Merge tag 'v4.6-rc3' into perf/core, to refresh the tree
[cascardo/linux.git] / include / linux / perf_event.h
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
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 <linux/workqueue.h>
56 #include <linux/cgroup.h>
57 #include <asm/local.h>
58
59 struct perf_callchain_entry {
60         __u64                           nr;
61         __u64                           ip[PERF_MAX_STACK_DEPTH];
62 };
63
64 struct perf_raw_record {
65         u32                             size;
66         void                            *data;
67 };
68
69 /*
70  * branch stack layout:
71  *  nr: number of taken branches stored in entries[]
72  *
73  * Note that nr can vary from sample to sample
74  * branches (to, from) are stored from most recent
75  * to least recent, i.e., entries[0] contains the most
76  * recent branch.
77  */
78 struct perf_branch_stack {
79         __u64                           nr;
80         struct perf_branch_entry        entries[0];
81 };
82
83 struct task_struct;
84
85 /*
86  * extra PMU register associated with an event
87  */
88 struct hw_perf_event_extra {
89         u64             config; /* register value */
90         unsigned int    reg;    /* register address or index */
91         int             alloc;  /* extra register already allocated */
92         int             idx;    /* index in shared_regs->regs[] */
93 };
94
95 /**
96  * struct hw_perf_event - performance event hardware details:
97  */
98 struct hw_perf_event {
99 #ifdef CONFIG_PERF_EVENTS
100         union {
101                 struct { /* hardware */
102                         u64             config;
103                         u64             last_tag;
104                         unsigned long   config_base;
105                         unsigned long   event_base;
106                         int             event_base_rdpmc;
107                         int             idx;
108                         int             last_cpu;
109                         int             flags;
110
111                         struct hw_perf_event_extra extra_reg;
112                         struct hw_perf_event_extra branch_reg;
113                 };
114                 struct { /* software */
115                         struct hrtimer  hrtimer;
116                 };
117                 struct { /* tracepoint */
118                         /* for tp_event->class */
119                         struct list_head        tp_list;
120                 };
121                 struct { /* intel_cqm */
122                         int                     cqm_state;
123                         u32                     cqm_rmid;
124                         int                     is_group_event;
125                         struct list_head        cqm_events_entry;
126                         struct list_head        cqm_groups_entry;
127                         struct list_head        cqm_group_entry;
128                 };
129                 struct { /* itrace */
130                         int                     itrace_started;
131                 };
132                 struct { /* amd_power */
133                         u64     pwr_acc;
134                         u64     ptsc;
135                 };
136 #ifdef CONFIG_HAVE_HW_BREAKPOINT
137                 struct { /* breakpoint */
138                         /*
139                          * Crufty hack to avoid the chicken and egg
140                          * problem hw_breakpoint has with context
141                          * creation and event initalization.
142                          */
143                         struct arch_hw_breakpoint       info;
144                         struct list_head                bp_list;
145                 };
146 #endif
147         };
148         /*
149          * If the event is a per task event, this will point to the task in
150          * question. See the comment in perf_event_alloc().
151          */
152         struct task_struct              *target;
153
154 /*
155  * hw_perf_event::state flags; used to track the PERF_EF_* state.
156  */
157 #define PERF_HES_STOPPED        0x01 /* the counter is stopped */
158 #define PERF_HES_UPTODATE       0x02 /* event->count up-to-date */
159 #define PERF_HES_ARCH           0x04
160
161         int                             state;
162
163         /*
164          * The last observed hardware counter value, updated with a
165          * local64_cmpxchg() such that pmu::read() can be called nested.
166          */
167         local64_t                       prev_count;
168
169         /*
170          * The period to start the next sample with.
171          */
172         u64                             sample_period;
173
174         /*
175          * The period we started this sample with.
176          */
177         u64                             last_period;
178
179         /*
180          * However much is left of the current period; note that this is
181          * a full 64bit value and allows for generation of periods longer
182          * than hardware might allow.
183          */
184         local64_t                       period_left;
185
186         /*
187          * State for throttling the event, see __perf_event_overflow() and
188          * perf_adjust_freq_unthr_context().
189          */
190         u64                             interrupts_seq;
191         u64                             interrupts;
192
193         /*
194          * State for freq target events, see __perf_event_overflow() and
195          * perf_adjust_freq_unthr_context().
