6be4a0f9137ce2d475f3d34e4f6a7c9ab0a6c77a
[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                                 precise        :  1, /* OoO invariant counter */
207
208                                 __reserved_1   : 48;
209
210         union {
211                 __u32           wakeup_events;    /* wakeup every n events */
212                 __u32           wakeup_watermark; /* bytes before wakeup   */
213         };
214
215         __u32                   bp_type;
216         __u64                   bp_addr;
217         __u64                   bp_len;
218 };
219
220 /*
221  * Ioctls that can be done on a perf event fd:
222  */
223 #define PERF_EVENT_IOC_ENABLE           _IO ('$', 0)
224 #define PERF_EVENT_IOC_DISABLE          _IO ('$', 1)
225 #define PERF_EVENT_IOC_REFRESH          _IO ('$', 2)
226 #define PERF_EVENT_IOC_RESET            _IO ('$', 3)
227 #define PERF_EVENT_IOC_PERIOD           _IOW('$', 4, __u64)
228 #define PERF_EVENT_IOC_SET_OUTPUT       _IO ('$', 5)
229 #define PERF_EVENT_IOC_SET_FILTER       _IOW('$', 6, char *)
230
231 enum perf_event_ioc_flags {
232         PERF_IOC_FLAG_GROUP             = 1U << 0,
233 };
234
235 /*
236  * Structure of the page that can be mapped via mmap
237  */
238 struct perf_event_mmap_page {
239         __u32   version;                /* version number of this structure */
240         __u32   compat_version;         /* lowest version this is compat with */
241
242         /*
243          * Bits needed to read the hw events in user-space.
244          *
245          *   u32 seq;
246          *   s64 count;
247          *
248          *   do {
249          *     seq = pc->lock;
250          *
251          *     barrier()
252          *     if (pc->index) {
253          *       count = pmc_read(pc->index - 1);
254          *       count += pc->offset;
255          *     } else
256          *       goto regular_read;
257          *
258          *     barrier();
259          *   } while (pc->lock != seq);
260          *
261          * NOTE: for obvious reason this only works on self-monitoring
262          *       processes.
263          */
264         __u32   lock;                   /* seqlock for synchronization */
265         __u32   index;                  /* hardware event identifier */
266         __s64   offset;                 /* add to hardware event value */
267         __u64   time_enabled;           /* time event active */
268         __u64   time_running;           /* time event on cpu */
269
270                 /*
271                  * Hole for extension of the self monitor capabilities
272                  */
273
274         __u64   __reserved[123];        /* align to 1k */
275
276         /*
277          * Control data for the mmap() data buffer.
278          *
279          * User-space reading the @data_head value should issue an rmb(), on
280          * SMP capable platforms, after reading this value -- see
281          * perf_event_wakeup().
282          *
283          * When the mapping is PROT_WRITE the @data_tail value should be
284          * written by userspace to reflect the last read data. In this case
285          * the kernel will not over-write unread data.
