmu {
511 	int (*enable)			(struct perf_event *event);
512 	void (*disable)			(struct perf_event *event);
513 	int (*start)			(struct perf_event *event);
514 	void (*stop)			(struct perf_event *event);
515 	void (*read)			(struct perf_event *event);
516 	void (*unthrottle)		(struct perf_event *event);
517 };
518 
519 /**
520  * enum perf_event_active_state - the states of a event
521  */
522 enum perf_event_active_state {
523 	PERF_EVENT_STATE_ERROR		= -2,
524 	PERF_EVENT_STATE_OFF		= -1,
525 	PERF_EVENT_STATE_INACTIVE	=  0,
526 	PERF_EVENT_STATE_ACTIVE		=  1,
527 };
528 
529 struct file;
530 
531 struct perf_mmap_data {
532 	struct rcu_head			rcu_head;
533 #ifdef CONFIG_PERF_USE_VMALLOC
534 	struct work_struct		work;
535 #endif
536 	int				data_order;
537 	int				nr_pages;	/* nr of data pages  */
538 	int				writable;	/* are we writable   */
539 	int				nr_locked;	/* nr pages mlocked  */
540 
541 	atomic_t			poll;		/* POLL_ for wakeups */
542 	atomic_t			events;		/* event_id limit       */
543 
544 	atomic_long_t			head;		/* write position    */
545 	atomic_long_t			done_head;	/* completed head    */
546 
547 	atomic_t			lock;		/* concurrent writes */
548 	atomic_t			wakeup;		/* needs a wakeup    */
549 	atomic_t			lost;		/* nr records lost   */
550 
551 	long				watermark;	/* wakeup watermark  */
552 
553 	struct perf_event_mmap_page	*user_page;
554 	void				*data_pages[0];
555 };
556 
557 struct perf_pending_entry {
558 	struct perf_pending_entry *next;
559 	void (*func)(struct perf_pending_entry *);
560 };
561 
562 struct perf_sample_data;
563 
564 typedef void (*perf_overflow_handler_t)(struct perf_event *, int,
565 					struct perf_sample_data *,
566 					struct pt_regs *regs);
567 
568 enum perf_group_flag {
569 	PERF_GROUP_SOFTWARE = 0x1,
570 };
571 
572 /**
573  * struct perf_event - performance event kernel representation:
574  */
575 struct perf_event {
576 #ifdef CONFIG_PERF_EVENTS
577 	struct list_head		group_entry;
578 	struct list_head		event_entry;
579 	struct list_head		sibling_list;
580 	int				nr_siblings;
581 	int				group_flags;
582 	struct perf_event		*group_leader;
583 	struct perf_event		*output;
584 	const struct pmu		*pmu;
585 
586 	enum perf_event_active_state	state;
587 	atomic64_t			count;
588 
589 	/*
590 	 * These are the total time in nanoseconds that the event
591 	 * has been enabled (i.e. eligible to run, and the task has
592 	 * been scheduled in, if this is a per-task event)
593 	 * and running (scheduled onto the CPU), respectively.
594 	 *
595 	 * They are computed from tstamp_enabled, tstamp_running and
596 	 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
597 	 */
598 	u64				total_time_enabled;
599 	u64				total_time_running;
600 
601 	/*
602 	 * These are timestamps used for computing total_time_enabled
603 	 * and total_time_running when the event is in INACTIVE or
604 	 * ACTIVE state, measured in nanoseconds from an arbitrary point
605 	 * in time.
606 	 * tstamp_enabled: the notional time when the event was enabled
607 	 * tstamp_running: the notional time when the event was scheduled on
608 	 * tstamp_stopped: in INACTIVE state, the notional time when the
609 	 *	event was scheduled off.
