xref: /linux-6.15/include/linux/perf_event.h (revision c145211d)
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 				 * precise_ip:
208 				 *
209 				 *  0 - SAMPLE_IP can have arbitrary skid
210 				 *  1 - SAMPLE_IP must have constant skid
211 				 *  2 - SAMPLE_IP requested to have 0 skid
212 				 *  3 - SAMPLE_IP must have 0 skid
213 				 *
214 				 *  See also PERF_RECORD_MISC_EXACT_IP
215 				 */
216 				precise_ip     :  2, /* skid constraint       */
217 
218 				__reserved_1   : 47;
219 
220 	union {
221 		__u32		wakeup_events;	  /* wakeup every n events */
222 		__u32		wakeup_watermark; /* bytes before wakeup   */
223 	};
224 
225 	__u32			bp_type;
226 	__u64			bp_addr;
227 	__u64			bp_len;
228 };
229 
230 /*
231  * Ioctls that can be done on a perf event fd:
232  */
233 #define PERF_EVENT_IOC_ENABLE		_IO ('$', 0)
234 #define PERF_EVENT_IOC_DISABLE		_IO ('$', 1)
235 #define PERF_EVENT_IOC_REFRESH		_IO ('$', 2)
236 #define PERF_EVENT_IOC_RESET		_IO ('$', 3)
237 #define PERF_EVENT_IOC_PERIOD		_IOW('$', 4, __u64)
238 #define PERF_EVENT_IOC_SET_OUTPUT	_IO ('$', 5)
239 #define PERF_EVENT_IOC_SET_FILTER	_IOW('$', 6, char *)
240 
241 enum perf_event_ioc_flags {
242 	PERF_IOC_FLAG_GROUP		= 1U << 0,
243 };
244 
245 /*
246  * Structure of the page that can be mapped via mmap
247  */
248 struct perf_event_mmap_page {
249 	__u32	version;		/* version number of this structure */
250 	__u32	compat_version;		/* lowest version this is compat with */
251 
252 	/*
253 	 * Bits needed to read the hw events in user-space.
254 	 *
255 	 *   u32 seq;
256 	 *   s64 count;
257 	 *
258 	 *   do {
259 	 *     seq = pc->lock;
260 	 *
261 	 *     barrier()
262 	 *     if (pc->index) {
263 	 *       count = pmc_read(pc->index - 1);
264 	 *       count += pc->offset;
265 	 *     } else
266 	 *       goto regular_read;
267 	 *
268 	 *     barrier();
269 	 *   } while (pc->lock != seq);
270 	 *
271 	 * NOTE: for obvious reason this only works on self-monitoring
272 	 *       processes.
273 	 */
274 	__u32	lock;			/* seqlock for synchronization */
275 	__u32	index;			/* hardware event identifier */
276 	__s64	offset;			/* add to hardware event value */
277 	__u64	time_enabled;		/* time event active */
278 	__u64	time_running;		/* time event on cpu */
279 
280 		/*
281 		 * Hole for extension of the self monitor capabilities
282 		 */
283 
284 	__u64	__reserved[123];	/* align to 1k */
285 
286 	/*
287 	 * Control data for the mmap() data buffer.
288 	 *
289 	 * User-space reading the @data_head value should issue an rmb(), on
290 	 * SMP capable platforms, after reading this value -- see
291 	 * perf_event_wakeup().
292 	 *
293 	 * When the mapping is PROT_WRITE the @data_tail value should be
294 	 * written by userspace to reflect the last read data. In this case
295 	 * the kernel will not over-write unread data.