196          */
197         u64                             freq_time_stamp;
198         u64                             freq_count_stamp;
199 #endif
200 };
201
202 struct perf_event;
203
204 /*
205  * Common implementation detail of pmu::{start,commit,cancel}_txn
206  */
207 #define PERF_PMU_TXN_ADD  0x1           /* txn to add/schedule event on PMU */
208 #define PERF_PMU_TXN_READ 0x2           /* txn to read event group from PMU */
209
210 /**
211  * pmu::capabilities flags
212  */
213 #define PERF_PMU_CAP_NO_INTERRUPT               0x01
214 #define PERF_PMU_CAP_NO_NMI                     0x02
215 #define PERF_PMU_CAP_AUX_NO_SG                  0x04
216 #define PERF_PMU_CAP_AUX_SW_DOUBLEBUF           0x08
217 #define PERF_PMU_CAP_EXCLUSIVE                  0x10
218 #define PERF_PMU_CAP_ITRACE                     0x20
219
220 /**
221  * struct pmu - generic performance monitoring unit
222  */
223 struct pmu {
224         struct list_head                entry;
225
226         struct module                   *module;
227         struct device                   *dev;
228         const struct attribute_group    **attr_groups;
229         const char                      *name;
230         int                             type;
231
232         /*
233          * various common per-pmu feature flags
234          */
235         int                             capabilities;
236
237         int * __percpu                  pmu_disable_count;
238         struct perf_cpu_context * __percpu pmu_cpu_context;
239         atomic_t                        exclusive_cnt; /* < 0: cpu; > 0: tsk */
240         int                             task_ctx_nr;
241         int                             hrtimer_interval_ms;
242
243         /*
244          * Fully disable/enable this PMU, can be used to protect from the PMI
245          * as well as for lazy/batch writing of the MSRs.
246          */
247         void (*pmu_enable)              (struct pmu *pmu); /* optional */
248         void (*pmu_disable)             (struct pmu *pmu); /* optional */
249
250         /*
251          * Try and initialize the event for this PMU.
252          *
253          * Returns:
254          *  -ENOENT     -- @event is not for this PMU
255          *
256          *  -ENODEV     -- @event is for this PMU but PMU not present
257          *  -EBUSY      -- @event is for this PMU but PMU temporarily unavailable
258          *  -EINVAL     -- @event is for this PMU but @event is not valid
259          *  -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
260          *  -EACCESS    -- @event is for this PMU, @event is valid, but no privilidges
261          *
262          *  0           -- @event is for this PMU and valid
263          *
264          * Other error return values are allowed.
265          */
266         int (*event_init)               (struct perf_event *event);
267
268         /*
269          * Notification that the event was mapped or unmapped.  Called
270          * in the context of the mapping task.
271          */
272         void (*event_mapped)            (struct perf_event *event); /*optional*/
273         void (*event_unmapped)          (struct perf_event *event); /*optional*/
274
275         /*
276          * Flags for ->add()/->del()/ ->start()/->stop(). There are
277          * matching hw_perf_event::state flags.
278          */
279 #define PERF_EF_START   0x01            /* start the counter when adding    */
280 #define PERF_EF_RELOAD  0x02            /* reload the counter when starting */
281 #define PERF_EF_UPDATE  0x04            /* update the counter when stopping */
282
283         /*
284          * Adds/Removes a counter to/from the PMU, can be done inside a
285          * transaction, see the ->*_txn() methods.
286          *
287          * The add/del callbacks will reserve all hardware resources required
288          * to service the event, this includes any counter constraint
289          * scheduling etc.
290          *
291          * Called with IRQs disabled and the PMU disabled on the CPU the event
292          * is on.
293          *
294          * ->add() called without PERF_EF_START should result in the same state
295          *  as ->add() followed by ->stop().
296          *
297          * ->del() must always PERF_EF_UPDATE stop an event. If it calls
298          *  ->stop() that must deal with already being stopped without
299          *  PERF_EF_UPDATE.
300          */
301         int  (*add)                     (struct perf_event *event, int flags);
302         void (*del)                     (struct perf_event *event, int flags);
303
304         /*
305          * Starts/Stops a counter present on the PMU.
306          *
307          * The PMI handler should stop the counter when perf_event_overflow()
308          * returns !0. ->start() will be used to continue.
309          *
310          * Also used to change the sample period.
311          *
312          * Called with IRQs disabled and the PMU disabled on the CPU the event
313          * is on -- will be called from NMI context with the PMU generates
314          * NMIs.
315          *
316          * ->stop() with PERF_EF_UPDATE will read the counter and update
317          *  period/count values like ->read() would.
318          *
319          * ->start() with PERF_EF_RELOAD will reprogram the the counter
320          *  value, must be preceded by a ->stop() with PERF_EF_UPDATE.
321          */
322         void (*start)                   (struct perf_event *event, int flags);
323         void (*stop)                    (struct perf_event *event, int flags);
324
325         /*
326          * Updates the counter value of the event.
327          *
328          * For sampling capable PMUs this will also update the software period
329          * hw_perf_event::period_left field.
330          */
331         void (*read)                    (struct perf_event *event);
332
333         /*
334          * Group events scheduling is treated as a transaction, add
335          * group events as a whole and perform one schedulability test.
336          * If the test fails, roll back the whole group
337          *
338          * Start the transaction, after this ->add() doesn't need to
339          * do schedulability tests.
340          *
341          * Optional.
