perf_event: Optimize perf_output_lock()
[linux-2.6.git] / include / linux / perf_event.h
1 /*
2  * Performance events:
3  *
4  *    Copyright (C) 2008-2009, Thomas Gleixner <tglx@linutronix.de>
5  *    Copyright (C) 2008-2009, Red Hat, Inc., Ingo Molnar
6  *    Copyright (C) 2008-2009, Red Hat, Inc., Peter Zijlstra
7  *
8  * Data type definitions, declarations, prototypes.
9  *
10  *    Started by: Thomas Gleixner and Ingo Molnar
11  *
12  * For licencing details see kernel-base/COPYING
13  */
14 #ifndef _LINUX_PERF_EVENT_H
15 #define _LINUX_PERF_EVENT_H
16
17 #include <linux/types.h>
18 #include <linux/ioctl.h>
19 #include <asm/byteorder.h>
20
21 /*
22  * User-space ABI bits:
23  */
24
25 /*
26  * attr.type
27  */
28 enum perf_type_id {
29         PERF_TYPE_HARDWARE                      = 0,
30         PERF_TYPE_SOFTWARE                      = 1,
31         PERF_TYPE_TRACEPOINT                    = 2,
32         PERF_TYPE_HW_CACHE                      = 3,
33         PERF_TYPE_RAW                           = 4,
34
35         PERF_TYPE_MAX,                          /* non-ABI */
36 };
37
38 /*
39  * Generalized performance event event_id types, used by the
40  * attr.event_id parameter of the sys_perf_event_open()
41  * syscall:
42  */
43 enum perf_hw_id {
44         /*
45          * Common hardware events, generalized by the kernel:
46          */
47         PERF_COUNT_HW_CPU_CYCLES                = 0,
48         PERF_COUNT_HW_INSTRUCTIONS              = 1,
49         PERF_COUNT_HW_CACHE_REFERENCES          = 2,
50         PERF_COUNT_HW_CACHE_MISSES              = 3,
51         PERF_COUNT_HW_BRANCH_INSTRUCTIONS       = 4,
52         PERF_COUNT_HW_BRANCH_MISSES             = 5,
53         PERF_COUNT_HW_BUS_CYCLES                = 6,
54
55         PERF_COUNT_HW_MAX,                      /* non-ABI */
56 };
57
58 /*
59  * Generalized hardware cache events:
60  *
61  *       { L1-D, L1-I, LLC, ITLB, DTLB, BPU } x
62  *       { read, write, prefetch } x
63  *       { accesses, misses }
64  */
65 enum perf_hw_cache_id {
66         PERF_COUNT_HW_CACHE_L1D                 = 0,
67         PERF_COUNT_HW_CACHE_L1I                 = 1,
68         PERF_COUNT_HW_CACHE_LL                  = 2,
69         PERF_COUNT_HW_CACHE_DTLB                = 3,
70         PERF_COUNT_HW_CACHE_ITLB                = 4,
71         PERF_COUNT_HW_CACHE_BPU                 = 5,
72
73         PERF_COUNT_HW_CACHE_MAX,                /* non-ABI */
74 };
75
76 enum perf_hw_cache_op_id {
77         PERF_COUNT_HW_CACHE_OP_READ             = 0,
78         PERF_COUNT_HW_CACHE_OP_WRITE            = 1,
79         PERF_COUNT_HW_CACHE_OP_PREFETCH         = 2,
80
81         PERF_COUNT_HW_CACHE_OP_MAX,             /* non-ABI */
82 };
83
84 enum perf_hw_cache_op_result_id {
85         PERF_COUNT_HW_CACHE_RESULT_ACCESS       = 0,
86         PERF_COUNT_HW_CACHE_RESULT_MISS         = 1,
87
88         PERF_COUNT_HW_CACHE_RESULT_MAX,         /* non-ABI */
89 };
90
91 /*
92  * Special "software" events provided by the kernel, even if the hardware
93  * does not support performance events. These events measure various
94  * physical and sw events of the kernel (and allow the profiling of them as
95  * well):
96  */
97 enum perf_sw_ids {
98         PERF_COUNT_SW_CPU_CLOCK                 = 0,
99         PERF_COUNT_SW_TASK_CLOCK                = 1,
100         PERF_COUNT_SW_PAGE_FAULTS               = 2,
101         PERF_COUNT_SW_CONTEXT_SWITCHES          = 3,
102         PERF_COUNT_SW_CPU_MIGRATIONS            = 4,
103         PERF_COUNT_SW_PAGE_FAULTS_MIN           = 5,
104         PERF_COUNT_SW_PAGE_FAULTS_MAJ           = 6,
105         PERF_COUNT_SW_ALIGNMENT_FAULTS          = 7,
106         PERF_COUNT_SW_EMULATION_FAULTS          = 8,
107
108         PERF_COUNT_SW_MAX,                      /* non-ABI */
109 };
110
111 /*
112  * Bits that can be set in attr.sample_type to request information
113  * in the overflow packets.
