Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/sparc-2.6
[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         PERF_TYPE_BREAKPOINT                    = 5,
35
36         PERF_TYPE_MAX,                          /* non-ABI */
37 };
38
39 /*
40  * Generalized performance event event_id types, used by the
41  * attr.event_id parameter of the sys_perf_event_open()
42  * syscall:
43  */
44 enum perf_hw_id {
45         /*
46          * Common hardware events, generalized by the kernel:
47          */
48         PERF_COUNT_HW_CPU_CYCLES                = 0,
49         PERF_COUNT_HW_INSTRUCTIONS              = 1,
50         PERF_COUNT_HW_CACHE_REFERENCES          = 2,
51         PERF_COUNT_HW_CACHE_MISSES              = 3,
52         PERF_COUNT_HW_BRANCH_INSTRUCTIONS       = 4,
53         PERF_COUNT_HW_BRANCH_MISSES             = 5,
54         PERF_COUNT_HW_BUS_CYCLES                = 6,
55
56         PERF_COUNT_HW_MAX,                      /* non-ABI */
57 };
58
59 /*
60  * Generalized hardware cache events:
61  *
62  *       { L1-D, L1-I, LLC, ITLB, DTLB, BPU } x
63  *       { read, write, prefetch } x
64  *       { accesses, misses }
65  */
66 enum perf_hw_cache_id {
67         PERF_COUNT_HW_CACHE_L1D                 = 0,
68         PERF_COUNT_HW_CACHE_L1I                 = 1,
69         PERF_COUNT_HW_CACHE_LL                  = 2,
70         PERF_COUNT_HW_CACHE_DTLB                = 3,
71         PERF_COUNT_HW_CACHE_ITLB                = 4,
72         PERF_COUNT_HW_CACHE_BPU                 = 5,
73
74         PERF_COUNT_HW_CACHE_MAX,                /* non-ABI */
75 };
76
77 enum perf_hw_cache_op_id {
78         PERF_COUNT_HW_CACHE_OP_READ             = 0,
79         PERF_COUNT_HW_CACHE_OP_WRITE            = 1,
80         PERF_COUNT_HW_CACHE_OP_PREFETCH         = 2,
81
82         PERF_COUNT_HW_CACHE_OP_MAX,             /* non-ABI */
83 };
84
85 enum perf_hw_cache_op_result_id {
86         PERF_COUNT_HW_CACHE_RESULT_ACCESS       = 0,
87         PERF_COUNT_HW_CACHE_RESULT_MISS         = 1,
88
89         PERF_COUNT_HW_CACHE_RESULT_MAX,         /* non-ABI */
90 };
91
92 /*
93  * Special "software" events provided by the kernel, even if the hardware
94  * does not support performance events. These events measure various
95  * physical and sw events of the kernel (and allow the profiling of them as
96  * well):
97  */
98 enum perf_sw_ids {
99         PERF_COUNT_SW_CPU_CLOCK                 = 0,
100         PERF_COUNT_SW_TASK_CLOCK                = 1,
101         PERF_COUNT_SW_PAGE_FAULTS               = 2,
102         PERF_COUNT_SW_CONTEXT_SWITCHES          = 3,
103         PERF_COUNT_SW_CPU_MIGRATIONS            = 4,
104         PERF_COUNT_SW_PAGE_FAULTS_MIN           = 5,
105         PERF_COUNT_SW_PAGE_FAULTS_MAJ           = 6,
106         PERF_COUNT_SW_ALIGNMENT_FAULTS          = 7,
107         PERF_COUNT_SW_EMULATION_FAULTS          = 8,
108
109         PERF_COUNT_SW_MAX,                      /* non-ABI */
110 };
111
112 /*
113  * Bits that can be set in attr.sample_type to request information
114  * in the overflow packets.
