sparc: Convert naked unsigned uses to unsigned int
[linux-2.6-block.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                         struct list_head        cqm_events_entry;
125                         struct list_head        cqm_groups_entry;
126                         struct list_head        cqm_group_entry;
127                 };
128                 struct { /* itrace */
129                         int                     itrace_started;
130                 };
131 #ifdef CONFIG_HAVE_HW_BREAKPOINT
132                 struct { /* breakpoint */
133                         /*
134                          * Crufty hack to avoid the chicken and egg
135                          * problem hw_breakpoint has with context
136                          * creation and event initalization.
137                          */
138                         struct arch_hw_breakpoint       info;
139                         struct list_head                bp_list;
140                 };
141 #endif
142         };
143         /*
144          * If the event is a per task event, this will point to the task in
145          * question. See the comment in perf_event_alloc().
146          */
147         struct task_struct              *target;
148
149 /*
150  * hw_perf_event::state flags; used to track the PERF_EF_* state.
151  */
152 #define PERF_HES_STOPPED        0x01 /* the counter is stopped */
153 #define PERF_HES_UPTODATE       0x02 /* event->count up-to-date */
154 #define PERF_HES_ARCH           0x04
155
156         int                             state;
157
158         /*
159          * The last observed hardware counter value, updated with a
160          * local64_cmpxchg() such that pmu::read() can be called nested.
161          */
162         local64_t                       prev_count;
163
164         /*
165          * The period to start the next sample with.
166          */
167         u64                             sample_period;
168
169         /*
170          * The period we started this sample with.
171          */
172         u64                             last_period;
173
174         /*
175          * However much is left of the current period; note that this is
176          * a full 64bit value and allows for generation of periods longer
177          * than hardware might allow.
178          */
179         local64_t                       period_left;
180
181         /*
182          * State for throttling the event, see __perf_event_overflow() and
183          * perf_adjust_freq_unthr_context().
184          */
185         u64                             interrupts_seq;
186         u64                             interrupts;
187
188         /*
189          * State for freq target events, see __perf_event_overflow() and
190          * perf_adjust_freq_unthr_context().
191          */
192         u64                             freq_time_stamp;
193         u64                             freq_count_stamp;
194 #endif
195 };
196
197 struct perf_event;
198
199 /*
200  * Common implementation detail of pmu::{start,commit,cancel}_txn
201  */
202 #define PERF_PMU_TXN_ADD  0x1           /* txn to add/schedule event on PMU */
203 #define PERF_PMU_TXN_READ 0x2           /* txn to read event group from PMU */
204
205 /**
206  * pmu::capabilities flags
207  */
208 #define PERF_PMU_CAP_NO_INTERRUPT               0x01
209 #define PERF_PMU_CAP_NO_NMI                     0x02
210 #define PERF_PMU_CAP_AUX_NO_SG                  0x04
211 #define PERF_PMU_CAP_AUX_SW_DOUBLEBUF           0x08
212 #define PERF_PMU_CAP_EXCLUSIVE                  0x10
213 #define PERF_PMU_CAP_ITRACE                     0x20
214
215 /**
216  * struct pmu - generic performance monitoring unit
217  */
218 struct pmu {
219         struct list_head                entry;
220
221         struct module                   *module;
222         struct device                   *dev;
223         const struct attribute_group    **attr_groups;
224         const char                      *name;
225         int                             type;
226
227         /*
228          * various common per-pmu feature flags
229          */
230         int                             capabilities;
231
232         int * __percpu                  pmu_disable_count;
233         struct perf_cpu_context * __percpu pmu_cpu_context;
234         atomic_t                        exclusive_cnt; /* < 0: cpu; > 0: tsk */
235         int                             task_ctx_nr;
236         int                             hrtimer_interval_ms;
237
238         /*
239          * Fully disable/enable this PMU, can be used to protect from the PMI
240          * as well as for lazy/batch writing of the MSRs.
241          */
242         void (*pmu_enable)              (struct pmu *pmu); /* optional */
243         void (*pmu_disable)             (struct pmu *pmu); /* optional */
244
245         /*
246          * Try and initialize the event for this PMU.
247          *
248          * Returns:
249          *  -ENOENT     -- @event is not for this PMU
250          *
251          *  -ENODEV     -- @event is for this PMU but PMU not present
252          *  -EBUSY      -- @event is for this PMU but PMU temporarily unavailable
253          *  -EINVAL     -- @event is for this PMU but @event is not valid
254          *  -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
255          *  -EACCESS    -- @event is for this PMU, @event is valid, but no privilidges
256          *
257          *  0           -- @event is for this PMU and valid
258          *
259          * Other error return values are allowed.