286          */
287         __u64   data_head;              /* head in the data section */
288         __u64   data_tail;              /* user-space written tail */
289 };
290
291 #define PERF_RECORD_MISC_CPUMODE_MASK           (7 << 0)
292 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN        (0 << 0)
293 #define PERF_RECORD_MISC_KERNEL                 (1 << 0)
294 #define PERF_RECORD_MISC_USER                   (2 << 0)
295 #define PERF_RECORD_MISC_HYPERVISOR             (3 << 0)
296 #define PERF_RECORD_MISC_GUEST_KERNEL           (4 << 0)
297 #define PERF_RECORD_MISC_GUEST_USER             (5 << 0)
298
299 #define PERF_RECORD_MISC_EXACT                  (1 << 14)
300 /*
301  * Reserve the last bit to indicate some extended misc field
302  */
303 #define PERF_RECORD_MISC_EXT_RESERVED           (1 << 15)
304
305 struct perf_event_header {
306         __u32   type;
307         __u16   misc;
308         __u16   size;
309 };
310
311 enum perf_event_type {
312
313         /*
314          * The MMAP events record the PROT_EXEC mappings so that we can
315          * correlate userspace IPs to code. They have the following structure:
316          *
317          * struct {
318          *      struct perf_event_header        header;
319          *
320          *      u32                             pid, tid;
321          *      u64                             addr;
322          *      u64                             len;
323          *      u64                             pgoff;
324          *      char                            filename[];
325          * };
326          */
327         PERF_RECORD_MMAP                        = 1,
328
329         /*
330          * struct {
331          *      struct perf_event_header        header;
332          *      u64                             id;
333          *      u64                             lost;
334          * };
335          */
336         PERF_RECORD_LOST                        = 2,
337
338         /*
339          * struct {
340          *      struct perf_event_header        header;
341          *
342          *      u32                             pid, tid;
343          *      char                            comm[];
344          * };
345          */
346         PERF_RECORD_COMM                        = 3,
347
348         /*
349          * struct {
350          *      struct perf_event_header        header;
351          *      u32                             pid, ppid;
352          *      u32                             tid, ptid;
353          *      u64                             time;
354          * };
355          */
356         PERF_RECORD_EXIT                        = 4,
357
358         /*
359          * struct {
360          *      struct perf_event_header        header;
361          *      u64                             time;
362          *      u64                             id;
363          *      u64                             stream_id;
364          * };
365          */
366         PERF_RECORD_THROTTLE                    = 5,
367         PERF_RECORD_UNTHROTTLE                  = 6,
368
369         /*
370          * struct {
371          *      struct perf_event_header        header;
372          *      u32                             pid, ppid;
373          *      u32                             tid, ptid;
374          *      u64                             time;
375          * };
376          */
377         PERF_RECORD_FORK                        = 7,
378
379         /*
380          * struct {
381          *      struct perf_event_header        header;
382          *      u32                             pid, tid;
383          *
384          *      struct read_format              values;
385          * };
386          */
387         PERF_RECORD_READ                        = 8,
388
389         /*
390          * struct {
391          *      struct perf_event_header        header;
392          *
393          *      { u64                   ip;       } && PERF_SAMPLE_IP
394          *      { u32                   pid, tid; } && PERF_SAMPLE_TID
395          *      { u64                   time;     } && PERF_SAMPLE_TIME
396          *      { u64                   addr;     } && PERF_SAMPLE_ADDR
397          *      { u64                   id;       } && PERF_SAMPLE_ID
398          *      { u64                   stream_id;} && PERF_SAMPLE_STREAM_ID
399          *      { u32                   cpu, res; } && PERF_SAMPLE_CPU
400          *      { u64                   period;   } && PERF_SAMPLE_PERIOD
401          *
402          *      { struct read_format    values;   } && PERF_SAMPLE_READ
403          *
404          *      { u64                   nr,
405          *        u64                   ips[nr];  } && PERF_SAMPLE_CALLCHAIN
406          *
407          *      #
408          *      # The RAW record below is opaque data wrt the ABI
409          *      #
410          *      # That is, the ABI doesn't make any promises wrt to
411          *      # the stability of its content, it may vary depending
412          *      # on event, hardware, kernel version and phase of
413          *      # the moon.
414          *      #
415          *      # In other words, PERF_SAMPLE_RAW contents are not an ABI.