610 	 */
611 	u64				tstamp_enabled;
612 	u64				tstamp_running;
613 	u64				tstamp_stopped;
614 
615 	struct perf_event_attr		attr;
616 	struct hw_perf_event		hw;
617 
618 	struct perf_event_context	*ctx;
619 	struct file			*filp;
620 
621 	/*
622 	 * These accumulate total time (in nanoseconds) that children
623 	 * events have been enabled and running, respectively.
624 	 */
625 	atomic64_t			child_total_time_enabled;
626 	atomic64_t			child_total_time_running;
627 
628 	/*
629 	 * Protect attach/detach and child_list:
630 	 */
631 	struct mutex			child_mutex;
632 	struct list_head		child_list;
633 	struct perf_event		*parent;
634 
635 	int				oncpu;
636 	int				cpu;
637 
638 	struct list_head		owner_entry;
639 	struct task_struct		*owner;
640 
641 	/* mmap bits */
642 	struct mutex			mmap_mutex;
643 	atomic_t			mmap_count;
644 	struct perf_mmap_data		*data;
645 
646 	/* poll related */
647 	wait_queue_head_t		waitq;
648 	struct fasync_struct		*fasync;
649 
650 	/* delayed work for NMIs and such */
651 	int				pending_wakeup;
652 	int				pending_kill;
653 	int				pending_disable;
654 	struct perf_pending_entry	pending;
655 
656 	atomic_t			event_limit;
657 
658 	void (*destroy)(struct perf_event *);
659 	struct rcu_head			rcu_head;
660 
661 	struct pid_namespace		*ns;
662 	u64				id;
663 
664 	perf_overflow_handler_t		overflow_handler;
665 
666 #ifdef CONFIG_EVENT_TRACING
667 	struct event_filter		*filter;
668 #endif
669 
670 #endif /* CONFIG_PERF_EVENTS */
671 };
672 
673 /**
674  * struct perf_event_context - event context structure
675  *
676  * Used as a container for task events and CPU events as well:
677  */
678 struct perf_event_context {
679 	/*
680 	 * Protect the states of the events in the list,
681 	 * nr_active, and the list:
682 	 */
683 	raw_spinlock_t			lock;
684 	/*
685 	 * Protect the list of events.  Locking either mutex or lock
686 	 * is sufficient to ensure the list doesn't change; to change
687 	 * the list you need to lock both the mutex and the spinlock.
688 	 */
689 	struct mutex			mutex;
690 
691 	struct list_head		pinned_groups;
692 	struct list_head		flexible_groups;
693 	struct list_head		event_list;
694 	int				nr_events;
695 	int				nr_active;
696 	int				is_active;
697 	int				nr_stat;
698 	atomic_t			refcount;
699 	struct task_struct		*task;
700 
701 	/*
702 	 * Context clock, runs when context enabled.
703 	 */
704 	u64				time;
705 	u64				timestamp;
706 
707 	/*
708 	 * These fields let us detect when two contexts have both
709 	 * been cloned (inherited) from a common ancestor.