260          */
261         int (*event_init)               (struct perf_event *event);
262
263         /*
264          * Notification that the event was mapped or unmapped.  Called
265          * in the context of the mapping task.
266          */
267         void (*event_mapped)            (struct perf_event *event); /*optional*/
268         void (*event_unmapped)          (struct perf_event *event); /*optional*/
269
270         /*
271          * Flags for ->add()/->del()/ ->start()/->stop(). There are
272          * matching hw_perf_event::state flags.
273          */
274 #define PERF_EF_START   0x01            /* start the counter when adding    */
275 #define PERF_EF_RELOAD  0x02            /* reload the counter when starting */
276 #define PERF_EF_UPDATE  0x04            /* update the counter when stopping */
277
278         /*
279          * Adds/Removes a counter to/from the PMU, can be done inside a
280          * transaction, see the ->*_txn() methods.
281          *
282          * The add/del callbacks will reserve all hardware resources required
283          * to service the event, this includes any counter constraint
284          * scheduling etc.
285          *
286          * Called with IRQs disabled and the PMU disabled on the CPU the event
287          * is on.
288          *
289          * ->add() called without PERF_EF_START should result in the same state
290          *  as ->add() followed by ->stop().
291          *
292          * ->del() must always PERF_EF_UPDATE stop an event. If it calls
293          *  ->stop() that must deal with already being stopped without
294          *  PERF_EF_UPDATE.
295          */
296         int  (*add)                     (struct perf_event *event, int flags);
297         void (*del)                     (struct perf_event *event, int flags);
298
299         /*
300          * Starts/Stops a counter present on the PMU.
301          *
302          * The PMI handler should stop the counter when perf_event_overflow()
303          * returns !0. ->start() will be used to continue.
304          *
305          * Also used to change the sample period.
306          *
307          * Called with IRQs disabled and the PMU disabled on the CPU the event
308          * is on -- will be called from NMI context with the PMU generates
309          * NMIs.
310          *
311          * ->stop() with PERF_EF_UPDATE will read the counter and update
312          *  period/count values like ->read() would.
313          *
314          * ->start() with PERF_EF_RELOAD will reprogram the the counter
315          *  value, must be preceded by a ->stop() with PERF_EF_UPDATE.
316          */
317         void (*start)                   (struct perf_event *event, int flags);
318         void (*stop)                    (struct perf_event *event, int flags);
319
320         /*
321          * Updates the counter value of the event.
322          *
323          * For sampling capable PMUs this will also update the software period
324          * hw_perf_event::period_left field.
325          */
326         void (*read)                    (struct perf_event *event);
327
328         /*
329          * Group events scheduling is treated as a transaction, add
330          * group events as a whole and perform one schedulability test.
331          * If the test fails, roll back the whole group
332          *
333          * Start the transaction, after this ->add() doesn't need to
334          * do schedulability tests.
335          *
336          * Optional.
337          */
338         void (*start_txn)               (struct pmu *pmu, unsigned int txn_flags);
339         /*
340          * If ->start_txn() disabled the ->add() schedulability test
341          * then ->commit_txn() is required to perform one. On success
342          * the transaction is closed. On error the transaction is kept
343          * open until ->cancel_txn() is called.
344          *
345          * Optional.
346          */
347         int  (*commit_txn)              (struct pmu *pmu);
348         /*
349          * Will cancel the transaction, assumes ->del() is called
350          * for each successful ->add() during the transaction.
351          *
352          * Optional.
353          */
354         void (*cancel_txn)              (struct pmu *pmu);
355
356         /*
357          * Will return the value for perf_event_mmap_page::index for this event,
358          * if no implementation is provided it will default to: event->hw.idx + 1.
359          */
360         int (*event_idx)                (struct perf_event *event); /*optional */
361
362         /*
363          * context-switches callback
364          */
365         void (*sched_task)              (struct perf_event_context *ctx,
366                                         bool sched_in);
367         /*
368          * PMU specific data size
369          */
370         size_t                          task_ctx_size;
371
372
373         /*
374          * Return the count value for a counter.
375          */
376         u64 (*count)                    (struct perf_event *event); /*optional*/
377
378         /*
379          * Set up pmu-private data structures for an AUX area
380          */
381         void *(*setup_aux)              (int cpu, void **pages,
382                                          int nr_pages, bool overwrite);
383                                         /* optional */
384
385         /*
386          * Free pmu-private AUX data structures
387          */
388         void (*free_aux)                (void *aux); /* optional */
389
390         /*
391          * Filter events for PMU-specific reasons.
