xref: /linux-6.15/include/linux/perf_event.h (revision 3e4cd073)
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 <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 	PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	= 7,
56 	PERF_COUNT_HW_STALLED_CYCLES_BACKEND	= 8,
57 
58 	PERF_COUNT_HW_MAX,			/* non-ABI */
59 };
60 
61 /*
62  * Generalized hardware cache events:
63  *
64  *       { L1-D, L1-I, LLC, ITLB, DTLB, BPU } x
65  *       { read, write, prefetch } x
66  *       { accesses, misses }
67  */
68 enum perf_hw_cache_id {
69 	PERF_COUNT_HW_CACHE_L1D			= 0,
70 	PERF_COUNT_HW_CACHE_L1I			= 1,
71 	PERF_COUNT_HW_CACHE_LL			= 2,
72 	PERF_COUNT_HW_CACHE_DTLB		= 3,
73 	PERF_COUNT_HW_CACHE_ITLB		= 4,
74 	PERF_COUNT_HW_CACHE_BPU			= 5,
75 
76 	PERF_COUNT_HW_CACHE_MAX,		/* non-ABI */
77 };
78 
79 enum perf_hw_cache_op_id {
80 	PERF_COUNT_HW_CACHE_OP_READ		= 0,
81 	PERF_COUNT_HW_CACHE_OP_WRITE		= 1,
82 	PERF_COUNT_HW_CACHE_OP_PREFETCH		= 2,
83 
84 	PERF_COUNT_HW_CACHE_OP_MAX,		/* non-ABI */
85 };
86 
87 enum perf_hw_cache_op_result_id {
88 	PERF_COUNT_HW_CACHE_RESULT_ACCESS	= 0,
89 	PERF_COUNT_HW_CACHE_RESULT_MISS		= 1,
90 
91 	PERF_COUNT_HW_CACHE_RESULT_MAX,		/* non-ABI */
92 };
93 
94 /*
95  * Special "software" events provided by the kernel, even if the hardware
96  * does not support performance events. These events measure various
97  * physical and sw events of the kernel (and allow the profiling of them as
98  * well):
99  */
100 enum perf_sw_ids {
101 	PERF_COUNT_SW_CPU_CLOCK			= 0,
102 	PERF_COUNT_SW_TASK_CLOCK		= 1,
103 	PERF_COUNT_SW_PAGE_FAULTS		= 2,
104 	PERF_COUNT_SW_CONTEXT_SWITCHES		= 3,
105 	PERF_COUNT_SW_CPU_MIGRATIONS		= 4,
106 	PERF_COUNT_SW_PAGE_FAULTS_MIN		= 5,
107 	PERF_COUNT_SW_PAGE_FAULTS_MAJ		= 6,
108 	PERF_COUNT_SW_ALIGNMENT_FAULTS		= 7,
109 	PERF_COUNT_SW_EMULATION_FAULTS		= 8,
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 				 * precise_ip:
210 				 *
211 				 *  0 - SAMPLE_IP can have arbitrary skid
212 				 *  1 - SAMPLE_IP must have constant skid
213 				 *  2 - SAMPLE_IP requested to have 0 skid
214 				 *  3 - SAMPLE_IP must have 0 skid
215 				 *
216 				 *  See also PERF_RECORD_MISC_EXACT_IP
217 				 */
218 				precise_ip     :  2, /* skid constraint       */
219 				mmap_data      :  1, /* non-exec mmap data    */
220 				sample_id_all  :  1, /* sample_type all events */
221 
222 				__reserved_1   : 45;
223 
224 	union {
225 		__u32		wakeup_events;	  /* wakeup every n events */
226 		__u32		wakeup_watermark; /* bytes before wakeup   */
227 	};
228 
229 	__u32			bp_type;
230 	union {
231 		__u64		bp_addr;
232 		__u64		config1; /* extension of config */
233 	};
234 	union {
235 		__u64		bp_len;
236 		__u64		config2; /* extension of config1 */
237 	};
238 };
239 
240 /*
241  * Ioctls that can be done on a perf event fd:
242  */
243 #define PERF_EVENT_IOC_ENABLE		_IO ('$', 0)
244 #define PERF_EVENT_IOC_DISABLE		_IO ('$', 1)
245 #define PERF_EVENT_IOC_REFRESH		_IO ('$', 2)
246 #define PERF_EVENT_IOC_RESET		_IO ('$', 3)
247 #define PERF_EVENT_IOC_PERIOD		_IOW('$', 4, __u64)
248 #define PERF_EVENT_IOC_SET_OUTPUT	_IO ('$', 5)
249 #define PERF_EVENT_IOC_SET_FILTER	_IOW('$', 6, char *)
250 
251 enum perf_event_ioc_flags {
252 	PERF_IOC_FLAG_GROUP		= 1U << 0,
253 };
254 
255 /*
256  * Structure of the page that can be mapped via mmap
257  */
258 struct perf_event_mmap_page {
259 	__u32	version;		/* version number of this structure */
260 	__u32	compat_version;		/* lowest version this is compat with */
261 
262 	/*
263 	 * Bits needed to read the hw events in user-space.
