perf_events: Fix races in group composition
[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                                  * precise_ip:
208                                  *
209                                  *  0 - SAMPLE_IP can have arbitrary skid
210                                  *  1 - SAMPLE_IP must have constant skid
211                                  *  2 - SAMPLE_IP requested to have 0 skid
212                                  *  3 - SAMPLE_IP must have 0 skid
213                                  *
214                                  *  See also PERF_RECORD_MISC_EXACT_IP
215                                  */
216                                 precise_ip     :  2, /* skid constraint       */
217
218                                 __reserved_1   : 47;
219
220         union {
221                 __u32           wakeup_events;    /* wakeup every n events */
222                 __u32           wakeup_watermark; /* bytes before wakeup   */
223         };
224
225         __u32                   bp_type;
226         __u64                   bp_addr;
227         __u64                   bp_len;
228 };
229
230 /*
231  * Ioctls that can be done on a perf event fd:
232  */
233 #define PERF_EVENT_IOC_ENABLE           _IO ('$', 0)
234 #define PERF_EVENT_IOC_DISABLE          _IO ('$', 1)
235 #define PERF_EVENT_IOC_REFRESH          _IO ('$', 2)
236 #define PERF_EVENT_IOC_RESET            _IO ('$', 3)
237 #define PERF_EVENT_IOC_PERIOD           _IOW('$', 4, __u64)
238 #define PERF_EVENT_IOC_SET_OUTPUT       _IO ('$', 5)
239 #define PERF_EVENT_IOC_SET_FILTER       _IOW('$', 6, char *)
240
241 enum perf_event_ioc_flags {
242         PERF_IOC_FLAG_GROUP             = 1U << 0,
243 };
244
245 /*
246  * Structure of the page that can be mapped via mmap
247  */
248 struct perf_event_mmap_page {
249         __u32   version;                /* version number of this structure */
250         __u32   compat_version;         /* lowest version this is compat with */
251
252         /*
253          * Bits needed to read the hw events in user-space.
254          *
255          *   u32 seq;
256          *   s64 count;
257          *
258          *   do {
259          *     seq = pc->lock;
260          *
261          *     barrier()
262          *     if (pc->index) {
263          *       count = pmc_read(pc->index - 1);
264          *       count += pc->offset;
265          *     } else
266          *       goto regular_read;
267          *
268          *     barrier();
269          *   } while (pc->lock != seq);
270          *
271          * NOTE: for obvious reason this only works on self-monitoring
272          *       processes.
273          */
274         __u32   lock;                   /* seqlock for synchronization */
275         __u32   index;                  /* hardware event identifier */
276         __s64   offset;                 /* add to hardware event value */
277         __u64   time_enabled;           /* time event active */
278         __u64   time_running;           /* time event on cpu */
279
280                 /*
281                  * Hole for extension of the self monitor capabilities
282                  */
283
284         __u64   __reserved[123];        /* align to 1k */
285
286         /*
287          * Control data for the mmap() data buffer.
288          *
289          * User-space reading the @data_head value should issue an rmb(), on
290          * SMP capable platforms, after reading this value -- see
291          * perf_event_wakeup().
292          *
293          * When the mapping is PROT_WRITE the @data_tail value should be
294          * written by userspace to reflect the last read data. In this case
295          * the kernel will not over-write unread data.
296          */
297         __u64   data_head;              /* head in the data section */
298         __u64   data_tail;              /* user-space written tail */
299 };
300
301 #define PERF_RECORD_MISC_CPUMODE_MASK           (7 << 0)
302 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN        (0 << 0)
303 #define PERF_RECORD_MISC_KERNEL                 (1 << 0)
304 #define PERF_RECORD_MISC_USER                   (2 << 0)
305 #define PERF_RECORD_MISC_HYPERVISOR             (3 << 0)
306 #define PERF_RECORD_MISC_GUEST_KERNEL           (4 << 0)
307 #define PERF_RECORD_MISC_GUEST_USER             (5 << 0)
308
309 /*
310  * Indicates that the content of PERF_SAMPLE_IP points to
311  * the actual instruction that triggered the event. See also
312  * perf_event_attr::precise_ip.