114  */
115 enum perf_event_sample_format {
116         PERF_SAMPLE_IP                          = 1U << 0,
117         PERF_SAMPLE_TID                         = 1U << 1,
118         PERF_SAMPLE_TIME                        = 1U << 2,
119         PERF_SAMPLE_ADDR                        = 1U << 3,
120         PERF_SAMPLE_READ                        = 1U << 4,
121         PERF_SAMPLE_CALLCHAIN                   = 1U << 5,
122         PERF_SAMPLE_ID                          = 1U << 6,
123         PERF_SAMPLE_CPU                         = 1U << 7,
124         PERF_SAMPLE_PERIOD                      = 1U << 8,
125         PERF_SAMPLE_STREAM_ID                   = 1U << 9,
126         PERF_SAMPLE_RAW                         = 1U << 10,
127
128         PERF_SAMPLE_MAX = 1U << 11,             /* non-ABI */
129 };
130
131 /*
132  * The format of the data returned by read() on a perf event fd,
133  * as specified by attr.read_format:
134  *
135  * struct read_format {
136  *      { u64           value;
137  *        { u64         time_enabled; } && PERF_FORMAT_ENABLED
138  *        { u64         time_running; } && PERF_FORMAT_RUNNING
139  *        { u64         id;           } && PERF_FORMAT_ID
140  *      } && !PERF_FORMAT_GROUP
141  *
142  *      { u64           nr;
143  *        { u64         time_enabled; } && PERF_FORMAT_ENABLED
144  *        { u64         time_running; } && PERF_FORMAT_RUNNING
145  *        { u64         value;
146  *          { u64       id;           } && PERF_FORMAT_ID
147  *        }             cntr[nr];
148  *      } && PERF_FORMAT_GROUP
149  * };
150  */
151 enum perf_event_read_format {
152         PERF_FORMAT_TOTAL_TIME_ENABLED          = 1U << 0,
153         PERF_FORMAT_TOTAL_TIME_RUNNING          = 1U << 1,
154         PERF_FORMAT_ID                          = 1U << 2,
155         PERF_FORMAT_GROUP                       = 1U << 3,
156
157         PERF_FORMAT_MAX = 1U << 4,              /* non-ABI */
158 };
159
160 #define PERF_ATTR_SIZE_VER0     64      /* sizeof first published struct */
161
162 /*
163  * Hardware event_id to monitor via a performance monitoring event:
164  */
165 struct perf_event_attr {
166
167         /*
168          * Major type: hardware/software/tracepoint/etc.
169          */
170         __u32                   type;
171
172         /*
173          * Size of the attr structure, for fwd/bwd compat.
174          */
175         __u32                   size;
176
177         /*
178          * Type specific configuration information.