296 	 */
297 	__u64   data_head;		/* head in the data section */
298 	__u64	data_tail;		/* user-space written tail */
299 };
300 
301 #define PERF_RECORD_MISC_CPUMODE_MASK		(7 << 0)
302 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN	(0 << 0)
303 #define PERF_RECORD_MISC_KERNEL			(1 << 0)
304 #define PERF_RECORD_MISC_USER			(2 << 0)
305 #define PERF_RECORD_MISC_HYPERVISOR		(3 << 0)
306 #define PERF_RECORD_MISC_GUEST_KERNEL		(4 << 0)
307 #define PERF_RECORD_MISC_GUEST_USER		(5 << 0)
308 
309 /*
310  * Indicates that the content of PERF_SAMPLE_IP points to
311  * the actual instruction that triggered the event. See also
312  * perf_event_attr::precise_ip.
313  */
314 #define PERF_RECORD_MISC_EXACT_IP		(1 << 14)
315 /*
316  * Reserve the last bit to indicate some extended misc field
317  */
318 #define PERF_RECORD_MISC_EXT_RESERVED		(1 << 15)
319 
320 struct perf_event_header {
321 	__u32	type;
322 	__u16	misc;
323 	__u16	size;
324 };
325 
326 enum perf_event_type {
327 
328 	/*
329 	 * The MMAP events record the PROT_EXEC mappings so that we can
330 	 * correlate userspace IPs to code. They have the following structure:
331 	 *
332 	 * struct {
333 	 *	struct perf_event_header	header;
334 	 *
335 	 *	u32				pid, tid;
336 	 *	u64				addr;
337 	 *	u64				len;
338 	 *	u64				pgoff;
339 	 *	char				filename[];
340 	 * };
341 	 */
342 	PERF_RECORD_MMAP			= 1,
343 
344 	/*
345 	 * struct {
346 	 *	struct perf_event_header	header;
347 	 *	u64				id;
348 	 *	u64				lost;
349 	 * };
350 	 */
351 	PERF_RECORD_LOST			= 2,
352 
353 	/*
354 	 * struct {
355 	 *	struct perf_event_header	header;
356 	 *
357 	 *	u32				pid, tid;
358 	 *	char				comm[];
359 	 * };
360 	 */
361 	PERF_RECORD_COMM			= 3,
362 
363 	/*
364 	 * struct {
365 	 *	struct perf_event_header	header;
366 	 *	u32				pid, ppid;
367 	 *	u32				tid, ptid;
368 	 *	u64				time;
369 	 * };
370 	 */
371 	PERF_RECORD_EXIT			= 4,
372 
373 	/*
374 	 * struct {
375 	 *	struct perf_event_header	header;
376 	 *	u64				time;
377 	 *	u64				id;
378 	 *	u64				stream_id;
379 	 * };
380 	 */
381 	PERF_RECORD_THROTTLE			= 5,
382 	PERF_RECORD_UNTHROTTLE			= 6,
383 
384 	/*
385 	 * struct {
386 	 *	struct perf_event_header	header;
387 	 *	u32				pid, ppid;
388 	 *	u32				tid, ptid;
389 	 *	u64				time;
390 	 * };
391 	 */
392 	PERF_RECORD_FORK			= 7,
393 
394 	/*
395 	 * struct {
396 	 *	struct perf_event_header	header;
397 	 *	u32				pid, tid;
398 	 *
399 	 *	struct read_format		values;
400 	 * };
401 	 */
402 	PERF_RECORD_READ			= 8,
403 
404 	/*
405 	 * struct {
406 	 *	struct perf_event_header	header;
407 	 *
408 	 *	{ u64			ip;	  } && PERF_SAMPLE_IP
409 	 *	{ u32			pid, tid; } && PERF_SAMPLE_TID
410 	 *	{ u64			time;     } && PERF_SAMPLE_TIME
411 	 *	{ u64			addr;     } && PERF_SAMPLE_ADDR
412 	 *	{ u64			id;	  } && PERF_SAMPLE_ID
413 	 *	{ u64			stream_id;} && PERF_SAMPLE_STREAM_ID
414 	 *	{ u32			cpu, res; } && PERF_SAMPLE_CPU
415 	 *	{ u64			period;   } && PERF_SAMPLE_PERIOD
416 	 *
417 	 *	{ struct read_format	values;	  } && PERF_SAMPLE_READ
418 	 *
419 	 *	{ u64			nr,
420 	 *	  u64			ips[nr];  } && PERF_SAMPLE_CALLCHAIN
421 	 *
422 	 *	#
423 	 *	# The RAW record below is opaque data wrt the ABI
424 	 *	#
425 	 *	# That is, the ABI doesn't make any promises wrt to
426 	 *	# the stability of its content, it may vary depending
427 	 *	# on event, hardware, kernel version and phase of
428 	 *	# the moon.