313  */
314 #define PERF_RECORD_MISC_EXACT_IP               (1 << 14)
315 /*
316  * Reserve the last bit to indicate some extended misc field
317  */
318 #define PERF_RECORD_MISC_EXT_RESERVED           (1 << 15)
319
320 struct perf_event_header {
321         __u32   type;
322         __u16   misc;
323         __u16   size;
324 };
325
326 enum perf_event_type {
327
328         /*
329          * The MMAP events record the PROT_EXEC mappings so that we can
330          * correlate userspace IPs to code. They have the following structure:
331          *
332          * struct {
333          *      struct perf_event_header        header;
334          *
335          *      u32                             pid, tid;
336          *      u64                             addr;
337          *      u64                             len;
338          *      u64                             pgoff;
339          *      char                            filename[];
340          * };
341          */
342         PERF_RECORD_MMAP                        = 1,
343
344         /*
345          * struct {
346          *      struct perf_event_header        header;
347          *      u64                             id;
348          *      u64                             lost;
349          * };
350          */
351         PERF_RECORD_LOST                        = 2,
352
353         /*
354          * struct {
355          *      struct perf_event_header        header;
356          *
357          *      u32                             pid, tid;
358          *      char                            comm[];
359          * };
360          */
361         PERF_RECORD_COMM                        = 3,
362
363         /*
364          * struct {
365          *      struct perf_event_header        header;
366          *      u32                             pid, ppid;
367          *      u32                             tid, ptid;
368          *      u64                             time;
369          * };
370          */
371         PERF_RECORD_EXIT                        = 4,
372
373         /*
374          * struct {
375          *      struct perf_event_header        header;
376          *      u64                             time;
377          *      u64                             id;
378          *      u64                             stream_id;
379          * };
380          */
381         PERF_RECORD_THROTTLE                    = 5,
382         PERF_RECORD_UNTHROTTLE                  = 6,
383
384         /*
385          * struct {
386          *      struct perf_event_header        header;
387          *      u32                             pid, ppid;
388          *      u32                             tid, ptid;
389          *      u64                             time;
390          * };
391          */
392         PERF_RECORD_FORK                        = 7,
393
394         /*
395          * struct {
396          *      struct perf_event_header        header;
397          *      u32                             pid, tid;
398          *
399          *      struct read_format              values;
400          * };
401          */
402         PERF_RECORD_READ                        = 8,
403
404         /*
405          * struct {
406          *      struct perf_event_header        header;
407          *
408          *      { u64                   ip;       } && PERF_SAMPLE_IP
409          *      { u32                   pid, tid; } && PERF_SAMPLE_TID
410          *      { u64                   time;     } && PERF_SAMPLE_TIME
411          *      { u64                   addr;     } && PERF_SAMPLE_ADDR
412          *      { u64                   id;       } && PERF_SAMPLE_ID
413          *      { u64                   stream_id;} && PERF_SAMPLE_STREAM_ID
414          *      { u32                   cpu, res; } && PERF_SAMPLE_CPU
415          *      { u64                   period;   } && PERF_SAMPLE_PERIOD
416          *
417          *      { struct read_format    values;   } && PERF_SAMPLE_READ
418          *
419          *      { u64                   nr,
420          *        u64                   ips[nr];  } && PERF_SAMPLE_CALLCHAIN
421          *
422          *      #
423          *      # The RAW record below is opaque data wrt the ABI
424          *      #
425          *      # That is, the ABI doesn't make any promises wrt to
426          *      # the stability of its content, it may vary depending
427          *      # on event, hardware, kernel version and phase of
428          *      # the moon.
429          *      #
430          *      # In other words, PERF_SAMPLE_RAW contents are not an ABI.