115  */
116 enum perf_event_sample_format {
117         PERF_SAMPLE_IP                          = 1U << 0,
118         PERF_SAMPLE_TID                         = 1U << 1,
119         PERF_SAMPLE_TIME                        = 1U << 2,
120         PERF_SAMPLE_ADDR                        = 1U << 3,
121         PERF_SAMPLE_READ                        = 1U << 4,
122         PERF_SAMPLE_CALLCHAIN                   = 1U << 5,
123         PERF_SAMPLE_ID                          = 1U << 6,
124         PERF_SAMPLE_CPU                         = 1U << 7,
125         PERF_SAMPLE_PERIOD                      = 1U << 8,
126         PERF_SAMPLE_STREAM_ID                   = 1U << 9,
127         PERF_SAMPLE_RAW                         = 1U << 10,
128
129         PERF_SAMPLE_MAX = 1U << 11,             /* non-ABI */
130 };
131
132 /*
133  * The format of the data returned by read() on a perf event fd,
134  * as specified by attr.read_format:
135  *
136  * struct read_format {
137  *      { u64           value;
138  *        { u64         time_enabled; } && PERF_FORMAT_ENABLED
139  *        { u64         time_running; } && PERF_FORMAT_RUNNING
140  *        { u64         id;           } && PERF_FORMAT_ID
141  *      } && !PERF_FORMAT_GROUP
142  *
143  *      { u64           nr;
144  *        { u64         time_enabled; } && PERF_FORMAT_ENABLED
145  *        { u64         time_running; } && PERF_FORMAT_RUNNING
146  *        { u64         value;
147  *          { u64       id;           } && PERF_FORMAT_ID
148  *        }             cntr[nr];
149  *      } && PERF_FORMAT_GROUP
150  * };
151  */
152 enum perf_event_read_format {
153         PERF_FORMAT_TOTAL_TIME_ENABLED          = 1U << 0,
154         PERF_FORMAT_TOTAL_TIME_RUNNING          = 1U << 1,
155         PERF_FORMAT_ID                          = 1U << 2,
156         PERF_FORMAT_GROUP                       = 1U << 3,
157
158         PERF_FORMAT_MAX = 1U << 4,              /* non-ABI */
159 };
160
161 #define PERF_ATTR_SIZE_VER0     64      /* sizeof first published struct */
162
163 /*
164  * Hardware event_id to monitor via a performance monitoring event:
165  */
166 struct perf_event_attr {
167
168         /*
169          * Major type: hardware/software/tracepoint/etc.
170          */
171         __u32                   type;
172
173         /*
174          * Size of the attr structure, for fwd/bwd compat.
175          */
176         __u32                   size;
177
178         /*
179          * Type specific configuration information.
180          */
181         __u64                   config;
182
183         union {
184                 __u64           sample_period;
185                 __u64           sample_freq;
186         };
187
188         __u64                   sample_type;
189         __u64                   read_format;
190
191         __u64                   disabled       :  1, /* off by default        */
192                                 inherit        :  1, /* children inherit it   */
193                                 pinned         :  1, /* must always be on PMU */
194                                 exclusive      :  1, /* only group on PMU     */
195                                 exclude_user   :  1, /* don't count user      */
196                                 exclude_kernel :  1, /* ditto kernel          */
197                                 exclude_hv     :  1, /* ditto hypervisor      */
198                                 exclude_idle   :  1, /* don't count when idle */
199                                 mmap           :  1, /* include mmap data     */
200                                 comm           :  1, /* include comm data     */
201                                 freq           :  1, /* use freq, not period  */
202                                 inherit_stat   :  1, /* per task counts       */
203                                 enable_on_exec :  1, /* next exec enables     */
204                                 task           :  1, /* trace fork/exit       */
205                                 watermark      :  1, /* wakeup_watermark      */
206
207                                 __reserved_1   : 49;
208
209         union {
210                 __u32           wakeup_events;    /* wakeup every n events */
211                 __u32           wakeup_watermark; /* bytes before wakeup   */
212         };
213
214         __u32                   bp_type;
215         __u64                   bp_addr;
216         __u64                   bp_len;
217 };
218
219 /*
220  * Ioctls that can be done on a perf event fd:
221  */
222 #define PERF_EVENT_IOC_ENABLE           _IO ('$', 0)
223 #define PERF_EVENT_IOC_DISABLE          _IO ('$', 1)
224 #define PERF_EVENT_IOC_REFRESH          _IO ('$', 2)
225 #define PERF_EVENT_IOC_RESET            _IO ('$', 3)
226 #define PERF_EVENT_IOC_PERIOD           _IOW('$', 4, __u64)
227 #define PERF_EVENT_IOC_SET_OUTPUT       _IO ('$', 5)
228 #define PERF_EVENT_IOC_SET_FILTER       _IOW('$', 6, char *)
229
230 enum perf_event_ioc_flags {
231         PERF_IOC_FLAG_GROUP             = 1U << 0,
232 };
233
234 /*
235  * Structure of the page that can be mapped via mmap
236  */
237 struct perf_event_mmap_page {
238         __u32   version;                /* version number of this structure */
239         __u32   compat_version;         /* lowest version this is compat with */
240
241         /*
242          * Bits needed to read the hw events in user-space.