264 	 *
265 	 *   u32 seq;
266 	 *   s64 count;
267 	 *
268 	 *   do {
269 	 *     seq = pc->lock;
270 	 *
271 	 *     barrier()
272 	 *     if (pc->index) {
273 	 *       count = pmc_read(pc->index - 1);
274 	 *       count += pc->offset;
275 	 *     } else
276 	 *       goto regular_read;
277 	 *
278 	 *     barrier();
279 	 *   } while (pc->lock != seq);
280 	 *
281 	 * NOTE: for obvious reason this only works on self-monitoring
282 	 *       processes.
283 	 */
284 	__u32	lock;			/* seqlock for synchronization */
285 	__u32	index;			/* hardware event identifier */
286 	__s64	offset;			/* add to hardware event value */
287 	__u64	time_enabled;		/* time event active */
288 	__u64	time_running;		/* time event on cpu */
289 
290 		/*
291 		 * Hole for extension of the self monitor capabilities
292 		 */
293 
294 	__u64	__reserved[123];	/* align to 1k */
295 
296 	/*
297 	 * Control data for the mmap() data buffer.
298 	 *
299 	 * User-space reading the @data_head value should issue an rmb(), on
300 	 * SMP capable platforms, after reading this value -- see
301 	 * perf_event_wakeup().
302 	 *
303 	 * When the mapping is PROT_WRITE the @data_tail value should be
304 	 * written by userspace to reflect the last read data. In this case
305 	 * the kernel will not over-write unread data.
306 	 */
307 	__u64   data_head;		/* head in the data section */
308 	__u64	data_tail;		/* user-space written tail */
309 };
310 
311 #define PERF_RECORD_MISC_CPUMODE_MASK		(7 << 0)
312 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN	(0 << 0)
313 #define PERF_RECORD_MISC_KERNEL			(1 << 0)
314 #define PERF_RECORD_MISC_USER			(2 << 0)
315 #define PERF_RECORD_MISC_HYPERVISOR		(3 << 0)
316 #define PERF_RECORD_MISC_GUEST_KERNEL		(4 << 0)
317 #define PERF_RECORD_MISC_GUEST_USER		(5 << 0)
318 
319 /*
320  * Indicates that the content of PERF_SAMPLE_IP points to
321  * the actual instruction that triggered the event. See also
322  * perf_event_attr::precise_ip.
323  */
324 #define PERF_RECORD_MISC_EXACT_IP		(1 << 14)
325 /*
326  * Reserve the last bit to indicate some extended misc field
327  */
328 #define PERF_RECORD_MISC_EXT_RESERVED		(1 << 15)
329 
330 struct perf_event_header {
331 	__u32	type;
332 	__u16	misc;
333 	__u16	size;
334 };
335 
336 enum perf_event_type {
337 
338 	/*
339 	 * If perf_event_attr.sample_id_all is set then all event types will
340 	 * have the sample_type selected fields related to where/when
341 	 * (identity) an event took place (TID, TIME, ID, CPU, STREAM_ID)
342 	 * described in PERF_RECORD_SAMPLE below, it will be stashed just after
343 	 * the perf_event_header and the fields already present for the existing
344 	 * fields, i.e. at the end of the payload. That way a newer perf.data
345 	 * file will be supported by older perf tools, with these new optional
346 	 * fields being ignored.
347 	 *
348 	 * The MMAP events record the PROT_EXEC mappings so that we can
349 	 * correlate userspace IPs to code. They have the following structure:
350 	 *
351 	 * struct {
352 	 *	struct perf_event_header	header;
353 	 *
354 	 *	u32				pid, tid;
355 	 *	u64				addr;
356 	 *	u64				len;
357 	 *	u64				pgoff;
358 	 *	char				filename[];
359 	 * };
360 	 */
361 	PERF_RECORD_MMAP			= 1,
362 
363 	/*
364 	 * struct {
365 	 *	struct perf_event_header	header;
366 	 *	u64				id;
367 	 *	u64				lost;
368 	 * };
369 	 */
370 	PERF_RECORD_LOST			= 2,
371 
372 	/*
373 	 * struct {
374 	 *	struct perf_event_header	header;
375 	 *
376 	 *	u32				pid, tid;
377 	 *	char				comm[];
378 	 * };
379 	 */
380 	PERF_RECORD_COMM			= 3,
381 
382 	/*
383 	 * struct {
384 	 *	struct perf_event_header	header;
385 	 *	u32				pid, ppid;
386 	 *	u32				tid, ptid;
387 	 *	u64				time;
388 	 * };
389 	 */
390 	PERF_RECORD_EXIT			= 4,
391 
392 	/*
393 	 * struct {
394 	 *	struct perf_event_header	header;
395 	 *	u64				time;
396 	 *	u64				id;
397 	 *	u64				stream_id;
398 	 * };
399 	 */
400 	PERF_RECORD_THROTTLE			= 5,
401 	PERF_RECORD_UNTHROTTLE			= 6,
402 
403 	/*
404 	 * struct {
405 	 *	struct perf_event_header	header;
406 	 *	u32				pid, ppid;
407 	 *	u32				tid, ptid;
408 	 *	u64				time;
