Home | History | Annotate | Download | only in linux
      1 /*
      2  * Performance events:
      3  *
      4  *    Copyright (C) 2008-2009, Thomas Gleixner <tglx (at) 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 
     19 /*
     20  * Kernel-internal data types and definitions:
     21  */
     22 
     23 #ifdef CONFIG_PERF_EVENTS
     24 # include <asm/perf_event.h>
     25 # include <asm/local64.h>
     26 #endif
     27 
     28 struct perf_guest_info_callbacks {
     29 	int				(*is_in_guest)(void);
     30 	int				(*is_user_mode)(void);
     31 	unsigned long			(*get_guest_ip)(void);
     32 };
     33 
     34 #ifdef CONFIG_HAVE_HW_BREAKPOINT
     35 #include <asm/hw_breakpoint.h>
     36 #endif
     37 
     38 #include <linux/list.h>
     39 #include <linux/mutex.h>
     40 #include <linux/rculist.h>
     41 #include <linux/rcupdate.h>
     42 #include <linux/spinlock.h>
     43 #include <linux/hrtimer.h>
     44 #include <linux/fs.h>
     45 #include <linux/pid_namespace.h>
     46 #include <linux/workqueue.h>
     47 #include <linux/ftrace.h>
     48 #include <linux/cpu.h>
     49 #include <linux/irq_work.h>
     50 #include <linux/static_key.h>
     51 #include <linux/jump_label_ratelimit.h>
     52 #include <linux/atomic.h>
     53 #include <linux/sysfs.h>
     54 #include <linux/perf_regs.h>
     55 #include <asm/local.h>
     56 
     57 struct perf_callchain_entry {
     58 	__u64				nr;
     59 	__u64				ip[PERF_MAX_STACK_DEPTH];
     60 };
     61 
     62 struct perf_raw_record {
     63 	u32				size;
     64 	void				*data;
     65 };
     66 
     67 /*
     68  * branch stack layout:
     69  *  nr: number of taken branches stored in entries[]
     70  *
     71  * Note that nr can vary from sample to sample
     72  * branches (to, from) are stored from most recent
     73  * to least recent, i.e., entries[0] contains the most
     74  * recent branch.
     75  */
     76 struct perf_branch_stack {
     77 	__u64				nr;
     78 	struct perf_branch_entry	entries[0];
     79 };
     80 
     81 struct perf_regs_user {
     82 	__u64		abi;
     83 	struct pt_regs	*regs;
     84 };
     85 
     86 struct task_struct;
     87 
     88 /*
     89  * extra PMU register associated with an event
     90  */
     91 struct hw_perf_event_extra {
     92 	u64		config;	/* register value */
     93 	unsigned int	reg;	/* register address or index */
     94 	int		alloc;	/* extra register already allocated */
     95 	int		idx;	/* index in shared_regs->regs[] */
     96 };
     97 
     98 struct event_constraint;
     99 
    100 /**
    101  * struct hw_perf_event - performance event hardware details:
    102  */
    103 struct hw_perf_event {
    104 #ifdef CONFIG_PERF_EVENTS
    105 	union {
    106 		struct { /* hardware */
    107 			u64		config;
    108 			u64		last_tag;
    109 			unsigned long	config_base;
    110 			unsigned long	event_base;
    111 			int		event_base_rdpmc;
    112 			int		idx;
    113 			int		last_cpu;
    114 			int		flags;
    115 
    116 			struct hw_perf_event_extra extra_reg;
    117 			struct hw_perf_event_extra branch_reg;
    118 
    119 			struct event_constraint *constraint;
    120 		};
    121 		struct { /* software */
    122 			struct hrtimer	hrtimer;
    123 		};
    124 		struct { /* tracepoint */
    125 			struct task_struct	*tp_target;
    126 			/* for tp_event->class */
    127 			struct list_head	tp_list;
    128 		};
    129 #ifdef CONFIG_HAVE_HW_BREAKPOINT
    130 		struct { /* breakpoint */
    131 			/*
    132 			 * Crufty hack to avoid the chicken and egg
    133 			 * problem hw_breakpoint has with context
    134 			 * creation and event initalization.