710 	 */
711 	struct perf_event_context	*parent_ctx;
712 	u64				parent_gen;
713 	u64				generation;
714 	int				pin_count;
715 	struct rcu_head			rcu_head;
716 };
717 
718 /**
719  * struct perf_event_cpu_context - per cpu event context structure
720  */
721 struct perf_cpu_context {
722 	struct perf_event_context	ctx;
723 	struct perf_event_context	*task_ctx;
724 	int				active_oncpu;
725 	int				max_pertask;
726 	int				exclusive;
727 
728 	/*
729 	 * Recursion avoidance:
730 	 *
731 	 * task, softirq, irq, nmi context
732 	 */
733 	int				recursion[4];
734 };
735 
736 struct perf_output_handle {
737 	struct perf_event		*event;
738 	struct perf_mmap_data		*data;
739 	unsigned long			head;
740 	unsigned long			offset;
741 	int				nmi;
742 	int				sample;
743 	int				locked;
744 };
745 
746 #ifdef CONFIG_PERF_EVENTS
747 
748 /*
749  * Set by architecture code:
750  */
751 extern int perf_max_events;
752 
753 extern const struct pmu *hw_perf_event_init(struct perf_event *event);
754 
755 extern void perf_event_task_sched_in(struct task_struct *task);
756 extern void perf_event_task_sched_out(struct task_struct *task, struct task_struct *next);
757 extern void perf_event_task_tick(struct task_struct *task);
758 extern int perf_event_init_task(struct task_struct *child);
759 extern void perf_event_exit_task(struct task_struct *child);
760 extern void perf_event_free_task(struct task_struct *task);
761 extern void set_perf_event_pending(void);
762 extern void perf_event_do_pending(void);
763 extern void perf_event_print_debug(void);
764 extern void __perf_disable(void);
765 extern bool __perf_enable(void);
766 extern void perf_disable(void);
767 extern void perf_enable(void);
768 extern int perf_event_task_disable(void);
769 extern int perf_event_task_enable(void);
770 extern int hw_perf_group_sched_in(struct perf_event *group_leader,
771 	       struct perf_cpu_context *cpuctx,
772 	       struct perf_event_context *ctx);
773 extern void perf_event_update_userpage(struct perf_event *event);
774 extern int perf_event_release_kernel(struct perf_event *event);
775 extern struct perf_event *
776 perf_event_create_kernel_counter(struct perf_event_attr *attr,
777 				int cpu,
778 				pid_t pid,
779 				perf_overflow_handler_t callback);
780 extern u64 perf_event_read_value(struct perf_event *event,
781 				 u64 *enabled, u64 *running);
782 
783 struct perf_sample_data {
784 	u64				type;
785 
786 	u64				ip;
787 	struct {
788 		u32	pid;
789 		u32	tid;
790 	}				tid_entry;
791 	u64				time;
792 	u64				addr;
793 	u64				id;
794 	u64				stream_id;
795 	struct {
796 		u32	cpu;
797 		u32	reserved;
798 	}				cpu_entry;
799 	u64				period;
800 	struct perf_callchain_entry	*callchain;
801 	struct perf_raw_record		*raw;
802 };
803 
804 static inline
805 void perf_sample_data_init(struct perf_sample_data *data, u64 addr)
806 {
807 	data->addr = addr;
808 	data->raw  = NULL;
809 }
810 
811 extern void perf_output_sample(struct perf_output_handle *handle,
812 			       struct perf_event_header *header,
813 			       struct perf_sample_data *data,
814 			       struct perf_event *event);
815 extern void perf_prepare_sample(struct perf_event_header *header,
816 				struct perf_sample_data *data,
817 				struct perf_event *event,
818 				struct pt_regs *regs);
819 
820 extern int perf_event_overflow(struct perf_event *event, int nmi,
821 				 struct perf_sample_data *data,
822 				 struct pt_regs *regs);
823 
824 /*
825  * Return 1 for a software event, 0 for a hardware event
826  */
827 static inline int is_software_event(struct perf_event *event)
828 {
829 	switch (event->attr.type) {
830 	case PERF_TYPE_SOFTWARE:
831 	case PERF_TYPE_TRACEPOINT:
832 	/* for now the breakpoint stuff also works as software event */
833 	case PERF_TYPE_BREAKPOINT:
834 		return 1;
835 	}
836 	return 0;
837 }
838 
839 extern atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
840 
841 extern void __perf_sw_event(u32, u64, int, struct pt_regs *, u64);