416          *      #
417          *
418          *      { u32                   size;
419          *        char                  data[size];}&& PERF_SAMPLE_RAW
420          * };
421          */
422         PERF_RECORD_SAMPLE                      = 9,
423
424         PERF_RECORD_MAX,                        /* non-ABI */
425 };
426
427 enum perf_callchain_context {
428         PERF_CONTEXT_HV                 = (__u64)-32,
429         PERF_CONTEXT_KERNEL             = (__u64)-128,
430         PERF_CONTEXT_USER               = (__u64)-512,
431
432         PERF_CONTEXT_GUEST              = (__u64)-2048,
433         PERF_CONTEXT_GUEST_KERNEL       = (__u64)-2176,
434         PERF_CONTEXT_GUEST_USER         = (__u64)-2560,
435
436         PERF_CONTEXT_MAX                = (__u64)-4095,
437 };
438
439 #define PERF_FLAG_FD_NO_GROUP   (1U << 0)
440 #define PERF_FLAG_FD_OUTPUT     (1U << 1)
441
442 #ifdef __KERNEL__
443 /*
444  * Kernel-internal data types and definitions:
445  */
446
447 #ifdef CONFIG_PERF_EVENTS
448 # include <asm/perf_event.h>
449 #endif
450
451 struct perf_guest_info_callbacks {
452         int (*is_in_guest) (void);
453         int (*is_user_mode) (void);
454         unsigned long (*get_guest_ip) (void);
455 };
456
457 #ifdef CONFIG_HAVE_HW_BREAKPOINT
458 #include <asm/hw_breakpoint.h>
459 #endif
460
461 #include <linux/list.h>
462 #include <linux/mutex.h>
463 #include <linux/rculist.h>
464 #include <linux/rcupdate.h>
465 #include <linux/spinlock.h>
466 #include <linux/hrtimer.h>
467 #include <linux/fs.h>
468 #include <linux/pid_namespace.h>
469 #include <linux/workqueue.h>
470 #include <linux/ftrace.h>
471 #include <linux/cpu.h>
472 #include <asm/atomic.h>
473
474 #define PERF_MAX_STACK_DEPTH            255
475
476 struct perf_callchain_entry {
477         __u64                           nr;
478         __u64                           ip[PERF_MAX_STACK_DEPTH];
479 };
480
481 struct perf_raw_record {
482         u32                             size;
483         void                            *data;
484 };
485
486 struct perf_branch_entry {
487         __u64                           from;
488         __u64                           to;
489         __u64                           flags;
490 };
491
492 struct perf_branch_stack {
493         __u64                           nr;
494         struct perf_branch_entry        entries[0];
495 };
496
497 struct task_struct;
498
499 /**
500  * struct hw_perf_event - performance event hardware details:
501  */
502 struct hw_perf_event {
503 #ifdef CONFIG_PERF_EVENTS
504         union {
505                 struct { /* hardware */
506                         u64             config;
507                         u64             last_tag;
508                         unsigned long   config_base;
509                         unsigned long   event_base;
510                         int             idx;
511                         int             last_cpu;
512                 };
513                 struct { /* software */
514                         s64             remaining;
515                         struct hrtimer  hrtimer;
516                 };
517 #ifdef CONFIG_HAVE_HW_BREAKPOINT
518                 /* breakpoint */
519                 struct arch_hw_breakpoint       info;
520 #endif
521         };
522         atomic64_t                      prev_count;
523         u64                             sample_period;
524         u64                             last_period;
525         atomic64_t                      period_left;
526         u64                             interrupts;
527
528         u64                             freq_time_stamp;
529         u64                             freq_count_stamp;
530 #endif
531 };
532
533 struct perf_event;
534
535 /**
536  * struct pmu - generic performance monitoring unit
537  */
538 struct pmu {
539         int (*enable)                   (struct perf_event *event);
540         void (*disable)                 (struct perf_event *event);
541         int (*start)                    (struct perf_event *event);
542         void (*stop)                    (struct perf_event *event);
543         void (*read)                    (struct perf_event *event);
544         void (*unthrottle)              (struct perf_event *event);
545 };
546
547 /**
548  * enum perf_event_active_state - the states of a event
549  */
550 enum perf_event_active_state {
551         PERF_EVENT_STATE_FREE           = -3,
552         PERF_EVENT_STATE_ERROR          = -2,
553         PERF_EVENT_STATE_OFF            = -1,
