• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Performance events:
3  *
4  *    Copyright (C) 2008-2009, Thomas Gleixner <tglx@linutronix.de>
5  *    Copyright (C) 2008-2011, Red Hat, Inc., Ingo Molnar
6  *    Copyright (C) 2008-2011, Red Hat, Inc., Peter Zijlstra
7  *
8  * Data type definitions, declarations, prototypes.
9  *
10  *    Started by: Thomas Gleixner and Ingo Molnar
11  *
12  * For licencing details see kernel-base/COPYING
13  */
14 #ifndef _LINUX_PERF_EVENT_H
15 #define _LINUX_PERF_EVENT_H
16 
17 #include <uapi/linux/perf_event.h>
18 #include <uapi/linux/bpf_perf_event.h>
19 
20 /*
21  * Kernel-internal data types and definitions:
22  */
23 
24 #ifdef CONFIG_PERF_EVENTS
25 # include <asm/perf_event.h>
26 # include <asm/local64.h>
27 #endif
28 
29 #define PERF_GUEST_ACTIVE	0x01
30 #define PERF_GUEST_USER	0x02
31 
32 struct perf_guest_info_callbacks {
33 	unsigned int			(*state)(void);
34 	unsigned long			(*get_ip)(void);
35 	unsigned int			(*handle_intel_pt_intr)(void);
36 };
37 
38 #ifdef CONFIG_HAVE_HW_BREAKPOINT
39 #include <linux/rhashtable-types.h>
40 #include <asm/hw_breakpoint.h>
41 #endif
42 
43 #include <linux/list.h>
44 #include <linux/mutex.h>
45 #include <linux/rculist.h>
46 #include <linux/rcupdate.h>
47 #include <linux/spinlock.h>
48 #include <linux/hrtimer.h>
49 #include <linux/fs.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/workqueue.h>
52 #include <linux/ftrace.h>
53 #include <linux/cpu.h>
54 #include <linux/irq_work.h>
55 #include <linux/static_key.h>
56 #include <linux/jump_label_ratelimit.h>
57 #include <linux/atomic.h>
58 #include <linux/sysfs.h>
59 #include <linux/perf_regs.h>
60 #include <linux/cgroup.h>
61 #include <linux/refcount.h>
62 #include <linux/security.h>
63 #include <linux/static_call.h>
64 #include <linux/lockdep.h>
65 #include <asm/local.h>
66 
67 struct perf_callchain_entry {
68 	__u64				nr;
69 	__u64				ip[]; /* /proc/sys/kernel/perf_event_max_stack */
70 };
71 
72 struct perf_callchain_entry_ctx {
73 	struct perf_callchain_entry *entry;
74 	u32			    max_stack;
75 	u32			    nr;
76 	short			    contexts;
77 	bool			    contexts_maxed;
78 };
79 
80 typedef unsigned long (*perf_copy_f)(void *dst, const void *src,
81 				     unsigned long off, unsigned long len);
82 
83 struct perf_raw_frag {
84 	union {
85 		struct perf_raw_frag	*next;
86 		unsigned long		pad;
87 	};
88 	perf_copy_f			copy;
89 	void				*data;
90 	u32				size;
91 } __packed;
92 
93 struct perf_raw_record {
94 	struct perf_raw_frag		frag;
95 	u32				size;
96 };
97 
98 /*
99  * branch stack layout:
100  *  nr: number of taken branches stored in entries[]
101  *  hw_idx: The low level index of raw branch records
102  *          for the most recent branch.
103  *          -1ULL means invalid/unknown.
104  *
105  * Note that nr can vary from sample to sample
106  * branches (to, from) are stored from most recent
107  * to least recent, i.e., entries[0] contains the most
108  * recent branch.
109  * The entries[] is an abstraction of raw branch records,
110  * which may not be stored in age order in HW, e.g. Intel LBR.
111  * The hw_idx is to expose the low level index of raw
112  * branch record for the most recent branch aka entries[0].
113  * The hw_idx index is between -1 (unknown) and max depth,
114  * which can be retrieved in /sys/devices/cpu/caps/branches.
115  * For the architectures whose raw branch records are
116  * already stored in age order, the hw_idx should be 0.
117  */
118 struct perf_branch_stack {
119 	__u64				nr;
120 	__u64				hw_idx;
121 	struct perf_branch_entry	entries[];
122 };
123 
124 struct task_struct;
125 
126 /*
127  * extra PMU register associated with an event
128  */
129 struct hw_perf_event_extra {
130 	u64		config;	/* register value */
131 	unsigned int	reg;	/* register address or index */
132 	int		alloc;	/* extra register already allocated */
133 	int		idx;	/* index in shared_regs->regs[] */
134 };
135 
136 /**
137  * hw_perf_event::flag values
138  *
139  * PERF_EVENT_FLAG_ARCH bits are reserved for architecture-specific
140  * usage.
141  */
142 #define PERF_EVENT_FLAG_ARCH			0x000fffff
143 #define PERF_EVENT_FLAG_USER_READ_CNT		0x80000000
144 
145 static_assert((PERF_EVENT_FLAG_USER_READ_CNT & PERF_EVENT_FLAG_ARCH) == 0);
146 
147 /**
148  * struct hw_perf_event - performance event hardware details:
149  */
150 struct hw_perf_event {
151 #ifdef CONFIG_PERF_EVENTS
152 	union {
153 		struct { /* hardware */
154 			u64		config;
155 			u64		last_tag;
156 			unsigned long	config_base;
157 			unsigned long	event_base;
158 			int		event_base_rdpmc;
159 			int		idx;
160 			int		last_cpu;
161 			int		flags;
162 
163 			struct hw_perf_event_extra extra_reg;
164 			struct hw_perf_event_extra branch_reg;
165 		};
166 		struct { /* software */
167 			struct hrtimer	hrtimer;
168 		};
169 		struct { /* tracepoint */
170 			/* for tp_event->class */
171 			struct list_head	tp_list;
172 		};
173 		struct { /* amd_power */
174 			u64	pwr_acc;
175 			u64	ptsc;
176 		};
177 #ifdef CONFIG_HAVE_HW_BREAKPOINT
178 		struct { /* breakpoint */
179 			/*
180 			 * Crufty hack to avoid the chicken and egg
181 			 * problem hw_breakpoint has with context
182 			 * creation and event initalization.
183 			 */
184 			struct arch_hw_breakpoint	info;
185 			struct rhlist_head		bp_list;
186 		};
187 #endif
188 		struct { /* amd_iommu */
189 			u8	iommu_bank;
190 			u8	iommu_cntr;
191 			u16	padding;
192 			u64	conf;
193 			u64	conf1;
194 		};
195 	};
196 	/*
197 	 * If the event is a per task event, this will point to the task in
198 	 * question. See the comment in perf_event_alloc().
199 	 */
200 	struct task_struct		*target;
201 
202 	/*
203 	 * PMU would store hardware filter configuration
204 	 * here.
205 	 */
206 	void				*addr_filters;
207 
208 	/* Last sync'ed generation of filters */
209 	unsigned long			addr_filters_gen;
210 
211 /*
212  * hw_perf_event::state flags; used to track the PERF_EF_* state.
213  */
214 #define PERF_HES_STOPPED	0x01 /* the counter is stopped */
215 #define PERF_HES_UPTODATE	0x02 /* event->count up-to-date */
216 #define PERF_HES_ARCH		0x04
217 
218 	int				state;
219 
220 	/*
221 	 * The last observed hardware counter value, updated with a
222 	 * local64_cmpxchg() such that pmu::read() can be called nested.
223 	 */
224 	local64_t			prev_count;
225 
226 	/*
227 	 * The period to start the next sample with.
228 	 */
229 	u64				sample_period;
230 
231 	union {
232 		struct { /* Sampling */
233 			/*
234 			 * The period we started this sample with.
235 			 */
236 			u64				last_period;
237 
238 			/*
239 			 * However much is left of the current period;
240 			 * note that this is a full 64bit value and
241 			 * allows for generation of periods longer
242 			 * than hardware might allow.
243 			 */
244 			local64_t			period_left;
245 		};
246 		struct { /* Topdown events counting for context switch */
247 			u64				saved_metric;
248 			u64				saved_slots;
249 		};
250 	};
251 
252 	/*
253 	 * State for throttling the event, see __perf_event_overflow() and
254 	 * perf_adjust_freq_unthr_context().
255 	 */
256 	u64                             interrupts_seq;
257 	u64				interrupts;
258 
259 	/*
260 	 * State for freq target events, see __perf_event_overflow() and
261 	 * perf_adjust_freq_unthr_context().