179          */
180         __u64                   config;
181
182         union {
183                 __u64           sample_period;
184                 __u64           sample_freq;
185         };
186
187         __u64                   sample_type;
188         __u64                   read_format;
189
190         __u64                   disabled       :  1, /* off by default        */
191                                 inherit        :  1, /* children inherit it   */
192                                 pinned         :  1, /* must always be on PMU */
193                                 exclusive      :  1, /* only group on PMU     */
194                                 exclude_user   :  1, /* don't count user      */
195                                 exclude_kernel :  1, /* ditto kernel          */
196                                 exclude_hv     :  1, /* ditto hypervisor      */
197                                 exclude_idle   :  1, /* don't count when idle */
198                                 mmap           :  1, /* include mmap data     */
199                                 comm           :  1, /* include comm data     */
200                                 freq           :  1, /* use freq, not period  */
201                                 inherit_stat   :  1, /* per task counts       */
202                                 enable_on_exec :  1, /* next exec enables     */
203                                 task           :  1, /* trace fork/exit       */
204                                 watermark      :  1, /* wakeup_watermark      */
205
206                                 __reserved_1   : 49;
207
208         union {
209                 __u32           wakeup_events;    /* wakeup every n events */
210                 __u32           wakeup_watermark; /* bytes before wakeup   */
211         };
212         __u32                   __reserved_2;
213
214         __u64                   __reserved_3;
215 };
216
217 /*
218  * Ioctls that can be done on a perf event fd:
219  */
220 #define PERF_EVENT_IOC_ENABLE           _IO ('$', 0)
221 #define PERF_EVENT_IOC_DISABLE          _IO ('$', 1)
222 #define PERF_EVENT_IOC_REFRESH          _IO ('$', 2)
223 #define PERF_EVENT_IOC_RESET            _IO ('$', 3)
224 #define PERF_EVENT_IOC_PERIOD           _IOW('$', 4, __u64)
225 #define PERF_EVENT_IOC_SET_OUTPUT       _IO ('$', 5)
226 #define PERF_EVENT_IOC_SET_FILTER       _IOW('$', 6, char *)
227
228 enum perf_event_ioc_flags {
229         PERF_IOC_FLAG_GROUP             = 1U << 0,
230 };
231
232 /*
233  * Structure of the page that can be mapped via mmap
234  */
235 struct perf_event_mmap_page {
236         __u32   version;                /* version number of this structure */
237         __u32   compat_version;         /* lowest version this is compat with */
238
239         /*
240          * Bits needed to read the hw events in user-space.
241          *
242          *   u32 seq;
243          *   s64 count;
244          *
245          *   do {
246          *     seq = pc->lock;
247          *
248          *     barrier()
249          *     if (pc->index) {
250          *       count = pmc_read(pc->index - 1);
251          *       count += pc->offset;
252          *     } else
253          *       goto regular_read;
254          *
255          *     barrier();
256          *   } while (pc->lock != seq);
257          *
258          * NOTE: for obvious reason this only works on self-monitoring
259          *       processes.
260          */
261         __u32   lock;                   /* seqlock for synchronization */
262         __u32   index;                  /* hardware event identifier */
263         __s64   offset;                 /* add to hardware event value */
264         __u64   time_enabled;           /* time event active */
265         __u64   time_running;           /* time event on cpu */
266
267                 /*
268                  * Hole for extension of the self monitor capabilities
269                  */
270
271         __u64   __reserved[123];        /* align to 1k */
272
273         /*
274          * Control data for the mmap() data buffer.
275          *
276          * User-space reading the @data_head value should issue an rmb(), on
277          * SMP capable platforms, after reading this value -- see
278          * perf_event_wakeup().
279          *
280          * When the mapping is PROT_WRITE the @data_tail value should be
281          * written by userspace to reflect the last read data. In this case
282          * the kernel will not over-write unread data.
283          */
284         __u64   data_head;              /* head in the data section */
285         __u64   data_tail;              /* user-space written tail */
286 };
287
288 #define PERF_RECORD_MISC_CPUMODE_MASK           (3 << 0)
289 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN                (0 << 0)
290 #define PERF_RECORD_MISC_KERNEL                 (1 << 0)
291 #define PERF_RECORD_MISC_USER                   (2 << 0)
292 #define PERF_RECORD_MISC_HYPERVISOR             (3 << 0)
293
294 struct perf_event_header {
295         __u32   type;
296         __u16   misc;
297         __u16   size;
298 };
299
300 enum perf_event_type {
301
302         /*
303          * The MMAP events record the PROT_EXEC mappings so that we can
304          * correlate userspace IPs to code. They have the following structure:
305          *
306          * struct {
307          *      struct perf_event_header        header;
308          *
309          *      u32                             pid, tid;
310          *      u64                             addr;