342          */
343         void (*start_txn)               (struct pmu *pmu, unsigned int txn_flags);
344         /*
345          * If ->start_txn() disabled the ->add() schedulability test
346          * then ->commit_txn() is required to perform one. On success
347          * the transaction is closed. On error the transaction is kept
348          * open until ->cancel_txn() is called.
349          *
350          * Optional.
351          */
352         int  (*commit_txn)              (struct pmu *pmu);
353         /*
354          * Will cancel the transaction, assumes ->del() is called
355          * for each successful ->add() during the transaction.
356          *
357          * Optional.
358          */
359         void (*cancel_txn)              (struct pmu *pmu);
360
361         /*
362          * Will return the value for perf_event_mmap_page::index for this event,
363          * if no implementation is provided it will default to: event->hw.idx + 1.
364          */
365         int (*event_idx)                (struct perf_event *event); /*optional */
366
367         /*
368          * context-switches callback
369          */
370         void (*sched_task)              (struct perf_event_context *ctx,
371                                         bool sched_in);
372         /*
373          * PMU specific data size
374          */
375         size_t                          task_ctx_size;
376
377
378         /*
379          * Return the count value for a counter.
380          */
381         u64 (*count)                    (struct perf_event *event); /*optional*/
382
383         /*
384          * Set up pmu-private data structures for an AUX area
385          */
386         void *(*setup_aux)              (int cpu, void **pages,
387                                          int nr_pages, bool overwrite);
388                                         /* optional */
389
390         /*
391          * Free pmu-private AUX data structures
392          */
393         void (*free_aux)                (void *aux); /* optional */
394
395         /*
396          * Filter events for PMU-specific reasons.
397          */
398         int (*filter_match)             (struct perf_event *event); /* optional */
399 };
400
401 /**
402  * enum perf_event_active_state - the states of a event
403  */
404 enum perf_event_active_state {
405         PERF_EVENT_STATE_DEAD           = -4,
406         PERF_EVENT_STATE_EXIT           = -3,
407         PERF_EVENT_STATE_ERROR          = -2,
408         PERF_EVENT_STATE_OFF            = -1,
409         PERF_EVENT_STATE_INACTIVE       =  0,
410         PERF_EVENT_STATE_ACTIVE         =  1,
411 };
412
413 struct file;
414 struct perf_sample_data;
415
416 typedef void (*perf_overflow_handler_t)(struct perf_event *,
417                                         struct perf_sample_data *,
418                                         struct pt_regs *regs);
419
420 enum perf_group_flag {
421         PERF_GROUP_SOFTWARE             = 0x1,
422 };
423
424 #define SWEVENT_HLIST_BITS              8
425 #define SWEVENT_HLIST_SIZE              (1 << SWEVENT_HLIST_BITS)
426
427 struct swevent_hlist {
428         struct hlist_head               heads[SWEVENT_HLIST_SIZE];
429         struct rcu_head                 rcu_head;
430 };
431
432 #define PERF_ATTACH_CONTEXT     0x01
433 #define PERF_ATTACH_GROUP       0x02
434 #define PERF_ATTACH_TASK        0x04
435 #define PERF_ATTACH_TASK_DATA   0x08
436
437 struct perf_cgroup;
438 struct ring_buffer;
439
440 /**
441  * struct perf_event - performance event kernel representation:
442  */
443 struct perf_event {
444 #ifdef CONFIG_PERF_EVENTS
445         /*
446          * entry onto perf_event_context::event_list;
447          *   modifications require ctx->lock
448          *   RCU safe iterations.
449          */
450         struct list_head                event_entry;
451
452         /*
453          * XXX: group_entry and sibling_list should be mutually exclusive;
454          * either you're a sibling on a group, or you're the group leader.
455          * Rework the code to always use the same list element.
456          *
457          * Locked for modification by both ctx->mutex and ctx->lock; holding
458          * either sufficies for read.
459          */
460         struct list_head                group_entry;
461         struct list_head                sibling_list;
462
463         /*
464          * We need storage to track the entries in perf_pmu_migrate_context; we
465          * cannot use the event_entry because of RCU and we want to keep the
466          * group in tact which avoids us using the other two entries.
467          */
468         struct list_head                migrate_entry;
469
470         struct hlist_node               hlist_entry;
471         struct list_head                active_entry;
472         int                             nr_siblings;
473         int                             group_flags;
474         struct perf_event               *group_leader;
475         struct pmu                      *pmu;
476         void                            *pmu_private;
477
478         enum perf_event_active_state    state;
479         unsigned int                    attach_state;
480         local64_t                       count;
481         atomic64_t                      child_count;
482
483         /*
484          * These are the total time in nanoseconds that the event
485          * has been enabled (i.e. eligible to run, and the task has
486          * been scheduled in, if this is a per-task event)
487          * and running (scheduled onto the CPU), respectively.