262 	 */
263 	u64				freq_time_stamp;
264 	u64				freq_count_stamp;
265 #endif
266 };
267 
268 struct perf_event;
269 
270 /*
271  * Common implementation detail of pmu::{start,commit,cancel}_txn
272  */
273 #define PERF_PMU_TXN_ADD  0x1		/* txn to add/schedule event on PMU */
274 #define PERF_PMU_TXN_READ 0x2		/* txn to read event group from PMU */
275 
276 /**
277  * pmu::capabilities flags
278  */
279 #define PERF_PMU_CAP_NO_INTERRUPT		0x0001
280 #define PERF_PMU_CAP_NO_NMI			0x0002
281 #define PERF_PMU_CAP_AUX_NO_SG			0x0004
282 #define PERF_PMU_CAP_EXTENDED_REGS		0x0008
283 #define PERF_PMU_CAP_EXCLUSIVE			0x0010
284 #define PERF_PMU_CAP_ITRACE			0x0020
285 #define PERF_PMU_CAP_HETEROGENEOUS_CPUS		0x0040
286 #define PERF_PMU_CAP_NO_EXCLUDE			0x0080
287 #define PERF_PMU_CAP_AUX_OUTPUT			0x0100
288 #define PERF_PMU_CAP_EXTENDED_HW_TYPE		0x0200
289 
290 struct perf_output_handle;
291 
292 /**
293  * struct pmu - generic performance monitoring unit
294  */
295 struct pmu {
296 	struct list_head		entry;
297 
298 	struct module			*module;
299 	struct device			*dev;
300 	const struct attribute_group	**attr_groups;
301 	const struct attribute_group	**attr_update;
302 	const char			*name;
303 	int				type;
304 
305 	/*
306 	 * various common per-pmu feature flags
307 	 */
308 	int				capabilities;
309 
310 	int __percpu			*pmu_disable_count;
311 	struct perf_cpu_context __percpu *pmu_cpu_context;
312 	atomic_t			exclusive_cnt; /* < 0: cpu; > 0: tsk */
313 	int				task_ctx_nr;
314 	int				hrtimer_interval_ms;
315 
316 	/* number of address filters this PMU can do */
317 	unsigned int			nr_addr_filters;
318 
319 	/*
320 	 * Fully disable/enable this PMU, can be used to protect from the PMI
321 	 * as well as for lazy/batch writing of the MSRs.
322 	 */
323 	void (*pmu_enable)		(struct pmu *pmu); /* optional */
324 	void (*pmu_disable)		(struct pmu *pmu); /* optional */
325 
326 	/*
327 	 * Try and initialize the event for this PMU.
328 	 *
329 	 * Returns:
330 	 *  -ENOENT	-- @event is not for this PMU
331 	 *
332 	 *  -ENODEV	-- @event is for this PMU but PMU not present
333 	 *  -EBUSY	-- @event is for this PMU but PMU temporarily unavailable
334 	 *  -EINVAL	-- @event is for this PMU but @event is not valid
335 	 *  -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
336 	 *  -EACCES	-- @event is for this PMU, @event is valid, but no privileges
337 	 *
338 	 *  0		-- @event is for this PMU and valid
339 	 *
340 	 * Other error return values are allowed.
341 	 */
342 	int (*event_init)		(struct perf_event *event);
343 
344 	/*
345 	 * Notification that the event was mapped or unmapped.  Called
346 	 * in the context of the mapping task.
347 	 */
348 	void (*event_mapped)		(struct perf_event *event, struct mm_struct *mm); /* optional */
349 	void (*event_unmapped)		(struct perf_event *event, struct mm_struct *mm); /* optional */
350 
351 	/*
352 	 * Flags for ->add()/->del()/ ->start()/->stop(). There are
353 	 * matching hw_perf_event::state flags.
354 	 */
355 #define PERF_EF_START	0x01		/* start the counter when adding    */
356 #define PERF_EF_RELOAD	0x02		/* reload the counter when starting */
357 #define PERF_EF_UPDATE	0x04		/* update the counter when stopping */
358 
359 	/*
360 	 * Adds/Removes a counter to/from the PMU, can be done inside a
361 	 * transaction, see the ->*_txn() methods.
362 	 *
363 	 * The add/del callbacks will reserve all hardware resources required
364 	 * to service the event, this includes any counter constraint
365 	 * scheduling etc.
366 	 *
367 	 * Called with IRQs disabled and the PMU disabled on the CPU the event
368 	 * is on.
369 	 *
370 	 * ->add() called without PERF_EF_START should result in the same state
371 	 *  as ->add() followed by ->stop().
372 	 *
373 	 * ->del() must always PERF_EF_UPDATE stop an event. If it calls
374 	 *  ->stop() that must deal with already being stopped without
375 	 *  PERF_EF_UPDATE.
376 	 */
377 	int  (*add)			(struct perf_event *event, int flags);
378 	void (*del)			(struct perf_event *event, int flags);
379 
380 	/*
381 	 * Starts/Stops a counter present on the PMU.
382 	 *
383 	 * The PMI handler should stop the counter when perf_event_overflow()
384 	 * returns !0. ->start() will be used to continue.
385 	 *
386 	 * Also used to change the sample period.
387 	 *
388 	 * Called with IRQs disabled and the PMU disabled on the CPU the event
389 	 * is on -- will be called from NMI context with the PMU generates
390 	 * NMIs.
391 	 *
392 	 * ->stop() with PERF_EF_UPDATE will read the counter and update
393 	 *  period/count values like ->read() would.
394 	 *
395 	 * ->start() with PERF_EF_RELOAD will reprogram the counter
396 	 *  value, must be preceded by a ->stop() with PERF_EF_UPDATE.
397 	 */
398 	void (*start)			(struct perf_event *event, int flags);
399 	void (*stop)			(struct perf_event *event, int flags);
400 
401 	/*
402 	 * Updates the counter value of the event.
403 	 *
404 	 * For sampling capable PMUs this will also update the software period
405 	 * hw_perf_event::period_left field.
406 	 */
407 	void (*read)			(struct perf_event *event);
408 
409 	/*
410 	 * Group events scheduling is treated as a transaction, add
411 	 * group events as a whole and perform one schedulability test.
412 	 * If the test fails, roll back the whole group
413 	 *
414 	 * Start the transaction, after this ->add() doesn't need to
415 	 * do schedulability tests.
416 	 *
417 	 * Optional.
418 	 */
419 	void (*start_txn)		(struct pmu *pmu, unsigned int txn_flags);
420 	/*
421 	 * If ->start_txn() disabled the ->add() schedulability test
422 	 * then ->commit_txn() is required to perform one. On success
423 	 * the transaction is closed. On error the transaction is kept
424 	 * open until ->cancel_txn() is called.
425 	 *
426 	 * Optional.
427 	 */
428 	int  (*commit_txn)		(struct pmu *pmu);
429 	/*
430 	 * Will cancel the transaction, assumes ->del() is called
431 	 * for each successful ->add() during the transaction.
432 	 *
433 	 * Optional.
434 	 */
435 	void (*cancel_txn)		(struct pmu *pmu);
436 
437 	/*
438 	 * Will return the value for perf_event_mmap_page::index for this event,
439 	 * if no implementation is provided it will default to: event->hw.idx + 1.
440 	 */
441 	int (*event_idx)		(struct perf_event *event); /*optional */
442 
443 	/*
444 	 * context-switches callback
445 	 */
446 	void (*sched_task)		(struct perf_event_context *ctx,
447 					bool sched_in);
448 
449 	/*
450 	 * Kmem cache of PMU specific data
451 	 */
452 	struct kmem_cache		*task_ctx_cache;
453 
454 	/*
455 	 * PMU specific parts of task perf event context (i.e. ctx->task_ctx_data)
456 	 * can be synchronized using this function. See Intel LBR callstack support
457 	 * implementation and Perf core context switch handling callbacks for usage
458 	 * examples.
459 	 */
460 	void (*swap_task_ctx)		(struct perf_event_context *prev,
461 					 struct perf_event_context *next);
462 					/* optional */
463 
464 	/*
465 	 * Set up pmu-private data structures for an AUX area
466 	 */
467 	void *(*setup_aux)		(struct perf_event *event, void **pages,
468 					 int nr_pages, bool overwrite);
469 					/* optional */
470 
471 	/*
472 	 * Free pmu-private AUX data structures
473 	 */
474 	void (*free_aux)		(void *aux); /* optional */
475 
476 	/*
477 	 * Take a snapshot of the AUX buffer without touching the event
478 	 * state, so that preempting ->start()/->stop() callbacks does
479 	 * not interfere with their logic. Called in PMI context.
480 	 *
481 	 * Returns the size of AUX data copied to the output handle.
482 	 *
483 	 * Optional.
484 	 */
485 	long (*snapshot_aux)		(struct perf_event *event,
486 					 struct perf_output_handle *handle,
487 					 unsigned long size);
488 
489 	/*
490 	 * Validate address range filters: make sure the HW supports the
491 	 * requested configuration and number of filters; return 0 if the
492 	 * supplied filters are valid, -errno otherwise.
493 	 *
494 	 * Runs in the context of the ioctl()ing process and is not serialized
495 	 * with the rest of the PMU callbacks.
496 	 */
497 	int (*addr_filters_validate)	(struct list_head *filters);
498 					/* optional */
499 
500 	/*
501 	 * Synchronize address range filter configuration:
502 	 * translate hw-agnostic filters into hardware configuration in
503 	 * event::hw::addr_filters.