    135 			 */
    136 			struct task_struct		*bp_target;
    137 			struct arch_hw_breakpoint	info;
    138 			struct list_head		bp_list;
    139 		};
    140 #endif
    141 	};
    142 	int				state;
    143 	local64_t			prev_count;
    144 	u64				sample_period;
    145 	u64				last_period;
    146 	local64_t			period_left;
    147 	u64                             interrupts_seq;
    148 	u64				interrupts;
    149 
    150 	u64				freq_time_stamp;
    151 	u64				freq_count_stamp;
    152 #endif
    153 };
    154 
    155 /*
    156  * hw_perf_event::state flags
    157  */
    158 #define PERF_HES_STOPPED	0x01 /* the counter is stopped */
    159 #define PERF_HES_UPTODATE	0x02 /* event->count up-to-date */
    160 #define PERF_HES_ARCH		0x04
    161 
    162 struct perf_event;
    163 
    164 /*
    165  * Common implementation detail of pmu::{start,commit,cancel}_txn
    166  */
    167 #define PERF_EVENT_TXN 0x1
    168 
    169 /**
    170  * struct pmu - generic performance monitoring unit
    171  */
    172 struct pmu {
    173 	struct list_head		entry;
    174 
    175 	struct device			*dev;
    176 	const struct attribute_group	**attr_groups;
    177 	const char			*name;
    178 	int				type;
    179 
    180 	int * __percpu			pmu_disable_count;
    181 	struct perf_cpu_context * __percpu pmu_cpu_context;
    182 	int				task_ctx_nr;
    183 	int				hrtimer_interval_ms;
    184 
    185 	/*
    186 	 * Fully disable/enable this PMU, can be used to protect from the PMI
    187 	 * as well as for lazy/batch writing of the MSRs.
    188 	 */
    189 	void (*pmu_enable)		(struct pmu *pmu); /* optional */
    190 	void (*pmu_disable)		(struct pmu *pmu); /* optional */
    191 
    192 	/*
    193 	 * Try and initialize the event for this PMU.
    194 	 * Should return -ENOENT when the @event doesn't match this PMU.
    195 	 */
    196 	int (*event_init)		(struct perf_event *event);
    197 
    198 #define PERF_EF_START	0x01		/* start the counter when adding    */
    199 #define PERF_EF_RELOAD	0x02		/* reload the counter when starting */
    200 #define PERF_EF_UPDATE	0x04		/* update the counter when stopping */
    201 
    202 	/*
    203 	 * Adds/Removes a counter to/from the PMU, can be done inside
    204 	 * a transaction, see the ->*_txn() methods.
    205 	 */
    206 	int  (*add)			(struct perf_event *event, int flags);
    207 	void (*del)			(struct perf_event *event, int flags);
    208 
    209 	/*
    210 	 * Starts/Stops a counter present on the PMU. The PMI handler
    211 	 * should stop the counter when perf_event_overflow() returns
    212 	 * !0. ->start() will be used to continue.
    213 	 */
    214 	void (*start)			(struct perf_event *event, int flags);
    215 	void (*stop)			(struct perf_event *event, int flags);
    216 
    217 	/*
    218 	 * Updates the counter value of the event.
    219 	 */
    220 	void (*read)			(struct perf_event *event);
    221 
    222 	/*
    223 	 * Group events scheduling is treated as a transaction, add
    224 	 * group events as a whole and perform one schedulability test.
    225 	 * If the test fails, roll back the whole group
    226 	 *
    227 	 * Start the transaction, after this ->add() doesn't need to
    228 	 * do schedulability tests.
    229 	 */
    230 	void (*start_txn)		(struct pmu *pmu); /* optional */
    231 	/*
    232 	 * If ->start_txn() disabled the ->add() schedulability test
    233 	 * then ->commit_txn() is required to perform one. On success
    234 	 * the transaction is closed. On error the transaction is kept
    235 	 * open until ->cancel_txn() is called.