488          *
489          * They are computed from tstamp_enabled, tstamp_running and
490          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
491          */
492         u64                             total_time_enabled;
493         u64                             total_time_running;
494
495         /*
496          * These are timestamps used for computing total_time_enabled
497          * and total_time_running when the event is in INACTIVE or
498          * ACTIVE state, measured in nanoseconds from an arbitrary point
499          * in time.
500          * tstamp_enabled: the notional time when the event was enabled
501          * tstamp_running: the notional time when the event was scheduled on
502          * tstamp_stopped: in INACTIVE state, the notional time when the
503          *      event was scheduled off.
504          */
505         u64                             tstamp_enabled;
506         u64                             tstamp_running;
507         u64                             tstamp_stopped;
508
509         /*
510          * timestamp shadows the actual context timing but it can
511          * be safely used in NMI interrupt context. It reflects the
512          * context time as it was when the event was last scheduled in.
513          *
514          * ctx_time already accounts for ctx->timestamp. Therefore to
515          * compute ctx_time for a sample, simply add perf_clock().
516          */
517         u64                             shadow_ctx_time;
518
519         struct perf_event_attr          attr;
520         u16                             header_size;
521         u16                             id_header_size;
522         u16                             read_size;
523         struct hw_perf_event            hw;
524
525         struct perf_event_context       *ctx;
526         atomic_long_t                   refcount;
527
528         /*
529          * These accumulate total time (in nanoseconds) that children
530          * events have been enabled and running, respectively.
531          */
532         atomic64_t                      child_total_time_enabled;
533         atomic64_t                      child_total_time_running;
534
535         /*
536          * Protect attach/detach and child_list:
537          */
538         struct mutex                    child_mutex;
539         struct list_head                child_list;
540         struct perf_event               *parent;
541
542         int                             oncpu;
543         int                             cpu;
544
545         struct list_head                owner_entry;
546         struct task_struct              *owner;
547
548         /* mmap bits */
549         struct mutex                    mmap_mutex;
550         atomic_t                        mmap_count;
551
552         struct ring_buffer              *rb;
553         struct list_head                rb_entry;
554         unsigned long                   rcu_batches;
555         int                             rcu_pending;
556
557         /* poll related */
558         wait_queue_head_t               waitq;
559         struct fasync_struct            *fasync;
560
561         /* delayed work for NMIs and such */
562         int                             pending_wakeup;
563         int                             pending_kill;
564         int                             pending_disable;
565         struct irq_work                 pending;
566
567         atomic_t                        event_limit;
568
569         void (*destroy)(struct perf_event *);
570         struct rcu_head                 rcu_head;
571
572         struct pid_namespace            *ns;
573         u64                             id;
574
575         u64                             (*clock)(void);
576         perf_overflow_handler_t         overflow_handler;
577         void                            *overflow_handler_context;
578
579 #ifdef CONFIG_EVENT_TRACING
580         struct trace_event_call         *tp_event;
581         struct event_filter             *filter;
582 #ifdef CONFIG_FUNCTION_TRACER
583         struct ftrace_ops               ftrace_ops;
584 #endif
585 #endif
586
587 #ifdef CONFIG_CGROUP_PERF
588         struct perf_cgroup              *cgrp; /* cgroup event is attach to */
589         int                             cgrp_defer_enabled;
590 #endif
591
592 #endif /* CONFIG_PERF_EVENTS */
593 };
594
595 /**
596  * struct perf_event_context - event context structure
597  *
598  * Used as a container for task events and CPU events as well:
599  */
600 struct perf_event_context {
601         struct pmu                      *pmu;
602         /*
603          * Protect the states of the events in the list,
604          * nr_active, and the list:
605          */
606         raw_spinlock_t                  lock;
607         /*
608          * Protect the list of events.  Locking either mutex or lock
609          * is sufficient to ensure the list doesn't change; to change
610          * the list you need to lock both the mutex and the spinlock.
611          */
612         struct mutex                    mutex;
613
614         struct list_head                active_ctx_list;
615         struct list_head                pinned_groups;
616         struct list_head                flexible_groups;
617         struct list_head                event_list;
618         int                             nr_events;
619         int                             nr_active;
620         int                             is_active;
621         int                             nr_stat;
622         int                             nr_freq;
623         int                             rotate_disable;
624         atomic_t                        refcount;
625         struct task_struct              *task;
626
627         /*
628          * Context clock, runs when context enabled.
629          */
630         u64                             time;
631         u64                             timestamp;
632
633         /*
634          * These fields let us detect when two contexts have both
635          * been cloned (inherited) from a common ancestor.
636          */
637         struct perf_event_context       *parent_ctx;
638         u64                             parent_gen;
639         u64                             generation;
640         int                             pin_count;
641         int                             nr_cgroups;      /* cgroup evts */
642         void                            *task_ctx_data; /* pmu specific data */
643         struct rcu_head                 rcu_head;
644 };
645
646 /*
647  * Number of contexts where an event can trigger:
648  *      task, softirq, hardirq, nmi.