409 	 * };
410 	 */
411 	PERF_RECORD_FORK			= 7,
412 
413 	/*
414 	 * struct {
415 	 *	struct perf_event_header	header;
416 	 *	u32				pid, tid;
417 	 *
418 	 *	struct read_format		values;
419 	 * };
420 	 */
421 	PERF_RECORD_READ			= 8,
422 
423 	/*
424 	 * struct {
425 	 *	struct perf_event_header	header;
426 	 *
427 	 *	{ u64			ip;	  } && PERF_SAMPLE_IP
428 	 *	{ u32			pid, tid; } && PERF_SAMPLE_TID
429 	 *	{ u64			time;     } && PERF_SAMPLE_TIME
430 	 *	{ u64			addr;     } && PERF_SAMPLE_ADDR
431 	 *	{ u64			id;	  } && PERF_SAMPLE_ID
432 	 *	{ u64			stream_id;} && PERF_SAMPLE_STREAM_ID
433 	 *	{ u32			cpu, res; } && PERF_SAMPLE_CPU
434 	 *	{ u64			period;   } && PERF_SAMPLE_PERIOD
435 	 *
436 	 *	{ struct read_format	values;	  } && PERF_SAMPLE_READ
437 	 *
438 	 *	{ u64			nr,
439 	 *	  u64			ips[nr];  } && PERF_SAMPLE_CALLCHAIN
440 	 *
441 	 *	#
442 	 *	# The RAW record below is opaque data wrt the ABI
443 	 *	#
444 	 *	# That is, the ABI doesn't make any promises wrt to
445 	 *	# the stability of its content, it may vary depending
446 	 *	# on event, hardware, kernel version and phase of
447 	 *	# the moon.
448 	 *	#
449 	 *	# In other words, PERF_SAMPLE_RAW contents are not an ABI.
450 	 *	#
451 	 *
452 	 *	{ u32			size;
453 	 *	  char                  data[size];}&& PERF_SAMPLE_RAW
454 	 * };
455 	 */
456 	PERF_RECORD_SAMPLE			= 9,
457 
458 	PERF_RECORD_MAX,			/* non-ABI */
459 };
460 
461 enum perf_callchain_context {
462 	PERF_CONTEXT_HV			= (__u64)-32,
463 	PERF_CONTEXT_KERNEL		= (__u64)-128,
464 	PERF_CONTEXT_USER		= (__u64)-512,
465 
466 	PERF_CONTEXT_GUEST		= (__u64)-2048,
467 	PERF_CONTEXT_GUEST_KERNEL	= (__u64)-2176,
468 	PERF_CONTEXT_GUEST_USER		= (__u64)-2560,
469 
470 	PERF_CONTEXT_MAX		= (__u64)-4095,
471 };
472 
473 #define PERF_FLAG_FD_NO_GROUP		(1U << 0)
474 #define PERF_FLAG_FD_OUTPUT		(1U << 1)
475 #define PERF_FLAG_PID_CGROUP		(1U << 2) /* pid=cgroup id, per-cpu mode only */
476 
477 #ifdef __KERNEL__
478 /*
479  * Kernel-internal data types and definitions:
480  */
481 
482 #ifdef CONFIG_PERF_EVENTS
483 # include <linux/cgroup.h>
484 # include <asm/perf_event.h>
485 # include <asm/local64.h>
486 #endif
487 
488 struct perf_guest_info_callbacks {
489 	int				(*is_in_guest)(void);
490 	int				(*is_user_mode)(void);
491 	unsigned long			(*get_guest_ip)(void);
492 };
493 
494 #ifdef CONFIG_HAVE_HW_BREAKPOINT
495 #include <asm/hw_breakpoint.h>
496 #endif
497 
498 #include <linux/list.h>
499 #include <linux/mutex.h>
500 #include <linux/rculist.h>
501 #include <linux/rcupdate.h>
502 #include <linux/spinlock.h>
503 #include <linux/hrtimer.h>
504 #include <linux/fs.h>
505 #include <linux/pid_namespace.h>
506 #include <linux/workqueue.h>
507 #include <linux/ftrace.h>
508 #include <linux/cpu.h>
509 #include <linux/irq_work.h>
510 #include <linux/jump_label.h>
511 #include <asm/atomic.h>
512 #include <asm/local.h>
513 
514 #define PERF_MAX_STACK_DEPTH		255
515 
516 struct perf_callchain_entry {
517 	__u64				nr;
518 	__u64				ip[PERF_MAX_STACK_DEPTH];
519 };
520 
521 struct perf_raw_record {
522 	u32				size;
523 	void				*data;
524 };
525 
526 struct perf_branch_entry {
527 	__u64				from;
528 	__u64				to;
529 	__u64				flags;
530 };
531 
532 struct perf_branch_stack {
533 	__u64				nr;
534 	struct perf_branch_entry	entries[0];
535 };
536 
537 struct task_struct;
538 
539 /**
540  * struct hw_perf_event - performance event hardware details:
541  */
542 struct hw_perf_event {
543 #ifdef CONFIG_PERF_EVENTS
544 	union {
545 		struct { /* hardware */
546 			u64		config;
547 			u64		last_tag;
548 			unsigned long	config_base;
549 			unsigned long	event_base;
550 			int		idx;
551 			int		last_cpu;
552 			unsigned int	extra_reg;
553 			u64		extra_config;
554 			int		extra_alloc;
555 		};
556 		struct { /* software */
557 			struct hrtimer	hrtimer;
558 		};
559 #ifdef CONFIG_HAVE_HW_BREAKPOINT
560 		struct { /* breakpoint */
561 			struct arch_hw_breakpoint	info;
562 			struct list_head		bp_list;
563 			/*
564 			 * Crufty hack to avoid the chicken and egg
565 			 * problem hw_breakpoint has with context
566 			 * creation and event initalization.