504 	 *
505 	 * Runs as a part of filter sync sequence that is done in ->start()
506 	 * callback by calling perf_event_addr_filters_sync().
507 	 *
508 	 * May (and should) traverse event::addr_filters::list, for which its
509 	 * caller provides necessary serialization.
510 	 */
511 	void (*addr_filters_sync)	(struct perf_event *event);
512 					/* optional */
513 
514 	/*
515 	 * Check if event can be used for aux_output purposes for
516 	 * events of this PMU.
517 	 *
518 	 * Runs from perf_event_open(). Should return 0 for "no match"
519 	 * or non-zero for "match".
520 	 */
521 	int (*aux_output_match)		(struct perf_event *event);
522 					/* optional */
523 
524 	/*
525 	 * Filter events for PMU-specific reasons.
526 	 */
527 	int (*filter_match)		(struct perf_event *event); /* optional */
528 
529 	/*
530 	 * Check period value for PERF_EVENT_IOC_PERIOD ioctl.
531 	 */
532 	int (*check_period)		(struct perf_event *event, u64 value); /* optional */
533 };
534 
535 enum perf_addr_filter_action_t {
536 	PERF_ADDR_FILTER_ACTION_STOP = 0,
537 	PERF_ADDR_FILTER_ACTION_START,
538 	PERF_ADDR_FILTER_ACTION_FILTER,
539 };
540 
541 /**
542  * struct perf_addr_filter - address range filter definition
543  * @entry:	event's filter list linkage
544  * @path:	object file's path for file-based filters
545  * @offset:	filter range offset
546  * @size:	filter range size (size==0 means single address trigger)
547  * @action:	filter/start/stop
548  *
549  * This is a hardware-agnostic filter configuration as specified by the user.
550  */
551 struct perf_addr_filter {
552 	struct list_head	entry;
553 	struct path		path;
554 	unsigned long		offset;
555 	unsigned long		size;
556 	enum perf_addr_filter_action_t	action;
557 };
558 
559 /**
560  * struct perf_addr_filters_head - container for address range filters
561  * @list:	list of filters for this event
562  * @lock:	spinlock that serializes accesses to the @list and event's
563  *		(and its children's) filter generations.
564  * @nr_file_filters:	number of file-based filters
565  *
566  * A child event will use parent's @list (and therefore @lock), so they are
567  * bundled together; see perf_event_addr_filters().
568  */
569 struct perf_addr_filters_head {
570 	struct list_head	list;
571 	raw_spinlock_t		lock;
572 	unsigned int		nr_file_filters;
573 };
574 
575 struct perf_addr_filter_range {
576 	unsigned long		start;
577 	unsigned long		size;
578 };
579 
580 /**
581  * enum perf_event_state - the states of an event:
582  */
583 enum perf_event_state {
584 	PERF_EVENT_STATE_DEAD		= -4,
585 	PERF_EVENT_STATE_EXIT		= -3,
586 	PERF_EVENT_STATE_ERROR		= -2,
587 	PERF_EVENT_STATE_OFF		= -1,
588 	PERF_EVENT_STATE_INACTIVE	=  0,
589 	PERF_EVENT_STATE_ACTIVE		=  1,
590 };
591 
592 struct file;
593 struct perf_sample_data;
594 
595 typedef void (*perf_overflow_handler_t)(struct perf_event *,
596 					struct perf_sample_data *,
597 					struct pt_regs *regs);
598 
599 /*
600  * Event capabilities. For event_caps and groups caps.
601  *
602  * PERF_EV_CAP_SOFTWARE: Is a software event.
603  * PERF_EV_CAP_READ_ACTIVE_PKG: A CPU event (or cgroup event) that can be read
604  * from any CPU in the package where it is active.
605  * PERF_EV_CAP_SIBLING: An event with this flag must be a group sibling and
606  * cannot be a group leader. If an event with this flag is detached from the
607  * group it is scheduled out and moved into an unrecoverable ERROR state.
608  */
609 #define PERF_EV_CAP_SOFTWARE		BIT(0)
610 #define PERF_EV_CAP_READ_ACTIVE_PKG	BIT(1)
611 #define PERF_EV_CAP_SIBLING		BIT(2)
612 
613 #define SWEVENT_HLIST_BITS		8
614 #define SWEVENT_HLIST_SIZE		(1 << SWEVENT_HLIST_BITS)
615 
616 struct swevent_hlist {
617 	struct hlist_head		heads[SWEVENT_HLIST_SIZE];
618 	struct rcu_head			rcu_head;
619 };
620 
621 #define PERF_ATTACH_CONTEXT	0x01
622 #define PERF_ATTACH_GROUP	0x02
623 #define PERF_ATTACH_TASK	0x04
624 #define PERF_ATTACH_TASK_DATA	0x08
625 #define PERF_ATTACH_ITRACE	0x10
626 #define PERF_ATTACH_SCHED_CB	0x20
627 #define PERF_ATTACH_CHILD	0x40
628 
629 struct bpf_prog;
630 struct perf_cgroup;
631 struct perf_buffer;
632 
633 struct pmu_event_list {
634 	raw_spinlock_t		lock;
635 	struct list_head	list;
636 };
637 
638 /*
639  * event->sibling_list is modified whole holding both ctx->lock and ctx->mutex
640  * as such iteration must hold either lock. However, since ctx->lock is an IRQ
641  * safe lock, and is only held by the CPU doing the modification, having IRQs
642  * disabled is sufficient since it will hold-off the IPIs.
643  */
644 #ifdef CONFIG_PROVE_LOCKING
645 #define lockdep_assert_event_ctx(event)				\
646 	WARN_ON_ONCE(__lockdep_enabled &&			\
647 		     (this_cpu_read(hardirqs_enabled) &&	\
648 		      lockdep_is_held(&(event)->ctx->mutex) != LOCK_STATE_HELD))
649 #else
650 #define lockdep_assert_event_ctx(event)
651 #endif
652 
653 #define for_each_sibling_event(sibling, event)			\
654 	lockdep_assert_event_ctx(event);			\
655 	if ((event)->group_leader == (event))			\
656 		list_for_each_entry((sibling), &(event)->sibling_list, sibling_list)
657 
658 /**
659  * struct perf_event - performance event kernel representation:
660  */
661 struct perf_event {
662 #ifdef CONFIG_PERF_EVENTS
663 	/*
664 	 * entry onto perf_event_context::event_list;
665 	 *   modifications require ctx->lock
666 	 *   RCU safe iterations.
667 	 */
668 	struct list_head		event_entry;
669 
670 	/*
671 	 * Locked for modification by both ctx->mutex and ctx->lock; holding
672 	 * either sufficies for read.
673 	 */
674 	struct list_head		sibling_list;
675 	struct list_head		active_list;
676 	/*
677 	 * Node on the pinned or flexible tree located at the event context;
678 	 */
679 	struct rb_node			group_node;
680 	u64				group_index;
681 	/*
682 	 * We need storage to track the entries in perf_pmu_migrate_context; we
683 	 * cannot use the event_entry because of RCU and we want to keep the
684 	 * group in tact which avoids us using the other two entries.
685 	 */
686 	struct list_head		migrate_entry;
687 
688 	struct hlist_node		hlist_entry;
689 	struct list_head		active_entry;
690 	int				nr_siblings;
691 
692 	/* Not serialized. Only written during event initialization. */
693 	int				event_caps;
694 	/* The cumulative AND of all event_caps for events in this group. */
695 	int				group_caps;
696 
697 #ifndef __GENKSYMS__
698 	unsigned int			group_generation;
699 #endif
700 	struct perf_event		*group_leader;
701 	struct pmu			*pmu;
702 	void				*pmu_private;
703 
704 	enum perf_event_state		state;
705 	unsigned int			attach_state;
706 	local64_t			count;
707 	atomic64_t			child_count;
708 
709 	/*
710 	 * These are the total time in nanoseconds that the event
711 	 * has been enabled (i.e. eligible to run, and the task has
712 	 * been scheduled in, if this is a per-task event)
713 	 * and running (scheduled onto the CPU), respectively.
714 	 */
715 	u64				total_time_enabled;
716 	u64				total_time_running;
717 	u64				tstamp;
718 
719 	struct perf_event_attr		attr;
720 	u16				header_size;
721 	u16				id_header_size;
722 	u16				read_size;
723 	struct hw_perf_event		hw;
724 
725 	struct perf_event_context	*ctx;
726 	atomic_long_t			refcount;
727 
728 	/*
729 	 * These accumulate total time (in nanoseconds) that children
730 	 * events have been enabled and running, respectively.