649  */
650 #define PERF_NR_CONTEXTS        4
651
652 /**
653  * struct perf_event_cpu_context - per cpu event context structure
654  */
655 struct perf_cpu_context {
656         struct perf_event_context       ctx;
657         struct perf_event_context       *task_ctx;
658         int                             active_oncpu;
659         int                             exclusive;
660
661         raw_spinlock_t                  hrtimer_lock;
662         struct hrtimer                  hrtimer;
663         ktime_t                         hrtimer_interval;
664         unsigned int                    hrtimer_active;
665
666         struct pmu                      *unique_pmu;
667         struct perf_cgroup              *cgrp;
668 };
669
670 struct perf_output_handle {
671         struct perf_event               *event;
672         struct ring_buffer              *rb;
673         unsigned long                   wakeup;
674         unsigned long                   size;
675         union {
676                 void                    *addr;
677                 unsigned long           head;
678         };
679         int                             page;
680 };
681
682 #ifdef CONFIG_CGROUP_PERF
683
684 /*
685  * perf_cgroup_info keeps track of time_enabled for a cgroup.
686  * This is a per-cpu dynamically allocated data structure.
687  */
688 struct perf_cgroup_info {
689         u64                             time;
690         u64                             timestamp;
691 };
692
693 struct perf_cgroup {
694         struct cgroup_subsys_state      css;
695         struct perf_cgroup_info __percpu *info;
696 };
697
698 /*
699  * Must ensure cgroup is pinned (css_get) before calling
700  * this function. In other words, we cannot call this function
701  * if there is no cgroup event for the current CPU context.
702  */
703 static inline struct perf_cgroup *
704 perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
705 {
706         return container_of(task_css_check(task, perf_event_cgrp_id,
707                                            ctx ? lockdep_is_held(&ctx->lock)
708                                                : true),
709                             struct perf_cgroup, css);
710 }
711 #endif /* CONFIG_CGROUP_PERF */
712
713 #ifdef CONFIG_PERF_EVENTS
714
715 extern void *perf_aux_output_begin(struct perf_output_handle *handle,
716                                    struct perf_event *event);
717 extern void perf_aux_output_end(struct perf_output_handle *handle,
718                                 unsigned long size, bool truncated);
719 extern int perf_aux_output_skip(struct perf_output_handle *handle,
720                                 unsigned long size);
721 extern void *perf_get_aux(struct perf_output_handle *handle);
722
723 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
724 extern void perf_pmu_unregister(struct pmu *pmu);
725
726 extern int perf_num_counters(void);
727 extern const char *perf_pmu_name(void);
728 extern void __perf_event_task_sched_in(struct task_struct *prev,
729                                        struct task_struct *task);
730 extern void __perf_event_task_sched_out(struct task_struct *prev,
731                                         struct task_struct *next);
732 extern int perf_event_init_task(struct task_struct *child);
733 extern void perf_event_exit_task(struct task_struct *child);
734 extern void perf_event_free_task(struct task_struct *task);
735 extern void perf_event_delayed_put(struct task_struct *task);
736 extern struct file *perf_event_get(unsigned int fd);
737 extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
738 extern void perf_event_print_debug(void);
739 extern void perf_pmu_disable(struct pmu *pmu);
740 extern void perf_pmu_enable(struct pmu *pmu);
741 extern void perf_sched_cb_dec(struct pmu *pmu);
742 extern void perf_sched_cb_inc(struct pmu *pmu);
743 extern int perf_event_task_disable(void);
744 extern int perf_event_task_enable(void);
745 extern int perf_event_refresh(struct perf_event *event, int refresh);
746 extern void perf_event_update_userpage(struct perf_event *event);
747 extern int perf_event_release_kernel(struct perf_event *event);
748 extern struct perf_event *
749 perf_event_create_kernel_counter(struct perf_event_attr *attr,
750                                 int cpu,
751                                 struct task_struct *task,
752                                 perf_overflow_handler_t callback,
753                                 void *context);
754 extern void perf_pmu_migrate_context(struct pmu *pmu,
755                                 int src_cpu, int dst_cpu);
756 extern u64 perf_event_read_local(struct perf_event *event);
757 extern u64 perf_event_read_value(struct perf_event *event,
758                                  u64 *enabled, u64 *running);
759
760
761 struct perf_sample_data {
762         /*
763          * Fields set by perf_sample_data_init(), group so as to
764          * minimize the cachelines touched.
765          */
766         u64                             addr;
767         struct perf_raw_record          *raw;
768         struct perf_branch_stack        *br_stack;
769         u64                             period;
770         u64                             weight;
771         u64                             txn;
772         union  perf_mem_data_src        data_src;
773
774         /*
775          * The other fields, optionally {set,used} by
776          * perf_{prepare,output}_sample().