429 	 *	#
430 	 *	# In other words, PERF_SAMPLE_RAW contents are not an ABI.
431 	 *	#
432 	 *
433 	 *	{ u32			size;
434 	 *	  char                  data[size];}&& PERF_SAMPLE_RAW
435 	 * };
436 	 */
437 	PERF_RECORD_SAMPLE			= 9,
438 
439 	PERF_RECORD_MAX,			/* non-ABI */
440 };
441 
442 enum perf_callchain_context {
443 	PERF_CONTEXT_HV			= (__u64)-32,
444 	PERF_CONTEXT_KERNEL		= (__u64)-128,
445 	PERF_CONTEXT_USER		= (__u64)-512,
446 
447 	PERF_CONTEXT_GUEST		= (__u64)-2048,
448 	PERF_CONTEXT_GUEST_KERNEL	= (__u64)-2176,
449 	PERF_CONTEXT_GUEST_USER		= (__u64)-2560,
450 
451 	PERF_CONTEXT_MAX		= (__u64)-4095,
452 };
453 
454 #define PERF_FLAG_FD_NO_GROUP	(1U << 0)
455 #define PERF_FLAG_FD_OUTPUT	(1U << 1)
456 
457 #ifdef __KERNEL__
458 /*
459  * Kernel-internal data types and definitions:
460  */
461 
462 #ifdef CONFIG_PERF_EVENTS
463 # include <asm/perf_event.h>
464 #endif
465 
466 struct perf_guest_info_callbacks {
467 	int (*is_in_guest) (void);
468 	int (*is_user_mode) (void);
469 	unsigned long (*get_guest_ip) (void);
470 };
471 
472 #ifdef CONFIG_HAVE_HW_BREAKPOINT
473 #include <asm/hw_breakpoint.h>
474 #endif
475 
476 #include <linux/list.h>
477 #include <linux/mutex.h>
478 #include <linux/rculist.h>
479 #include <linux/rcupdate.h>
480 #include <linux/spinlock.h>
481 #include <linux/hrtimer.h>
482 #include <linux/fs.h>
483 #include <linux/pid_namespace.h>
484 #include <linux/workqueue.h>
485 #include <linux/ftrace.h>
486 #include <linux/cpu.h>
487 #include <asm/atomic.h>
488 
489 #define PERF_MAX_STACK_DEPTH		255
490 
491 struct perf_callchain_entry {
492 	__u64				nr;
493 	__u64				ip[PERF_MAX_STACK_DEPTH];
494 };
495 
496 struct perf_raw_record {
497 	u32				size;
498 	void				*data;
499 };
500 
501 struct perf_branch_entry {
502 	__u64				from;
503 	__u64				to;
504 	__u64				flags;
505 };
506 
507 struct perf_branch_stack {
508 	__u64				nr;
509 	struct perf_branch_entry	entries[0];
510 };
511 
512 struct task_struct;
513 
514 /**
515  * struct hw_perf_event - performance event hardware details:
516  */
517 struct hw_perf_event {
518 #ifdef CONFIG_PERF_EVENTS
519 	union {
520 		struct { /* hardware */
521 			u64		config;
522 			u64		last_tag;
523 			unsigned long	config_base;
524 			unsigned long	event_base;
525 			int		idx;
526 			int		last_cpu;
527 		};
528 		struct { /* software */
529 			s64		remaining;
530 			struct hrtimer	hrtimer;
531 		};
532 #ifdef CONFIG_HAVE_HW_BREAKPOINT
533 		/* breakpoint */
534 		struct arch_hw_breakpoint	info;
535 #endif
536 	};
537 	atomic64_t			prev_count;
538 	u64				sample_period;
539 	u64				last_period;
540 	atomic64_t			period_left;
541 	u64				interrupts;
542 
543 	u64				freq_time_stamp;
544 	u64				freq_count_stamp;
545 #endif
546 };
547 
548 struct perf_event;
549 
550 #define PERF_EVENT_TXN_STARTED 1
551 
552 /**
553  * struct pmu - generic performance monitoring unit
554  */
555 struct pmu {
556 	int (*enable)			(struct perf_event *event);
557 	void (*disable)			(struct perf_event *event);
