perf: Per PMU disable
[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                                 mmap_data      :  1, /* non-exec mmap data    */
218
219                                 __reserved_1   : 46;
220
221         union {
222                 __u32           wakeup_events;    /* wakeup every n events */
223                 __u32           wakeup_watermark; /* bytes before wakeup   */
224         };
225
226         __u32                   bp_type;
227         __u64                   bp_addr;
228         __u64                   bp_len;
229 };
230
231 /*
232  * Ioctls that can be done on a perf event fd:
233  */
234 #define PERF_EVENT_IOC_ENABLE           _IO ('$', 0)
235 #define PERF_EVENT_IOC_DISABLE          _IO ('$', 1)
236 #define PERF_EVENT_IOC_REFRESH          _IO ('$', 2)
237 #define PERF_EVENT_IOC_RESET            _IO ('$', 3)
238 #define PERF_EVENT_IOC_PERIOD           _IOW('$', 4, __u64)
239 #define PERF_EVENT_IOC_SET_OUTPUT       _IO ('$', 5)
240 #define PERF_EVENT_IOC_SET_FILTER       _IOW('$', 6, char *)
241
242 enum perf_event_ioc_flags {
243         PERF_IOC_FLAG_GROUP             = 1U << 0,
244 };
245
246 /*
247  * Structure of the page that can be mapped via mmap
248  */
249 struct perf_event_mmap_page {
250         __u32   version;                /* version number of this structure */
251         __u32   compat_version;         /* lowest version this is compat with */
252
253         /*
254          * Bits needed to read the hw events in user-space.
255          *
256          *   u32 seq;
257          *   s64 count;
258          *
259          *   do {
260          *     seq = pc->lock;
261          *
262          *     barrier()
263          *     if (pc->index) {
264          *       count = pmc_read(pc->index - 1);
265          *       count += pc->offset;
266          *     } else
267          *       goto regular_read;
268          *
269          *     barrier();
270          *   } while (pc->lock != seq);
271          *
272          * NOTE: for obvious reason this only works on self-monitoring
273          *       processes.
274          */
275         __u32   lock;                   /* seqlock for synchronization */
276         __u32   index;                  /* hardware event identifier */
277         __s64   offset;                 /* add to hardware event value */
278         __u64   time_enabled;           /* time event active */
279         __u64   time_running;           /* time event on cpu */
280
281                 /*
282                  * Hole for extension of the self monitor capabilities
283                  */
284
285         __u64   __reserved[123];        /* align to 1k */
286
287         /*
288          * Control data for the mmap() data buffer.
289          *
290          * User-space reading the @data_head value should issue an rmb(), on
291          * SMP capable platforms, after reading this value -- see
292          * perf_event_wakeup().
293          *
294          * When the mapping is PROT_WRITE the @data_tail value should be
295          * written by userspace to reflect the last read data. In this case
296          * the kernel will not over-write unread data.
297          */
298         __u64   data_head;              /* head in the data section */
299         __u64   data_tail;              /* user-space written tail */
300 };
301
302 #define PERF_RECORD_MISC_CPUMODE_MASK           (7 << 0)
303 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN        (0 << 0)
304 #define PERF_RECORD_MISC_KERNEL                 (1 << 0)
305 #define PERF_RECORD_MISC_USER                   (2 << 0)
306 #define PERF_RECORD_MISC_HYPERVISOR             (3 << 0)
307 #define PERF_RECORD_MISC_GUEST_KERNEL           (4 << 0)
308 #define PERF_RECORD_MISC_GUEST_USER             (5 << 0)
309
310 /*
311  * Indicates that the content of PERF_SAMPLE_IP points to
312  * the actual instruction that triggered the event. See also
313  * perf_event_attr::precise_ip.