777          */
778         u64                             type;
779         u64                             ip;
780         struct {
781                 u32     pid;
782                 u32     tid;
783         }                               tid_entry;
784         u64                             time;
785         u64                             id;
786         u64                             stream_id;
787         struct {
788                 u32     cpu;
789                 u32     reserved;
790         }                               cpu_entry;
791         struct perf_callchain_entry     *callchain;
792
793         /*
794          * regs_user may point to task_pt_regs or to regs_user_copy, depending
795          * on arch details.
796          */
797         struct perf_regs                regs_user;
798         struct pt_regs                  regs_user_copy;
799
800         struct perf_regs                regs_intr;
801         u64                             stack_user_size;
802 } ____cacheline_aligned;
803
804 /* default value for data source */
805 #define PERF_MEM_NA (PERF_MEM_S(OP, NA)   |\
806                     PERF_MEM_S(LVL, NA)   |\
807                     PERF_MEM_S(SNOOP, NA) |\
808                     PERF_MEM_S(LOCK, NA)  |\
809                     PERF_MEM_S(TLB, NA))
810
811 static inline void perf_sample_data_init(struct perf_sample_data *data,
812                                          u64 addr, u64 period)
813 {
814         /* remaining struct members initialized in perf_prepare_sample() */
815         data->addr = addr;
816         data->raw  = NULL;
817         data->br_stack = NULL;
818         data->period = period;
819         data->weight = 0;
820         data->data_src.val = PERF_MEM_NA;
821         data->txn = 0;
822 }
823
824 extern void perf_output_sample(struct perf_output_handle *handle,
825                                struct perf_event_header *header,
826                                struct perf_sample_data *data,
827                                struct perf_event *event);
828 extern void perf_prepare_sample(struct perf_event_header *header,
829                                 struct perf_sample_data *data,
830                                 struct perf_event *event,
831                                 struct pt_regs *regs);
832
833 extern int perf_event_overflow(struct perf_event *event,
834                                  struct perf_sample_data *data,
835                                  struct pt_regs *regs);
836
837 extern void perf_event_output(struct perf_event *event,
838                                 struct perf_sample_data *data,
839                                 struct pt_regs *regs);
840
841 static inline bool
842 is_default_overflow_handler(struct perf_event *event)
843 {
844         return (event->overflow_handler == perf_event_output);
845 }
846
847 extern void
848 perf_event_header__init_id(struct perf_event_header *header,
849                            struct perf_sample_data *data,
850                            struct perf_event *event);
851 extern void
852 perf_event__output_id_sample(struct perf_event *event,
853                              struct perf_output_handle *handle,
854                              struct perf_sample_data *sample);
855
856 extern void
857 perf_log_lost_samples(struct perf_event *event, u64 lost);
858
859 static inline bool is_sampling_event(struct perf_event *event)
860 {
861         return event->attr.sample_period != 0;
862 }
863
864 /*
865  * Return 1 for a software event, 0 for a hardware event
866  */
867 static inline int is_software_event(struct perf_event *event)
868 {
869         return event->pmu->task_ctx_nr == perf_sw_context;
870 }
871
872 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
873
874 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
875 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
876
877 #ifndef perf_arch_fetch_caller_regs
878 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
879 #endif
880
881 /*
882  * Take a snapshot of the regs. Skip ip and frame pointer to
883  * the nth caller. We only need a few of the regs:
884  * - ip for PERF_SAMPLE_IP
885  * - cs for user_mode() tests
886  * - bp for callchains
887  * - eflags, for future purposes, just in case
888  */
889 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
890 {
891         memset(regs, 0, sizeof(*regs));
892
893         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
894 }
895
896 static __always_inline void
897 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
898 {
899         if (static_key_false(&perf_swevent_enabled[event_id]))
900                 __perf_sw_event(event_id, nr, regs, addr);
901 }
902
903 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
904
905 /*
906  * 'Special' version for the scheduler, it hard assumes no recursion,
907  * which is guaranteed by us not actually scheduling inside other swevents
908  * because those disable preemption.