731 	 */
732 	atomic64_t			child_total_time_enabled;
733 	atomic64_t			child_total_time_running;
734 
735 	/*
736 	 * Protect attach/detach and child_list:
737 	 */
738 	struct mutex			child_mutex;
739 	struct list_head		child_list;
740 	struct perf_event		*parent;
741 
742 	int				oncpu;
743 	int				cpu;
744 
745 	struct list_head		owner_entry;
746 	struct task_struct		*owner;
747 
748 	/* mmap bits */
749 	struct mutex			mmap_mutex;
750 	atomic_t			mmap_count;
751 
752 	struct perf_buffer		*rb;
753 	struct list_head		rb_entry;
754 	unsigned long			rcu_batches;
755 	int				rcu_pending;
756 
757 	/* poll related */
758 	wait_queue_head_t		waitq;
759 	struct fasync_struct		*fasync;
760 
761 	/* delayed work for NMIs and such */
762 	unsigned int			pending_wakeup;
763 	unsigned int			pending_kill;
764 	unsigned int			pending_disable;
765 	unsigned int			pending_sigtrap;
766 	unsigned long			pending_addr;	/* SIGTRAP */
767 	struct irq_work			pending_irq;
768 	struct callback_head		pending_task;
769 	unsigned int			pending_work;
770 
771 	atomic_t			event_limit;
772 
773 	/* address range filters */
774 	struct perf_addr_filters_head	addr_filters;
775 	/* vma address array for file-based filders */
776 	struct perf_addr_filter_range	*addr_filter_ranges;
777 	unsigned long			addr_filters_gen;
778 
779 	/* for aux_output events */
780 	struct perf_event		*aux_event;
781 
782 	void (*destroy)(struct perf_event *);
783 	struct rcu_head			rcu_head;
784 
785 	struct pid_namespace		*ns;
786 	u64				id;
787 
788 	atomic64_t			lost_samples;
789 
790 	u64				(*clock)(void);
791 	perf_overflow_handler_t		overflow_handler;
792 	void				*overflow_handler_context;
793 #ifdef CONFIG_BPF_SYSCALL
794 	perf_overflow_handler_t		orig_overflow_handler;
795 	struct bpf_prog			*prog;
796 	u64				bpf_cookie;
797 #endif
798 
799 #ifdef CONFIG_EVENT_TRACING
800 	struct trace_event_call		*tp_event;
801 	struct event_filter		*filter;
802 #ifdef CONFIG_FUNCTION_TRACER
803 	struct ftrace_ops               ftrace_ops;
804 #endif
805 #endif
806 
807 #ifdef CONFIG_CGROUP_PERF
808 	struct perf_cgroup		*cgrp; /* cgroup event is attach to */
809 #endif
810 
811 #ifdef CONFIG_SECURITY
812 	void *security;
813 #endif
814 	struct list_head		sb_list;
815 #endif /* CONFIG_PERF_EVENTS */
816 };
817 
818 
819 struct perf_event_groups {
820 	struct rb_root	tree;
821 	u64		index;
822 };
823 
824 /**
825  * struct perf_event_context - event context structure
826  *
827  * Used as a container for task events and CPU events as well:
828  */
829 struct perf_event_context {
830 	struct pmu			*pmu;
831 	/*
832 	 * Protect the states of the events in the list,
833 	 * nr_active, and the list:
834 	 */
835 	raw_spinlock_t			lock;
836 	/*
837 	 * Protect the list of events.  Locking either mutex or lock
838 	 * is sufficient to ensure the list doesn't change; to change
839 	 * the list you need to lock both the mutex and the spinlock.
840 	 */
841 	struct mutex			mutex;
842 
843 	struct list_head		active_ctx_list;
844 	struct perf_event_groups	pinned_groups;
845 	struct perf_event_groups	flexible_groups;
846 	struct list_head		event_list;
847 
848 	struct list_head		pinned_active;
849 	struct list_head		flexible_active;
850 
851 	int				nr_events;
852 	int				nr_active;
853 	int				nr_user;
854 	int				is_active;
855 	int				nr_stat;
856 	int				nr_freq;
857 	int				rotate_disable;
858 	/*
859 	 * Set when nr_events != nr_active, except tolerant to events not
860 	 * necessary to be active due to scheduling constraints, such as cgroups.
861 	 */
862 	int				rotate_necessary;
863 	refcount_t			refcount;
864 	struct task_struct		*task;
865 
866 	/*
867 	 * Context clock, runs when context enabled.
868 	 */
869 	u64				time;
870 	u64				timestamp;
871 	u64				timeoffset;
872 
873 	/*
874 	 * These fields let us detect when two contexts have both
875 	 * been cloned (inherited) from a common ancestor.
876 	 */
877 	struct perf_event_context	*parent_ctx;
878 	u64				parent_gen;
879 	u64				generation;
880 	int				pin_count;
881 #ifdef CONFIG_CGROUP_PERF
882 	int				nr_cgroups;	 /* cgroup evts */
883 #endif
884 	void				*task_ctx_data; /* pmu specific data */
885 	struct rcu_head			rcu_head;
886 
887 	/*
888 	 * Sum (event->pending_sigtrap + event->pending_work)
889 	 *
890 	 * The SIGTRAP is targeted at ctx->task, as such it won't do changing
891 	 * that until the signal is delivered.
892 	 */
893 	local_t				nr_pending;
894 };
895 
896 /*
897  * Number of contexts where an event can trigger:
898  *	task, softirq, hardirq, nmi.
899  */
900 #define PERF_NR_CONTEXTS	4
901 
902 /**
903  * struct perf_cpu_context - per cpu event context structure
904  */
905 struct perf_cpu_context {
906 	struct perf_event_context	ctx;
907 	struct perf_event_context	*task_ctx;
908 	int				active_oncpu;
909 	int				exclusive;
910 
911 	raw_spinlock_t			hrtimer_lock;
912 	struct hrtimer			hrtimer;
913 	ktime_t				hrtimer_interval;
914 	unsigned int			hrtimer_active;
915 
916 #ifdef CONFIG_CGROUP_PERF
917 	struct perf_cgroup		*cgrp;
918 	struct list_head		cgrp_cpuctx_entry;
919 #endif
920 
921 	struct list_head		sched_cb_entry;
922 	int				sched_cb_usage;
923 
924 	int				online;
925 	/*
926 	 * Per-CPU storage for iterators used in visit_groups_merge. The default
927 	 * storage is of size 2 to hold the CPU and any CPU event iterators.
928 	 */
929 	int				heap_size;
930 	struct perf_event		**heap;
931 	struct perf_event		*heap_default[2];
932 };
933 
934 struct perf_output_handle {
935 	struct perf_event		*event;
936 	struct perf_buffer		*rb;
937 	unsigned long			wakeup;
938 	unsigned long			size;
939 	u64				aux_flags;
940 	union {
941 		void			*addr;
942 		unsigned long		head;
943 	};
944 	int				page;
945 };
946 
947 struct bpf_perf_event_data_kern {
948 	bpf_user_pt_regs_t *regs;
949 	struct perf_sample_data *data;
950 	struct perf_event *event;
951 };
952 
953 #ifdef CONFIG_CGROUP_PERF
954 
955 /*
956  * perf_cgroup_info keeps track of time_enabled for a cgroup.
957  * This is a per-cpu dynamically allocated data structure.
958  */
959 struct perf_cgroup_info {
960 	u64				time;
961 	u64				timestamp;
962 	u64				timeoffset;
963 	int				active;
964 };
965 
966 struct perf_cgroup {
967 	struct cgroup_subsys_state	css;
968 	struct perf_cgroup_info	__percpu *info;
969 };
970 
971 /*
972  * Must ensure cgroup is pinned (css_get) before calling
973  * this function. In other words, we cannot call this function
974  * if there is no cgroup event for the current CPU context.