431          *      #
432          *
433          *      { u32                   size;
434          *        char                  data[size];}&& PERF_SAMPLE_RAW
435          * };
436          */
437         PERF_RECORD_SAMPLE                      = 9,
438
439         PERF_RECORD_MAX,                        /* non-ABI */
440 };
441
442 enum perf_callchain_context {
443         PERF_CONTEXT_HV                 = (__u64)-32,
444         PERF_CONTEXT_KERNEL             = (__u64)-128,
445         PERF_CONTEXT_USER               = (__u64)-512,
446
447         PERF_CONTEXT_GUEST              = (__u64)-2048,
448         PERF_CONTEXT_GUEST_KERNEL       = (__u64)-2176,
449         PERF_CONTEXT_GUEST_USER         = (__u64)-2560,
450
451         PERF_CONTEXT_MAX                = (__u64)-4095,
452 };
453
454 #define PERF_FLAG_FD_NO_GROUP   (1U << 0)
455 #define PERF_FLAG_FD_OUTPUT     (1U << 1)
456
457 #ifdef __KERNEL__
458 /*
459  * Kernel-internal data types and definitions:
460  */
461
462 #ifdef CONFIG_PERF_EVENTS
463 # include <asm/perf_event.h>
464 #endif
465
466 struct perf_guest_info_callbacks {
467         int (*is_in_guest) (void);
468         int (*is_user_mode) (void);
469         unsigned long (*get_guest_ip) (void);
470 };
471
472 #ifdef CONFIG_HAVE_HW_BREAKPOINT
473 #include <asm/hw_breakpoint.h>
474 #endif
475
476 #include <linux/list.h>
477 #include <linux/mutex.h>
478 #include <linux/rculist.h>
479 #include <linux/rcupdate.h>
480 #include <linux/spinlock.h>
481 #include <linux/hrtimer.h>
482 #include <linux/fs.h>
483 #include <linux/pid_namespace.h>
484 #include <linux/workqueue.h>
485 #include <linux/ftrace.h>
486 #include <linux/cpu.h>
487 #include <asm/atomic.h>
488 #include <asm/local.h>
489
490 #define PERF_MAX_STACK_DEPTH            255
491
492 struct perf_callchain_entry {
493         __u64                           nr;
494         __u64                           ip[PERF_MAX_STACK_DEPTH];
495 };
496
497 struct perf_raw_record {
498         u32                             size;
499         void                            *data;
500 };
501
502 struct perf_branch_entry {
503         __u64                           from;
504         __u64                           to;
505         __u64                           flags;
506 };
507
508 struct perf_branch_stack {
509         __u64                           nr;
510         struct perf_branch_entry        entries[0];
511 };
512
513 struct task_struct;
514
515 /**
516  * struct hw_perf_event - performance event hardware details:
517  */
518 struct hw_perf_event {
519 #ifdef CONFIG_PERF_EVENTS
520         union {
521                 struct { /* hardware */
522                         u64             config;
523                         u64             last_tag;
524                         unsigned long   config_base;
525                         unsigned long   event_base;
526                         int             idx;
527                         int             last_cpu;
528                 };
529                 struct { /* software */
530                         s64             remaining;
531                         struct hrtimer  hrtimer;
532                 };
533 #ifdef CONFIG_HAVE_HW_BREAKPOINT
534                 /* breakpoint */
535                 struct arch_hw_breakpoint       info;
536 #endif
537         };
538         atomic64_t                      prev_count;
539         u64                             sample_period;
540         u64                             last_period;
541         atomic64_t                      period_left;
542         u64                             interrupts;
543
544         u64                             freq_time_stamp;
545         u64                             freq_count_stamp;
546 #endif
547 };
548
549 struct perf_event;
550
551 #define PERF_EVENT_TXN_STARTED 1
552
553 /**
554  * struct pmu - generic performance monitoring unit
555  */
556 struct pmu {
557         int (*enable)                   (struct perf_event *event);
558         void (*disable)                 (struct perf_event *event);
559         int (*start)                    (struct perf_event *event);
560         void (*stop)                    (struct perf_event *event);
561         void (*read)                    (struct perf_event *event);
562         void (*unthrottle)              (struct perf_event *event);
563
564         /*
565          * group events scheduling is treated as a transaction,
566          * add group events as a whole and perform one schedulability test.