567 			 */
568 			struct task_struct		*bp_target;
569 		};
570 #endif
571 	};
572 	int				state;
573 	local64_t			prev_count;
574 	u64				sample_period;
575 	u64				last_period;
576 	local64_t			period_left;
577 	u64				interrupts;
578 
579 	u64				freq_time_stamp;
580 	u64				freq_count_stamp;
581 #endif
582 };
583 
584 /*
585  * hw_perf_event::state flags
586  */
587 #define PERF_HES_STOPPED	0x01 /* the counter is stopped */
588 #define PERF_HES_UPTODATE	0x02 /* event->count up-to-date */
589 #define PERF_HES_ARCH		0x04
590 
591 struct perf_event;
592 
593 /*
594  * Common implementation detail of pmu::{start,commit,cancel}_txn
595  */
596 #define PERF_EVENT_TXN 0x1
597 
598 /**
599  * struct pmu - generic performance monitoring unit
600  */
601 struct pmu {
602 	struct list_head		entry;
603 
604 	struct device			*dev;
605 	char				*name;
606 	int				type;
607 
608 	int * __percpu			pmu_disable_count;
609 	struct perf_cpu_context * __percpu pmu_cpu_context;
610 	int				task_ctx_nr;
611 
612 	/*
613 	 * Fully disable/enable this PMU, can be used to protect from the PMI
614 	 * as well as for lazy/batch writing of the MSRs.
615 	 */
616 	void (*pmu_enable)		(struct pmu *pmu); /* optional */
617 	void (*pmu_disable)		(struct pmu *pmu); /* optional */
618 
619 	/*
620 	 * Try and initialize the event for this PMU.
621 	 * Should return -ENOENT when the @event doesn't match this PMU.
622 	 */
623 	int (*event_init)		(struct perf_event *event);
624 
625 #define PERF_EF_START	0x01		/* start the counter when adding    */
626 #define PERF_EF_RELOAD	0x02		/* reload the counter when starting */
627 #define PERF_EF_UPDATE	0x04		/* update the counter when stopping */
628 
629 	/*
630 	 * Adds/Removes a counter to/from the PMU, can be done inside
631 	 * a transaction, see the ->*_txn() methods.
632 	 */
633 	int  (*add)			(struct perf_event *event, int flags);
634 	void (*del)			(struct perf_event *event, int flags);
635 
636 	/*
637 	 * Starts/Stops a counter present on the PMU. The PMI handler
638 	 * should stop the counter when perf_event_overflow() returns
639 	 * !0. ->start() will be used to continue.
640 	 */
641 	void (*start)			(struct perf_event *event, int flags);
642 	void (*stop)			(struct perf_event *event, int flags);
643 
644 	/*
645 	 * Updates the counter value of the event.
646 	 */
647 	void (*read)			(struct perf_event *event);
648 
649 	/*
650 	 * Group events scheduling is treated as a transaction, add
651 	 * group events as a whole and perform one schedulability test.
652 	 * If the test fails, roll back the whole group
653 	 *
654 	 * Start the transaction, after this ->add() doesn't need to
655 	 * do schedulability tests.
656 	 */
657 	void (*start_txn)		(struct pmu *pmu); /* optional */
658 	/*
659 	 * If ->start_txn() disabled the ->add() schedulability test
660 	 * then ->commit_txn() is required to perform one. On success
661 	 * the transaction is closed. On error the transaction is kept
662 	 * open until ->cancel_txn() is called.
663 	 */
664 	int  (*commit_txn)		(struct pmu *pmu); /* optional */
665 	/*
666 	 * Will cancel the transaction, assumes ->del() is called
667 	 * for each successful ->add() during the transaction.