975  */
976 static inline struct perf_cgroup *
perf_cgroup_from_task(struct task_struct * task,struct perf_event_context * ctx)977 perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
978 {
979 	return container_of(task_css_check(task, perf_event_cgrp_id,
980 					   ctx ? lockdep_is_held(&ctx->lock)
981 					       : true),
982 			    struct perf_cgroup, css);
983 }
984 #endif /* CONFIG_CGROUP_PERF */
985 
986 #ifdef CONFIG_PERF_EVENTS
987 
988 extern void *perf_aux_output_begin(struct perf_output_handle *handle,
989 				   struct perf_event *event);
990 extern void perf_aux_output_end(struct perf_output_handle *handle,
991 				unsigned long size);
992 extern int perf_aux_output_skip(struct perf_output_handle *handle,
993 				unsigned long size);
994 extern void *perf_get_aux(struct perf_output_handle *handle);
995 extern void perf_aux_output_flag(struct perf_output_handle *handle, u64 flags);
996 extern void perf_event_itrace_started(struct perf_event *event);
997 
998 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
999 extern void perf_pmu_unregister(struct pmu *pmu);
1000 
1001 extern void __perf_event_task_sched_in(struct task_struct *prev,
1002 				       struct task_struct *task);
1003 extern void __perf_event_task_sched_out(struct task_struct *prev,
1004 					struct task_struct *next);
1005 extern int perf_event_init_task(struct task_struct *child, u64 clone_flags);
1006 extern void perf_event_exit_task(struct task_struct *child);
1007 extern void perf_event_free_task(struct task_struct *task);
1008 extern void perf_event_delayed_put(struct task_struct *task);
1009 extern struct file *perf_event_get(unsigned int fd);
1010 extern const struct perf_event *perf_get_event(struct file *file);
1011 extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
1012 extern void perf_event_print_debug(void);
1013 extern void perf_pmu_disable(struct pmu *pmu);
1014 extern void perf_pmu_enable(struct pmu *pmu);
1015 extern void perf_sched_cb_dec(struct pmu *pmu);
1016 extern void perf_sched_cb_inc(struct pmu *pmu);
1017 extern int perf_event_task_disable(void);
1018 extern int perf_event_task_enable(void);
1019 
1020 extern void perf_pmu_resched(struct pmu *pmu);
1021 
1022 extern int perf_event_refresh(struct perf_event *event, int refresh);
1023 extern void perf_event_update_userpage(struct perf_event *event);
1024 extern int perf_event_release_kernel(struct perf_event *event);
1025 extern struct perf_event *
1026 perf_event_create_kernel_counter(struct perf_event_attr *attr,
1027 				int cpu,
1028 				struct task_struct *task,
1029 				perf_overflow_handler_t callback,
1030 				void *context);
1031 extern void perf_pmu_migrate_context(struct pmu *pmu,
1032 				int src_cpu, int dst_cpu);
1033 int perf_event_read_local(struct perf_event *event, u64 *value,
1034 			  u64 *enabled, u64 *running);
1035 extern u64 perf_event_read_value(struct perf_event *event,
1036 				 u64 *enabled, u64 *running);
1037 
1038 
1039 struct perf_sample_data {
1040 	/*
1041 	 * Fields set by perf_sample_data_init(), group so as to
1042 	 * minimize the cachelines touched.
1043 	 */
1044 	u64				sample_flags;
1045 	u64				period;
1046 
1047 	/*
1048 	 * The other fields, optionally {set,used} by
1049 	 * perf_{prepare,output}_sample().
1050 	 */
1051 	struct perf_branch_stack	*br_stack;
1052 	union perf_sample_weight	weight;
1053 	union  perf_mem_data_src	data_src;
1054 	u64				txn;
1055 	u64				addr;
1056 	struct perf_raw_record		*raw;
1057 
1058 	u64				type;
1059 	u64				ip;
1060 	struct {
1061 		u32	pid;
1062 		u32	tid;
1063 	}				tid_entry;
1064 	u64				time;
1065 	u64				id;
1066 	u64				stream_id;
1067 	struct {
1068 		u32	cpu;
1069 		u32	reserved;
1070 	}				cpu_entry;
1071 	struct perf_callchain_entry	*callchain;
1072 	u64				aux_size;
1073 
1074 	struct perf_regs		regs_user;
1075 	struct perf_regs		regs_intr;
1076 	u64				stack_user_size;
1077 
1078 	u64				phys_addr;
1079 	u64				cgroup;
1080 	u64				data_page_size;
1081 	u64				code_page_size;
1082 } ____cacheline_aligned;
1083 
1084 /* default value for data source */
1085 #define PERF_MEM_NA (PERF_MEM_S(OP, NA)   |\
1086 		    PERF_MEM_S(LVL, NA)   |\
1087 		    PERF_MEM_S(SNOOP, NA) |\
1088 		    PERF_MEM_S(LOCK, NA)  |\
1089 		    PERF_MEM_S(TLB, NA))
1090 
perf_sample_data_init(struct perf_sample_data * data,u64 addr,u64 period)1091 static inline void perf_sample_data_init(struct perf_sample_data *data,
1092 					 u64 addr, u64 period)
1093 {
1094 	/* remaining struct members initialized in perf_prepare_sample() */
1095 	data->sample_flags = PERF_SAMPLE_PERIOD;
1096 	data->period = period;
1097 
1098 	if (addr) {
1099 		data->addr = addr;
1100 		data->sample_flags |= PERF_SAMPLE_ADDR;
1101 	}
1102 }
1103 
1104 /*
1105  * Clear all bitfields in the perf_branch_entry.
1106  * The to and from fields are not cleared because they are
1107  * systematically modified by caller.
1108  */
perf_clear_branch_entry_bitfields(struct perf_branch_entry * br)1109 static inline void perf_clear_branch_entry_bitfields(struct perf_branch_entry *br)
1110 {
1111 	br->mispred = 0;
1112 	br->predicted = 0;
1113 	br->in_tx = 0;
1114 	br->abort = 0;
1115 	br->cycles = 0;
1116 	br->type = 0;
1117 	br->spec = PERF_BR_SPEC_NA;
1118 	br->reserved = 0;
1119 }
1120 
1121 extern void perf_output_sample(struct perf_output_handle *handle,
1122 			       struct perf_event_header *header,
1123 			       struct perf_sample_data *data,
1124 			       struct perf_event *event);
1125 extern void perf_prepare_sample(struct perf_event_header *header,
1126 				struct perf_sample_data *data,
1127 				struct perf_event *event,
1128 				struct pt_regs *regs);
1129 
1130 extern int perf_event_overflow(struct perf_event *event,
1131 				 struct perf_sample_data *data,
1132 				 struct pt_regs *regs);
1133 
1134 extern void perf_event_output_forward(struct perf_event *event,
1135 				     struct perf_sample_data *data,
1136 				     struct pt_regs *regs);
1137 extern void perf_event_output_backward(struct perf_event *event,
1138 				       struct perf_sample_data *data,
1139 				       struct pt_regs *regs);
1140 extern int perf_event_output(struct perf_event *event,
1141 			     struct perf_sample_data *data,
1142 			     struct pt_regs *regs);
1143 
1144 static inline bool
__is_default_overflow_handler(perf_overflow_handler_t overflow_handler)1145 __is_default_overflow_handler(perf_overflow_handler_t overflow_handler)
1146 {
1147 	if (likely(overflow_handler == perf_event_output_forward))
1148 		return true;
1149 	if (unlikely(overflow_handler == perf_event_output_backward))
1150 		return true;
1151 	return false;
1152 }
1153 
1154 #define is_default_overflow_handler(event) \
1155 	__is_default_overflow_handler((event)->overflow_handler)
1156 
1157 #ifdef CONFIG_BPF_SYSCALL
uses_default_overflow_handler(struct perf_event * event)1158 static inline bool uses_default_overflow_handler(struct perf_event *event)
1159 {
1160 	if (likely(is_default_overflow_handler(event)))
1161 		return true;
1162 
1163 	return __is_default_overflow_handler(event->orig_overflow_handler);
1164 }
1165 #else
1166 #define uses_default_overflow_handler(event) \
1167 	is_default_overflow_handler(event)
1168 #endif
1169 
1170 extern void
1171 perf_event_header__init_id(struct perf_event_header *header,
1172 			   struct perf_sample_data *data,
1173 			   struct perf_event *event);
1174 extern void
1175 perf_event__output_id_sample(struct perf_event *event,
1176 			     struct perf_output_handle *handle,
1177 			     struct perf_sample_data *sample);
1178 
1179 extern void
1180 perf_log_lost_samples(struct perf_event *event, u64 lost);
1181 
event_has_any_exclude_flag(struct perf_event * event)1182 static inline bool event_has_any_exclude_flag(struct perf_event *event)
1183 {
1184 	struct perf_event_attr *attr = &event->attr;
1185 
1186 	return attr->exclude_idle || attr->exclude_user ||
1187 	       attr->exclude_kernel || attr->exclude_hv ||
1188 	       attr->exclude_guest || attr->exclude_host;
1189 }
1190 
is_sampling_event(struct perf_event * event)1191 static inline bool is_sampling_event(struct perf_event *event)
1192 {
1193 	return event->attr.sample_period != 0;
1194 }
1195 
1196 /*
1197  * Return 1 for a software event, 0 for a hardware event
1198  */
is_software_event(struct perf_event * event)1199 static inline int is_software_event(struct perf_event *event)
1200 {
1201 	return event->event_caps & PERF_EV_CAP_SOFTWARE;
1202 }
1203 
1204 /*
1205  * Return 1 for event in sw context, 0 for event in hw context
1206  */
in_software_context(struct perf_event * event)1207 static inline int in_software_context(struct perf_event *event)
1208 {
1209 	return event->ctx->pmu->task_ctx_nr == perf_sw_context;
1210 }
1211 
is_exclusive_pmu(struct pmu * pmu)1212 static inline int is_exclusive_pmu(struct pmu *pmu)
1213 {
1214 	return pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE;
1215 }
1216 
1217 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
1218 
1219 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
1220 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
1221 
1222 #ifndef perf_arch_fetch_caller_regs
perf_arch_fetch_caller_regs(struct pt_regs * regs,unsigned long ip)1223 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
1224 #endif
1225 
1226 /*
1227  * When generating a perf sample in-line, instead of from an interrupt /
1228  * exception, we lack a pt_regs. This is typically used from software events
1229  * like: SW_CONTEXT_SWITCHES, SW_MIGRATIONS and the tie-in with tracepoints.