558 	int (*start)			(struct perf_event *event);
559 	void (*stop)			(struct perf_event *event);
560 	void (*read)			(struct perf_event *event);
561 	void (*unthrottle)		(struct perf_event *event);
562 
563 	/*
564 	 * group events scheduling is treated as a transaction,
565 	 * add group events as a whole and perform one schedulability test.
566 	 * If test fails, roll back the whole group
567 	 */
568 
569 	void (*start_txn)	(const struct pmu *pmu);
570 	void (*cancel_txn)	(const struct pmu *pmu);
571 	int  (*commit_txn)	(const struct pmu *pmu);
572 };
573 
574 /**
575  * enum perf_event_active_state - the states of a event
576  */
577 enum perf_event_active_state {
578 	PERF_EVENT_STATE_ERROR		= -2,
579 	PERF_EVENT_STATE_OFF		= -1,
580 	PERF_EVENT_STATE_INACTIVE	=  0,
581 	PERF_EVENT_STATE_ACTIVE		=  1,
582 };
583 
584 struct file;
585 
586 struct perf_mmap_data {
587 	struct rcu_head			rcu_head;
588 #ifdef CONFIG_PERF_USE_VMALLOC
589 	struct work_struct		work;
590 #endif
591 	int				data_order;
592 	int				nr_pages;	/* nr of data pages  */
593 	int				writable;	/* are we writable   */
594 	int				nr_locked;	/* nr pages mlocked  */
595 
596 	atomic_t			poll;		/* POLL_ for wakeups */
597 	atomic_t			events;		/* event_id limit       */
598 
599 	atomic_long_t			head;		/* write position    */
600 	atomic_long_t			done_head;	/* completed head    */
601 
602 	atomic_t			lock;		/* concurrent writes */
603 	atomic_t			wakeup;		/* needs a wakeup    */
604 	atomic_t			lost;		/* nr records lost   */
605 
606 	long				watermark;	/* wakeup watermark  */
607 
608 	struct perf_event_mmap_page	*user_page;
609 	void				*data_pages[0];
610 };
611 
612 struct perf_pending_entry {
613 	struct perf_pending_entry *next;
614 	void (*func)(struct perf_pending_entry *);
615 };
616 
617 struct perf_sample_data;
618 
619 typedef void (*perf_overflow_handler_t)(struct perf_event *, int,
620 					struct perf_sample_data *,
621 					struct pt_regs *regs);
622 
623 enum perf_group_flag {
624 	PERF_GROUP_SOFTWARE = 0x1,
625 };
626 
627 #define SWEVENT_HLIST_BITS	8
628 #define SWEVENT_HLIST_SIZE	(1 << SWEVENT_HLIST_BITS)
629 
630 struct swevent_hlist {
631 	struct hlist_head	heads[SWEVENT_HLIST_SIZE];
632 	struct rcu_head		rcu_head;
633 };
634 
635 /**
636  * struct perf_event - performance event kernel representation:
637  */
638 struct perf_event {
639 #ifdef CONFIG_PERF_EVENTS
640 	struct list_head		group_entry;
641 	struct list_head		event_entry;
642 	struct list_head		sibling_list;
643 	struct hlist_node		hlist_entry;
644 	int				nr_siblings;
645 	int				group_flags;
646 	struct perf_event		*group_leader;
647 	struct perf_event		*output;
648 	const struct pmu		*pmu;
649 
650 	enum perf_event_active_state	state;
651 	atomic64_t			count;
652 
653 	/*
654 	 * These are the total time in nanoseconds that the event
655 	 * has been enabled (i.e. eligible to run, and the task has
656 	 * been scheduled in, if this is a per-task event)
657 	 * and running (scheduled onto the CPU), respectively.