    236 	 */
    237 	int  (*commit_txn)		(struct pmu *pmu); /* optional */
    238 	/*
    239 	 * Will cancel the transaction, assumes ->del() is called
    240 	 * for each successful ->add() during the transaction.
    241 	 */
    242 	void (*cancel_txn)		(struct pmu *pmu); /* optional */
    243 
    244 	/*
    245 	 * Will return the value for perf_event_mmap_page::index for this event,
    246 	 * if no implementation is provided it will default to: event->hw.idx + 1.
    247 	 */
    248 	int (*event_idx)		(struct perf_event *event); /*optional */
    249 
    250 	/*
    251 	 * flush branch stack on context-switches (needed in cpu-wide mode)
    252 	 */
    253 	void (*flush_branch_stack)	(void);
    254 };
    255 
    256 /**
    257  * enum perf_event_active_state - the states of a event
    258  */
    259 enum perf_event_active_state {
    260 	PERF_EVENT_STATE_ERROR		= -2,
    261 	PERF_EVENT_STATE_OFF		= -1,
    262 	PERF_EVENT_STATE_INACTIVE	=  0,
    263 	PERF_EVENT_STATE_ACTIVE		=  1,
    264 };
    265 
    266 struct file;
    267 struct perf_sample_data;
    268 
    269 typedef void (*perf_overflow_handler_t)(struct perf_event *,
    270 					struct perf_sample_data *,
    271 					struct pt_regs *regs);
    272 
    273 enum perf_group_flag {
    274 	PERF_GROUP_SOFTWARE		= 0x1,
    275 };
    276 
    277 #define SWEVENT_HLIST_BITS		8
    278 #define SWEVENT_HLIST_SIZE		(1 << SWEVENT_HLIST_BITS)
    279 
    280 struct swevent_hlist {
    281 	struct hlist_head		heads[SWEVENT_HLIST_SIZE];
    282 	struct rcu_head			rcu_head;
    283 };
    284 
    285 #define PERF_ATTACH_CONTEXT	0x01
    286 #define PERF_ATTACH_GROUP	0x02
    287 #define PERF_ATTACH_TASK	0x04
    288 
    289 struct perf_cgroup;
    290 struct ring_buffer;
    291 
    292 /**
    293  * struct perf_event - performance event kernel representation:
    294  */
    295 struct perf_event {
    296 #ifdef CONFIG_PERF_EVENTS
    297 	/*
    298 	 * entry onto perf_event_context::event_list;
    299 	 *   modifications require ctx->lock
    300 	 *   RCU safe iterations.
    301 	 */
    302 	struct list_head		event_entry;
    303 
    304 	/*
    305 	 * XXX: group_entry and sibling_list should be mutually exclusive;
    306 	 * either you're a sibling on a group, or you're the group leader.
    307 	 * Rework the code to always use the same list element.
    308 	 *
    309 	 * Locked for modification by both ctx->mutex and ctx->lock; holding
    310 	 * either sufficies for read.
    311 	 */
    312 	struct list_head		group_entry;
    313 	struct list_head		sibling_list;
    314 
    315 	/*
    316 	 * We need storage to track the entries in perf_pmu_migrate_context; we
    317 	 * cannot use the event_entry because of RCU and we want to keep the
    318 	 * group in tact which avoids us using the other two entries.
    319 	 */
    320 	struct list_head		migrate_entry;
    321 
    322 	struct hlist_node		hlist_entry;
    323 	int				nr_siblings;
    324 	int				group_flags;
    325 	struct perf_event		*group_leader;
    326 	struct pmu			*pmu;
    327 
    328 	enum perf_event_active_state	state;
    329 	unsigned int			attach_state;
    330 	local64_t			count;
    331 	atomic64_t			child_count;
    332 
    333 	/*
    334 	 * These are the total time in nanoseconds that the event
    335 	 * has been enabled (i.e. eligible to run, and the task has
    336 	 * been scheduled in, if this is a per-task event)
    337 	 * and running (scheduled onto the CPU), respectively.