909  */
910 static __always_inline void
911 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
912 {
913         if (static_key_false(&perf_swevent_enabled[event_id])) {
914                 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
915
916                 perf_fetch_caller_regs(regs);
917                 ___perf_sw_event(event_id, nr, regs, addr);
918         }
919 }
920
921 extern struct static_key_false perf_sched_events;
922
923 static __always_inline bool
924 perf_sw_migrate_enabled(void)
925 {
926         if (static_key_false(&perf_swevent_enabled[PERF_COUNT_SW_CPU_MIGRATIONS]))
927                 return true;
928         return false;
929 }
930
931 static inline void perf_event_task_migrate(struct task_struct *task)
932 {
933         if (perf_sw_migrate_enabled())
934                 task->sched_migrated = 1;
935 }
936
937 static inline void perf_event_task_sched_in(struct task_struct *prev,
938                                             struct task_struct *task)
939 {
940         if (static_branch_unlikely(&perf_sched_events))
941                 __perf_event_task_sched_in(prev, task);
942
943         if (perf_sw_migrate_enabled() && task->sched_migrated) {
944                 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
945
946                 perf_fetch_caller_regs(regs);
947                 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, regs, 0);
948                 task->sched_migrated = 0;
949         }
950 }
951
952 static inline void perf_event_task_sched_out(struct task_struct *prev,
953                                              struct task_struct *next)
954 {
955         perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
956
957         if (static_branch_unlikely(&perf_sched_events))
958                 __perf_event_task_sched_out(prev, next);
959 }
960
961 static inline u64 __perf_event_count(struct perf_event *event)
962 {
963         return local64_read(&event->count) + atomic64_read(&event->child_count);
964 }
965
966 extern void perf_event_mmap(struct vm_area_struct *vma);
967 extern struct perf_guest_info_callbacks *perf_guest_cbs;
968 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
969 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
970
971 extern void perf_event_exec(void);
972 extern void perf_event_comm(struct task_struct *tsk, bool exec);
973 extern void perf_event_fork(struct task_struct *tsk);
974
975 /* Callchains */
976 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
977
978 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
979 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
980 extern struct perf_callchain_entry *
981 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
982                    bool crosstask, bool add_mark);
983 extern int get_callchain_buffers(void);
984 extern void put_callchain_buffers(void);
985
986 static inline int perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
987 {
988         if (entry->nr < PERF_MAX_STACK_DEPTH) {
989                 entry->ip[entry->nr++] = ip;
990                 return 0;
991         } else {
992                 return -1; /* no more room, stop walking the stack */
993         }
994 }
995
996 extern int sysctl_perf_event_paranoid;
997 extern int sysctl_perf_event_mlock;
998 extern int sysctl_perf_event_sample_rate;
999 extern int sysctl_perf_cpu_time_max_percent;
1000
1001 extern void perf_sample_event_took(u64 sample_len_ns);
1002
1003 extern int perf_proc_update_handler(struct ctl_table *table, int write,
1004                 void __user *buffer, size_t *lenp,
1005                 loff_t *ppos);
1006 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
1007                 void __user *buffer, size_t *lenp,
1008                 loff_t *ppos);
1009
1010
1011 static inline bool perf_paranoid_tracepoint_raw(void)
1012 {
1013         return sysctl_perf_event_paranoid > -1;
1014 }
1015
1016 static inline bool perf_paranoid_cpu(void)
1017 {
1018         return sysctl_perf_event_paranoid > 0;
1019 }
1020
1021 static inline bool perf_paranoid_kernel(void)
1022 {
1023         return sysctl_perf_event_paranoid > 1;
1024 }
1025
1026 extern void perf_event_init(void);
1027 extern void perf_tp_event(u64 addr, u64 count, void *record,
1028                           int entry_size, struct pt_regs *regs,
1029                           struct hlist_head *head, int rctx,
1030                           struct task_struct *task);
1031 extern void perf_bp_event(struct perf_event *event, void *data);
1032
1033 #ifndef perf_misc_flags
1034 # define perf_misc_flags(regs) \
1035                 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1036 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1037 #endif
1038
1039 static inline bool has_branch_stack(struct perf_event *event)
1040 {
1041         return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1042 }
1043
1044 static inline bool needs_branch_stack(struct perf_event *event)
1045 {
1046         return event->attr.branch_sample_type != 0;
1047 }
1048
1049 static inline bool has_aux(struct perf_event *event)
1050 {
1051         return event->pmu->setup_aux;
1052 }
1053
1054 extern int perf_output_begin(struct perf_output_handle *handle,
1055                              struct perf_event *event, unsigned int size);
1056 extern void perf_output_end(struct perf_output_handle *handle);
1057 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
1058                              const void *buf, unsigned int len);
1059 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
1060                                      unsigned int len);
1061 extern int perf_swevent_get_recursion_context(void);
1062 extern void perf_swevent_put_recursion_context(int rctx);
1063 extern u64 perf_swevent_set_period(struct perf_event *event);
1064 extern void perf_event_enable(struct perf_event *event);
1065 extern void perf_event_disable(struct perf_event *event);
1066 extern void perf_event_disable_local(struct perf_event *event);
1067 extern void perf_event_task_tick(void);
1068 #else /* !CONFIG_PERF_EVENTS: */
1069 static inline void *
1070 perf_aux_output_begin(struct perf_output_handle *handle,
1071                       struct perf_event *event)                         { return NULL; }
1072 static inline void
1073 perf_aux_output_end(struct perf_output_handle *handle, unsigned long size,
1074                     bool truncated)                                     { }
1075 static inline int
1076 perf_aux_output_skip(struct perf_output_handle *handle,
1077                      unsigned long size)                                { return -EINVAL; }
1078 static inline void *
1079 perf_get_aux(struct perf_output_handle *handle)                         { return NULL; }
1080 static inline void
1081 perf_event_task_migrate(struct task_struct *task)                       { }
1082 static inline void
1083 perf_event_task_sched_in(struct task_struct *prev,
1084                          struct task_struct *task)                      { }
1085 static inline void