668 	 */
669 	void (*cancel_txn)		(struct pmu *pmu); /* optional */
670 };
671 
672 /**
673  * enum perf_event_active_state - the states of a event
674  */
675 enum perf_event_active_state {
676 	PERF_EVENT_STATE_ERROR		= -2,
677 	PERF_EVENT_STATE_OFF		= -1,
678 	PERF_EVENT_STATE_INACTIVE	=  0,
679 	PERF_EVENT_STATE_ACTIVE		=  1,
680 };
681 
682 struct file;
683 
684 #define PERF_BUFFER_WRITABLE		0x01
685 
686 struct perf_buffer {
687 	atomic_t			refcount;
688 	struct rcu_head			rcu_head;
689 #ifdef CONFIG_PERF_USE_VMALLOC
690 	struct work_struct		work;
691 	int				page_order;	/* allocation order  */
692 #endif
693 	int				nr_pages;	/* nr of data pages  */
694 	int				writable;	/* are we writable   */
695 
696 	atomic_t			poll;		/* POLL_ for wakeups */
697 
698 	local_t				head;		/* write position    */
699 	local_t				nest;		/* nested writers    */
700 	local_t				events;		/* event limit       */
701 	local_t				wakeup;		/* wakeup stamp      */
702 	local_t				lost;		/* nr records lost   */
703 
704 	long				watermark;	/* wakeup watermark  */
705 
706 	struct perf_event_mmap_page	*user_page;
707 	void				*data_pages[0];
708 };
709 
710 struct perf_sample_data;
711 
712 typedef void (*perf_overflow_handler_t)(struct perf_event *, int,
713 					struct perf_sample_data *,
714 					struct pt_regs *regs);
715 
716 enum perf_group_flag {
717 	PERF_GROUP_SOFTWARE		= 0x1,
718 };
719 
720 #define SWEVENT_HLIST_BITS		8
721 #define SWEVENT_HLIST_SIZE		(1 << SWEVENT_HLIST_BITS)
722 
723 struct swevent_hlist {
724 	struct hlist_head		heads[SWEVENT_HLIST_SIZE];
725 	struct rcu_head			rcu_head;
726 };
727 
728 #define PERF_ATTACH_CONTEXT	0x01
729 #define PERF_ATTACH_GROUP	0x02
730 #define PERF_ATTACH_TASK	0x04
731 
732 #ifdef CONFIG_CGROUP_PERF
733 /*
734  * perf_cgroup_info keeps track of time_enabled for a cgroup.
735  * This is a per-cpu dynamically allocated data structure.
736  */
737 struct perf_cgroup_info {
738 	u64				time;
739 	u64				timestamp;
740 };
741 
742 struct perf_cgroup {
743 	struct				cgroup_subsys_state css;
744 	struct				perf_cgroup_info *info;	/* timing info, one per cpu */
745 };
746 #endif
747 
748 /**
749  * struct perf_event - performance event kernel representation:
750  */
751 struct perf_event {
752 #ifdef CONFIG_PERF_EVENTS
753 	struct list_head		group_entry;
754 	struct list_head		event_entry;
755 	struct list_head		sibling_list;
756 	struct hlist_node		hlist_entry;
757 	int				nr_siblings;
758 	int				group_flags;
759 	struct perf_event		*group_leader;
760 	struct pmu			*pmu;
761 
762 	enum perf_event_active_state	state;
763 	unsigned int			attach_state;
764 	local64_t			count;
765 	atomic64_t			child_count;
766 
767 	/*
768 	 * These are the total time in nanoseconds that the event
769 	 * has been enabled (i.e. eligible to run, and the task has
770 	 * been scheduled in, if this is a per-task event)
771 	 * and running (scheduled onto the CPU), respectively.
772 	 *
773 	 * They are computed from tstamp_enabled, tstamp_running and
774 	 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
775 	 */
776 	u64				total_time_enabled;
777 	u64				total_time_running;
778 
779 	/*
780 	 * These are timestamps used for computing total_time_enabled
781 	 * and total_time_running when the event is in INACTIVE or
782 	 * ACTIVE state, measured in nanoseconds from an arbitrary point
783 	 * in time.
784 	 * tstamp_enabled: the notional time when the event was enabled
785 	 * tstamp_running: the notional time when the event was scheduled on
786 	 * tstamp_stopped: in INACTIVE state, the notional time when the
787 	 *	event was scheduled off.
788 	 */
789 	u64				tstamp_enabled;
790 	u64				tstamp_running;
791 	u64				tstamp_stopped;
792 
793 	/*
794 	 * timestamp shadows the actual context timing but it can
795 	 * be safely used in NMI interrupt context. It reflects the
796 	 * context time as it was when the event was last scheduled in.
797 	 *
798 	 * ctx_time already accounts for ctx->timestamp. Therefore to
799 	 * compute ctx_time for a sample, simply add perf_clock().