311          *      u64                             len;
312          *      u64                             pgoff;
313          *      char                            filename[];
314          * };
315          */
316         PERF_RECORD_MMAP                        = 1,
317
318         /*
319          * struct {
320          *      struct perf_event_header        header;
321          *      u64                             id;
322          *      u64                             lost;
323          * };
324          */
325         PERF_RECORD_LOST                        = 2,
326
327         /*
328          * struct {
329          *      struct perf_event_header        header;
330          *
331          *      u32                             pid, tid;
332          *      char                            comm[];
333          * };
334          */
335         PERF_RECORD_COMM                        = 3,
336
337         /*
338          * struct {
339          *      struct perf_event_header        header;
340          *      u32                             pid, ppid;
341          *      u32                             tid, ptid;
342          *      u64                             time;
343          * };
344          */
345         PERF_RECORD_EXIT                        = 4,
346
347         /*
348          * struct {
349          *      struct perf_event_header        header;
350          *      u64                             time;
351          *      u64                             id;
352          *      u64                             stream_id;
353          * };
354          */
355         PERF_RECORD_THROTTLE            = 5,
356         PERF_RECORD_UNTHROTTLE          = 6,
357
358         /*
359          * struct {
360          *      struct perf_event_header        header;
361          *      u32                             pid, ppid;
362          *      u32                             tid, ptid;
363          *      u64                             time;
364          * };
365          */
366         PERF_RECORD_FORK                        = 7,
367
368         /*
369          * struct {
370          *      struct perf_event_header        header;
371          *      u32                             pid, tid;
372          *
373          *      struct read_format              values;
374          * };
375          */
376         PERF_RECORD_READ                        = 8,
377
378         /*
379          * struct {
380          *      struct perf_event_header        header;
381          *
382          *      { u64                   ip;       } && PERF_SAMPLE_IP
383          *      { u32                   pid, tid; } && PERF_SAMPLE_TID
384          *      { u64                   time;     } && PERF_SAMPLE_TIME
385          *      { u64                   addr;     } && PERF_SAMPLE_ADDR
386          *      { u64                   id;       } && PERF_SAMPLE_ID
387          *      { u64                   stream_id;} && PERF_SAMPLE_STREAM_ID
388          *      { u32                   cpu, res; } && PERF_SAMPLE_CPU
389          *      { u64                   period;   } && PERF_SAMPLE_PERIOD
390          *
391          *      { struct read_format    values;   } && PERF_SAMPLE_READ
392          *
393          *      { u64                   nr,
394          *        u64                   ips[nr];  } && PERF_SAMPLE_CALLCHAIN
395          *
396          *      #
397          *      # The RAW record below is opaque data wrt the ABI
398          *      #
399          *      # That is, the ABI doesn't make any promises wrt to
400          *      # the stability of its content, it may vary depending
401          *      # on event, hardware, kernel version and phase of
402          *      # the moon.
403          *      #
404          *      # In other words, PERF_SAMPLE_RAW contents are not an ABI.
405          *      #
406          *
407          *      { u32                   size;
408          *        char                  data[size];}&& PERF_SAMPLE_RAW
409          * };
410          */
411         PERF_RECORD_SAMPLE              = 9,
412
413         PERF_RECORD_MAX,                        /* non-ABI */
414 };
415
416 enum perf_callchain_context {
417         PERF_CONTEXT_HV                 = (__u64)-32,
418         PERF_CONTEXT_KERNEL             = (__u64)-128,
419         PERF_CONTEXT_USER               = (__u64)-512,
420
421         PERF_CONTEXT_GUEST              = (__u64)-2048,
422         PERF_CONTEXT_GUEST_KERNEL       = (__u64)-2176,
423         PERF_CONTEXT_GUEST_USER         = (__u64)-2560,
424
425         PERF_CONTEXT_MAX                = (__u64)-4095,
426 };
427
428 #define PERF_FLAG_FD_NO_GROUP   (1U << 0)
429 #define PERF_FLAG_FD_OUTPUT     (1U << 1)
430
431 #ifdef __KERNEL__
432 /*
433  * Kernel-internal data types and definitions:
434  */
435
436 #ifdef CONFIG_PERF_EVENTS
437 # include <asm/perf_event.h>
438 #endif
439
440 #include <linux/list.h>
441 #include <linux/mutex.h>
442 #include <linux/rculist.h>
443 #include <linux/rcupdate.h>
444 #include <linux/spinlock.h>
445 #include <linux/hrtimer.h>
446 #include <linux/fs.h>
447 #include <linux/pid_namespace.h>
448 #include <linux/workqueue.h>
449 #include <asm/atomic.h>
450
451 #define PERF_MAX_STACK_DEPTH            255
452
453 struct perf_callchain_entry {
454         __u64                           nr;
455         __u64                           ip[PERF_MAX_STACK_DEPTH];
456 };
457
458 struct perf_raw_record {
459         u32                             size;
460         void                            *data;
461 };
462
463 struct task_struct;
464
465 /**