1230  *
1231  * We typically don't need a full set, but (for x86) do require:
1232  * - ip for PERF_SAMPLE_IP
1233  * - cs for user_mode() tests
1234  * - sp for PERF_SAMPLE_CALLCHAIN
1235  * - eflags for MISC bits and CALLCHAIN (see: perf_hw_regs())
1236  *
1237  * NOTE: assumes @regs is otherwise already 0 filled; this is important for
1238  * things like PERF_SAMPLE_REGS_INTR.
1239  */
perf_fetch_caller_regs(struct pt_regs * regs)1240 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1241 {
1242 	perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1243 }
1244 
1245 static __always_inline void
perf_sw_event(u32 event_id,u64 nr,struct pt_regs * regs,u64 addr)1246 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1247 {
1248 	if (static_key_false(&perf_swevent_enabled[event_id]))
1249 		__perf_sw_event(event_id, nr, regs, addr);
1250 }
1251 
1252 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
1253 
1254 /*
1255  * 'Special' version for the scheduler, it hard assumes no recursion,
1256  * which is guaranteed by us not actually scheduling inside other swevents
1257  * because those disable preemption.
1258  */
__perf_sw_event_sched(u32 event_id,u64 nr,u64 addr)1259 static __always_inline void __perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
1260 {
1261 	struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1262 
1263 	perf_fetch_caller_regs(regs);
1264 	___perf_sw_event(event_id, nr, regs, addr);
1265 }
1266 
1267 extern struct static_key_false perf_sched_events;
1268 
__perf_sw_enabled(int swevt)1269 static __always_inline bool __perf_sw_enabled(int swevt)
1270 {
1271 	return static_key_false(&perf_swevent_enabled[swevt]);
1272 }
1273 
perf_event_task_migrate(struct task_struct * task)1274 static inline void perf_event_task_migrate(struct task_struct *task)
1275 {
1276 	if (__perf_sw_enabled(PERF_COUNT_SW_CPU_MIGRATIONS))
1277 		task->sched_migrated = 1;
1278 }
1279 
perf_event_task_sched_in(struct task_struct * prev,struct task_struct * task)1280 static inline void perf_event_task_sched_in(struct task_struct *prev,
1281 					    struct task_struct *task)
1282 {
1283 	if (static_branch_unlikely(&perf_sched_events))
1284 		__perf_event_task_sched_in(prev, task);
1285 
1286 	if (__perf_sw_enabled(PERF_COUNT_SW_CPU_MIGRATIONS) &&
1287 	    task->sched_migrated) {
1288 		__perf_sw_event_sched(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 0);
1289 		task->sched_migrated = 0;
1290 	}
1291 }
1292 
perf_event_task_sched_out(struct task_struct * prev,struct task_struct * next)1293 static inline void perf_event_task_sched_out(struct task_struct *prev,
1294 					     struct task_struct *next)
1295 {
1296 	if (__perf_sw_enabled(PERF_COUNT_SW_CONTEXT_SWITCHES))
1297 		__perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
1298 
1299 #ifdef CONFIG_CGROUP_PERF
1300 	if (__perf_sw_enabled(PERF_COUNT_SW_CGROUP_SWITCHES) &&
1301 	    perf_cgroup_from_task(prev, NULL) !=
1302 	    perf_cgroup_from_task(next, NULL))
1303 		__perf_sw_event_sched(PERF_COUNT_SW_CGROUP_SWITCHES, 1, 0);
1304 #endif
1305 
1306 	if (static_branch_unlikely(&perf_sched_events))
1307 		__perf_event_task_sched_out(prev, next);
1308 }
1309 
1310 extern void perf_event_mmap(struct vm_area_struct *vma);
1311 
1312 extern void perf_event_ksymbol(u16 ksym_type, u64 addr, u32 len,
1313 			       bool unregister, const char *sym);
1314 extern void perf_event_bpf_event(struct bpf_prog *prog,
1315 				 enum perf_bpf_event_type type,
1316 				 u16 flags);
1317 
1318 #ifdef CONFIG_GUEST_PERF_EVENTS
1319 extern struct perf_guest_info_callbacks __rcu *perf_guest_cbs;
perf_get_guest_cbs(void)1320 static inline struct perf_guest_info_callbacks *perf_get_guest_cbs(void)
1321 {
1322 	/*
1323 	 * Callbacks are RCU-protected and must be READ_ONCE to avoid reloading
1324 	 * the callbacks between a !NULL check and dereferences, to ensure
1325 	 * pending stores/changes to the callback pointers are visible before a
1326 	 * non-NULL perf_guest_cbs is visible to readers, and to prevent a
1327 	 * module from unloading callbacks while readers are active.
1328 	 */
1329 	return rcu_dereference(perf_guest_cbs);
1330 }
perf_guest_state(void)1331 static inline unsigned int perf_guest_state(void)
1332 {
1333 	struct perf_guest_info_callbacks *guest_cbs = perf_get_guest_cbs();
1334 
1335 	return guest_cbs ? guest_cbs->state() : 0;
1336 }
perf_guest_get_ip(void)1337 static inline unsigned long perf_guest_get_ip(void)
1338 {
1339 	struct perf_guest_info_callbacks *guest_cbs = perf_get_guest_cbs();
1340 
1341 	/*
1342 	 * Arbitrarily return '0' in the unlikely scenario that the callbacks
1343 	 * are unregistered between checking guest state and getting the IP.
1344 	 */
1345 	return guest_cbs ? guest_cbs->get_ip() : 0;
1346 }
perf_guest_handle_intel_pt_intr(void)1347 static inline unsigned int perf_guest_handle_intel_pt_intr(void)
1348 {
1349 	struct perf_guest_info_callbacks *guest_cbs = perf_get_guest_cbs();
1350 
1351 	if (guest_cbs && guest_cbs->handle_intel_pt_intr)
1352 		return guest_cbs->handle_intel_pt_intr();
1353 	return 0;
1354 }
1355 extern void perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs);
1356 extern void perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs);
1357 #else
perf_guest_state(void)1358 static inline unsigned int perf_guest_state(void)		 { return 0; }
perf_guest_get_ip(void)1359 static inline unsigned long perf_guest_get_ip(void)		 { return 0; }
perf_guest_handle_intel_pt_intr(void)1360 static inline unsigned int perf_guest_handle_intel_pt_intr(void) { return 0; }
1361 #endif /* CONFIG_GUEST_PERF_EVENTS */
1362 
1363 extern void perf_event_exec(void);
1364 extern void perf_event_comm(struct task_struct *tsk, bool exec);
1365 extern void perf_event_namespaces(struct task_struct *tsk);
1366 extern void perf_event_fork(struct task_struct *tsk);
1367 extern void perf_event_text_poke(const void *addr,
1368 				 const void *old_bytes, size_t old_len,
1369 				 const void *new_bytes, size_t new_len);
1370 
1371 /* Callchains */
1372 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1373 
1374 extern void perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1375 extern void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1376 extern struct perf_callchain_entry *
1377 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
1378 		   u32 max_stack, bool crosstask, bool add_mark);
1379 extern struct perf_callchain_entry *perf_callchain(struct perf_event *event, struct pt_regs *regs);
1380 extern int get_callchain_buffers(int max_stack);
1381 extern void put_callchain_buffers(void);
1382 extern struct perf_callchain_entry *get_callchain_entry(int *rctx);
1383 extern void put_callchain_entry(int rctx);
1384 
1385 extern int sysctl_perf_event_max_stack;
1386 extern int sysctl_perf_event_max_contexts_per_stack;
1387 
perf_callchain_store_context(struct perf_callchain_entry_ctx * ctx,u64 ip)1388 static inline int perf_callchain_store_context(struct perf_callchain_entry_ctx *ctx, u64 ip)
1389 {
1390 	if (ctx->contexts < sysctl_perf_event_max_contexts_per_stack) {
1391 		struct perf_callchain_entry *entry = ctx->entry;
1392 		entry->ip[entry->nr++] = ip;
1393 		++ctx->contexts;
1394 		return 0;
1395 	} else {
1396 		ctx->contexts_maxed = true;
1397 		return -1; /* no more room, stop walking the stack */
1398 	}
1399 }
1400 
perf_callchain_store(struct perf_callchain_entry_ctx * ctx,u64 ip)1401 static inline int perf_callchain_store(struct perf_callchain_entry_ctx *ctx, u64 ip)
1402 {
1403 	if (ctx->nr < ctx->max_stack && !ctx->contexts_maxed) {
1404 		struct perf_callchain_entry *entry = ctx->entry;
1405 		entry->ip[entry->nr++] = ip;
1406 		++ctx->nr;
1407 		return 0;
1408 	} else {
1409 		return -1; /* no more room, stop walking the stack */
1410 	}
1411 }
1412 
1413 extern int sysctl_perf_event_paranoid;
1414 extern int sysctl_perf_event_mlock;
1415 extern int sysctl_perf_event_sample_rate;
1416 extern int sysctl_perf_cpu_time_max_percent;
1417 
1418 extern void perf_sample_event_took(u64 sample_len_ns);
1419 
1420 int perf_proc_update_handler(struct ctl_table *table, int write,
1421 		void *buffer, size_t *lenp, loff_t *ppos);
1422 int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
1423 		void *buffer, size_t *lenp, loff_t *ppos);