800 	 */
801 	u64				shadow_ctx_time;
802 
803 	struct perf_event_attr		attr;
804 	u16				header_size;
805 	u16				id_header_size;
806 	u16				read_size;
807 	struct hw_perf_event		hw;
808 
809 	struct perf_event_context	*ctx;
810 	struct file			*filp;
811 
812 	/*
813 	 * These accumulate total time (in nanoseconds) that children
814 	 * events have been enabled and running, respectively.
815 	 */
816 	atomic64_t			child_total_time_enabled;
817 	atomic64_t			child_total_time_running;
818 
819 	/*
820 	 * Protect attach/detach and child_list:
821 	 */
822 	struct mutex			child_mutex;
823 	struct list_head		child_list;
824 	struct perf_event		*parent;
825 
826 	int				oncpu;
827 	int				cpu;
828 
829 	struct list_head		owner_entry;
830 	struct task_struct		*owner;
831 
832 	/* mmap bits */
833 	struct mutex			mmap_mutex;
834 	atomic_t			mmap_count;
835 	int				mmap_locked;
836 	struct user_struct		*mmap_user;
837 	struct perf_buffer		*buffer;
838 
839 	/* poll related */
840 	wait_queue_head_t		waitq;
841 	struct fasync_struct		*fasync;
842 
843 	/* delayed work for NMIs and such */
844 	int				pending_wakeup;
845 	int				pending_kill;
846 	int				pending_disable;
847 	struct irq_work			pending;
848 
849 	atomic_t			event_limit;
850 
851 	void (*destroy)(struct perf_event *);
852 	struct rcu_head			rcu_head;
853 
854 	struct pid_namespace		*ns;
855 	u64				id;
856 
857 	perf_overflow_handler_t		overflow_handler;
858 
859 #ifdef CONFIG_EVENT_TRACING
860 	struct ftrace_event_call	*tp_event;
861 	struct event_filter		*filter;
862 #endif
863 
864 #ifdef CONFIG_CGROUP_PERF
865 	struct perf_cgroup		*cgrp; /* cgroup event is attach to */
866 	int				cgrp_defer_enabled;
867 #endif
868 
869 #endif /* CONFIG_PERF_EVENTS */
870 };
871 
872 enum perf_event_context_type {
873 	task_context,
874 	cpu_context,
875 };
876 
877 /**
878  * struct perf_event_context - event context structure
879  *
880  * Used as a container for task events and CPU events as well:
881  */
882 struct perf_event_context {
883 	struct pmu			*pmu;
884 	enum perf_event_context_type	type;
885 	/*
886 	 * Protect the states of the events in the list,
887 	 * nr_active, and the list:
888 	 */
889 	raw_spinlock_t			lock;
890 	/*
891 	 * Protect the list of events.  Locking either mutex or lock
892 	 * is sufficient to ensure the list doesn't change; to change
893 	 * the list you need to lock both the mutex and the spinlock.
894 	 */
895 	struct mutex			mutex;
896 
897 	struct list_head		pinned_groups;
898 	struct list_head		flexible_groups;
899 	struct list_head		event_list;
900 	int				nr_events;
901 	int				nr_active;
902 	int				is_active;
903 	int				nr_stat;
904 	int				rotate_disable;
905 	atomic_t			refcount;
906 	struct task_struct		*task;
907 
908 	/*
909 	 * Context clock, runs when context enabled.
910 	 */
911 	u64				time;
912 	u64				timestamp;
913 
914 	/*
915 	 * These fields let us detect when two contexts have both
916 	 * been cloned (inherited) from a common ancestor.
917 	 */
918 	struct perf_event_context	*parent_ctx;
919 	u64				parent_gen;
920 	u64				generation;
921 	int				pin_count;
922 	struct rcu_head			rcu_head;
923 	int				nr_cgroups; /* cgroup events present */
924 };
925 
926 /*
927  * Number of contexts where an event can trigger:
928  *	task, softirq, hardirq, nmi.