392          */
393         int (*filter_match)             (struct perf_event *event); /* optional */
394 };
395
396 /**
397  * enum perf_event_active_state - the states of a event
398  */
399 enum perf_event_active_state {
400         PERF_EVENT_STATE_DEAD           = -4,
401         PERF_EVENT_STATE_EXIT           = -3,
402         PERF_EVENT_STATE_ERROR          = -2,
403         PERF_EVENT_STATE_OFF            = -1,
404         PERF_EVENT_STATE_INACTIVE       =  0,
405         PERF_EVENT_STATE_ACTIVE         =  1,
406 };
407
408 struct file;
409 struct perf_sample_data;
410
411 typedef void (*perf_overflow_handler_t)(struct perf_event *,
412                                         struct perf_sample_data *,
413                                         struct pt_regs *regs);
414
415 enum perf_group_flag {
416         PERF_GROUP_SOFTWARE             = 0x1,
417 };
418
419 #define SWEVENT_HLIST_BITS              8
420 #define SWEVENT_HLIST_SIZE              (1 << SWEVENT_HLIST_BITS)
421
422 struct swevent_hlist {
423         struct hlist_head               heads[SWEVENT_HLIST_SIZE];
424         struct rcu_head                 rcu_head;
425 };
426
427 #define PERF_ATTACH_CONTEXT     0x01
428 #define PERF_ATTACH_GROUP       0x02
429 #define PERF_ATTACH_TASK        0x04
430 #define PERF_ATTACH_TASK_DATA   0x08
431
432 struct perf_cgroup;
433 struct ring_buffer;
434
435 /**
436  * struct perf_event - performance event kernel representation:
437  */
438 struct perf_event {
439 #ifdef CONFIG_PERF_EVENTS
440         /*
441          * entry onto perf_event_context::event_list;
442          *   modifications require ctx->lock
443          *   RCU safe iterations.
444          */
445         struct list_head                event_entry;
446
447         /*
448          * XXX: group_entry and sibling_list should be mutually exclusive;
449          * either you're a sibling on a group, or you're the group leader.
450          * Rework the code to always use the same list element.
451          *
452          * Locked for modification by both ctx->mutex and ctx->lock; holding
453          * either sufficies for read.
454          */
455         struct list_head                group_entry;
456         struct list_head                sibling_list;
457
458         /*
459          * We need storage to track the entries in perf_pmu_migrate_context; we
460          * cannot use the event_entry because of RCU and we want to keep the
461          * group in tact which avoids us using the other two entries.
462          */
463         struct list_head                migrate_entry;
464
465         struct hlist_node               hlist_entry;
466         struct list_head                active_entry;
467         int                             nr_siblings;
468         int                             group_flags;
469         struct perf_event               *group_leader;
470         struct pmu                      *pmu;
471         void                            *pmu_private;
472
473         enum perf_event_active_state    state;
474         unsigned int                    attach_state;
475         local64_t                       count;
476         atomic64_t                      child_count;
477
478         /*
479          * These are the total time in nanoseconds that the event
480          * has been enabled (i.e. eligible to run, and the task has
481          * been scheduled in, if this is a per-task event)
482          * and running (scheduled onto the CPU), respectively.
483          *
484          * They are computed from tstamp_enabled, tstamp_running and
485          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
486          */
487         u64                             total_time_enabled;
488         u64                             total_time_running;
489
490         /*
491          * These are timestamps used for computing total_time_enabled
492          * and total_time_running when the event is in INACTIVE or
493          * ACTIVE state, measured in nanoseconds from an arbitrary point
494          * in time.
495          * tstamp_enabled: the notional time when the event was enabled
496          * tstamp_running: the notional time when the event was scheduled on
497          * tstamp_stopped: in INACTIVE state, the notional time when the
498          *      event was scheduled off.
499          */
500         u64                             tstamp_enabled;
501         u64                             tstamp_running;
502         u64                             tstamp_stopped;
503
504         /*
505          * timestamp shadows the actual context timing but it can
506          * be safely used in NMI interrupt context. It reflects the
507          * context time as it was when the event was last scheduled in.
508          *
509          * ctx_time already accounts for ctx->timestamp. Therefore to
510          * compute ctx_time for a sample, simply add perf_clock().
511          */
512         u64                             shadow_ctx_time;
513
514         struct perf_event_attr          attr;
515         u16                             header_size;
516         u16                             id_header_size;
517         u16                             read_size;
518         struct hw_perf_event            hw;
519
520         struct perf_event_context       *ctx;
521         atomic_long_t                   refcount;
522
523         /*
524          * These accumulate total time (in nanoseconds) that children
525          * events have been enabled and running, respectively.