1424 int perf_event_max_stack_handler(struct ctl_table *table, int write,
1425 		void *buffer, size_t *lenp, loff_t *ppos);
1426 
1427 /* Access to perf_event_open(2) syscall. */
1428 #define PERF_SECURITY_OPEN		0
1429 
1430 /* Finer grained perf_event_open(2) access control. */
1431 #define PERF_SECURITY_CPU		1
1432 #define PERF_SECURITY_KERNEL		2
1433 #define PERF_SECURITY_TRACEPOINT	3
1434 
perf_is_paranoid(void)1435 static inline int perf_is_paranoid(void)
1436 {
1437 	return sysctl_perf_event_paranoid > -1;
1438 }
1439 
perf_allow_kernel(struct perf_event_attr * attr)1440 static inline int perf_allow_kernel(struct perf_event_attr *attr)
1441 {
1442 	if (sysctl_perf_event_paranoid > 1 && !perfmon_capable())
1443 		return -EACCES;
1444 
1445 	return security_perf_event_open(attr, PERF_SECURITY_KERNEL);
1446 }
1447 
perf_allow_cpu(struct perf_event_attr * attr)1448 static inline int perf_allow_cpu(struct perf_event_attr *attr)
1449 {
1450 	if (sysctl_perf_event_paranoid > 0 && !perfmon_capable())
1451 		return -EACCES;
1452 
1453 	return security_perf_event_open(attr, PERF_SECURITY_CPU);
1454 }
1455 
perf_allow_tracepoint(struct perf_event_attr * attr)1456 static inline int perf_allow_tracepoint(struct perf_event_attr *attr)
1457 {
1458 	if (sysctl_perf_event_paranoid > -1 && !perfmon_capable())
1459 		return -EPERM;
1460 
1461 	return security_perf_event_open(attr, PERF_SECURITY_TRACEPOINT);
1462 }
1463 
1464 extern void perf_event_init(void);
1465 extern void perf_tp_event(u16 event_type, u64 count, void *record,
1466 			  int entry_size, struct pt_regs *regs,
1467 			  struct hlist_head *head, int rctx,
1468 			  struct task_struct *task);
1469 extern void perf_bp_event(struct perf_event *event, void *data);
1470 
1471 #ifndef perf_misc_flags
1472 # define perf_misc_flags(regs) \
1473 		(user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1474 # define perf_instruction_pointer(regs)	instruction_pointer(regs)
1475 #endif
1476 #ifndef perf_arch_bpf_user_pt_regs
1477 # define perf_arch_bpf_user_pt_regs(regs) regs
1478 #endif
1479 
has_branch_stack(struct perf_event * event)1480 static inline bool has_branch_stack(struct perf_event *event)
1481 {
1482 	return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1483 }
1484 
needs_branch_stack(struct perf_event * event)1485 static inline bool needs_branch_stack(struct perf_event *event)
1486 {
1487 	return event->attr.branch_sample_type != 0;
1488 }
1489 
has_aux(struct perf_event * event)1490 static inline bool has_aux(struct perf_event *event)
1491 {
1492 	return event->pmu->setup_aux;
1493 }
1494 
is_write_backward(struct perf_event * event)1495 static inline bool is_write_backward(struct perf_event *event)
1496 {
1497 	return !!event->attr.write_backward;
1498 }
1499 
has_addr_filter(struct perf_event * event)1500 static inline bool has_addr_filter(struct perf_event *event)
1501 {
1502 	return event->pmu->nr_addr_filters;
1503 }
1504 
1505 /*
1506  * An inherited event uses parent's filters
1507  */
1508 static inline struct perf_addr_filters_head *
perf_event_addr_filters(struct perf_event * event)1509 perf_event_addr_filters(struct perf_event *event)
1510 {
1511 	struct perf_addr_filters_head *ifh = &event->addr_filters;
1512 
1513 	if (event->parent)
1514 		ifh = &event->parent->addr_filters;
1515 
1516 	return ifh;
1517 }
1518 
1519 extern void perf_event_addr_filters_sync(struct perf_event *event);
1520 extern void perf_report_aux_output_id(struct perf_event *event, u64 hw_id);
1521 
1522 extern int perf_output_begin(struct perf_output_handle *handle,
1523 			     struct perf_sample_data *data,
1524 			     struct perf_event *event, unsigned int size);
1525 extern int perf_output_begin_forward(struct perf_output_handle *handle,
1526 				     struct perf_sample_data *data,
1527 				     struct perf_event *event,
1528 				     unsigned int size);
1529 extern int perf_output_begin_backward(struct perf_output_handle *handle,
1530 				      struct perf_sample_data *data,
1531 				      struct perf_event *event,
1532 				      unsigned int size);
1533 
1534 extern void perf_output_end(struct perf_output_handle *handle);
1535 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
1536 			     const void *buf, unsigned int len);
1537 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
1538 				     unsigned int len);
1539 extern long perf_output_copy_aux(struct perf_output_handle *aux_handle,
1540 				 struct perf_output_handle *handle,
1541 				 unsigned long from, unsigned long to);
1542 extern int perf_swevent_get_recursion_context(void);
1543 extern void perf_swevent_put_recursion_context(int rctx);
1544 extern u64 perf_swevent_set_period(struct perf_event *event);
1545 extern void perf_event_enable(struct perf_event *event);
1546 extern void perf_event_disable(struct perf_event *event);
1547 extern void perf_event_disable_local(struct perf_event *event);
1548 extern void perf_event_disable_inatomic(struct perf_event *event);
1549 extern void perf_event_task_tick(void);
1550 extern int perf_event_account_interrupt(struct perf_event *event);
1551 extern int perf_event_period(struct perf_event *event, u64 value);
1552 extern u64 perf_event_pause(struct perf_event *event, bool reset);
1553 #else /* !CONFIG_PERF_EVENTS: */
1554 static inline void *
perf_aux_output_begin(struct perf_output_handle * handle,struct perf_event * event)1555 perf_aux_output_begin(struct perf_output_handle *handle,
1556 		      struct perf_event *event)				{ return NULL; }
1557 static inline void
perf_aux_output_end(struct perf_output_handle * handle,unsigned long size)1558 perf_aux_output_end(struct perf_output_handle *handle, unsigned long size)
1559 									{ }
1560 static inline int
perf_aux_output_skip(struct perf_output_handle * handle,unsigned long size)1561 perf_aux_output_skip(struct perf_output_handle *handle,
1562 		     unsigned long size)				{ return -EINVAL; }
1563 static inline void *
perf_get_aux(struct perf_output_handle * handle)1564 perf_get_aux(struct perf_output_handle *handle)				{ return NULL; }
1565 static inline void
perf_event_task_migrate(struct task_struct * task)1566 perf_event_task_migrate(struct task_struct *task)			{ }
1567 static inline void
perf_event_task_sched_in(struct task_struct * prev,struct task_struct * task)1568 perf_event_task_sched_in(struct task_struct *prev,
1569 			 struct task_struct *task)			{ }
1570 static inline void
perf_event_task_sched_out(struct task_struct * prev,struct task_struct * next)1571 perf_event_task_sched_out(struct task_struct *prev,
1572 			  struct task_struct *next)			{ }
perf_event_init_task(struct task_struct * child,u64 clone_flags)1573 static inline int perf_event_init_task(struct task_struct *child,
1574 				       u64 clone_flags)			{ return 0; }
perf_event_exit_task(struct task_struct * child)1575 static inline void perf_event_exit_task(struct task_struct *child)	{ }
perf_event_free_task(struct task_struct * task)1576 static inline void perf_event_free_task(struct task_struct *task)	{ }
perf_event_delayed_put(struct task_struct * task)1577 static inline void perf_event_delayed_put(struct task_struct *task)	{ }
perf_event_get(unsigned int fd)1578 static inline struct file *perf_event_get(unsigned int fd)	{ return ERR_PTR(-EINVAL); }
perf_get_event(struct file * file)1579 static inline const struct perf_event *perf_get_event(struct file *file)
1580 {
1581 	return ERR_PTR(-EINVAL);
1582 }
perf_event_attrs(struct perf_event * event)1583 static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
1584 {
1585 	return ERR_PTR(-EINVAL);
1586 }
perf_event_read_local(struct perf_event * event,u64 * value,u64 * enabled,u64 * running)1587 static inline int perf_event_read_local(struct perf_event *event, u64 *value,
1588 					u64 *enabled, u64 *running)
1589 {
1590 	return -EINVAL;
1591 }
perf_event_print_debug(void)1592 static inline void perf_event_print_debug(void)				{ }
perf_event_task_disable(void)1593 static inline int perf_event_task_disable(void)				{ return -EINVAL; }
perf_event_task_enable(void)1594 static inline int perf_event_task_enable(void)				{ return -EINVAL; }
perf_event_refresh(struct perf_event * event,int refresh)1595 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1596 {
1597 	return -EINVAL;
1598 }
1599 