567          * If test fails, roll back the whole group
568          */
569
570         void (*start_txn)       (const struct pmu *pmu);
571         void (*cancel_txn)      (const struct pmu *pmu);
572         int  (*commit_txn)      (const struct pmu *pmu);
573 };
574
575 /**
576  * enum perf_event_active_state - the states of a event
577  */
578 enum perf_event_active_state {
579         PERF_EVENT_STATE_ERROR          = -2,
580         PERF_EVENT_STATE_OFF            = -1,
581         PERF_EVENT_STATE_INACTIVE       =  0,
582         PERF_EVENT_STATE_ACTIVE         =  1,
583 };
584
585 struct file;
586
587 struct perf_mmap_data {
588         atomic_t                        refcount;
589         struct rcu_head                 rcu_head;
590 #ifdef CONFIG_PERF_USE_VMALLOC
591         struct work_struct              work;
592         int                             page_order;     /* allocation order  */
593 #endif
594         int                             nr_pages;       /* nr of data pages  */
595         int                             writable;       /* are we writable   */
596
597         atomic_t                        poll;           /* POLL_ for wakeups */
598
599         local_t                         head;           /* write position    */
600         local_t                         nest;           /* nested writers    */
601         local_t                         events;         /* event limit       */
602         local_t                         wakeup;         /* wakeup stamp      */
603         local_t                         lost;           /* nr records lost   */
604
605         long                            watermark;      /* wakeup watermark  */
606
607         struct perf_event_mmap_page     *user_page;
608         void                            *data_pages[0];
609 };
610
611 struct perf_pending_entry {
612         struct perf_pending_entry *next;
613         void (*func)(struct perf_pending_entry *);
614 };
615
616 struct perf_sample_data;
617
618 typedef void (*perf_overflow_handler_t)(struct perf_event *, int,
619                                         struct perf_sample_data *,
620                                         struct pt_regs *regs);
621
622 enum perf_group_flag {
623         PERF_GROUP_SOFTWARE = 0x1,
624 };
625
626 #define SWEVENT_HLIST_BITS      8
627 #define SWEVENT_HLIST_SIZE      (1 << SWEVENT_HLIST_BITS)
628
629 struct swevent_hlist {
630         struct hlist_head       heads[SWEVENT_HLIST_SIZE];
631         struct rcu_head         rcu_head;
632 };
633
634 #define PERF_ATTACH_CONTEXT     0x01
635 #define PERF_ATTACH_GROUP       0x02
636
637 /**
638  * struct perf_event - performance event kernel representation:
639  */
640 struct perf_event {
641 #ifdef CONFIG_PERF_EVENTS
642         struct list_head                group_entry;
643         struct list_head                event_entry;
644         struct list_head                sibling_list;
645         struct hlist_node               hlist_entry;
646         int                             nr_siblings;
647         int                             group_flags;
648         struct perf_event               *group_leader;
649         const struct pmu                *pmu;
650
651         enum perf_event_active_state    state;
652         unsigned int                    attach_state;
653         atomic64_t                      count;
654
655         /*
656          * These are the total time in nanoseconds that the event
657          * has been enabled (i.e. eligible to run, and the task has
658          * been scheduled in, if this is a per-task event)
659          * and running (scheduled onto the CPU), respectively.
660          *
661          * They are computed from tstamp_enabled, tstamp_running and
662          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
663          */
664         u64                             total_time_enabled;
665         u64                             total_time_running;
666
667         /*
668          * These are timestamps used for computing total_time_enabled
669          * and total_time_running when the event is in INACTIVE or
670          * ACTIVE state, measured in nanoseconds from an arbitrary point
671          * in time.
672          * tstamp_enabled: the notional time when the event was enabled
673          * tstamp_running: the notional time when the event was scheduled on
674          * tstamp_stopped: in INACTIVE state, the notional time when the
675          *      event was scheduled off.
676          */
677         u64                             tstamp_enabled;
678         u64                             tstamp_running;
679         u64                             tstamp_stopped;
680
681         struct perf_event_attr          attr;
682         struct hw_perf_event            hw;
683
684         struct perf_event_context       *ctx;
685         struct file                     *filp;
686
687         /*
688          * These accumulate total time (in nanoseconds) that children
689          * events have been enabled and running, respectively.