658 	 *
659 	 * They are computed from tstamp_enabled, tstamp_running and
660 	 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
661 	 */
662 	u64				total_time_enabled;
663 	u64				total_time_running;
664 
665 	/*
666 	 * These are timestamps used for computing total_time_enabled
667 	 * and total_time_running when the event is in INACTIVE or
668 	 * ACTIVE state, measured in nanoseconds from an arbitrary point
669 	 * in time.
670 	 * tstamp_enabled: the notional time when the event was enabled
671 	 * tstamp_running: the notional time when the event was scheduled on
672 	 * tstamp_stopped: in INACTIVE state, the notional time when the
673 	 *	event was scheduled off.
674 	 */
675 	u64				tstamp_enabled;
676 	u64				tstamp_running;
677 	u64				tstamp_stopped;
678 
679 	struct perf_event_attr		attr;
680 	struct hw_perf_event		hw;
681 
682 	struct perf_event_context	*ctx;
683 	struct file			*filp;
684 
685 	/*
686 	 * These accumulate total time (in nanoseconds) that children
687 	 * events have been enabled and running, respectively.
688 	 */
689 	atomic64_t			child_total_time_enabled;
690 	atomic64_t			child_total_time_running;
691 
692 	/*
693 	 * Protect attach/detach and child_list:
694 	 */
695 	struct mutex			child_mutex;
696 	struct list_head		child_list;
697 	struct perf_event		*parent;
698 
699 	int				oncpu;
700 	int				cpu;
701 
702 	struct list_head		owner_entry;
703 	struct task_struct		*owner;
704 
705 	/* mmap bits */
706 	struct mutex			mmap_mutex;
707 	atomic_t			mmap_count;
708 	struct perf_mmap_data		*data;
709 
710 	/* poll related */
711 	wait_queue_head_t		waitq;
712 	struct fasync_struct		*fasync;
713 
714 	/* delayed work for NMIs and such */
715 	int				pending_wakeup;
716 	int				pending_kill;
717 	int				pending_disable;
718 	struct perf_pending_entry	pending;
719 
720 	atomic_t			event_limit;
721 
722 	void (*destroy)(struct perf_event *);
723 	struct rcu_head			rcu_head;
724 
725 	struct pid_namespace		*ns;
726 	u64				id;
727 
728 	perf_overflow_handler_t		overflow_handler;
729 
730 #ifdef CONFIG_EVENT_TRACING
731 	struct event_filter		*filter;
732 #endif
733 
734 #endif /* CONFIG_PERF_EVENTS */
735 };
736 
737 /**
738  * struct perf_event_context - event context structure
739  *
740  * Used as a container for task events and CPU events as well:
741  */
742 struct perf_event_context {
743 	/*
744 	 * Protect the states of the events in the list,
745 	 * nr_active, and the list:
746 	 */
747 	raw_spinlock_t			lock;
748 	/*
749 	 * Protect the list of events.  Locking either mutex or lock
750 	 * is sufficient to ensure the list doesn't change; to change
751 	 * the list you need to lock both the mutex and the spinlock.