243          *
244          *   u32 seq;
245          *   s64 count;
246          *
247          *   do {
248          *     seq = pc->lock;
249          *
250          *     barrier()
251          *     if (pc->index) {
252          *       count = pmc_read(pc->index - 1);
253          *       count += pc->offset;
254          *     } else
255          *       goto regular_read;
256          *
257          *     barrier();
258          *   } while (pc->lock != seq);
259          *
260          * NOTE: for obvious reason this only works on self-monitoring
261          *       processes.
262          */
263         __u32   lock;                   /* seqlock for synchronization */
264         __u32   index;                  /* hardware event identifier */
265         __s64   offset;                 /* add to hardware event value */
266         __u64   time_enabled;           /* time event active */
267         __u64   time_running;           /* time event on cpu */
268
269                 /*
270                  * Hole for extension of the self monitor capabilities
271                  */
272
273         __u64   __reserved[123];        /* align to 1k */
274
275         /*
276          * Control data for the mmap() data buffer.
277          *
278          * User-space reading the @data_head value should issue an rmb(), on
279          * SMP capable platforms, after reading this value -- see
280          * perf_event_wakeup().
281          *
282          * When the mapping is PROT_WRITE the @data_tail value should be
283          * written by userspace to reflect the last read data. In this case
284          * the kernel will not over-write unread data.
285          */
286         __u64   data_head;              /* head in the data section */
287         __u64   data_tail;              /* user-space written tail */
288 };
289
290 #define PERF_RECORD_MISC_CPUMODE_MASK           (3 << 0)
291 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN        (0 << 0)
292 #define PERF_RECORD_MISC_KERNEL                 (1 << 0)
293 #define PERF_RECORD_MISC_USER                   (2 << 0)
294 #define PERF_RECORD_MISC_HYPERVISOR             (3 << 0)
295
296 struct perf_event_header {
297         __u32   type;
298         __u16   misc;
299         __u16   size;
300 };
301
302 enum perf_event_type {
303
304         /*
305          * The MMAP events record the PROT_EXEC mappings so that we can
306          * correlate userspace IPs to code. They have the following structure:
307          *
308          * struct {
309          *      struct perf_event_header        header;
310          *
311          *      u32                             pid, tid;
312          *      u64                             addr;
313          *      u64                             len;
314          *      u64                             pgoff;
315          *      char                            filename[];
316          * };
317          */
318         PERF_RECORD_MMAP                        = 1,
319
320         /*
321          * struct {
322          *      struct perf_event_header        header;
323          *      u64                             id;
324          *      u64                             lost;
325          * };
326          */
327         PERF_RECORD_LOST                        = 2,
328
329         /*
330          * struct {
331          *      struct perf_event_header        header;
332          *
333          *      u32                             pid, tid;
334          *      char                            comm[];
335          * };
336          */
337         PERF_RECORD_COMM                        = 3,
338
339         /*
340          * struct {
341          *      struct perf_event_header        header;
342          *      u32                             pid, ppid;
343          *      u32                             tid, ptid;
344          *      u64                             time;
345          * };
346          */
347         PERF_RECORD_EXIT                        = 4,
348
349         /*
350          * struct {
351          *      struct perf_event_header        header;
352          *      u64                             time;
353          *      u64                             id;
354          *      u64                             stream_id;
355          * };
356          */
357         PERF_RECORD_THROTTLE                    = 5,
358         PERF_RECORD_UNTHROTTLE                  = 6,
359
360         /*
361          * struct {
362          *      struct perf_event_header        header;
363          *      u32                             pid, ppid;
364          *      u32                             tid, ptid;
365          *      u64                             time;
366          * };
367          */
368         PERF_RECORD_FORK                        = 7,
369
370         /*
371          * struct {
372          *      struct perf_event_header        header;