526          */
527         atomic64_t                      child_total_time_enabled;
528         atomic64_t                      child_total_time_running;
529
530         /*
531          * Protect attach/detach and child_list:
532          */
533         struct mutex                    child_mutex;
534         struct list_head                child_list;
535         struct perf_event               *parent;
536
537         int                             oncpu;
538         int                             cpu;
539
540         struct list_head                owner_entry;
541         struct task_struct              *owner;
542
543         /* mmap bits */
544         struct mutex                    mmap_mutex;
545         atomic_t                        mmap_count;
546
547         struct ring_buffer              *rb;
548         struct list_head                rb_entry;
549         unsigned long                   rcu_batches;
550         int                             rcu_pending;
551
552         /* poll related */
553         wait_queue_head_t               waitq;
554         struct fasync_struct            *fasync;
555
556         /* delayed work for NMIs and such */
557         int                             pending_wakeup;
558         int                             pending_kill;
559         int                             pending_disable;
560         struct irq_work                 pending;
561
562         atomic_t                        event_limit;
563
564         void (*destroy)(struct perf_event *);
565         struct rcu_head                 rcu_head;
566
567         struct pid_namespace            *ns;
568         u64                             id;
569
570         u64                             (*clock)(void);
571         perf_overflow_handler_t         overflow_handler;
572         void                            *overflow_handler_context;
573
574 #ifdef CONFIG_EVENT_TRACING
575         struct trace_event_call         *tp_event;
576         struct event_filter             *filter;
577 #ifdef CONFIG_FUNCTION_TRACER
578         struct ftrace_ops               ftrace_ops;
579 #endif
580 #endif
581
582 #ifdef CONFIG_CGROUP_PERF
583         struct perf_cgroup              *cgrp; /* cgroup event is attach to */
584         int                             cgrp_defer_enabled;
585 #endif
586
587 #endif /* CONFIG_PERF_EVENTS */
588 };
589
590 /**
591  * struct perf_event_context - event context structure
592  *
593  * Used as a container for task events and CPU events as well:
594  */
595 struct perf_event_context {
596         struct pmu                      *pmu;
597         /*
598          * Protect the states of the events in the list,
599          * nr_active, and the list:
600          */
601         raw_spinlock_t                  lock;
602         /*
603          * Protect the list of events.  Locking either mutex or lock
604          * is sufficient to ensure the list doesn't change; to change
605          * the list you need to lock both the mutex and the spinlock.
606          */
607         struct mutex                    mutex;
608
609         struct list_head                active_ctx_list;
610         struct list_head                pinned_groups;
611         struct list_head                flexible_groups;
612         struct list_head                event_list;
613         int                             nr_events;
614         int                             nr_active;
615         int                             is_active;
616         int                             nr_stat;
617         int                             nr_freq;
618         int                             rotate_disable;
619         atomic_t                        refcount;
620         struct task_struct              *task;
621
622         /*
623          * Context clock, runs when context enabled.
624          */
625         u64                             time;
626         u64                             timestamp;
627
628         /*
629          * These fields let us detect when two contexts have both
630          * been cloned (inherited) from a common ancestor.
631          */
632         struct perf_event_context       *parent_ctx;
633         u64                             parent_gen;
634         u64                             generation;
635         int                             pin_count;
636         int                             nr_cgroups;      /* cgroup evts */
637         void                            *task_ctx_data; /* pmu specific data */
638         struct rcu_head                 rcu_head;
639 };
640
641 /*
642  * Number of contexts where an event can trigger:
643  *      task, softirq, hardirq, nmi.
644  */
645 #define PERF_NR_CONTEXTS        4
646
647 /**
648  * struct perf_event_cpu_context - per cpu event context structure
649  */
650 struct perf_cpu_context {
651         struct perf_event_context       ctx;
652         struct perf_event_context       *task_ctx;
653         int                             active_oncpu;
654         int                             exclusive;
655
656         raw_spinlock_t                  hrtimer_lock;
657         struct hrtimer                  hrtimer;
658         ktime_t                         hrtimer_interval;
659         unsigned int                    hrtimer_active;
660
661         struct pmu                      *unique_pmu;
662         struct perf_cgroup              *cgrp;
663 };
664
665 struct perf_output_handle {
666         struct perf_event               *event;
667         struct ring_buffer              *rb;
668         unsigned long                   wakeup;
669         unsigned long                   size;
670         union {
671                 void                    *addr;
672                 unsigned long           head;
673         };
674         int                             page;
675 };
676
677 #ifdef CONFIG_CGROUP_PERF
678
679 /*
680  * perf_cgroup_info keeps track of time_enabled for a cgroup.