929  */
930 #define PERF_NR_CONTEXTS	4
931 
932 /**
933  * struct perf_event_cpu_context - per cpu event context structure
934  */
935 struct perf_cpu_context {
936 	struct perf_event_context	ctx;
937 	struct perf_event_context	*task_ctx;
938 	int				active_oncpu;
939 	int				exclusive;
940 	struct list_head		rotation_list;
941 	int				jiffies_interval;
942 	struct pmu			*active_pmu;
943 	struct perf_cgroup		*cgrp;
944 };
945 
946 struct perf_output_handle {
947 	struct perf_event		*event;
948 	struct perf_buffer		*buffer;
949 	unsigned long			wakeup;
950 	unsigned long			size;
951 	void				*addr;
952 	int				page;
953 	int				nmi;
954 	int				sample;
955 };
956 
957 #ifdef CONFIG_PERF_EVENTS
958 
959 extern int perf_pmu_register(struct pmu *pmu, char *name, int type);
960 extern void perf_pmu_unregister(struct pmu *pmu);
961 
962 extern int perf_num_counters(void);
963 extern const char *perf_pmu_name(void);
964 extern void __perf_event_task_sched_in(struct task_struct *task);
965 extern void __perf_event_task_sched_out(struct task_struct *task, struct task_struct *next);
966 extern int perf_event_init_task(struct task_struct *child);
967 extern void perf_event_exit_task(struct task_struct *child);
968 extern void perf_event_free_task(struct task_struct *task);
969 extern void perf_event_delayed_put(struct task_struct *task);
970 extern void perf_event_print_debug(void);
971 extern void perf_pmu_disable(struct pmu *pmu);
972 extern void perf_pmu_enable(struct pmu *pmu);
973 extern int perf_event_task_disable(void);
974 extern int perf_event_task_enable(void);
975 extern void perf_event_update_userpage(struct perf_event *event);
976 extern int perf_event_release_kernel(struct perf_event *event);
977 extern struct perf_event *
978 perf_event_create_kernel_counter(struct perf_event_attr *attr,
979 				int cpu,
980 				struct task_struct *task,
981 				perf_overflow_handler_t callback);
982 extern u64 perf_event_read_value(struct perf_event *event,
983 				 u64 *enabled, u64 *running);
984 
985 struct perf_sample_data {
986 	u64				type;
987 
988 	u64				ip;
989 	struct {
990 		u32	pid;
991 		u32	tid;
992 	}				tid_entry;
993 	u64				time;
994 	u64				addr;
995 	u64				id;
996 	u64				stream_id;
997 	struct {
998 		u32	cpu;
999 		u32	reserved;
1000 	}				cpu_entry;
1001 	u64				period;
1002 	struct perf_callchain_entry	*callchain;
1003 	struct perf_raw_record		*raw;
1004 };
1005 
1006 static inline void perf_sample_data_init(struct perf_sample_data *data, u64 addr)
1007 {
1008 	data->addr = addr;
1009 	data->raw  = NULL;
1010 }
1011 
1012 extern void perf_output_sample(struct perf_output_handle *handle,
1013 			       struct perf_event_header *header,
1014 			       struct perf_sample_data *data,
1015 			       struct perf_event *event);
1016 extern void perf_prepare_sample(struct perf_event_header *header,
1017 				struct perf_sample_data *data,
1018 				struct perf_event *event,
1019 				struct pt_regs *regs);
1020 
1021 extern int perf_event_overflow(struct perf_event *event, int nmi,
1022 				 struct perf_sample_data *data,
1023 				 struct pt_regs *regs);
1024 
1025 static inline bool is_sampling_event(struct perf_event *event)
1026 {
1027 	return event->attr.sample_period != 0;
1028 }
1029 
1030 /*
1031  * Return 1 for a software event, 0 for a hardware event
1032  */
1033 static inline int is_software_event(struct perf_event *event)
1034 {
1035 	return event->pmu->task_ctx_nr == perf_sw_context;
1036 }
1037 
1038 extern struct jump_label_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
1039 
1040 extern void __perf_sw_event(u32, u64, int, struct pt_regs *, u64);
1041 
1042 #ifndef perf_arch_fetch_caller_regs
1043 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
1044 #endif
1045 
1046 /*
1047  * Take a snapshot of the regs. Skip ip and frame pointer to
1048  * the nth caller. We only need a few of the regs:
1049  * - ip for PERF_SAMPLE_IP
1050  * - cs for user_mode() tests
1051  * - bp for callchains
1052  * - eflags, for future purposes, just in case
1053  */
1054 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1055 {
1056 	memset(regs, 0, sizeof(*regs));
1057 
1058 	perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1059 }
1060 
1061 static __always_inline void
1062 perf_sw_event(u32 event_id, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
1063 {
1064 	struct pt_regs hot_regs;
1065 
1066 	if (static_branch(&perf_swevent_enabled[event_id])) {
1067 		if (!regs) {
1068 			perf_fetch_caller_regs(&hot_regs);
1069 			regs = &hot_regs;
1070 		}
1071 		__perf_sw_event(event_id, nr, nmi, regs, addr);
1072 	}
1073 }
1074 
1075 extern struct jump_label_key perf_sched_events;
1076 
1077 static inline void perf_event_task_sched_in(struct task_struct *task)
1078 {
1079 	if (static_branch(&perf_sched_events))
1080 		__perf_event_task_sched_in(task);
1081 }
1082 
1083 static inline void perf_event_task_sched_out(struct task_struct *task, struct task_struct *next)
1084 {
1085 	perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0);
1086 
1087 	__perf_event_task_sched_out(task, next);
1088 }
1089 
1090 extern void perf_event_mmap(struct vm_area_struct *vma);