373          *      u32                             pid, tid;
374          *
375          *      struct read_format              values;
376          * };
377          */
378         PERF_RECORD_READ                        = 8,
379
380         /*
381          * struct {
382          *      struct perf_event_header        header;
383          *
384          *      { u64                   ip;       } && PERF_SAMPLE_IP
385          *      { u32                   pid, tid; } && PERF_SAMPLE_TID
386          *      { u64                   time;     } && PERF_SAMPLE_TIME
387          *      { u64                   addr;     } && PERF_SAMPLE_ADDR
388          *      { u64                   id;       } && PERF_SAMPLE_ID
389          *      { u64                   stream_id;} && PERF_SAMPLE_STREAM_ID
390          *      { u32                   cpu, res; } && PERF_SAMPLE_CPU
391          *      { u64                   period;   } && PERF_SAMPLE_PERIOD
392          *
393          *      { struct read_format    values;   } && PERF_SAMPLE_READ
394          *
395          *      { u64                   nr,
396          *        u64                   ips[nr];  } && PERF_SAMPLE_CALLCHAIN
397          *
398          *      #
399          *      # The RAW record below is opaque data wrt the ABI
400          *      #
401          *      # That is, the ABI doesn't make any promises wrt to
402          *      # the stability of its content, it may vary depending
403          *      # on event, hardware, kernel version and phase of
404          *      # the moon.
405          *      #
406          *      # In other words, PERF_SAMPLE_RAW contents are not an ABI.
407          *      #
408          *
409          *      { u32                   size;
410          *        char                  data[size];}&& PERF_SAMPLE_RAW
411          * };
412          */
413         PERF_RECORD_SAMPLE                      = 9,
414
415         PERF_RECORD_MAX,                        /* non-ABI */
416 };
417
418 enum perf_callchain_context {
419         PERF_CONTEXT_HV                 = (__u64)-32,
420         PERF_CONTEXT_KERNEL             = (__u64)-128,
421         PERF_CONTEXT_USER               = (__u64)-512,
422
423         PERF_CONTEXT_GUEST              = (__u64)-2048,
424         PERF_CONTEXT_GUEST_KERNEL       = (__u64)-2176,
425         PERF_CONTEXT_GUEST_USER         = (__u64)-2560,
426
427         PERF_CONTEXT_MAX                = (__u64)-4095,
428 };
429
430 #define PERF_FLAG_FD_NO_GROUP   (1U << 0)
431 #define PERF_FLAG_FD_OUTPUT     (1U << 1)
432
433 #ifdef __KERNEL__
434 /*
435  * Kernel-internal data types and definitions:
436  */
437
438 #ifdef CONFIG_PERF_EVENTS
439 # include <asm/perf_event.h>
440 #endif
441
442 #ifdef CONFIG_HAVE_HW_BREAKPOINT
443 #include <asm/hw_breakpoint.h>
444 #endif
445
446 #include <linux/list.h>
447 #include <linux/mutex.h>
448 #include <linux/rculist.h>
449 #include <linux/rcupdate.h>
450 #include <linux/spinlock.h>
451 #include <linux/hrtimer.h>
452 #include <linux/fs.h>
453 #include <linux/pid_namespace.h>
454 #include <linux/workqueue.h>
455 #include <linux/ftrace.h>
456 #include <linux/cpu.h>
457 #include <asm/atomic.h>
458
459 #define PERF_MAX_STACK_DEPTH            255
460
461 struct perf_callchain_entry {
462         __u64                           nr;
463         __u64                           ip[PERF_MAX_STACK_DEPTH];
464 };
465
466 struct perf_raw_record {
467         u32                             size;
468         void                            *data;
469 };
470
471 struct task_struct;
472
473 /**
474  * struct hw_perf_event - performance event hardware details:
475  */
476 struct hw_perf_event {
477 #ifdef CONFIG_PERF_EVENTS
478         union {
479                 struct { /* hardware */
480                         u64             config;
481                         u64             last_tag;
482                         unsigned long   config_base;
483                         unsigned long   event_base;
484                         int             idx;
485                         int             last_cpu;
486                 };
487                 struct { /* software */
488                         s64             remaining;
489                         struct hrtimer  hrtimer;
490                 };
491 #ifdef CONFIG_HAVE_HW_BREAKPOINT
492                 /* breakpoint */
493                 struct arch_hw_breakpoint       info;
494 #endif
495         };
496         atomic64_t                      prev_count;
497         u64                             sample_period;
498         u64                             last_period;
499         atomic64_t                      period_left;
500         u64                             interrupts;
501
502         u64                             freq_time_stamp;
503         u64                             freq_count_stamp;
504 #endif
505 };
506
507 struct perf_event;
508
509 /**
510  * struct pmu - generic performance monitoring unit
511  */