314  */
315 #define PERF_RECORD_MISC_EXACT_IP               (1 << 14)
316 /*
317  * Reserve the last bit to indicate some extended misc field
318  */
319 #define PERF_RECORD_MISC_EXT_RESERVED           (1 << 15)
320
321 struct perf_event_header {
322         __u32   type;
323         __u16   misc;
324         __u16   size;
325 };
326
327 enum perf_event_type {
328
329         /*
330          * The MMAP events record the PROT_EXEC mappings so that we can
331          * correlate userspace IPs to code. They have the following structure:
332          *
333          * struct {
334          *      struct perf_event_header        header;
335          *
336          *      u32                             pid, tid;
337          *      u64                             addr;
338          *      u64                             len;
339          *      u64                             pgoff;
340          *      char                            filename[];
341          * };
342          */
343         PERF_RECORD_MMAP                        = 1,
344
345         /*
346          * struct {
347          *      struct perf_event_header        header;
348          *      u64                             id;
349          *      u64                             lost;
350          * };
351          */
352         PERF_RECORD_LOST                        = 2,
353
354         /*
355          * struct {
356          *      struct perf_event_header        header;
357          *
358          *      u32                             pid, tid;
359          *      char                            comm[];
360          * };
361          */
362         PERF_RECORD_COMM                        = 3,
363
364         /*
365          * struct {
366          *      struct perf_event_header        header;
367          *      u32                             pid, ppid;
368          *      u32                             tid, ptid;
369          *      u64                             time;
370          * };
371          */
372         PERF_RECORD_EXIT                        = 4,
373
374         /*
375          * struct {
376          *      struct perf_event_header        header;
377          *      u64                             time;
378          *      u64                             id;
379          *      u64                             stream_id;
380          * };
381          */
382         PERF_RECORD_THROTTLE                    = 5,
383         PERF_RECORD_UNTHROTTLE                  = 6,
384
385         /*
386          * struct {
387          *      struct perf_event_header        header;
388          *      u32                             pid, ppid;
389          *      u32                             tid, ptid;
390          *      u64                             time;
391          * };
392          */
393         PERF_RECORD_FORK                        = 7,
394
395         /*
396          * struct {
397          *      struct perf_event_header        header;
398          *      u32                             pid, tid;
399          *
400          *      struct read_format              values;
401          * };
402          */
403         PERF_RECORD_READ                        = 8,
404
405         /*
406          * struct {
407          *      struct perf_event_header        header;
408          *
409          *      { u64                   ip;       } && PERF_SAMPLE_IP
410          *      { u32                   pid, tid; } && PERF_SAMPLE_TID
411          *      { u64                   time;     } && PERF_SAMPLE_TIME
412          *      { u64                   addr;     } && PERF_SAMPLE_ADDR
413          *      { u64                   id;       } && PERF_SAMPLE_ID
414          *      { u64                   stream_id;} && PERF_SAMPLE_STREAM_ID
415          *      { u32                   cpu, res; } && PERF_SAMPLE_CPU
416          *      { u64                   period;   } && PERF_SAMPLE_PERIOD
417          *
418          *      { struct read_format    values;   } && PERF_SAMPLE_READ
419          *
420          *      { u64                   nr,
421          *        u64                   ips[nr];  } && PERF_SAMPLE_CALLCHAIN
422          *
423          *      #
424          *      # The RAW record below is opaque data wrt the ABI
425          *      #
426          *      # That is, the ABI doesn't make any promises wrt to
427          *      # the stability of its content, it may vary depending
428          *      # on event, hardware, kernel version and phase of
429          *      # the moon.
430          *      #
431          *      # In other words, PERF_SAMPLE_RAW contents are not an ABI.