681  * This is a per-cpu dynamically allocated data structure.
682  */
683 struct perf_cgroup_info {
684         u64                             time;
685         u64                             timestamp;
686 };
687
688 struct perf_cgroup {
689         struct cgroup_subsys_state      css;
690         struct perf_cgroup_info __percpu *info;
691 };
692
693 /*
694  * Must ensure cgroup is pinned (css_get) before calling
695  * this function. In other words, we cannot call this function
696  * if there is no cgroup event for the current CPU context.
697  */
698 static inline struct perf_cgroup *
699 perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
700 {
701         return container_of(task_css_check(task, perf_event_cgrp_id,
702                                            ctx ? lockdep_is_held(&ctx->lock)
703                                                : true),
704                             struct perf_cgroup, css);
705 }
706 #endif /* CONFIG_CGROUP_PERF */
707
708 #ifdef CONFIG_PERF_EVENTS
709
710 extern void *perf_aux_output_begin(struct perf_output_handle *handle,
711                                    struct perf_event *event);
712 extern void perf_aux_output_end(struct perf_output_handle *handle,
713                                 unsigned long size, bool truncated);
714 extern int perf_aux_output_skip(struct perf_output_handle *handle,
715                                 unsigned long size);
716 extern void *perf_get_aux(struct perf_output_handle *handle);
717
718 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
719 extern void perf_pmu_unregister(struct pmu *pmu);
720
721 extern int perf_num_counters(void);
722 extern const char *perf_pmu_name(void);
723 extern void __perf_event_task_sched_in(struct task_struct *prev,
724                                        struct task_struct *task);
725 extern void __perf_event_task_sched_out(struct task_struct *prev,
726                                         struct task_struct *next);
727 extern int perf_event_init_task(struct task_struct *child);
728 extern void perf_event_exit_task(struct task_struct *child);
729 extern void perf_event_free_task(struct task_struct *task);
730 extern void perf_event_delayed_put(struct task_struct *task);
731 extern struct file *perf_event_get(unsigned int fd);
732 extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
733 extern void perf_event_print_debug(void);
734 extern void perf_pmu_disable(struct pmu *pmu);
735 extern void perf_pmu_enable(struct pmu *pmu);
736 extern void perf_sched_cb_dec(struct pmu *pmu);
737 extern void perf_sched_cb_inc(struct pmu *pmu);
738 extern int perf_event_task_disable(void);
739 extern int perf_event_task_enable(void);
740 extern int perf_event_refresh(struct perf_event *event, int refresh);
741 extern void perf_event_update_userpage(struct perf_event *event);
742 extern int perf_event_release_kernel(struct perf_event *event);
743 extern struct perf_event *
744 perf_event_create_kernel_counter(struct perf_event_attr *attr,
745                                 int cpu,
746                                 struct task_struct *task,
747                                 perf_overflow_handler_t callback,
748                                 void *context);
749 extern void perf_pmu_migrate_context(struct pmu *pmu,
750                                 int src_cpu, int dst_cpu);
751 extern u64 perf_event_read_local(struct perf_event *event);
752 extern u64 perf_event_read_value(struct perf_event *event,
753                                  u64 *enabled, u64 *running);
754
755
756 struct perf_sample_data {
757         /*
758          * Fields set by perf_sample_data_init(), group so as to
759          * minimize the cachelines touched.
760          */
761         u64                             addr;
762         struct perf_raw_record          *raw;
763         struct perf_branch_stack        *br_stack;
764         u64                             period;
765         u64                             weight;
766         u64                             txn;
767         union  perf_mem_data_src        data_src;
768
769         /*
770          * The other fields, optionally {set,used} by
771          * perf_{prepare,output}_sample().
772          */
773         u64                             type;
774         u64                             ip;
775         struct {
776                 u32     pid;
777                 u32     tid;
778         }                               tid_entry;
779         u64                             time;
780         u64                             id;
781         u64                             stream_id;
782         struct {
783                 u32     cpu;
784                 u32     reserved;
785         }                               cpu_entry;
786         struct perf_callchain_entry     *callchain;
787
788         /*
789          * regs_user may point to task_pt_regs or to regs_user_copy, depending
790          * on arch details.