512 struct pmu {
513         int (*enable)                   (struct perf_event *event);
514         void (*disable)                 (struct perf_event *event);
515         int (*start)                    (struct perf_event *event);
516         void (*stop)                    (struct perf_event *event);
517         void (*read)                    (struct perf_event *event);
518         void (*unthrottle)              (struct perf_event *event);
519 };
520
521 /**
522  * enum perf_event_active_state - the states of a event
523  */
524 enum perf_event_active_state {
525         PERF_EVENT_STATE_ERROR          = -2,
526         PERF_EVENT_STATE_OFF            = -1,
527         PERF_EVENT_STATE_INACTIVE       =  0,
528         PERF_EVENT_STATE_ACTIVE         =  1,
529 };
530
531 struct file;
532
533 struct perf_mmap_data {
534         struct rcu_head                 rcu_head;
535 #ifdef CONFIG_PERF_USE_VMALLOC
536         struct work_struct              work;
537 #endif
538         int                             data_order;
539         int                             nr_pages;       /* nr of data pages  */
540         int                             writable;       /* are we writable   */
541         int                             nr_locked;      /* nr pages mlocked  */
542
543         atomic_t                        poll;           /* POLL_ for wakeups */
544         atomic_t                        events;         /* event_id limit       */
545
546         atomic_long_t                   head;           /* write position    */
547         atomic_long_t                   done_head;      /* completed head    */
548
549         atomic_t                        lock;           /* concurrent writes */
550         atomic_t                        wakeup;         /* needs a wakeup    */
551         atomic_t                        lost;           /* nr records lost   */
552
553         long                            watermark;      /* wakeup watermark  */
554
555         struct perf_event_mmap_page     *user_page;
556         void                            *data_pages[0];
557 };
558
559 struct perf_pending_entry {
560         struct perf_pending_entry *next;
561         void (*func)(struct perf_pending_entry *);
562 };
563
564 struct perf_sample_data;
565
566 typedef void (*perf_overflow_handler_t)(struct perf_event *, int,
567                                         struct perf_sample_data *,
568                                         struct pt_regs *regs);
569
570 enum perf_group_flag {
571         PERF_GROUP_SOFTWARE = 0x1,
572 };
573
574 /**
575  * struct perf_event - performance event kernel representation:
576  */
577 struct perf_event {
578 #ifdef CONFIG_PERF_EVENTS
579         struct list_head                group_entry;
580         struct list_head                event_entry;
581         struct list_head                sibling_list;
582         int                             nr_siblings;
583         int                             group_flags;
584         struct perf_event               *group_leader;
585         struct perf_event               *output;
586         const struct pmu                *pmu;
587
588         enum perf_event_active_state    state;
589         atomic64_t                      count;
590
591         /*
592          * These are the total time in nanoseconds that the event
593          * has been enabled (i.e. eligible to run, and the task has
594          * been scheduled in, if this is a per-task event)
595          * and running (scheduled onto the CPU), respectively.
596          *
597          * They are computed from tstamp_enabled, tstamp_running and
598          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
599          */
600         u64                             total_time_enabled;
601         u64                             total_time_running;
602
603         /*
604          * These are timestamps used for computing total_time_enabled
605          * and total_time_running when the event is in INACTIVE or
606          * ACTIVE state, measured in nanoseconds from an arbitrary point
607          * in time.
608          * tstamp_enabled: the notional time when the event was enabled
609          * tstamp_running: the notional time when the event was scheduled on
610          * tstamp_stopped: in INACTIVE state, the notional time when the
611          *      event was scheduled off.
612          */
613         u64                             tstamp_enabled;
614         u64                             tstamp_running;
615         u64                             tstamp_stopped;
616
617         struct perf_event_attr          attr;
618         struct hw_perf_event            hw;
619
620         struct perf_event_context       *ctx;
621         struct file                     *filp;
622
623         /*
624          * These accumulate total time (in nanoseconds) that children
625          * events have been enabled and running, respectively.