752 	 */
753 	struct mutex			mutex;
754 
755 	struct list_head		pinned_groups;
756 	struct list_head		flexible_groups;
757 	struct list_head		event_list;
758 	int				nr_events;
759 	int				nr_active;
760 	int				is_active;
761 	int				nr_stat;
762 	atomic_t			refcount;
763 	struct task_struct		*task;
764 
765 	/*
766 	 * Context clock, runs when context enabled.
767 	 */
768 	u64				time;
769 	u64				timestamp;
770 
771 	/*
772 	 * These fields let us detect when two contexts have both
773 	 * been cloned (inherited) from a common ancestor.
774 	 */
775 	struct perf_event_context	*parent_ctx;
776 	u64				parent_gen;
777 	u64				generation;
778 	int				pin_count;
779 	struct rcu_head			rcu_head;
780 };
781 
782 /**
783  * struct perf_event_cpu_context - per cpu event context structure
784  */
785 struct perf_cpu_context {
786 	struct perf_event_context	ctx;
787 	struct perf_event_context	*task_ctx;
788 	int				active_oncpu;
789 	int				max_pertask;
790 	int				exclusive;
791 	struct swevent_hlist		*swevent_hlist;
792 	struct mutex			hlist_mutex;
793 	int				hlist_refcount;
794 
795 	/*
796 	 * Recursion avoidance:
797 	 *
798 	 * task, softirq, irq, nmi context
799 	 */
800 	int				recursion[4];
801 };
802 
803 struct perf_output_handle {
804 	struct perf_event		*event;
805 	struct perf_mmap_data		*data;
806 	unsigned long			head;
807 	unsigned long			offset;
808 	int				nmi;
809 	int				sample;
810 	int				locked;
811 };
812 
813 #ifdef CONFIG_PERF_EVENTS
814 
815 /*
816  * Set by architecture code:
817  */
818 extern int perf_max_events;
819 
820 extern const struct pmu *hw_perf_event_init(struct perf_event *event);
821 
822 extern void perf_event_task_sched_in(struct task_struct *task);
823 extern void perf_event_task_sched_out(struct task_struct *task, struct task_struct *next);
824 extern void perf_event_task_tick(struct task_struct *task);
825 extern int perf_event_init_task(struct task_struct *child);
826 extern void perf_event_exit_task(struct task_struct *child);
827 extern void perf_event_free_task(struct task_struct *task);
828 extern void set_perf_event_pending(void);
829 extern void perf_event_do_pending(void);
830 extern void perf_event_print_debug(void);
831 extern void __perf_disable(void);
832 extern bool __perf_enable(void);
833 extern void perf_disable(void);
834 extern void perf_enable(void);
835 extern int perf_event_task_disable(void);
836 extern int perf_event_task_enable(void);
837 extern void perf_event_update_userpage(struct perf_event *event);
838 extern int perf_event_release_kernel(struct perf_event *event);
839 extern struct perf_event *
840 perf_event_create_kernel_counter(struct perf_event_attr *attr,
841 				int cpu,
842 				pid_t pid,
843 				perf_overflow_handler_t callback);
844 extern u64 perf_event_read_value(struct perf_event *event,
845 				 u64 *enabled, u64 *running);
846 
847 struct perf_sample_data {
848 	u64				type;
849 
850 	u64				ip;
851 	struct {
852 		u32	pid;
853 		u32	tid;
854 	}				tid_entry;
855 	u64				time;
856 	u64				addr;
857 	u64				id;
858 	u64				stream_id;
859 	struct {
860 		u32	cpu;
861 		u32	reserved;
862 	}				cpu_entry;
863 	u64				period;
864 	struct perf_callchain_entry	*callchain;
865 	struct perf_raw_record		*raw;
866 };
867 
868 static inline
869 void perf_sample_data_init(struct perf_sample_data *data, u64 addr)
870 {
871 	data->addr = addr;
872 	data->raw  = NULL;
873 }
874 
875 extern void perf_output_sample(struct perf_output_handle *handle,
876 			       struct perf_event_header *header,
877 			       struct perf_sample_data *data,