690          */
691         atomic64_t                      child_total_time_enabled;
692         atomic64_t                      child_total_time_running;
693
694         /*
695          * Protect attach/detach and child_list:
696          */
697         struct mutex                    child_mutex;
698         struct list_head                child_list;
699         struct perf_event               *parent;
700
701         int                             oncpu;
702         int                             cpu;
703
704         struct list_head                owner_entry;
705         struct task_struct              *owner;
706
707         /* mmap bits */
708         struct mutex                    mmap_mutex;
709         atomic_t                        mmap_count;
710         int                             mmap_locked;
711         struct user_struct              *mmap_user;
712         struct perf_mmap_data           *data;
713
714         /* poll related */
715         wait_queue_head_t               waitq;
716         struct fasync_struct            *fasync;
717
718         /* delayed work for NMIs and such */
719         int                             pending_wakeup;
720         int                             pending_kill;
721         int                             pending_disable;
722         struct perf_pending_entry       pending;
723
724         atomic_t                        event_limit;
725
726         void (*destroy)(struct perf_event *);
727         struct rcu_head                 rcu_head;
728
729         struct pid_namespace            *ns;
730         u64                             id;
731
732         perf_overflow_handler_t         overflow_handler;
733
734 #ifdef CONFIG_EVENT_TRACING
735         struct ftrace_event_call        *tp_event;
736         struct event_filter             *filter;
737 #endif
738
739 #endif /* CONFIG_PERF_EVENTS */
740 };
741
742 /**
743  * struct perf_event_context - event context structure
744  *
745  * Used as a container for task events and CPU events as well:
746  */
747 struct perf_event_context {
748         /*
749          * Protect the states of the events in the list,
750          * nr_active, and the list:
751          */
752         raw_spinlock_t                  lock;
753         /*
754          * Protect the list of events.  Locking either mutex or lock
755          * is sufficient to ensure the list doesn't change; to change
756          * the list you need to lock both the mutex and the spinlock.
757          */
758         struct mutex                    mutex;
759
760         struct list_head                pinned_groups;
761         struct list_head                flexible_groups;
762         struct list_head                event_list;
763         int                             nr_events;
764         int                             nr_active;
765         int                             is_active;
766         int                             nr_stat;
767         atomic_t                        refcount;
768         struct task_struct              *task;
769
770         /*
771          * Context clock, runs when context enabled.
772          */
773         u64                             time;
774         u64                             timestamp;
775
776         /*
777          * These fields let us detect when two contexts have both
778          * been cloned (inherited) from a common ancestor.
779          */
780         struct perf_event_context       *parent_ctx;
781         u64                             parent_gen;
782         u64                             generation;
783         int                             pin_count;
784         struct rcu_head                 rcu_head;
785 };
786
787 /**
788  * struct perf_event_cpu_context - per cpu event context structure
789  */
790 struct perf_cpu_context {
791         struct perf_event_context       ctx;
792         struct perf_event_context       *task_ctx;
793         int                             active_oncpu;
794         int                             max_pertask;
795         int                             exclusive;
796         struct swevent_hlist            *swevent_hlist;
797         struct mutex                    hlist_mutex;
798         int                             hlist_refcount;
799
800         /*
801          * Recursion avoidance:
802          *
803          * task, softirq, irq, nmi context
804          */
805         int                             recursion[4];
806 };
807
808 struct perf_output_handle {
809         struct perf_event               *event;
810         struct perf_mmap_data           *data;
811         unsigned long                   wakeup;
812         unsigned long                   size;
813         void                            *addr;
814         int                             page;
815         int                             nmi;
816         int                             sample;
817 };
818
819 #ifdef CONFIG_PERF_EVENTS
820
821 /*
822  * Set by architecture code:
823  */
824 extern int perf_max_events;
825
826 extern const struct pmu *hw_perf_event_init(struct perf_event *event);
827
828 extern void perf_event_task_sched_in(struct task_struct *task);
829 extern void perf_event_task_sched_out(struct task_struct *task, struct task_struct *next);