554         PERF_EVENT_STATE_INACTIVE       =  0,
555         PERF_EVENT_STATE_ACTIVE         =  1,
556 };
557
558 struct file;
559
560 struct perf_mmap_data {
561         struct rcu_head                 rcu_head;
562 #ifdef CONFIG_PERF_USE_VMALLOC
563         struct work_struct              work;
564 #endif
565         int                             data_order;
566         int                             nr_pages;       /* nr of data pages  */
567         int                             writable;       /* are we writable   */
568         int                             nr_locked;      /* nr pages mlocked  */
569
570         atomic_t                        poll;           /* POLL_ for wakeups */
571         atomic_t                        events;         /* event_id limit       */
572
573         atomic_long_t                   head;           /* write position    */
574         atomic_long_t                   done_head;      /* completed head    */
575
576         atomic_t                        lock;           /* concurrent writes */
577         atomic_t                        wakeup;         /* needs a wakeup    */
578         atomic_t                        lost;           /* nr records lost   */
579
580         long                            watermark;      /* wakeup watermark  */
581
582         struct perf_event_mmap_page     *user_page;
583         void                            *data_pages[0];
584 };
585
586 struct perf_pending_entry {
587         struct perf_pending_entry *next;
588         void (*func)(struct perf_pending_entry *);
589 };
590
591 struct perf_sample_data;
592
593 typedef void (*perf_overflow_handler_t)(struct perf_event *, int,
594                                         struct perf_sample_data *,
595                                         struct pt_regs *regs);
596
597 enum perf_group_flag {
598         PERF_GROUP_SOFTWARE = 0x1,
599 };
600
601 #define SWEVENT_HLIST_BITS      8
602 #define SWEVENT_HLIST_SIZE      (1 << SWEVENT_HLIST_BITS)
603
604 struct swevent_hlist {
605         struct hlist_head       heads[SWEVENT_HLIST_SIZE];
606         struct rcu_head         rcu_head;
607 };
608
609 /**
610  * struct perf_event - performance event kernel representation:
611  */
612 struct perf_event {
613 #ifdef CONFIG_PERF_EVENTS
614         struct list_head                group_entry;
615         struct list_head                event_entry;
616         struct list_head                sibling_list;
617         struct hlist_node               hlist_entry;
618         int                             nr_siblings;
619         int                             group_flags;
620         struct perf_event               *group_leader;
621         struct perf_event               *output;
622         const struct pmu                *pmu;
623
624         enum perf_event_active_state    state;
625         atomic64_t                      count;
626
627         /*
628          * These are the total time in nanoseconds that the event
629          * has been enabled (i.e. eligible to run, and the task has
630          * been scheduled in, if this is a per-task event)
631          * and running (scheduled onto the CPU), respectively.
632          *
633          * They are computed from tstamp_enabled, tstamp_running and
634          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
635          */
636         u64                             total_time_enabled;
637         u64                             total_time_running;
638
639         /*
640          * These are timestamps used for computing total_time_enabled
641          * and total_time_running when the event is in INACTIVE or
642          * ACTIVE state, measured in nanoseconds from an arbitrary point
643          * in time.
644          * tstamp_enabled: the notional time when the event was enabled
645          * tstamp_running: the notional time when the event was scheduled on
646          * tstamp_stopped: in INACTIVE state, the notional time when the
647          *      event was scheduled off.
648          */
649         u64                             tstamp_enabled;
650         u64                             tstamp_running;
651         u64                             tstamp_stopped;
652
653         struct perf_event_attr          attr;
654         struct hw_perf_event            hw;
655
656         struct perf_event_context       *ctx;
657         struct file                     *filp;
658
659         /*
660          * These accumulate total time (in nanoseconds) that children
661          * events have been enabled and running, respectively.