878 			       struct perf_event *event);
879 extern void perf_prepare_sample(struct perf_event_header *header,
880 				struct perf_sample_data *data,
881 				struct perf_event *event,
882 				struct pt_regs *regs);
883 
884 extern int perf_event_overflow(struct perf_event *event, int nmi,
885 				 struct perf_sample_data *data,
886 				 struct pt_regs *regs);
887 
888 /*
889  * Return 1 for a software event, 0 for a hardware event
890  */
891 static inline int is_software_event(struct perf_event *event)
892 {
893 	switch (event->attr.type) {
894 	case PERF_TYPE_SOFTWARE:
895 	case PERF_TYPE_TRACEPOINT:
896 	/* for now the breakpoint stuff also works as software event */
897 	case PERF_TYPE_BREAKPOINT:
898 		return 1;
899 	}
900 	return 0;
901 }
902 
903 extern atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
904 
905 extern void __perf_sw_event(u32, u64, int, struct pt_regs *, u64);
906 
907 extern void
908 perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip, int skip);
909 
910 /*
911  * Take a snapshot of the regs. Skip ip and frame pointer to
912  * the nth caller. We only need a few of the regs:
913  * - ip for PERF_SAMPLE_IP
914  * - cs for user_mode() tests
915  * - bp for callchains
916  * - eflags, for future purposes, just in case
917  */
918 static inline void perf_fetch_caller_regs(struct pt_regs *regs, int skip)
919 {
920 	unsigned long ip;
921 
922 	memset(regs, 0, sizeof(*regs));
923 
924 	switch (skip) {
925 	case 1 :
926 		ip = CALLER_ADDR0;
927 		break;
928 	case 2 :
929 		ip = CALLER_ADDR1;
930 		break;
931 	case 3 :
932 		ip = CALLER_ADDR2;
933 		break;
934 	case 4:
935 		ip = CALLER_ADDR3;
936 		break;
937 	/* No need to support further for now */
938 	default:
939 		ip = 0;
940 	}
941 
942 	return perf_arch_fetch_caller_regs(regs, ip, skip);
943 }
944 
945 static inline void
946 perf_sw_event(u32 event_id, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
947 {
948 	if (atomic_read(&perf_swevent_enabled[event_id])) {
949 		struct pt_regs hot_regs;
950 
951 		if (!regs) {
952 			perf_fetch_caller_regs(&hot_regs, 1);
953 			regs = &hot_regs;
954 		}
955 		__perf_sw_event(event_id, nr, nmi, regs, addr);
956 	}
957 }
958 
959 extern void __perf_event_mmap(struct vm_area_struct *vma);
960 
961 static inline void perf_event_mmap(struct vm_area_struct *vma)
962 {
963 	if (vma->vm_flags & VM_EXEC)
964 		__perf_event_mmap(vma);
965 }
966 
967 extern struct perf_guest_info_callbacks *perf_guest_cbs;
968 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
969 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
970 
971 extern void perf_event_comm(struct task_struct *tsk);
972 extern void perf_event_fork(struct task_struct *tsk);
973 
974 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
975 
976 extern int sysctl_perf_event_paranoid;
977 extern int sysctl_perf_event_mlock;
978 extern int sysctl_perf_event_sample_rate;
979 
980 static inline bool perf_paranoid_tracepoint_raw(void)
981 {
982 	return sysctl_perf_event_paranoid > -1;
983 }
984 
985 static inline bool perf_paranoid_cpu(void)
986 {
987 	return sysctl_perf_event_paranoid > 0;
988 }
989 
990 static inline bool perf_paranoid_kernel(void)
991 {
992 	return sysctl_perf_event_paranoid > 1;
993 }
994 
995 extern void perf_event_init(void);
996 extern void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
997 			  int entry_size, struct pt_regs *regs);
998 extern void perf_bp_event(struct perf_event *event, void *data);
999 
1000 #ifndef perf_misc_flags