466  * struct hw_perf_event - performance event hardware details:
467  */
468 struct hw_perf_event {
469 #ifdef CONFIG_PERF_EVENTS
470         union {
471                 struct { /* hardware */
472                         u64             config;
473                         unsigned long   config_base;
474                         unsigned long   event_base;
475                         int             idx;
476                 };
477                 struct { /* software */
478                         s64             remaining;
479                         struct hrtimer  hrtimer;
480                 };
481         };
482         atomic64_t                      prev_count;
483         u64                             sample_period;
484         u64                             last_period;
485         atomic64_t                      period_left;
486         u64                             interrupts;
487
488         u64                             freq_count;
489         u64                             freq_interrupts;
490         u64                             freq_stamp;
491 #endif
492 };
493
494 struct perf_event;
495
496 /**
497  * struct pmu - generic performance monitoring unit
498  */
499 struct pmu {
500         int (*enable)                   (struct perf_event *event);
501         void (*disable)                 (struct perf_event *event);
502         void (*read)                    (struct perf_event *event);
503         void (*unthrottle)              (struct perf_event *event);
504 };
505
506 /**
507  * enum perf_event_active_state - the states of a event
508  */
509 enum perf_event_active_state {
510         PERF_EVENT_STATE_ERROR          = -2,
511         PERF_EVENT_STATE_OFF            = -1,
512         PERF_EVENT_STATE_INACTIVE       =  0,
513         PERF_EVENT_STATE_ACTIVE         =  1,
514 };
515
516 struct file;
517
518 struct perf_mmap_data {
519         struct rcu_head                 rcu_head;
520 #ifdef CONFIG_PERF_USE_VMALLOC
521         struct work_struct              work;
522 #endif
523         int                             data_order;
524         int                             nr_pages;       /* nr of data pages  */
525         int                             writable;       /* are we writable   */
526         int                             nr_locked;      /* nr pages mlocked  */
527
528         atomic_t                        poll;           /* POLL_ for wakeups */
529         atomic_t                        events;         /* event_id limit       */
530
531         atomic_long_t                   head;           /* write position    */
532         atomic_long_t                   done_head;      /* completed head    */
533
534         atomic_t                        lock;           /* concurrent writes */
535         atomic_t                        wakeup;         /* needs a wakeup    */
536         atomic_t                        lost;           /* nr records lost   */
537
538         long                            watermark;      /* wakeup watermark  */
539
540         struct perf_event_mmap_page     *user_page;
541         void                            *data_pages[0];
542 };
543
544 struct perf_pending_entry {
545         struct perf_pending_entry *next;
546         void (*func)(struct perf_pending_entry *);
547 };
548
549 /**
550  * struct perf_event - performance event kernel representation:
551  */
552 struct perf_event {
553 #ifdef CONFIG_PERF_EVENTS
554         struct list_head                group_entry;
555         struct list_head                event_entry;
556         struct list_head                sibling_list;
557         int                             nr_siblings;
558         struct perf_event               *group_leader;
559         struct perf_event               *output;
560         const struct pmu                *pmu;
561
562         enum perf_event_active_state    state;
563         atomic64_t                      count;
564
565         /*
566          * These are the total time in nanoseconds that the event
567          * has been enabled (i.e. eligible to run, and the task has
568          * been scheduled in, if this is a per-task event)
569          * and running (scheduled onto the CPU), respectively.
570          *
571          * They are computed from tstamp_enabled, tstamp_running and
572          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
573          */
574         u64                             total_time_enabled;
575         u64                             total_time_running;
576
577         /*
578          * These are timestamps used for computing total_time_enabled
579          * and total_time_running when the event is in INACTIVE or
580          * ACTIVE state, measured in nanoseconds from an arbitrary point
581          * in time.
582          * tstamp_enabled: the notional time when the event was enabled
583          * tstamp_running: the notional time when the event was scheduled on
584          * tstamp_stopped: in INACTIVE state, the notional time when the
585          *      event was scheduled off.
586          */
587         u64                             tstamp_enabled;
588         u64                             tstamp_running;
589         u64                             tstamp_stopped;
590
591         struct perf_event_attr  attr;
592         struct hw_perf_event            hw;
593
594         struct perf_event_context       *ctx;
595         struct file                     *filp;
596
597         /*
598          * These accumulate total time (in nanoseconds) that children
599          * events have been enabled and running, respectively.