    338 	 *
    339 	 * They are computed from tstamp_enabled, tstamp_running and
    340 	 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
    341 	 */
    342 	u64				total_time_enabled;
    343 	u64				total_time_running;
    344 
    345 	/*
    346 	 * These are timestamps used for computing total_time_enabled
    347 	 * and total_time_running when the event is in INACTIVE or
    348 	 * ACTIVE state, measured in nanoseconds from an arbitrary point
    349 	 * in time.
    350 	 * tstamp_enabled: the notional time when the event was enabled
    351 	 * tstamp_running: the notional time when the event was scheduled on
    352 	 * tstamp_stopped: in INACTIVE state, the notional time when the
    353 	 *	event was scheduled off.
    354 	 */
    355 	u64				tstamp_enabled;
    356 	u64				tstamp_running;
    357 	u64				tstamp_stopped;
    358 
    359 	/*
    360 	 * timestamp shadows the actual context timing but it can
    361 	 * be safely used in NMI interrupt context. It reflects the
    362 	 * context time as it was when the event was last scheduled in.
    363 	 *
    364 	 * ctx_time already accounts for ctx->timestamp. Therefore to
    365 	 * compute ctx_time for a sample, simply add perf_clock().
    366 	 */
    367 	u64				shadow_ctx_time;
    368 
    369 	struct perf_event_attr		attr;
    370 	u16				header_size;
    371 	u16				id_header_size;
    372 	u16				read_size;
    373 	struct hw_perf_event		hw;
    374 
    375 	struct perf_event_context	*ctx;
    376 	atomic_long_t			refcount;
    377 
    378 	/*
    379 	 * These accumulate total time (in nanoseconds) that children
    380 	 * events have been enabled and running, respectively.
    381 	 */
    382 	atomic64_t			child_total_time_enabled;
    383 	atomic64_t			child_total_time_running;
    384 
    385 	/*
    386 	 * Protect attach/detach and child_list:
    387 	 */
    388 	struct mutex			child_mutex;
    389 	struct list_head		child_list;
    390 	struct perf_event		*parent;
    391 
    392 	int				oncpu;
    393 	int				cpu;
    394 
    395 	struct list_head		owner_entry;
    396 	struct task_struct		*owner;
    397 
    398 	/* mmap bits */
    399 	struct mutex			mmap_mutex;
    400 	atomic_t			mmap_count;
    401 
    402 	struct ring_buffer		*rb;
    403 	struct list_head		rb_entry;
    404 
    405 	/* poll related */
    406 	wait_queue_head_t		waitq;
    407 	struct fasync_struct		*fasync;
    408 
    409 	/* delayed work for NMIs and such */
    410 	int				pending_wakeup;
    411 	int				pending_kill;
    412 	int				pending_disable;
    413 	struct irq_work			pending;
    414 
    415 	atomic_t			event_limit;
    416 
    417 	void (*destroy)(struct perf_event *);
    418 	struct rcu_head			rcu_head;
    419 
    420 	struct pid_namespace		*ns;
    421 	u64				id;
    422 
    423 	perf_overflow_handler_t		overflow_handler;
    424 	void				*overflow_handler_context;
    425 
    426 #ifdef CONFIG_EVENT_TRACING
    427 	struct ftrace_event_call	*tp_event;
    428 	struct event_filter		*filter;
    429 #ifdef CONFIG_FUNCTION_TRACER
    430 	struct ftrace_ops               ftrace_ops;
    431 #endif
    432 #endif
    433 
    434 #ifdef CONFIG_CGROUP_PERF
    435 	struct perf_cgroup		*cgrp; /* cgroup event is attach to */
    436 	int				cgrp_defer_enabled;
    437 #endif
    438 
    439 #endif /* CONFIG_PERF_EVENTS */
    440 };
    441 
    442 enum perf_event_context_type {
    443 	task_context,
    444 	cpu_context,
    445 };
    446 
    447 /**
    448  * struct perf_event_context - event context structure
    449  *
    450  * Used as a container for task events and CPU events as well:
    451  */
    452 struct perf_event_context {
    453 	struct pmu			*pmu;
    454 	enum perf_event_context_type	type;
    455 	/*
    456 	 * Protect the states of the events in the list,
    457 	 * nr_active, and the list:
    458 	 */
    459 	raw_spinlock_t			lock;
    460 	/*
    461 	 * Protect the list of events.  Locking either mutex or lock
    462 	 * is sufficient to ensure the list doesn't change; to change
    463 	 * the list you need to lock both the mutex and the spinlock.