1091 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1092 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1093 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1094 
1095 extern void perf_event_comm(struct task_struct *tsk);
1096 extern void perf_event_fork(struct task_struct *tsk);
1097 
1098 /* Callchains */
1099 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1100 
1101 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
1102 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
1103 
1104 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
1105 {
1106 	if (entry->nr < PERF_MAX_STACK_DEPTH)
1107 		entry->ip[entry->nr++] = ip;
1108 }
1109 
1110 extern int sysctl_perf_event_paranoid;
1111 extern int sysctl_perf_event_mlock;
1112 extern int sysctl_perf_event_sample_rate;
1113 
1114 extern int perf_proc_update_handler(struct ctl_table *table, int write,
1115 		void __user *buffer, size_t *lenp,
1116 		loff_t *ppos);
1117 
1118 static inline bool perf_paranoid_tracepoint_raw(void)
1119 {
1120 	return sysctl_perf_event_paranoid > -1;
1121 }
1122 
1123 static inline bool perf_paranoid_cpu(void)
1124 {
1125 	return sysctl_perf_event_paranoid > 0;
1126 }
1127 
1128 static inline bool perf_paranoid_kernel(void)
1129 {
1130 	return sysctl_perf_event_paranoid > 1;
1131 }
1132 
1133 extern void perf_event_init(void);
1134 extern void perf_tp_event(u64 addr, u64 count, void *record,
1135 			  int entry_size, struct pt_regs *regs,
1136 			  struct hlist_head *head, int rctx);
1137 extern void perf_bp_event(struct perf_event *event, void *data);
1138 
1139 #ifndef perf_misc_flags
1140 # define perf_misc_flags(regs) \
1141 		(user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1142 # define perf_instruction_pointer(regs)	instruction_pointer(regs)
1143 #endif
1144 
1145 extern int perf_output_begin(struct perf_output_handle *handle,
1146 			     struct perf_event *event, unsigned int size,
1147 			     int nmi, int sample);
1148 extern void perf_output_end(struct perf_output_handle *handle);
1149 extern void perf_output_copy(struct perf_output_handle *handle,
1150 			     const void *buf, unsigned int len);
1151 extern int perf_swevent_get_recursion_context(void);
1152 extern void perf_swevent_put_recursion_context(int rctx);
1153 extern void perf_event_enable(struct perf_event *event);
1154 extern void perf_event_disable(struct perf_event *event);
1155 extern void perf_event_task_tick(void);
1156 #else
1157 static inline void
1158 perf_event_task_sched_in(struct task_struct *task)			{ }
1159 static inline void
1160 perf_event_task_sched_out(struct task_struct *task,
1161 			    struct task_struct *next)			{ }
1162 static inline int perf_event_init_task(struct task_struct *child)	{ return 0; }
1163 static inline void perf_event_exit_task(struct task_struct *child)	{ }
1164 static inline void perf_event_free_task(struct task_struct *task)	{ }
1165 static inline void perf_event_delayed_put(struct task_struct *task)	{ }
1166 static inline void perf_event_print_debug(void)				{ }
1167 static inline int perf_event_task_disable(void)				{ return -EINVAL; }
1168 static inline int perf_event_task_enable(void)				{ return -EINVAL; }
1169 
1170 static inline void
1171 perf_sw_event(u32 event_id, u64 nr, int nmi,
1172 		     struct pt_regs *regs, u64 addr)			{ }
1173 static inline void
1174 perf_bp_event(struct perf_event *event, void *data)			{ }
1175 
1176 static inline int perf_register_guest_info_callbacks
1177 (struct perf_guest_info_callbacks *callbacks)				{ return 0; }
1178 static inline int perf_unregister_guest_info_callbacks
1179 (struct perf_guest_info_callbacks *callbacks)				{ return 0; }
1180 
1181 static inline void perf_event_mmap(struct vm_area_struct *vma)		{ }
1182 static inline void perf_event_comm(struct task_struct *tsk)		{ }
1183 static inline void perf_event_fork(struct task_struct *tsk)		{ }
1184 static inline void perf_event_init(void)				{ }
1185 static inline int  perf_swevent_get_recursion_context(void)		{ return -1; }
1186 static inline void perf_swevent_put_recursion_context(int rctx)		{ }
1187 static inline void perf_event_enable(struct perf_event *event)		{ }
1188 static inline void perf_event_disable(struct perf_event *event)		{ }
1189 static inline void perf_event_task_tick(void)				{ }
1190 #endif
1191 
1192 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1193 
1194 /*
1195  * This has to have a higher priority than migration_notifier in sched.c.
1196  */
1197 #define perf_cpu_notifier(fn)						\
1198 do {									\
1199 	static struct notifier_block fn##_nb __cpuinitdata =		\
1200 		{ .notifier_call = fn, .priority = CPU_PRI_PERF };	\
1201 	fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,			\
1202 		(void *)(unsigned long)smp_processor_id());		\
1203 	fn(&fn##_nb, (unsigned long)CPU_STARTING,			\
1204 		(void *)(unsigned long)smp_processor_id());		\
1205 	fn(&fn##_nb, (unsigned long)CPU_ONLINE,				\
1206 		(void *)(unsigned long)smp_processor_id());		\
1207 	register_cpu_notifier(&fn##_nb);				\
1208 } while (0)
1209 
1210 #endif /* __KERNEL__ */
1211 #endif /* _LINUX_PERF_EVENT_H */
1212