600          */
601         atomic64_t                      child_total_time_enabled;
602         atomic64_t                      child_total_time_running;
603
604         /*
605          * Protect attach/detach and child_list:
606          */
607         struct mutex                    child_mutex;
608         struct list_head                child_list;
609         struct perf_event               *parent;
610
611         int                             oncpu;
612         int                             cpu;
613
614         struct list_head                owner_entry;
615         struct task_struct              *owner;
616
617         /* mmap bits */
618         struct mutex                    mmap_mutex;
619         atomic_t                        mmap_count;
620         struct perf_mmap_data           *data;
621
622         /* poll related */
623         wait_queue_head_t               waitq;
624         struct fasync_struct            *fasync;
625
626         /* delayed work for NMIs and such */
627         int                             pending_wakeup;
628         int                             pending_kill;
629         int                             pending_disable;
630         struct perf_pending_entry       pending;
631
632         atomic_t                        event_limit;
633
634         void (*destroy)(struct perf_event *);
635         struct rcu_head                 rcu_head;
636
637         struct pid_namespace            *ns;
638         u64                             id;
639
640 #ifdef CONFIG_EVENT_PROFILE
641         struct event_filter             *filter;
642 #endif
643
644 #endif /* CONFIG_PERF_EVENTS */
645 };
646
647 /**
648  * struct perf_event_context - event context structure
649  *
650  * Used as a container for task events and CPU events as well:
651  */
652 struct perf_event_context {
653         /*
654          * Protect the states of the events in the list,
655          * nr_active, and the list:
656          */
657         spinlock_t                      lock;
658         /*
659          * Protect the list of events.  Locking either mutex or lock
660          * is sufficient to ensure the list doesn't change; to change
661          * the list you need to lock both the mutex and the spinlock.
662          */
663         struct mutex                    mutex;
664
665         struct list_head                group_list;
666         struct list_head                event_list;
667         int                             nr_events;
668         int                             nr_active;
669         int                             is_active;
670         int                             nr_stat;
671         atomic_t                        refcount;
672         struct task_struct              *task;
673
674         /*
675          * Context clock, runs when context enabled.
676          */
677         u64                             time;
678         u64                             timestamp;
679
680         /*
681          * These fields let us detect when two contexts have both
682          * been cloned (inherited) from a common ancestor.
683          */
684         struct perf_event_context       *parent_ctx;
685         u64                             parent_gen;
686         u64                             generation;
687         int                             pin_count;
688         struct rcu_head                 rcu_head;
689 };
690
691 /**
692  * struct perf_event_cpu_context - per cpu event context structure
693  */
694 struct perf_cpu_context {
695         struct perf_event_context       ctx;
696         struct perf_event_context       *task_ctx;
697         int                             active_oncpu;
698         int                             max_pertask;
699         int                             exclusive;
700
701         /*
702          * Recursion avoidance:
703          *
704          * task, softirq, irq, nmi context
705          */
706         int                             recursion[4];
707 };
708
709 struct perf_output_handle {
710         struct perf_event               *event;
711         struct perf_mmap_data           *data;
712         unsigned long                   head;
713         unsigned long                   offset;
714         int                             nmi;
715         int                             sample;
716         int                             locked;
717 };
718
719 #ifdef CONFIG_PERF_EVENTS
720
721 /*
722  * Set by architecture code:
723  */
724 extern int perf_max_events;
725
726 extern const struct pmu *hw_perf_event_init(struct perf_event *event);
727
728 extern void perf_event_task_sched_in(struct task_struct *task, int cpu);
729 extern void perf_event_task_sched_out(struct task_struct *task,
730                                         struct task_struct *next, int cpu);
731 extern void perf_event_task_tick(struct task_struct *task, int cpu);
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 set_perf_event_pending(void);
736 extern void perf_event_do_pending(void);
737 extern void perf_event_print_debug(void);
738 extern void __perf_disable(void);
739 extern bool __perf_enable(void);
740 extern void perf_disable(void);
741 extern void perf_enable(void);
742 extern int perf_event_task_disable(void);
743 extern int perf_event_task_enable(void);
744 extern int hw_perf_group_sched_in(struct perf_event *group_leader,
745                struct perf_cpu_context *cpuctx,