    464 	 */
    465 	struct mutex			mutex;
    466 
    467 	struct list_head		pinned_groups;
    468 	struct list_head		flexible_groups;
    469 	struct list_head		event_list;
    470 	int				nr_events;
    471 	int				nr_active;
    472 	int				is_active;
    473 	int				nr_stat;
    474 	int				nr_freq;
    475 	int				rotate_disable;
    476 	atomic_t			refcount;
    477 	struct task_struct		*task;
    478 
    479 	/*
    480 	 * Context clock, runs when context enabled.
    481 	 */
    482 	u64				time;
    483 	u64				timestamp;
    484 
    485 	/*
    486 	 * These fields let us detect when two contexts have both
    487 	 * been cloned (inherited) from a common ancestor.
    488 	 */
    489 	struct perf_event_context	*parent_ctx;
    490 	u64				parent_gen;
    491 	u64				generation;
    492 	int				pin_count;
    493 	int				nr_cgroups;	 /* cgroup evts */
    494 	int				nr_branch_stack; /* branch_stack evt */
    495 	struct rcu_head			rcu_head;
    496 };
    497 
    498 /*
    499  * Number of contexts where an event can trigger:
    500  *	task, softirq, hardirq, nmi.
    501  */
    502 #define PERF_NR_CONTEXTS	4
    503 
    504 /**
    505  * struct perf_event_cpu_context - per cpu event context structure
    506  */
    507 struct perf_cpu_context {
    508 	struct perf_event_context	ctx;
    509 	struct perf_event_context	*task_ctx;
    510 	int				active_oncpu;
    511 	int				exclusive;
    512 	struct hrtimer			hrtimer;
    513 	ktime_t				hrtimer_interval;
    514 	struct list_head		rotation_list;
    515 	struct pmu			*unique_pmu;
    516 	struct perf_cgroup		*cgrp;
    517 };
    518 
    519 struct perf_output_handle {
    520 	struct perf_event		*event;
    521 	struct ring_buffer		*rb;
    522 	unsigned long			wakeup;
    523 	unsigned long			size;
    524 	void				*addr;
    525 	int				page;
    526 };
    527 
    528 #ifdef CONFIG_PERF_EVENTS
    529 
    530 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
    531 extern void perf_pmu_unregister(struct pmu *pmu);
    532 
    533 extern int perf_num_counters(void);
    534 extern const char *perf_pmu_name(void);
    535 extern void __perf_event_task_sched_in(struct task_struct *prev,
    536 				       struct task_struct *task);
    537 extern void __perf_event_task_sched_out(struct task_struct *prev,
    538 					struct task_struct *next);
    539 extern int perf_event_init_task(struct task_struct *child);
    540 extern void perf_event_exit_task(struct task_struct *child);
    541 extern void perf_event_free_task(struct task_struct *task);
    542 extern void perf_event_delayed_put(struct task_struct *task);
    543 extern void perf_event_print_debug(void);
    544 extern void perf_pmu_disable(struct pmu *pmu);
    545 extern void perf_pmu_enable(struct pmu *pmu);
    546 extern int perf_event_task_disable(void);
    547 extern int perf_event_task_enable(void);
    548 extern int perf_event_refresh(struct perf_event *event, int refresh);
    549 extern void perf_event_update_userpage(struct perf_event *event);
    550 extern int perf_event_release_kernel(struct perf_event *event);
    551 extern struct perf_event *
    552 perf_event_create_kernel_counter(struct perf_event_attr *attr,
    553 				int cpu,
    554 				struct task_struct *task,
    555 				perf_overflow_handler_t callback,
    556 				void *context);
    557 extern void perf_pmu_migrate_context(struct pmu *pmu,
    558 				int src_cpu, int dst_cpu);
    559 extern u64 perf_event_read_value(struct perf_event *event,
    560 				 u64 *enabled, u64 *running);
    561 
    562 
    563 struct perf_sample_data {
    564 	u64				type;
    565 
    566 	u64				ip;
    567 	struct {
    568 		u32	pid;
    569 		u32	tid;
    570 	}				tid_entry;
    571 	u64				time;
    572 	u64				addr;
    573 	u64				id;
    574 	u64				stream_id;
    575 	struct {
    576 		u32	cpu;