662          */
663         atomic64_t                      child_total_time_enabled;
664         atomic64_t                      child_total_time_running;
665
666         /*
667          * Protect attach/detach and child_list:
668          */
669         struct mutex                    child_mutex;
670         struct list_head                child_list;
671         struct perf_event               *parent;
672
673         int                             oncpu;
674         int                             cpu;
675
676         struct list_head                owner_entry;
677         struct task_struct              *owner;
678
679         /* mmap bits */
680         struct mutex                    mmap_mutex;
681         atomic_t                        mmap_count;
682         struct perf_mmap_data           *data;
683
684         /* poll related */
685         wait_queue_head_t               waitq;
686         struct fasync_struct            *fasync;
687
688         /* delayed work for NMIs and such */
689         int                             pending_wakeup;
690         int                             pending_kill;
691         int                             pending_disable;
692         struct perf_pending_entry       pending;
693
694         atomic_t                        event_limit;
695
696         void (*destroy)(struct perf_event *);
697         struct rcu_head                 rcu_head;
698
699         struct pid_namespace            *ns;
700         u64                             id;
701
702         perf_overflow_handler_t         overflow_handler;
703
704 #ifdef CONFIG_EVENT_TRACING
705         struct event_filter             *filter;
706 #endif
707
708 #endif /* CONFIG_PERF_EVENTS */
709 };
710
711 /**
712  * struct perf_event_context - event context structure
713  *
714  * Used as a container for task events and CPU events as well:
715  */
716 struct perf_event_context {
717         /*
718          * Protect the states of the events in the list,
719          * nr_active, and the list:
720          */
721         raw_spinlock_t                  lock;
722         /*
723          * Protect the list of events.  Locking either mutex or lock
724          * is sufficient to ensure the list doesn't change; to change
725          * the list you need to lock both the mutex and the spinlock.
726          */
727         struct mutex                    mutex;
728
729         struct list_head                pinned_groups;
730         struct list_head                flexible_groups;
731         struct list_head                event_list;
732         int                             nr_events;
733         int                             nr_active;
734         int                             is_active;
735         int                             nr_stat;
736         atomic_t                        refcount;
737         struct task_struct              *task;
738
739         /*
740          * Context clock, runs when context enabled.
741          */
742         u64                             time;
743         u64                             timestamp;
744
745         /*
746          * These fields let us detect when two contexts have both
747          * been cloned (inherited) from a common ancestor.
748          */
749         struct perf_event_context       *parent_ctx;
750         u64                             parent_gen;
751         u64                             generation;
752         int                             pin_count;
753         struct rcu_head                 rcu_head;
754 };
755
756 /**
757  * struct perf_event_cpu_context - per cpu event context structure
758  */
759 struct perf_cpu_context {
760         struct perf_event_context       ctx;
761         struct perf_event_context       *task_ctx;
762         int                             active_oncpu;
763         int                             max_pertask;
764         int                             exclusive;
765         struct swevent_hlist            *swevent_hlist;
766         struct mutex                    hlist_mutex;
767         int                             hlist_refcount;
768
769         /*
770          * Recursion avoidance:
771          *
772          * task, softirq, irq, nmi context
773          */
774         int                             recursion[4];
775 };
776
777 struct perf_output_handle {
778         struct perf_event               *event;
779         struct perf_mmap_data           *data;
780         unsigned long                   head;
781         unsigned long                   offset;
782         int                             nmi;
783         int                             sample;
784         int                             locked;
785 };
786
787 #ifdef CONFIG_PERF_EVENTS
788
789 /*
790  * Set by architecture code:
791  */
792 extern int perf_max_events;
793
794 extern const struct pmu *hw_perf_event_init(struct perf_event *event);
795
796 extern void perf_event_task_sched_in(struct task_struct *task);
797 extern void perf_event_task_sched_out(struct task_struct *task, struct task_struct *next);
798 extern void perf_event_task_tick(struct task_struct *task);
799 extern int perf_event_init_task(struct task_struct *child);
800 extern void perf_event_exit_task(struct task_struct *child);