791          */
792         struct perf_regs                regs_user;
793         struct pt_regs                  regs_user_copy;
794
795         struct perf_regs                regs_intr;
796         u64                             stack_user_size;
797 } ____cacheline_aligned;
798
799 /* default value for data source */
800 #define PERF_MEM_NA (PERF_MEM_S(OP, NA)   |\
801                     PERF_MEM_S(LVL, NA)   |\
802                     PERF_MEM_S(SNOOP, NA) |\
803                     PERF_MEM_S(LOCK, NA)  |\
804                     PERF_MEM_S(TLB, NA))
805
806 static inline void perf_sample_data_init(struct perf_sample_data *data,
807                                          u64 addr, u64 period)
808 {
809         /* remaining struct members initialized in perf_prepare_sample() */
810         data->addr = addr;
811         data->raw  = NULL;
812         data->br_stack = NULL;
813         data->period = period;
814         data->weight = 0;
815         data->data_src.val = PERF_MEM_NA;
816         data->txn = 0;
817 }
818
819 extern void perf_output_sample(struct perf_output_handle *handle,
820                                struct perf_event_header *header,
821                                struct perf_sample_data *data,
822                                struct perf_event *event);
823 extern void perf_prepare_sample(struct perf_event_header *header,
824                                 struct perf_sample_data *data,
825                                 struct perf_event *event,
826                                 struct pt_regs *regs);
827
828 extern int perf_event_overflow(struct perf_event *event,
829                                  struct perf_sample_data *data,
830                                  struct pt_regs *regs);
831
832 extern void perf_event_output(struct perf_event *event,
833                                 struct perf_sample_data *data,
834                                 struct pt_regs *regs);
835
836 extern void
837 perf_event_header__init_id(struct perf_event_header *header,
838                            struct perf_sample_data *data,
839                            struct perf_event *event);
840 extern void
841 perf_event__output_id_sample(struct perf_event *event,
842                              struct perf_output_handle *handle,
843                              struct perf_sample_data *sample);
844
845 extern void
846 perf_log_lost_samples(struct perf_event *event, u64 lost);
847
848 static inline bool is_sampling_event(struct perf_event *event)
849 {
850         return event->attr.sample_period != 0;
851 }
852
853 /*
854  * Return 1 for a software event, 0 for a hardware event
855  */
856 static inline int is_software_event(struct perf_event *event)
857 {
858         return event->pmu->task_ctx_nr == perf_sw_context;
859 }
860
861 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
862
863 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
864 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
865
866 #ifndef perf_arch_fetch_caller_regs
867 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
868 #endif
869
870 /*
871  * Take a snapshot of the regs. Skip ip and frame pointer to
872  * the nth caller. We only need a few of the regs:
873  * - ip for PERF_SAMPLE_IP
874  * - cs for user_mode() tests
875  * - bp for callchains
876  * - eflags, for future purposes, just in case
877  */
878 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
879 {
880         memset(regs, 0, sizeof(*regs));
881
882         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
883 }
884
885 static __always_inline void
886 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
887 {
888         if (static_key_false(&perf_swevent_enabled[event_id]))
889                 __perf_sw_event(event_id, nr, regs, addr);
890 }
891
892 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
893
894 /*
895  * 'Special' version for the scheduler, it hard assumes no recursion,
896  * which is guaranteed by us not actually scheduling inside other swevents
897  * because those disable preemption.
898  */
899 static __always_inline void
900 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
901 {
902         if (static_key_false(&perf_swevent_enabled[event_id])) {
903                 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
904
905                 perf_fetch_caller_regs(regs);
906                 ___perf_sw_event(event_id, nr, regs, addr);
907         }
908 }
909
910 extern struct static_key_false perf_sched_events;
911
912 static __always_inline bool
913 perf_sw_migrate_enabled(void)
914 {
915         if (static_key_false(&perf_swevent_enabled[PERF_COUNT_SW_CPU_MIGRATIONS]))
916                 return true;
917         return false;
918 }
919
920 static inline void perf_event_task_migrate(struct task_struct *task)
921 {
922         if (perf_sw_migrate_enabled())
923                 task->sched_migrated = 1;
924 }
925
926 static inline void perf_event_task_sched_in(struct task_struct *prev,
927                                             struct task_struct *task)
928 {
929         if (static_branch_unlikely(&perf_sched_events))
930                 __perf_event_task_sched_in(prev, task);
931
932         if (perf_sw_migrate_enabled() && task->sched_migrated) {
933                 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
934
935                 perf_fetch_caller_regs(regs);
936                 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, regs, 0);
937                 task->sched_migrated = 0;
938         }
939 }
940
941 static inline void perf_event_task_sched_out(struct task_struct *prev,
942                                              struct task_struct *next)
943 {
944         perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
945
946         if (static_branch_unlikely(&perf_sched_events))
947                 __perf_event_task_sched_out(prev, next);
948 }
949
950 static inline u64 __perf_event_count(struct perf_event *event)
951 {
952         return local64_read(&event->count) + atomic64_read(&event->child_count);