    577 		u32	reserved;
    578 	}				cpu_entry;
    579 	u64				period;
    580 	union  perf_mem_data_src	data_src;
    581 	struct perf_callchain_entry	*callchain;
    582 	struct perf_raw_record		*raw;
    583 	struct perf_branch_stack	*br_stack;
    584 	struct perf_regs_user		regs_user;
    585 	u64				stack_user_size;
    586 	u64				weight;
    587 };
    588 
    589 static inline void perf_sample_data_init(struct perf_sample_data *data,
    590 					 u64 addr, u64 period)
    591 {
    592 	/* remaining struct members initialized in perf_prepare_sample() */
    593 	data->addr = addr;
    594 	data->raw  = NULL;
    595 	data->br_stack = NULL;
    596 	data->period = period;
    597 	data->regs_user.abi = PERF_SAMPLE_REGS_ABI_NONE;
    598 	data->regs_user.regs = NULL;
    599 	data->stack_user_size = 0;
    600 	data->weight = 0;
    601 	data->data_src.val = 0;
    602 }
    603 
    604 extern void perf_output_sample(struct perf_output_handle *handle,
    605 			       struct perf_event_header *header,
    606 			       struct perf_sample_data *data,
    607 			       struct perf_event *event);
    608 extern void perf_prepare_sample(struct perf_event_header *header,
    609 				struct perf_sample_data *data,
    610 				struct perf_event *event,
    611 				struct pt_regs *regs);
    612 
    613 extern int perf_event_overflow(struct perf_event *event,
    614 				 struct perf_sample_data *data,
    615 				 struct pt_regs *regs);
    616 
    617 static inline bool is_sampling_event(struct perf_event *event)
    618 {
    619 	return event->attr.sample_period != 0;
    620 }
    621 
    622 /*
    623  * Return 1 for a software event, 0 for a hardware event
    624  */
    625 static inline int is_software_event(struct perf_event *event)
    626 {
    627 	return event->pmu->task_ctx_nr == perf_sw_context;
    628 }
    629 
    630 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
    631 
    632 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
    633 
    634 #ifndef perf_arch_fetch_caller_regs
    635 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
    636 #endif
    637 
    638 /*
    639  * Take a snapshot of the regs. Skip ip and frame pointer to
    640  * the nth caller. We only need a few of the regs:
    641  * - ip for PERF_SAMPLE_IP
    642  * - cs for user_mode() tests
    643  * - bp for callchains
    644  * - eflags, for future purposes, just in case
    645  */
    646 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
    647 {
    648 	memset(regs, 0, sizeof(*regs));
    649 
    650 	perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
    651 }
    652 
    653 static __always_inline void
    654 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
    655 {
    656 	struct pt_regs hot_regs;
    657 
    658 	if (static_key_false(&perf_swevent_enabled[event_id])) {
    659 		if (!regs) {
    660 			perf_fetch_caller_regs(&hot_regs);
    661 			regs = &hot_regs;
    662 		}
    663 		__perf_sw_event(event_id, nr, regs, addr);
    664 	}
    665 }
    666 
    667 extern struct static_key_deferred perf_sched_events;
    668 
    669 static inline void perf_event_task_sched_in(struct task_struct *prev,
    670 					    struct task_struct *task)
    671 {
    672 	if (static_key_false(&perf_sched_events.key))
    673 		__perf_event_task_sched_in(prev, task);
    674 }
    675 
    676 static inline void perf_event_task_sched_out(struct task_struct *prev,
    677 					     struct task_struct *next)
    678 {
    679 	perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, NULL, 0);
    680 
    681 	if (static_key_false(&perf_sched_events.key))
    682 		__perf_event_task_sched_out(prev, next);
    683 }
    684 
    685 extern void perf_event_mmap(struct vm_area_struct *vma);
    686 extern struct perf_guest_info_callbacks *perf_guest_cbs;