1600 static inline void
perf_sw_event(u32 event_id,u64 nr,struct pt_regs * regs,u64 addr)1601 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)	{ }
1602 static inline void
perf_bp_event(struct perf_event * event,void * data)1603 perf_bp_event(struct perf_event *event, void *data)			{ }
1604 
perf_event_mmap(struct vm_area_struct * vma)1605 static inline void perf_event_mmap(struct vm_area_struct *vma)		{ }
1606 
1607 typedef int (perf_ksymbol_get_name_f)(char *name, int name_len, void *data);
perf_event_ksymbol(u16 ksym_type,u64 addr,u32 len,bool unregister,const char * sym)1608 static inline void perf_event_ksymbol(u16 ksym_type, u64 addr, u32 len,
1609 				      bool unregister, const char *sym)	{ }
perf_event_bpf_event(struct bpf_prog * prog,enum perf_bpf_event_type type,u16 flags)1610 static inline void perf_event_bpf_event(struct bpf_prog *prog,
1611 					enum perf_bpf_event_type type,
1612 					u16 flags)			{ }
perf_event_exec(void)1613 static inline void perf_event_exec(void)				{ }
perf_event_comm(struct task_struct * tsk,bool exec)1614 static inline void perf_event_comm(struct task_struct *tsk, bool exec)	{ }
perf_event_namespaces(struct task_struct * tsk)1615 static inline void perf_event_namespaces(struct task_struct *tsk)	{ }
perf_event_fork(struct task_struct * tsk)1616 static inline void perf_event_fork(struct task_struct *tsk)		{ }
perf_event_text_poke(const void * addr,const void * old_bytes,size_t old_len,const void * new_bytes,size_t new_len)1617 static inline void perf_event_text_poke(const void *addr,
1618 					const void *old_bytes,
1619 					size_t old_len,
1620 					const void *new_bytes,
1621 					size_t new_len)			{ }
perf_event_init(void)1622 static inline void perf_event_init(void)				{ }
perf_swevent_get_recursion_context(void)1623 static inline int  perf_swevent_get_recursion_context(void)		{ return -1; }
perf_swevent_put_recursion_context(int rctx)1624 static inline void perf_swevent_put_recursion_context(int rctx)		{ }
perf_swevent_set_period(struct perf_event * event)1625 static inline u64 perf_swevent_set_period(struct perf_event *event)	{ return 0; }
perf_event_enable(struct perf_event * event)1626 static inline void perf_event_enable(struct perf_event *event)		{ }
perf_event_disable(struct perf_event * event)1627 static inline void perf_event_disable(struct perf_event *event)		{ }
__perf_event_disable(void * info)1628 static inline int __perf_event_disable(void *info)			{ return -1; }
perf_event_task_tick(void)1629 static inline void perf_event_task_tick(void)				{ }
perf_event_release_kernel(struct perf_event * event)1630 static inline int perf_event_release_kernel(struct perf_event *event)	{ return 0; }
perf_event_period(struct perf_event * event,u64 value)1631 static inline int perf_event_period(struct perf_event *event, u64 value)
1632 {
1633 	return -EINVAL;
1634 }
perf_event_pause(struct perf_event * event,bool reset)1635 static inline u64 perf_event_pause(struct perf_event *event, bool reset)
1636 {
1637 	return 0;
1638 }
1639 #endif
1640 
1641 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1642 extern void perf_restore_debug_store(void);
1643 #else
perf_restore_debug_store(void)1644 static inline void perf_restore_debug_store(void)			{ }
1645 #endif
1646 
perf_raw_frag_last(const struct perf_raw_frag * frag)1647 static __always_inline bool perf_raw_frag_last(const struct perf_raw_frag *frag)
1648 {
1649 	return frag->pad < sizeof(u64);
1650 }
1651 
1652 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1653 
1654 struct perf_pmu_events_attr {
1655 	struct device_attribute attr;
1656 	u64 id;
1657 	const char *event_str;
1658 };
1659 
1660 struct perf_pmu_events_ht_attr {
1661 	struct device_attribute			attr;
1662 	u64					id;
1663 	const char				*event_str_ht;
1664 	const char				*event_str_noht;
1665 };
1666 
1667 struct perf_pmu_events_hybrid_attr {
1668 	struct device_attribute			attr;
1669 	u64					id;
1670 	const char				*event_str;
1671 	u64					pmu_type;
1672 };
1673 
1674 struct perf_pmu_format_hybrid_attr {
1675 	struct device_attribute			attr;
1676 	u64					pmu_type;
1677 };
1678 
1679 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1680 			      char *page);
1681 
1682 #define PMU_EVENT_ATTR(_name, _var, _id, _show)				\
1683 static struct perf_pmu_events_attr _var = {				\
1684 	.attr = __ATTR(_name, 0444, _show, NULL),			\
1685 	.id   =  _id,							\
1686 };
1687 
1688 #define PMU_EVENT_ATTR_STRING(_name, _var, _str)			    \
1689 static struct perf_pmu_events_attr _var = {				    \
1690 	.attr		= __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1691 	.id		= 0,						    \
1692 	.event_str	= _str,						    \
1693 };
1694 
1695 #define PMU_EVENT_ATTR_ID(_name, _show, _id)				\
1696 	(&((struct perf_pmu_events_attr[]) {				\
1697 		{ .attr = __ATTR(_name, 0444, _show, NULL),		\
1698 		  .id = _id, }						\
1699 	})[0].attr.attr)
1700 
1701 #define PMU_FORMAT_ATTR(_name, _format)					\
1702 static ssize_t								\
1703 _name##_show(struct device *dev,					\
1704 			       struct device_attribute *attr,		\
1705 			       char *page)				\
1706 {									\
1707 	BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);			\
1708 	return sprintf(page, _format "\n");				\
1709 }									\
1710 									\
1711 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1712 
1713 /* Performance counter hotplug functions */
1714 #ifdef CONFIG_PERF_EVENTS
1715 int perf_event_init_cpu(unsigned int cpu);
1716 int perf_event_exit_cpu(unsigned int cpu);
1717 #else
1718 #define perf_event_init_cpu	NULL
1719 #define perf_event_exit_cpu	NULL
1720 #endif
1721 
1722 extern void __weak arch_perf_update_userpage(struct perf_event *event,
1723 					     struct perf_event_mmap_page *userpg,
1724 					     u64 now);
1725 
1726 #ifdef CONFIG_MMU
1727 extern __weak u64 arch_perf_get_page_size(struct mm_struct *mm, unsigned long addr);
1728 #endif
1729 
1730 /*
1731  * Snapshot branch stack on software events.
1732  *
1733  * Branch stack can be very useful in understanding software events. For
1734  * example, when a long function, e.g. sys_perf_event_open, returns an
1735  * errno, it is not obvious why the function failed. Branch stack could
1736  * provide very helpful information in this type of scenarios.
1737  *
1738  * On software event, it is necessary to stop the hardware branch recorder
1739  * fast. Otherwise, the hardware register/buffer will be flushed with
1740  * entries of the triggering event. Therefore, static call is used to
1741  * stop the hardware recorder.
1742  */
1743 
1744 /*
1745  * cnt is the number of entries allocated for entries.
1746  * Return number of entries copied to .
1747  */
1748 typedef int (perf_snapshot_branch_stack_t)(struct perf_branch_entry *entries,
1749 					   unsigned int cnt);
1750 DECLARE_STATIC_CALL(perf_snapshot_branch_stack, perf_snapshot_branch_stack_t);
1751 
1752 #ifndef PERF_NEEDS_LOPWR_CB
perf_lopwr_cb(bool mode)1753 static inline void perf_lopwr_cb(bool mode)
1754 {
1755 }
1756 #endif
1757 
1758 #ifdef CONFIG_PERF_EVENTS
branch_sample_no_flags(const struct perf_event * event)1759 static inline bool branch_sample_no_flags(const struct perf_event *event)
1760 {
1761 	return event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_NO_FLAGS;
1762 }
1763 
branch_sample_no_cycles(const struct perf_event * event)1764 static inline bool branch_sample_no_cycles(const struct perf_event *event)
1765 {
1766 	return event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_NO_CYCLES;
1767 }
1768 
branch_sample_type(const struct perf_event * event)1769 static inline bool branch_sample_type(const struct perf_event *event)
1770 {
1771 	return event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_TYPE_SAVE;
1772 }
1773 
branch_sample_hw_index(const struct perf_event * event)1774 static inline bool branch_sample_hw_index(const struct perf_event *event)
1775 {
1776 	return event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX;
1777 }
1778 
branch_sample_priv(const struct perf_event * event)1779 static inline bool branch_sample_priv(const struct perf_event *event)
1780 {
1781 	return event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_PRIV_SAVE;
1782 }
1783 #endif /* CONFIG_PERF_EVENTS */
1784 #endif /* _LINUX_PERF_EVENT_H */
1785