626          */
627         atomic64_t                      child_total_time_enabled;
628         atomic64_t                      child_total_time_running;
629
630         /*
631          * Protect attach/detach and child_list:
632          */
633         struct mutex                    child_mutex;
634         struct list_head                child_list;
635         struct perf_event               *parent;
636
637         int                             oncpu;
638         int                             cpu;
639
640         struct list_head                owner_entry;
641         struct task_struct              *owner;
642
643         /* mmap bits */
644         struct mutex                    mmap_mutex;
645         atomic_t                        mmap_count;
646         struct perf_mmap_data           *data;
647
648         /* poll related */
649         wait_queue_head_t               waitq;
650         struct fasync_struct            *fasync;
651
652         /* delayed work for NMIs and such */
653         int                             pending_wakeup;
654         int                             pending_kill;
655         int                             pending_disable;
656         struct perf_pending_entry       pending;
657
658         atomic_t                        event_limit;
659
660         void (*destroy)(struct perf_event *);
661         struct rcu_head                 rcu_head;
662
663         struct pid_namespace            *ns;
664         u64                             id;
665
666         perf_overflow_handler_t         overflow_handler;
667
668 #ifdef CONFIG_EVENT_TRACING
669         struct event_filter             *filter;
670 #endif
671
672 #endif /* CONFIG_PERF_EVENTS */
673 };
674
675 /**
676  * struct perf_event_context - event context structure
677  *
678  * Used as a container for task events and CPU events as well:
679  */
680 struct perf_event_context {
681         /*
682          * Protect the states of the events in the list,
683          * nr_active, and the list:
684          */
685         raw_spinlock_t                  lock;
686         /*
687          * Protect the list of events.  Locking either mutex or lock
688          * is sufficient to ensure the list doesn't change; to change
689          * the list you need to lock both the mutex and the spinlock.
690          */
691         struct mutex                    mutex;
692
693         struct list_head                pinned_groups;
694         struct list_head                flexible_groups;
695         struct list_head                event_list;
696         int                             nr_events;
697         int                             nr_active;
698         int                             is_active;
699         int                             nr_stat;
700         atomic_t                        refcount;
701         struct task_struct              *task;
702
703         /*
704          * Context clock, runs when context enabled.
705          */
706         u64                             time;
707         u64                             timestamp;
708
709         /*
710          * These fields let us detect when two contexts have both
711          * been cloned (inherited) from a common ancestor.
712          */
713         struct perf_event_context       *parent_ctx;
714         u64                             parent_gen;
715         u64                             generation;
716         int                             pin_count;
717         struct rcu_head                 rcu_head;
718 };
719
720 /**
721  * struct perf_event_cpu_context - per cpu event context structure
722  */
723 struct perf_cpu_context {
724         struct perf_event_context       ctx;
725         struct perf_event_context       *task_ctx;
726         int                             active_oncpu;
727         int                             max_pertask;
728         int                             exclusive;
729
730         /*
731          * Recursion avoidance:
732          *
733          * task, softirq, irq, nmi context
734          */
735         int                             recursion[4];
736 };
737
738 struct perf_output_handle {
739         struct perf_event               *event;
740         struct perf_mmap_data           *data;
741         unsigned long                   head;
742         unsigned long                   offset;
743         int                             nmi;
744         int                             sample;
745         int                             locked;
746 };
747
748 #ifdef CONFIG_PERF_EVENTS
749
750 /*
751  * Set by architecture code:
752  */
753 extern int perf_max_events;
754
755 extern const struct pmu *hw_perf_event_init(struct perf_event *event);
756
757 extern void perf_event_task_sched_in(struct task_struct *task);
758 extern void perf_event_task_sched_out(struct task_struct *task, struct task_struct *next);
759 extern void perf_event_task_tick(struct task_struct *task);
760 extern int perf_event_init_task(struct task_struct *child);