801 extern void perf_event_free_task(struct task_struct *task);
802 extern void set_perf_event_pending(void);
803 extern void perf_event_do_pending(void);
804 extern void perf_event_print_debug(void);
805 extern void __perf_disable(void);
806 extern bool __perf_enable(void);
807 extern void perf_disable(void);
808 extern void perf_enable(void);
809 extern int perf_event_task_disable(void);
810 extern int perf_event_task_enable(void);
811 extern int hw_perf_group_sched_in(struct perf_event *group_leader,
812                struct perf_cpu_context *cpuctx,
813                struct perf_event_context *ctx);
814 extern void perf_event_update_userpage(struct perf_event *event);
815 extern int perf_event_release_kernel(struct perf_event *event);
816 extern struct perf_event *
817 perf_event_create_kernel_counter(struct perf_event_attr *attr,
818                                 int cpu,
819                                 pid_t pid,
820                                 perf_overflow_handler_t callback);
821 extern u64 perf_event_read_value(struct perf_event *event,
822                                  u64 *enabled, u64 *running);
823
824 struct perf_sample_data {
825         u64                             type;
826
827         u64                             ip;
828         struct {
829                 u32     pid;
830                 u32     tid;
831         }                               tid_entry;
832         u64                             time;
833         u64                             addr;
834         u64                             id;
835         u64                             stream_id;
836         struct {
837                 u32     cpu;
838                 u32     reserved;
839         }                               cpu_entry;
840         u64                             period;
841         struct perf_callchain_entry     *callchain;
842         struct perf_raw_record          *raw;
843 };
844
845 static inline
846 void perf_sample_data_init(struct perf_sample_data *data, u64 addr)
847 {
848         data->addr = addr;
849         data->raw  = NULL;
850 }
851
852 extern void perf_output_sample(struct perf_output_handle *handle,
853                                struct perf_event_header *header,
854                                struct perf_sample_data *data,
855                                struct perf_event *event);
856 extern void perf_prepare_sample(struct perf_event_header *header,
857                                 struct perf_sample_data *data,
858                                 struct perf_event *event,
859                                 struct pt_regs *regs);
860
861 extern int perf_event_overflow(struct perf_event *event, int nmi,
862                                  struct perf_sample_data *data,
863                                  struct pt_regs *regs);
864
865 /*
866  * Return 1 for a software event, 0 for a hardware event
867  */
868 static inline int is_software_event(struct perf_event *event)
869 {
870         switch (event->attr.type) {
871         case PERF_TYPE_SOFTWARE:
872         case PERF_TYPE_TRACEPOINT:
873         /* for now the breakpoint stuff also works as software event */
874         case PERF_TYPE_BREAKPOINT:
875                 return 1;
876         }
877         return 0;
878 }
879
880 extern atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
881
882 extern void __perf_sw_event(u32, u64, int, struct pt_regs *, u64);
883
884 extern void
885 perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip, int skip);
886
887 /*
888  * Take a snapshot of the regs. Skip ip and frame pointer to
889  * the nth caller. We only need a few of the regs:
890  * - ip for PERF_SAMPLE_IP
891  * - cs for user_mode() tests
892  * - bp for callchains
893  * - eflags, for future purposes, just in case
894  */
895 static inline void perf_fetch_caller_regs(struct pt_regs *regs, int skip)
896 {
897         unsigned long ip;
898
899         memset(regs, 0, sizeof(*regs));
900
901         switch (skip) {
902         case 1 :
903                 ip = CALLER_ADDR0;
904                 break;
905         case 2 :
906                 ip = CALLER_ADDR1;
907                 break;
908         case 3 :
909                 ip = CALLER_ADDR2;
910                 break;
911         case 4:
912                 ip = CALLER_ADDR3;
913                 break;
914         /* No need to support further for now */
915         default:
916                 ip = 0;
917         }
918
919         return perf_arch_fetch_caller_regs(regs, ip, skip);
920 }
921
922 static inline void
923 perf_sw_event(u32 event_id, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
924 {
925         if (atomic_read(&perf_swevent_enabled[event_id])) {
926                 struct pt_regs hot_regs;
927
928                 if (!regs) {
929                         perf_fetch_caller_regs(&hot_regs, 1);
930                         regs = &hot_regs;
931                 }
932                 __perf_sw_event(event_id, nr, nmi, regs, addr);
933         }
934 }
935
936 extern void __perf_event_mmap(struct vm_area_struct *vma);
937
938 static inline void perf_event_mmap(struct vm_area_struct *vma)
939 {
940         if (vma->vm_flags & VM_EXEC)
941                 __perf_event_mmap(vma);
942 }