746                struct perf_event_context *ctx, int cpu);
747 extern void perf_event_update_userpage(struct perf_event *event);
748
749 struct perf_sample_data {
750         u64                             type;
751
752         u64                             ip;
753         struct {
754                 u32     pid;
755                 u32     tid;
756         }                               tid_entry;
757         u64                             time;
758         u64                             addr;
759         u64                             id;
760         u64                             stream_id;
761         struct {
762                 u32     cpu;
763                 u32     reserved;
764         }                               cpu_entry;
765         u64                             period;
766         struct perf_callchain_entry     *callchain;
767         struct perf_raw_record          *raw;
768 };
769
770 extern void perf_output_sample(struct perf_output_handle *handle,
771                                struct perf_event_header *header,
772                                struct perf_sample_data *data,
773                                struct perf_event *event);
774 extern void perf_prepare_sample(struct perf_event_header *header,
775                                 struct perf_sample_data *data,
776                                 struct perf_event *event,
777                                 struct pt_regs *regs);
778
779 extern int perf_event_overflow(struct perf_event *event, int nmi,
780                                  struct perf_sample_data *data,
781                                  struct pt_regs *regs);
782
783 /*
784  * Return 1 for a software event, 0 for a hardware event
785  */
786 static inline int is_software_event(struct perf_event *event)
787 {
788         return (event->attr.type != PERF_TYPE_RAW) &&
789                 (event->attr.type != PERF_TYPE_HARDWARE) &&
790                 (event->attr.type != PERF_TYPE_HW_CACHE);
791 }
792
793 extern atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
794
795 extern void __perf_sw_event(u32, u64, int, struct pt_regs *, u64);
796
797 static inline void
798 perf_sw_event(u32 event_id, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
799 {
800         if (atomic_read(&perf_swevent_enabled[event_id]))
801                 __perf_sw_event(event_id, nr, nmi, regs, addr);
802 }
803
804 extern void __perf_event_mmap(struct vm_area_struct *vma);
805
806 static inline void perf_event_mmap(struct vm_area_struct *vma)
807 {
808         if (vma->vm_flags & VM_EXEC)
809                 __perf_event_mmap(vma);
810 }
811
812 extern void perf_event_comm(struct task_struct *tsk);
813 extern void perf_event_fork(struct task_struct *tsk);
814
815 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
816
817 extern int sysctl_perf_event_paranoid;
818 extern int sysctl_perf_event_mlock;
819 extern int sysctl_perf_event_sample_rate;
820
821 extern void perf_event_init(void);
822 extern void perf_tp_event(int event_id, u64 addr, u64 count,
823                                  void *record, int entry_size);
824
825 #ifndef perf_misc_flags
826 #define perf_misc_flags(regs)   (user_mode(regs) ? PERF_RECORD_MISC_USER : \
827                                  PERF_RECORD_MISC_KERNEL)
828 #define perf_instruction_pointer(regs)  instruction_pointer(regs)
829 #endif
830
831 extern int perf_output_begin(struct perf_output_handle *handle,
832                              struct perf_event *event, unsigned int size,
833                              int nmi, int sample);
834 extern void perf_output_end(struct perf_output_handle *handle);
835 extern void perf_output_copy(struct perf_output_handle *handle,
836                              const void *buf, unsigned int len);
837 #else
838 static inline void
839 perf_event_task_sched_in(struct task_struct *task, int cpu)             { }
840 static inline void
841 perf_event_task_sched_out(struct task_struct *task,
842                             struct task_struct *next, int cpu)          { }
843 static inline void
844 perf_event_task_tick(struct task_struct *task, int cpu)                 { }
845 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
846 static inline void perf_event_exit_task(struct task_struct *child)      { }
847 static inline void perf_event_free_task(struct task_struct *task)       { }
848 static inline void perf_event_do_pending(void)                          { }
849 static inline void perf_event_print_debug(void)                         { }
850 static inline void perf_disable(void)                                   { }
851 static inline void perf_enable(void)                                    { }
852 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
853 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
854
855 static inline void
856 perf_sw_event(u32 event_id, u64 nr, int nmi,
857                      struct pt_regs *regs, u64 addr)                    { }
858
859 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
860 static inline void perf_event_comm(struct task_struct *tsk)             { }
861 static inline void perf_event_fork(struct task_struct *tsk)             { }
862 static inline void perf_event_init(void)                                { }
863
864 #endif
865
866 #define perf_output_put(handle, x) \
867         perf_output_copy((handle), &(x), sizeof(x))
868
869 #endif /* __KERNEL__ */
870 #endif /* _LINUX_PERF_EVENT_H */