    687 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
    688 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
    689 
    690 extern void perf_event_comm(struct task_struct *tsk);
    691 extern void perf_event_fork(struct task_struct *tsk);
    692 
    693 /* Callchains */
    694 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
    695 
    696 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
    697 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
    698 
    699 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
    700 {
    701 	if (entry->nr < PERF_MAX_STACK_DEPTH)
    702 		entry->ip[entry->nr++] = ip;
    703 }
    704 
    705 extern int sysctl_perf_event_paranoid;
    706 extern int sysctl_perf_event_mlock;
    707 extern int sysctl_perf_event_sample_rate;
    708 extern int sysctl_perf_cpu_time_max_percent;
    709 
    710 extern void perf_sample_event_took(u64 sample_len_ns);
    711 
    712 extern int perf_proc_update_handler(struct ctl_table *table, int write,
    713 		void __user *buffer, size_t *lenp,
    714 		loff_t *ppos);
    715 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
    716 		void __user *buffer, size_t *lenp,
    717 		loff_t *ppos);
    718 
    719 
    720 static inline bool perf_paranoid_tracepoint_raw(void)
    721 {
    722 	return sysctl_perf_event_paranoid > -1;
    723 }
    724 
    725 static inline bool perf_paranoid_cpu(void)
    726 {
    727 	return sysctl_perf_event_paranoid > 0;
    728 }
    729 
    730 static inline bool perf_paranoid_kernel(void)
    731 {
    732 	return sysctl_perf_event_paranoid > 1;
    733 }
    734 
    735 extern void perf_event_init(void);
    736 extern void perf_tp_event(u64 addr, u64 count, void *record,
    737 			  int entry_size, struct pt_regs *regs,
    738 			  struct hlist_head *head, int rctx,
    739 			  struct task_struct *task);
    740 extern void perf_bp_event(struct perf_event *event, void *data);
    741 
    742 #ifndef perf_misc_flags
    743 # define perf_misc_flags(regs) \
    744 		(user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
    745 # define perf_instruction_pointer(regs)	instruction_pointer(regs)
    746 #endif
    747 
    748 static inline bool has_branch_stack(struct perf_event *event)
    749 {
    750 	return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
    751 }
    752 
    753 extern int perf_output_begin(struct perf_output_handle *handle,
    754 			     struct perf_event *event, unsigned int size);
    755 extern void perf_output_end(struct perf_output_handle *handle);
    756 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
    757 			     const void *buf, unsigned int len);
    758 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
    759 				     unsigned int len);
    760 extern int perf_swevent_get_recursion_context(void);
    761 extern void perf_swevent_put_recursion_context(int rctx);
    762 extern u64 perf_swevent_set_period(struct perf_event *event);
    763 extern void perf_event_enable(struct perf_event *event);
    764 extern void perf_event_disable(struct perf_event *event);
    765 extern int __perf_event_disable(void *info);
    766 extern void perf_event_task_tick(void);
    767 #else
    768 static inline void
    769 perf_event_task_sched_in(struct task_struct *prev,
    770 			 struct task_struct *task)			{ }
    771 static inline void
    772 perf_event_task_sched_out(struct task_struct *prev,
    773 			  struct task_struct *next)			{ }
    774 static inline int perf_event_init_task(struct task_struct *child)	{ return 0; }
    775 static inline void perf_event_exit_task(struct task_struct *child)	{ }
    776 static inline void perf_event_free_task(struct task_struct *task)	{ }
    777 static inline void perf_event_delayed_put(struct task_struct *task)	{ }
    778 static inline void perf_event_print_debug(void)				{ }