943
944 extern struct perf_guest_info_callbacks *perf_guest_cbs;
945 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
946 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
947
948 extern void perf_event_comm(struct task_struct *tsk);
949 extern void perf_event_fork(struct task_struct *tsk);
950
951 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
952
953 extern int sysctl_perf_event_paranoid;
954 extern int sysctl_perf_event_mlock;
955 extern int sysctl_perf_event_sample_rate;
956
957 static inline bool perf_paranoid_tracepoint_raw(void)
958 {
959         return sysctl_perf_event_paranoid > -1;
960 }
961
962 static inline bool perf_paranoid_cpu(void)
963 {
964         return sysctl_perf_event_paranoid > 0;
965 }
966
967 static inline bool perf_paranoid_kernel(void)
968 {
969         return sysctl_perf_event_paranoid > 1;
970 }
971
972 extern void perf_event_init(void);
973 extern void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
974                           int entry_size, struct pt_regs *regs);
975 extern void perf_bp_event(struct perf_event *event, void *data);
976
977 #ifndef perf_misc_flags
978 #define perf_misc_flags(regs)   (user_mode(regs) ? PERF_RECORD_MISC_USER : \
979                                  PERF_RECORD_MISC_KERNEL)
980 #define perf_instruction_pointer(regs)  instruction_pointer(regs)
981 #endif
982
983 extern int perf_output_begin(struct perf_output_handle *handle,
984                              struct perf_event *event, unsigned int size,
985                              int nmi, int sample);
986 extern void perf_output_end(struct perf_output_handle *handle);
987 extern void perf_output_copy(struct perf_output_handle *handle,
988                              const void *buf, unsigned int len);
989 extern int perf_swevent_get_recursion_context(void);
990 extern void perf_swevent_put_recursion_context(int rctx);
991 extern void perf_event_enable(struct perf_event *event);
992 extern void perf_event_disable(struct perf_event *event);
993 #else
994 static inline void
995 perf_event_task_sched_in(struct task_struct *task)                      { }
996 static inline void
997 perf_event_task_sched_out(struct task_struct *task,
998                             struct task_struct *next)                   { }
999 static inline void
1000 perf_event_task_tick(struct task_struct *task)                          { }
1001 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
1002 static inline void perf_event_exit_task(struct task_struct *child)      { }
1003 static inline void perf_event_free_task(struct task_struct *task)       { }
1004 static inline void perf_event_do_pending(void)                          { }
1005 static inline void perf_event_print_debug(void)                         { }
1006 static inline void perf_disable(void)                                   { }
1007 static inline void perf_enable(void)                                    { }
1008 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
1009 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
1010
1011 static inline void
1012 perf_sw_event(u32 event_id, u64 nr, int nmi,
1013                      struct pt_regs *regs, u64 addr)                    { }
1014 static inline void
1015 perf_bp_event(struct perf_event *event, void *data)                     { }
1016
1017 static inline int perf_register_guest_info_callbacks
1018 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1019 static inline int perf_unregister_guest_info_callbacks
1020 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1021
1022 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
1023 static inline void perf_event_comm(struct task_struct *tsk)             { }
1024 static inline void perf_event_fork(struct task_struct *tsk)             { }
1025 static inline void perf_event_init(void)                                { }
1026 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
1027 static inline void perf_swevent_put_recursion_context(int rctx)         { }
1028 static inline void perf_event_enable(struct perf_event *event)          { }
1029 static inline void perf_event_disable(struct perf_event *event)         { }
1030 #endif
1031
1032 #define perf_output_put(handle, x) \
1033         perf_output_copy((handle), &(x), sizeof(x))
1034
1035 /*
1036  * This has to have a higher priority than migration_notifier in sched.c.
1037  */
1038 #define perf_cpu_notifier(fn)                                   \
1039 do {                                                            \
1040         static struct notifier_block fn##_nb __cpuinitdata =    \
1041                 { .notifier_call = fn, .priority = 20 };        \
1042         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,             \
1043                 (void *)(unsigned long)smp_processor_id());     \
1044         fn(&fn##_nb, (unsigned long)CPU_STARTING,               \
1045                 (void *)(unsigned long)smp_processor_id());     \
1046         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                 \
1047                 (void *)(unsigned long)smp_processor_id());     \
1048         register_cpu_notifier(&fn##_nb);                        \
1049 } while (0)
1050
1051 #endif /* __KERNEL__ */
1052 #endif /* _LINUX_PERF_EVENT_H */