953 }
954
955 extern void perf_event_mmap(struct vm_area_struct *vma);
956 extern struct perf_guest_info_callbacks *perf_guest_cbs;
957 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
958 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
959
960 extern void perf_event_exec(void);
961 extern void perf_event_comm(struct task_struct *tsk, bool exec);
962 extern void perf_event_fork(struct task_struct *tsk);
963
964 /* Callchains */
965 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
966
967 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
968 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
969 extern struct perf_callchain_entry *
970 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
971                    bool crosstask, bool add_mark);
972 extern int get_callchain_buffers(void);
973 extern void put_callchain_buffers(void);
974
975 static inline int perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
976 {
977         if (entry->nr < PERF_MAX_STACK_DEPTH) {
978                 entry->ip[entry->nr++] = ip;
979                 return 0;
980         } else {
981                 return -1; /* no more room, stop walking the stack */
982         }
983 }
984
985 extern int sysctl_perf_event_paranoid;
986 extern int sysctl_perf_event_mlock;
987 extern int sysctl_perf_event_sample_rate;
988 extern int sysctl_perf_cpu_time_max_percent;
989
990 extern void perf_sample_event_took(u64 sample_len_ns);
991
992 extern int perf_proc_update_handler(struct ctl_table *table, int write,
993                 void __user *buffer, size_t *lenp,
994                 loff_t *ppos);
995 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
996                 void __user *buffer, size_t *lenp,
997                 loff_t *ppos);
998
999
1000 static inline bool perf_paranoid_tracepoint_raw(void)
1001 {
1002         return sysctl_perf_event_paranoid > -1;
1003 }
1004
1005 static inline bool perf_paranoid_cpu(void)
1006 {
1007         return sysctl_perf_event_paranoid > 0;
1008 }
1009
1010 static inline bool perf_paranoid_kernel(void)
1011 {
1012         return sysctl_perf_event_paranoid > 1;
1013 }
1014
1015 extern void perf_event_init(void);
1016 extern void perf_tp_event(u64 addr, u64 count, void *record,
1017                           int entry_size, struct pt_regs *regs,
1018                           struct hlist_head *head, int rctx,
1019                           struct task_struct *task);
1020 extern void perf_bp_event(struct perf_event *event, void *data);
1021
1022 #ifndef perf_misc_flags
1023 # define perf_misc_flags(regs) \
1024                 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1025 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1026 #endif
1027
1028 static inline bool has_branch_stack(struct perf_event *event)
1029 {
1030         return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1031 }
1032
1033 static inline bool needs_branch_stack(struct perf_event *event)
1034 {
1035         return event->attr.branch_sample_type != 0;
1036 }
1037
1038 static inline bool has_aux(struct perf_event *event)
1039 {
1040         return event->pmu->setup_aux;
1041 }
1042
1043 extern int perf_output_begin(struct perf_output_handle *handle,
1044                              struct perf_event *event, unsigned int size);
1045 extern void perf_output_end(struct perf_output_handle *handle);
1046 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
1047                              const void *buf, unsigned int len);
1048 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
1049                                      unsigned int len);
1050 extern int perf_swevent_get_recursion_context(void);
1051 extern void perf_swevent_put_recursion_context(int rctx);
1052 extern u64 perf_swevent_set_period(struct perf_event *event);
1053 extern void perf_event_enable(struct perf_event *event);
1054 extern void perf_event_disable(struct perf_event *event);
1055 extern void perf_event_disable_local(struct perf_event *event);
1056 extern void perf_event_task_tick(void);
1057 #else /* !CONFIG_PERF_EVENTS: */
1058 static inline void *
1059 perf_aux_output_begin(struct perf_output_handle *handle,
1060                       struct perf_event *event)                         { return NULL; }
1061 static inline void
1062 perf_aux_output_end(struct perf_output_handle *handle, unsigned long size,
1063                     bool truncated)                                     { }
1064 static inline int
1065 perf_aux_output_skip(struct perf_output_handle *handle,
1066                      unsigned long size)                                { return -EINVAL; }
1067 static inline void *
1068 perf_get_aux(struct perf_output_handle *handle)                         { return NULL; }
1069 static inline void
1070 perf_event_task_migrate(struct task_struct *task)                       { }
1071 static inline void
1072 perf_event_task_sched_in(struct task_struct *prev,
1073                          struct task_struct *task)                      { }
1074 static inline void
1075 perf_event_task_sched_out(struct task_struct *prev,
1076                           struct task_struct *next)                     { }
1077 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
1078 static inline void perf_event_exit_task(struct task_struct *child)      { }
1079 static inline void perf_event_free_task(struct task_struct *task)       { }
1080 static inline void perf_event_delayed_put(struct task_struct *task)     { }
1081 static inline struct file *perf_event_get(unsigned int fd)      { return ERR_PTR(-EINVAL); }
1082 static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
1083 {
1084         return ERR_PTR(-EINVAL);
1085 }
1086 static inline u64 perf_event_read_local(struct perf_event *event)       { return -EINVAL; }
1087 static inline void perf_event_print_debug(void)                         { }
1088 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
1089 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
1090 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1091 {