761 extern void perf_event_exit_task(struct task_struct *child);
762 extern void perf_event_free_task(struct task_struct *task);
763 extern void set_perf_event_pending(void);
764 extern void perf_event_do_pending(void);
765 extern void perf_event_print_debug(void);
766 extern void __perf_disable(void);
767 extern bool __perf_enable(void);
768 extern void perf_disable(void);
769 extern void perf_enable(void);
770 extern int perf_event_task_disable(void);
771 extern int perf_event_task_enable(void);
772 extern int hw_perf_group_sched_in(struct perf_event *group_leader,
773                struct perf_cpu_context *cpuctx,
774                struct perf_event_context *ctx);
775 extern void perf_event_update_userpage(struct perf_event *event);
776 extern int perf_event_release_kernel(struct perf_event *event);
777 extern struct perf_event *
778 perf_event_create_kernel_counter(struct perf_event_attr *attr,
779                                 int cpu,
780                                 pid_t pid,
781                                 perf_overflow_handler_t callback);
782 extern u64 perf_event_read_value(struct perf_event *event,
783                                  u64 *enabled, u64 *running);
784
785 struct perf_sample_data {
786         u64                             type;
787
788         u64                             ip;
789         struct {
790                 u32     pid;
791                 u32     tid;
792         }                               tid_entry;
793         u64                             time;
794         u64                             addr;
795         u64                             id;
796         u64                             stream_id;
797         struct {
798                 u32     cpu;
799                 u32     reserved;
800         }                               cpu_entry;
801         u64                             period;
802         struct perf_callchain_entry     *callchain;
803         struct perf_raw_record          *raw;
804 };
805
806 static inline
807 void perf_sample_data_init(struct perf_sample_data *data, u64 addr)
808 {
809         data->addr = addr;
810         data->raw  = NULL;
811 }
812
813 extern void perf_output_sample(struct perf_output_handle *handle,
814                                struct perf_event_header *header,
815                                struct perf_sample_data *data,
816                                struct perf_event *event);
817 extern void perf_prepare_sample(struct perf_event_header *header,
818                                 struct perf_sample_data *data,
819                                 struct perf_event *event,
820                                 struct pt_regs *regs);
821
822 extern int perf_event_overflow(struct perf_event *event, int nmi,
823                                  struct perf_sample_data *data,
824                                  struct pt_regs *regs);
825
826 /*
827  * Return 1 for a software event, 0 for a hardware event
828  */
829 static inline int is_software_event(struct perf_event *event)
830 {
831         switch (event->attr.type) {
832         case PERF_TYPE_SOFTWARE:
833         case PERF_TYPE_TRACEPOINT:
834         /* for now the breakpoint stuff also works as software event */
835         case PERF_TYPE_BREAKPOINT:
836                 return 1;
837         }
838         return 0;
839 }
840
841 extern atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
842
843 extern void __perf_sw_event(u32, u64, int, struct pt_regs *, u64);
844
845 extern void
846 perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip, int skip);
847
848 /*
849  * Take a snapshot of the regs. Skip ip and frame pointer to
850  * the nth caller. We only need a few of the regs:
851  * - ip for PERF_SAMPLE_IP
852  * - cs for user_mode() tests
853  * - bp for callchains
854  * - eflags, for future purposes, just in case
855  */
856 static inline void perf_fetch_caller_regs(struct pt_regs *regs, int skip)
857 {
858         unsigned long ip;
859
860         memset(regs, 0, sizeof(*regs));
861
862         switch (skip) {
863         case 1 :
864                 ip = CALLER_ADDR0;
865                 break;
866         case 2 :
867                 ip = CALLER_ADDR1;
868                 break;
869         case 3 :
870                 ip = CALLER_ADDR2;
871                 break;
872         case 4:
873                 ip = CALLER_ADDR3;
874                 break;
875         /* No need to support further for now */
876         default:
877                 ip = 0;
878         }
879
880         return perf_arch_fetch_caller_regs(regs, ip, skip);
881 }
882
883 static inline void
884 perf_sw_event(u32 event_id, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
885 {
886         if (atomic_read(&perf_swevent_enabled[event_id])) {
887                 struct pt_regs hot_regs;
888
889                 if (!regs) {
890                         perf_fetch_caller_regs(&hot_regs, 1);
891                         regs = &hot_regs;
892                 }
893                 __perf_sw_event(event_id, nr, nmi, regs, addr);