    779 static inline int perf_event_task_disable(void)				{ return -EINVAL; }
    780 static inline int perf_event_task_enable(void)				{ return -EINVAL; }
    781 static inline int perf_event_refresh(struct perf_event *event, int refresh)
    782 {
    783 	return -EINVAL;
    784 }
    785 
    786 static inline void
    787 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)	{ }
    788 static inline void
    789 perf_bp_event(struct perf_event *event, void *data)			{ }
    790 
    791 static inline int perf_register_guest_info_callbacks
    792 (struct perf_guest_info_callbacks *callbacks)				{ return 0; }
    793 static inline int perf_unregister_guest_info_callbacks
    794 (struct perf_guest_info_callbacks *callbacks)				{ return 0; }
    795 
    796 static inline void perf_event_mmap(struct vm_area_struct *vma)		{ }
    797 static inline void perf_event_comm(struct task_struct *tsk)		{ }
    798 static inline void perf_event_fork(struct task_struct *tsk)		{ }
    799 static inline void perf_event_init(void)				{ }
    800 static inline int  perf_swevent_get_recursion_context(void)		{ return -1; }
    801 static inline void perf_swevent_put_recursion_context(int rctx)		{ }
    802 static inline u64 perf_swevent_set_period(struct perf_event *event)	{ return 0; }
    803 static inline void perf_event_enable(struct perf_event *event)		{ }
    804 static inline void perf_event_disable(struct perf_event *event)		{ }
    805 static inline int __perf_event_disable(void *info)			{ return -1; }
    806 static inline void perf_event_task_tick(void)				{ }
    807 #endif
    808 
    809 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_NO_HZ_FULL)
    810 extern bool perf_event_can_stop_tick(void);
    811 #else
    812 static inline bool perf_event_can_stop_tick(void)			{ return true; }
    813 #endif
    814 
    815 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
    816 extern void perf_restore_debug_store(void);
    817 #else
    818 static inline void perf_restore_debug_store(void)			{ }
    819 #endif
    820 
    821 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
    822 
    823 /*
    824  * This has to have a higher priority than migration_notifier in sched/core.c.
    825  */
    826 #define perf_cpu_notifier(fn)						\
    827 do {									\
    828 	static struct notifier_block fn##_nb =				\
    829 		{ .notifier_call = fn, .priority = CPU_PRI_PERF };	\
    830 	unsigned long cpu = smp_processor_id();				\
    831 	unsigned long flags;						\
    832 	fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,			\
    833 		(void *)(unsigned long)cpu);				\
    834 	local_irq_save(flags);						\
    835 	fn(&fn##_nb, (unsigned long)CPU_STARTING,			\
    836 		(void *)(unsigned long)cpu);				\
    837 	local_irq_restore(flags);					\
    838 	fn(&fn##_nb, (unsigned long)CPU_ONLINE,				\
    839 		(void *)(unsigned long)cpu);				\
    840 	register_cpu_notifier(&fn##_nb);				\
    841 } while (0)
    842 
    843 
    844 struct perf_pmu_events_attr {
    845 	struct device_attribute attr;
    846 	u64 id;
    847 	const char *event_str;
    848 };
    849 
    850 #define PMU_EVENT_ATTR(_name, _var, _id, _show)				\
    851 static struct perf_pmu_events_attr _var = {				\
    852 	.attr = __ATTR(_name, 0444, _show, NULL),			\
    853 	.id   =  _id,							\
    854 };
    855 
    856 #define PMU_FORMAT_ATTR(_name, _format)					\
    857 static ssize_t								\
    858 _name##_show(struct device *dev,					\
    859 			       struct device_attribute *attr,		\
    860 			       char *page)				\
    861 {									\
    862 	BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);			\
    863 	return sprintf(page, _format "\n");				\
    864 }									\
    865 									\
    866 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
    867 
    868 #endif /* _LINUX_PERF_EVENT_H */
    869