842 
843 static inline void
844 perf_sw_event(u32 event_id, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
845 {
846 	if (atomic_read(&perf_swevent_enabled[event_id]))
847 		__perf_sw_event(event_id, nr, nmi, regs, addr);
848 }
849 
850 extern void __perf_event_mmap(struct vm_area_struct *vma);
851 
852 static inline void perf_event_mmap(struct vm_area_struct *vma)
853 {
854 	if (vma->vm_flags & VM_EXEC)
855 		__perf_event_mmap(vma);
856 }
857 
858 extern void perf_event_comm(struct task_struct *tsk);
859 extern void perf_event_fork(struct task_struct *tsk);
860 
861 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
862 
863 extern int sysctl_perf_event_paranoid;
864 extern int sysctl_perf_event_mlock;
865 extern int sysctl_perf_event_sample_rate;
866 
867 static inline bool perf_paranoid_tracepoint_raw(void)
868 {
869 	return sysctl_perf_event_paranoid > -1;
870 }
871 
872 static inline bool perf_paranoid_cpu(void)
873 {
874 	return sysctl_perf_event_paranoid > 0;
875 }
876 
877 static inline bool perf_paranoid_kernel(void)
878 {
879 	return sysctl_perf_event_paranoid > 1;
880 }
881 
882 extern void perf_event_init(void);
883 extern void perf_tp_event(int event_id, u64 addr, u64 count, void *record, int entry_size);
884 extern void perf_bp_event(struct perf_event *event, void *data);
885 
886 #ifndef perf_misc_flags
887 #define perf_misc_flags(regs)	(user_mode(regs) ? PERF_RECORD_MISC_USER : \
888 				 PERF_RECORD_MISC_KERNEL)
889 #define perf_instruction_pointer(regs)	instruction_pointer(regs)
890 #endif
891 
892 extern int perf_output_begin(struct perf_output_handle *handle,
893 			     struct perf_event *event, unsigned int size,
894 			     int nmi, int sample);
895 extern void perf_output_end(struct perf_output_handle *handle);
896 extern void perf_output_copy(struct perf_output_handle *handle,
897 			     const void *buf, unsigned int len);
898 extern int perf_swevent_get_recursion_context(void);
899 extern void perf_swevent_put_recursion_context(int rctx);
900 extern void perf_event_enable(struct perf_event *event);
901 extern void perf_event_disable(struct perf_event *event);
902 #else
903 static inline void
904 perf_event_task_sched_in(struct task_struct *task)			{ }
905 static inline void
906 perf_event_task_sched_out(struct task_struct *task,
907 			    struct task_struct *next)			{ }
908 static inline void
909 perf_event_task_tick(struct task_struct *task)				{ }
910 static inline int perf_event_init_task(struct task_struct *child)	{ return 0; }
911 static inline void perf_event_exit_task(struct task_struct *child)	{ }
912 static inline void perf_event_free_task(struct task_struct *task)	{ }
913 static inline void perf_event_do_pending(void)				{ }
914 static inline void perf_event_print_debug(void)				{ }
915 static inline void perf_disable(void)					{ }
916 static inline void perf_enable(void)					{ }
917 static inline int perf_event_task_disable(void)				{ return -EINVAL; }
918 static inline int perf_event_task_enable(void)				{ return -EINVAL; }
919 
920 static inline void
921 perf_sw_event(u32 event_id, u64 nr, int nmi,
922 		     struct pt_regs *regs, u64 addr)			{ }
923 static inline void
924 perf_bp_event(struct perf_event *event, void *data)			{ }
925 
926 static inline void perf_event_mmap(struct vm_area_struct *vma)		{ }
927 static inline void perf_event_comm(struct task_struct *tsk)		{ }
928 static inline void perf_event_fork(struct task_struct *tsk)		{ }
929 static inline void perf_event_init(void)				{ }
930 static inline int  perf_swevent_get_recursion_context(void)		{ return -1; }
931 static inline void perf_swevent_put_recursion_context(int rctx)		{ }
932 static inline void perf_event_enable(struct perf_event *event)		{ }
933 static inline void perf_event_disable(struct perf_event *event)		{ }
934 #endif
935 
936 #define perf_output_put(handle, x) \
937 	perf_output_copy((handle), &(x), sizeof(x))
938 
939 #endif /* __KERNEL__ */
940 #endif /* _LINUX_PERF_EVENT_H */
941