830 extern void perf_event_task_tick(struct task_struct *task);
831 extern int perf_event_init_task(struct task_struct *child);
832 extern void perf_event_exit_task(struct task_struct *child);
833 extern void perf_event_free_task(struct task_struct *task);
834 extern void set_perf_event_pending(void);
835 extern void perf_event_do_pending(void);
836 extern void perf_event_print_debug(void);
837 extern void __perf_disable(void);
838 extern bool __perf_enable(void);
839 extern void perf_disable(void);
840 extern void perf_enable(void);
841 extern int perf_event_task_disable(void);
842 extern int perf_event_task_enable(void);
843 extern void perf_event_update_userpage(struct perf_event *event);
844 extern int perf_event_release_kernel(struct perf_event *event);
845 extern struct perf_event *
846 perf_event_create_kernel_counter(struct perf_event_attr *attr,
847                                 int cpu,
848                                 pid_t pid,
849                                 perf_overflow_handler_t callback);
850 extern u64 perf_event_read_value(struct perf_event *event,
851                                  u64 *enabled, u64 *running);
852
853 struct perf_sample_data {
854         u64                             type;
855
856         u64                             ip;
857         struct {
858                 u32     pid;
859                 u32     tid;
860         }                               tid_entry;
861         u64                             time;
862         u64                             addr;
863         u64                             id;
864         u64                             stream_id;
865         struct {
866                 u32     cpu;
867                 u32     reserved;
868         }                               cpu_entry;
869         u64                             period;
870         struct perf_callchain_entry     *callchain;
871         struct perf_raw_record          *raw;
872 };
873
874 static inline
875 void perf_sample_data_init(struct perf_sample_data *data, u64 addr)
876 {
877         data->addr = addr;
878         data->raw  = NULL;
879 }
880
881 extern void perf_output_sample(struct perf_output_handle *handle,
882                                struct perf_event_header *header,
883                                struct perf_sample_data *data,
884                                struct perf_event *event);
885 extern void perf_prepare_sample(struct perf_event_header *header,
886                                 struct perf_sample_data *data,
887                                 struct perf_event *event,
888                                 struct pt_regs *regs);
889
890 extern int perf_event_overflow(struct perf_event *event, int nmi,
891                                  struct perf_sample_data *data,
892                                  struct pt_regs *regs);
893
894 /*
895  * Return 1 for a software event, 0 for a hardware event
896  */
897 static inline int is_software_event(struct perf_event *event)
898 {
899         switch (event->attr.type) {
900         case PERF_TYPE_SOFTWARE:
901         case PERF_TYPE_TRACEPOINT:
902         /* for now the breakpoint stuff also works as software event */
903         case PERF_TYPE_BREAKPOINT:
904                 return 1;
905         }
906         return 0;
907 }
908
909 extern atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
910
911 extern void __perf_sw_event(u32, u64, int, struct pt_regs *, u64);
912
913 extern void
914 perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip, int skip);
915
916 /*
917  * Take a snapshot of the regs. Skip ip and frame pointer to
918  * the nth caller. We only need a few of the regs:
919  * - ip for PERF_SAMPLE_IP
920  * - cs for user_mode() tests
921  * - bp for callchains
922  * - eflags, for future purposes, just in case
923  */
924 static inline void perf_fetch_caller_regs(struct pt_regs *regs, int skip)
925 {
926         unsigned long ip;
927
928         memset(regs, 0, sizeof(*regs));
929
930         switch (skip) {
931         case 1 :
932                 ip = CALLER_ADDR0;
933                 break;
934         case 2 :
935                 ip = CALLER_ADDR1;
936                 break;
937         case 3 :
938                 ip = CALLER_ADDR2;
939                 break;
940         case 4:
941                 ip = CALLER_ADDR3;
942                 break;
943         /* No need to support further for now */
944         default:
945                 ip = 0;
946         }
947
948         return perf_arch_fetch_caller_regs(regs, ip, skip);
949 }
950
951 static inline void
952 perf_sw_event(u32 event_id, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
953 {
954         if (atomic_read(&perf_swevent_enabled[event_id])) {
955                 struct pt_regs hot_regs;
956
957                 if (!regs) {
958                         perf_fetch_caller_regs(&hot_regs, 1);
959                         regs = &hot_regs;
960                 }
961                 __perf_sw_event(event_id, nr, nmi, regs, addr);
962         }
963 }
964
965 extern void __perf_event_mmap(struct vm_area_struct *vma);
966
967 static inline void perf_event_mmap(struct vm_area_struct *vma)
968 {
969         if (vma->vm_flags & VM_EXEC)
970                 __perf_event_mmap(vma);
971 }
972
973 extern struct perf_guest_info_callbacks *perf_guest_cbs;