1001 #define perf_misc_flags(regs)	(user_mode(regs) ? PERF_RECORD_MISC_USER : \
1002 				 PERF_RECORD_MISC_KERNEL)
1003 #define perf_instruction_pointer(regs)	instruction_pointer(regs)
1004 #endif
1005 
1006 extern int perf_output_begin(struct perf_output_handle *handle,
1007 			     struct perf_event *event, unsigned int size,
1008 			     int nmi, int sample);
1009 extern void perf_output_end(struct perf_output_handle *handle);
1010 extern void perf_output_copy(struct perf_output_handle *handle,
1011 			     const void *buf, unsigned int len);
1012 extern int perf_swevent_get_recursion_context(void);
1013 extern void perf_swevent_put_recursion_context(int rctx);
1014 extern void perf_event_enable(struct perf_event *event);
1015 extern void perf_event_disable(struct perf_event *event);
1016 #else
1017 static inline void
1018 perf_event_task_sched_in(struct task_struct *task)			{ }
1019 static inline void
1020 perf_event_task_sched_out(struct task_struct *task,
1021 			    struct task_struct *next)			{ }
1022 static inline void
1023 perf_event_task_tick(struct task_struct *task)				{ }
1024 static inline int perf_event_init_task(struct task_struct *child)	{ return 0; }
1025 static inline void perf_event_exit_task(struct task_struct *child)	{ }
1026 static inline void perf_event_free_task(struct task_struct *task)	{ }
1027 static inline void perf_event_do_pending(void)				{ }
1028 static inline void perf_event_print_debug(void)				{ }
1029 static inline void perf_disable(void)					{ }
1030 static inline void perf_enable(void)					{ }
1031 static inline int perf_event_task_disable(void)				{ return -EINVAL; }
1032 static inline int perf_event_task_enable(void)				{ return -EINVAL; }
1033 
1034 static inline void
1035 perf_sw_event(u32 event_id, u64 nr, int nmi,
1036 		     struct pt_regs *regs, u64 addr)			{ }
1037 static inline void
1038 perf_bp_event(struct perf_event *event, void *data)			{ }
1039 
1040 static inline int perf_register_guest_info_callbacks
1041 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1042 static inline int perf_unregister_guest_info_callbacks
1043 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1044 
1045 static inline void perf_event_mmap(struct vm_area_struct *vma)		{ }
1046 static inline void perf_event_comm(struct task_struct *tsk)		{ }
1047 static inline void perf_event_fork(struct task_struct *tsk)		{ }
1048 static inline void perf_event_init(void)				{ }
1049 static inline int  perf_swevent_get_recursion_context(void)		{ return -1; }
1050 static inline void perf_swevent_put_recursion_context(int rctx)		{ }
1051 static inline void perf_event_enable(struct perf_event *event)		{ }
1052 static inline void perf_event_disable(struct perf_event *event)		{ }
1053 #endif
1054 
1055 #define perf_output_put(handle, x) \
1056 	perf_output_copy((handle), &(x), sizeof(x))
1057 
1058 /*
1059  * This has to have a higher priority than migration_notifier in sched.c.
1060  */
1061 #define perf_cpu_notifier(fn)					\
1062 do {								\
1063 	static struct notifier_block fn##_nb __cpuinitdata =	\
1064 		{ .notifier_call = fn, .priority = 20 };	\
1065 	fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,		\
1066 		(void *)(unsigned long)smp_processor_id());	\
1067 	fn(&fn##_nb, (unsigned long)CPU_STARTING,		\
1068 		(void *)(unsigned long)smp_processor_id());	\
1069 	fn(&fn##_nb, (unsigned long)CPU_ONLINE,			\
1070 		(void *)(unsigned long)smp_processor_id());	\
1071 	register_cpu_notifier(&fn##_nb);			\
1072 } while (0)
1073 
1074 #endif /* __KERNEL__ */
1075 #endif /* _LINUX_PERF_EVENT_H */
1076