1092         return -EINVAL;
1093 }
1094
1095 static inline void
1096 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)     { }
1097 static inline void
1098 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)                     { }
1099 static inline void
1100 perf_bp_event(struct perf_event *event, void *data)                     { }
1101
1102 static inline int perf_register_guest_info_callbacks
1103 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1104 static inline int perf_unregister_guest_info_callbacks
1105 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1106
1107 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
1108 static inline void perf_event_exec(void)                                { }
1109 static inline void perf_event_comm(struct task_struct *tsk, bool exec)  { }
1110 static inline void perf_event_fork(struct task_struct *tsk)             { }
1111 static inline void perf_event_init(void)                                { }
1112 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
1113 static inline void perf_swevent_put_recursion_context(int rctx)         { }
1114 static inline u64 perf_swevent_set_period(struct perf_event *event)     { return 0; }
1115 static inline void perf_event_enable(struct perf_event *event)          { }
1116 static inline void perf_event_disable(struct perf_event *event)         { }
1117 static inline int __perf_event_disable(void *info)                      { return -1; }
1118 static inline void perf_event_task_tick(void)                           { }
1119 static inline int perf_event_release_kernel(struct perf_event *event)   { return 0; }
1120 #endif
1121
1122 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1123 extern void perf_restore_debug_store(void);
1124 #else
1125 static inline void perf_restore_debug_store(void)                       { }
1126 #endif
1127
1128 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1129
1130 /*
1131  * This has to have a higher priority than migration_notifier in sched/core.c.
1132  */
1133 #define perf_cpu_notifier(fn)                                           \
1134 do {                                                                    \
1135         static struct notifier_block fn##_nb =                          \
1136                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
1137         unsigned long cpu = smp_processor_id();                         \
1138         unsigned long flags;                                            \
1139                                                                         \
1140         cpu_notifier_register_begin();                                  \
1141         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,                     \
1142                 (void *)(unsigned long)cpu);                            \
1143         local_irq_save(flags);                                          \
1144         fn(&fn##_nb, (unsigned long)CPU_STARTING,                       \
1145                 (void *)(unsigned long)cpu);                            \
1146         local_irq_restore(flags);                                       \
1147         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                         \
1148                 (void *)(unsigned long)cpu);                            \
1149         __register_cpu_notifier(&fn##_nb);                              \
1150         cpu_notifier_register_done();                                   \
1151 } while (0)
1152
1153 /*
1154  * Bare-bones version of perf_cpu_notifier(), which doesn't invoke the
1155  * callback for already online CPUs.
1156  */
1157 #define __perf_cpu_notifier(fn)                                         \
1158 do {                                                                    \
1159         static struct notifier_block fn##_nb =                          \
1160                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
1161                                                                         \
1162         __register_cpu_notifier(&fn##_nb);                              \
1163 } while (0)
1164
1165 struct perf_pmu_events_attr {
1166         struct device_attribute attr;
1167         u64 id;
1168         const char *event_str;
1169 };
1170
1171 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1172                               char *page);
1173
1174 #define PMU_EVENT_ATTR(_name, _var, _id, _show)                         \
1175 static struct perf_pmu_events_attr _var = {                             \
1176         .attr = __ATTR(_name, 0444, _show, NULL),                       \
1177         .id   =  _id,                                                   \
1178 };
1179
1180 #define PMU_EVENT_ATTR_STRING(_name, _var, _str)                            \
1181 static struct perf_pmu_events_attr _var = {                                 \
1182         .attr           = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1183         .id             = 0,                                                \
1184         .event_str      = _str,                                             \
1185 };
1186
1187 #define PMU_FORMAT_ATTR(_name, _format)                                 \
1188 static ssize_t                                                          \
1189 _name##_show(struct device *dev,                                        \
1190                                struct device_attribute *attr,           \
1191                                char *page)                              \
1192 {                                                                       \
1193         BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);                     \
1194         return sprintf(page, _format "\n");                             \
1195 }                                                                       \
1196                                                                         \
1197 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1198
1199 #endif /* _LINUX_PERF_EVENT_H */