974 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
975 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
976
977 extern void perf_event_comm(struct task_struct *tsk);
978 extern void perf_event_fork(struct task_struct *tsk);
979
980 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
981
982 extern int sysctl_perf_event_paranoid;
983 extern int sysctl_perf_event_mlock;
984 extern int sysctl_perf_event_sample_rate;
985
986 static inline bool perf_paranoid_tracepoint_raw(void)
987 {
988         return sysctl_perf_event_paranoid > -1;
989 }
990
991 static inline bool perf_paranoid_cpu(void)
992 {
993         return sysctl_perf_event_paranoid > 0;
994 }
995
996 static inline bool perf_paranoid_kernel(void)
997 {
998         return sysctl_perf_event_paranoid > 1;
999 }
1000
1001 extern void perf_event_init(void);
1002 extern void perf_tp_event(u64 addr, u64 count, void *record,
1003                           int entry_size, struct pt_regs *regs,
1004                           struct hlist_head *head);
1005 extern void perf_bp_event(struct perf_event *event, void *data);
1006
1007 #ifndef perf_misc_flags
1008 #define perf_misc_flags(regs)   (user_mode(regs) ? PERF_RECORD_MISC_USER : \
1009                                  PERF_RECORD_MISC_KERNEL)
1010 #define perf_instruction_pointer(regs)  instruction_pointer(regs)
1011 #endif
1012
1013 extern int perf_output_begin(struct perf_output_handle *handle,
1014                              struct perf_event *event, unsigned int size,
1015                              int nmi, int sample);
1016 extern void perf_output_end(struct perf_output_handle *handle);
1017 extern void perf_output_copy(struct perf_output_handle *handle,
1018                              const void *buf, unsigned int len);
1019 extern int perf_swevent_get_recursion_context(void);
1020 extern void perf_swevent_put_recursion_context(int rctx);
1021 extern void perf_event_enable(struct perf_event *event);
1022 extern void perf_event_disable(struct perf_event *event);
1023 #else
1024 static inline void
1025 perf_event_task_sched_in(struct task_struct *task)                      { }
1026 static inline void
1027 perf_event_task_sched_out(struct task_struct *task,
1028                             struct task_struct *next)                   { }
1029 static inline void
1030 perf_event_task_tick(struct task_struct *task)                          { }
1031 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
1032 static inline void perf_event_exit_task(struct task_struct *child)      { }
1033 static inline void perf_event_free_task(struct task_struct *task)       { }
1034 static inline void perf_event_do_pending(void)                          { }
1035 static inline void perf_event_print_debug(void)                         { }
1036 static inline void perf_disable(void)                                   { }
1037 static inline void perf_enable(void)                                    { }
1038 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
1039 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
1040
1041 static inline void
1042 perf_sw_event(u32 event_id, u64 nr, int nmi,
1043                      struct pt_regs *regs, u64 addr)                    { }
1044 static inline void
1045 perf_bp_event(struct perf_event *event, void *data)                     { }
1046
1047 static inline int perf_register_guest_info_callbacks
1048 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1049 static inline int perf_unregister_guest_info_callbacks
1050 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1051
1052 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
1053 static inline void perf_event_comm(struct task_struct *tsk)             { }
1054 static inline void perf_event_fork(struct task_struct *tsk)             { }
1055 static inline void perf_event_init(void)                                { }
1056 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
1057 static inline void perf_swevent_put_recursion_context(int rctx)         { }
1058 static inline void perf_event_enable(struct perf_event *event)          { }
1059 static inline void perf_event_disable(struct perf_event *event)         { }
1060 #endif
1061
1062 #define perf_output_put(handle, x) \
1063         perf_output_copy((handle), &(x), sizeof(x))
1064
1065 /*
1066  * This has to have a higher priority than migration_notifier in sched.c.
1067  */
1068 #define perf_cpu_notifier(fn)                                   \
1069 do {                                                            \
1070         static struct notifier_block fn##_nb __cpuinitdata =    \
1071                 { .notifier_call = fn, .priority = 20 };        \
1072         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,             \
1073                 (void *)(unsigned long)smp_processor_id());     \
1074         fn(&fn##_nb, (unsigned long)CPU_STARTING,               \
1075                 (void *)(unsigned long)smp_processor_id());     \
1076         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                 \
1077                 (void *)(unsigned long)smp_processor_id());     \
1078         register_cpu_notifier(&fn##_nb);                        \
1079 } while (0)
1080
1081 #endif /* __KERNEL__ */
1082 #endif /* _LINUX_PERF_EVENT_H */