894         }
895 }
896
897 extern void __perf_event_mmap(struct vm_area_struct *vma);
898
899 static inline void perf_event_mmap(struct vm_area_struct *vma)
900 {
901         if (vma->vm_flags & VM_EXEC)
902                 __perf_event_mmap(vma);
903 }
904
905 extern void perf_event_comm(struct task_struct *tsk);
906 extern void perf_event_fork(struct task_struct *tsk);
907
908 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
909
910 extern int sysctl_perf_event_paranoid;
911 extern int sysctl_perf_event_mlock;
912 extern int sysctl_perf_event_sample_rate;
913
914 static inline bool perf_paranoid_tracepoint_raw(void)
915 {
916         return sysctl_perf_event_paranoid > -1;
917 }
918
919 static inline bool perf_paranoid_cpu(void)
920 {
921         return sysctl_perf_event_paranoid > 0;
922 }
923
924 static inline bool perf_paranoid_kernel(void)
925 {
926         return sysctl_perf_event_paranoid > 1;
927 }
928
929 extern void perf_event_init(void);
930 extern void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
931                           int entry_size, struct pt_regs *regs);
932 extern void perf_bp_event(struct perf_event *event, void *data);
933
934 #ifndef perf_misc_flags
935 #define perf_misc_flags(regs)   (user_mode(regs) ? PERF_RECORD_MISC_USER : \
936                                  PERF_RECORD_MISC_KERNEL)
937 #define perf_instruction_pointer(regs)  instruction_pointer(regs)
938 #endif
939
940 extern int perf_output_begin(struct perf_output_handle *handle,
941                              struct perf_event *event, unsigned int size,
942                              int nmi, int sample);
943 extern void perf_output_end(struct perf_output_handle *handle);
944 extern void perf_output_copy(struct perf_output_handle *handle,
945                              const void *buf, unsigned int len);
946 extern int perf_swevent_get_recursion_context(void);
947 extern void perf_swevent_put_recursion_context(int rctx);
948 extern void perf_event_enable(struct perf_event *event);
949 extern void perf_event_disable(struct perf_event *event);
950 #else
951 static inline void
952 perf_event_task_sched_in(struct task_struct *task)                      { }
953 static inline void
954 perf_event_task_sched_out(struct task_struct *task,
955                             struct task_struct *next)                   { }
956 static inline void
957 perf_event_task_tick(struct task_struct *task)                          { }
958 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
959 static inline void perf_event_exit_task(struct task_struct *child)      { }
960 static inline void perf_event_free_task(struct task_struct *task)       { }
961 static inline void perf_event_do_pending(void)                          { }
962 static inline void perf_event_print_debug(void)                         { }
963 static inline void perf_disable(void)                                   { }
964 static inline void perf_enable(void)                                    { }
965 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
966 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
967
968 static inline void
969 perf_sw_event(u32 event_id, u64 nr, int nmi,
970                      struct pt_regs *regs, u64 addr)                    { }
971 static inline void
972 perf_bp_event(struct perf_event *event, void *data)                     { }
973
974 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
975 static inline void perf_event_comm(struct task_struct *tsk)             { }
976 static inline void perf_event_fork(struct task_struct *tsk)             { }
977 static inline void perf_event_init(void)                                { }
978 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
979 static inline void perf_swevent_put_recursion_context(int rctx)         { }
980 static inline void perf_event_enable(struct perf_event *event)          { }
981 static inline void perf_event_disable(struct perf_event *event)         { }
982 #endif
983
984 #define perf_output_put(handle, x) \
985         perf_output_copy((handle), &(x), sizeof(x))
986
987 /*
988  * This has to have a higher priority than migration_notifier in sched.c.
989  */
990 #define perf_cpu_notifier(fn)                                   \
991 do {                                                            \
992         static struct notifier_block fn##_nb __cpuinitdata =    \
993                 { .notifier_call = fn, .priority = 20 };        \
994         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,             \
995                 (void *)(unsigned long)smp_processor_id());     \
996         fn(&fn##_nb, (unsigned long)CPU_STARTING,               \
997                 (void *)(unsigned long)smp_processor_id());     \
998         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                 \
999                 (void *)(unsigned long)smp_processor_id());     \
1000         register_cpu_notifier(&fn##_nb);                        \
1001 } while (0)
1002
1003 #endif /* __KERNEL__ */
1004 #endif /* _LINUX_PERF_EVENT_H */