1086 perf_event_task_sched_out(struct task_struct *prev,
1087                           struct task_struct *next)                     { }
1088 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
1089 static inline void perf_event_exit_task(struct task_struct *child)      { }
1090 static inline void perf_event_free_task(struct task_struct *task)       { }
1091 static inline void perf_event_delayed_put(struct task_struct *task)     { }
1092 static inline struct file *perf_event_get(unsigned int fd)      { return ERR_PTR(-EINVAL); }
1093 static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
1094 {
1095         return ERR_PTR(-EINVAL);
1096 }
1097 static inline u64 perf_event_read_local(struct perf_event *event)       { return -EINVAL; }
1098 static inline void perf_event_print_debug(void)                         { }
1099 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
1100 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
1101 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1102 {
1103         return -EINVAL;
1104 }
1105
1106 static inline void
1107 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)     { }
1108 static inline void
1109 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)                     { }
1110 static inline void
1111 perf_bp_event(struct perf_event *event, void *data)                     { }
1112
1113 static inline int perf_register_guest_info_callbacks
1114 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1115 static inline int perf_unregister_guest_info_callbacks
1116 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1117
1118 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
1119 static inline void perf_event_exec(void)                                { }
1120 static inline void perf_event_comm(struct task_struct *tsk, bool exec)  { }
1121 static inline void perf_event_fork(struct task_struct *tsk)             { }
1122 static inline void perf_event_init(void)                                { }
1123 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
1124 static inline void perf_swevent_put_recursion_context(int rctx)         { }
1125 static inline u64 perf_swevent_set_period(struct perf_event *event)     { return 0; }
1126 static inline void perf_event_enable(struct perf_event *event)          { }
1127 static inline void perf_event_disable(struct perf_event *event)         { }
1128 static inline int __perf_event_disable(void *info)                      { return -1; }
1129 static inline void perf_event_task_tick(void)                           { }
1130 static inline int perf_event_release_kernel(struct perf_event *event)   { return 0; }
1131 #endif
1132
1133 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1134 extern void perf_restore_debug_store(void);
1135 #else
1136 static inline void perf_restore_debug_store(void)                       { }
1137 #endif
1138
1139 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1140
1141 /*
1142  * This has to have a higher priority than migration_notifier in sched/core.c.
1143  */
1144 #define perf_cpu_notifier(fn)                                           \
1145 do {                                                                    \
1146         static struct notifier_block fn##_nb =                          \
1147                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
1148         unsigned long cpu = smp_processor_id();                         \
1149         unsigned long flags;                                            \
1150                                                                         \
1151         cpu_notifier_register_begin();                                  \
1152         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,                     \
1153                 (void *)(unsigned long)cpu);                            \
1154         local_irq_save(flags);                                          \
1155         fn(&fn##_nb, (unsigned long)CPU_STARTING,                       \
1156                 (void *)(unsigned long)cpu);                            \
1157         local_irq_restore(flags);                                       \
1158         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                         \
1159                 (void *)(unsigned long)cpu);                            \
1160         __register_cpu_notifier(&fn##_nb);                              \
1161         cpu_notifier_register_done();                                   \
1162 } while (0)
1163
1164 /*
1165  * Bare-bones version of perf_cpu_notifier(), which doesn't invoke the
1166  * callback for already online CPUs.
1167  */
1168 #define __perf_cpu_notifier(fn)                                         \
1169 do {                                                                    \
1170         static struct notifier_block fn##_nb =                          \
1171                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
1172                                                                         \
1173         __register_cpu_notifier(&fn##_nb);                              \
1174 } while (0)
1175
1176 struct perf_pmu_events_attr {
1177         struct device_attribute attr;
1178         u64 id;
1179         const char *event_str;
1180 };
1181
1182 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1183                               char *page);
1184
1185 #define PMU_EVENT_ATTR(_name, _var, _id, _show)                         \
1186 static struct perf_pmu_events_attr _var = {                             \
1187         .attr = __ATTR(_name, 0444, _show, NULL),                       \
1188         .id   =  _id,                                                   \
1189 };
1190
1191 #define PMU_EVENT_ATTR_STRING(_name, _var, _str)                            \
1192 static struct perf_pmu_events_attr _var = {                                 \
1193         .attr           = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1194         .id             = 0,                                                \
1195         .event_str      = _str,                                             \
1196 };
1197
1198 #define PMU_FORMAT_ATTR(_name, _format)                                 \
1199 static ssize_t                                                          \
1200 _name##_show(struct device *dev,                                        \
1201                                struct device_attribute *attr,           \
1202                                char *page)                              \
1203 {                                                                       \
1204         BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);                     \
1205         return sprintf(page, _format "\n");                             \
1206 }                                                                       \
1207                                                                         \
1208 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1209
1210 #endif /* _LINUX_PERF_EVENT_H */