432          *      #
433          *
434          *      { u32                   size;
435          *        char                  data[size];}&& PERF_SAMPLE_RAW
436          * };
437          */
438         PERF_RECORD_SAMPLE                      = 9,
439
440         PERF_RECORD_MAX,                        /* non-ABI */
441 };
442
443 enum perf_callchain_context {
444         PERF_CONTEXT_HV                 = (__u64)-32,
445         PERF_CONTEXT_KERNEL             = (__u64)-128,
446         PERF_CONTEXT_USER               = (__u64)-512,
447
448         PERF_CONTEXT_GUEST              = (__u64)-2048,
449         PERF_CONTEXT_GUEST_KERNEL       = (__u64)-2176,
450         PERF_CONTEXT_GUEST_USER         = (__u64)-2560,
451
452         PERF_CONTEXT_MAX                = (__u64)-4095,
453 };
454
455 #define PERF_FLAG_FD_NO_GROUP   (1U << 0)
456 #define PERF_FLAG_FD_OUTPUT     (1U << 1)
457
458 #ifdef __KERNEL__
459 /*
460  * Kernel-internal data types and definitions:
461  */
462
463 #ifdef CONFIG_PERF_EVENTS
464 # include <asm/perf_event.h>
465 # include <asm/local64.h>
466 #endif
467
468 struct perf_guest_info_callbacks {
469         int (*is_in_guest) (void);
470         int (*is_user_mode) (void);
471         unsigned long (*get_guest_ip) (void);
472 };
473
474 #ifdef CONFIG_HAVE_HW_BREAKPOINT
475 #include <asm/hw_breakpoint.h>
476 #endif
477
478 #include <linux/list.h>
479 #include <linux/mutex.h>
480 #include <linux/rculist.h>
481 #include <linux/rcupdate.h>
482 #include <linux/spinlock.h>
483 #include <linux/hrtimer.h>
484 #include <linux/fs.h>
485 #include <linux/pid_namespace.h>
486 #include <linux/workqueue.h>
487 #include <linux/ftrace.h>
488 #include <linux/cpu.h>
489 #include <asm/atomic.h>
490 #include <asm/local.h>
491
492 #define PERF_MAX_STACK_DEPTH            255
493
494 struct perf_callchain_entry {
495         __u64                           nr;
496         __u64                           ip[PERF_MAX_STACK_DEPTH];
497 };
498
499 struct perf_raw_record {
500         u32                             size;
501         void                            *data;
502 };
503
504 struct perf_branch_entry {
505         __u64                           from;
506         __u64                           to;
507         __u64                           flags;
508 };
509
510 struct perf_branch_stack {
511         __u64                           nr;
512         struct perf_branch_entry        entries[0];
513 };
514
515 struct task_struct;
516
517 /**
518  * struct hw_perf_event - performance event hardware details:
519  */
520 struct hw_perf_event {
521 #ifdef CONFIG_PERF_EVENTS
522         union {
523                 struct { /* hardware */
524                         u64             config;
525                         u64             last_tag;
526                         unsigned long   config_base;
527                         unsigned long   event_base;
528                         int             idx;
529                         int             last_cpu;
530                 };
531                 struct { /* software */
532                         s64             remaining;
533                         struct hrtimer  hrtimer;
534                 };
535 #ifdef CONFIG_HAVE_HW_BREAKPOINT
536                 struct { /* breakpoint */
537                         struct arch_hw_breakpoint       info;
538                         struct list_head                bp_list;
539                 };
540 #endif
541         };
542         local64_t                       prev_count;
543         u64                             sample_period;
544         u64                             last_period;
545         local64_t                       period_left;
546         u64                             interrupts;
547
548         u64                             freq_time_stamp;
549         u64                             freq_count_stamp;
550 #endif
551 };
552
553 struct perf_event;
554
555 /*
556  * Common implementation detail of pmu::{start,commit,cancel}_txn
557  */
558 #define PERF_EVENT_TXN 0x1
559
560 /**
561  * struct pmu - generic performance monitoring unit
562  */
563 struct pmu {
564         struct list_head                entry;
565
566         int                             *pmu_disable_count;
567
568         void (*pmu_enable)              (struct pmu *pmu);
569         void (*pmu_disable)             (struct pmu *pmu);
570
571         /*
572          * Should return -ENOENT when the @event doesn't match this PMU.
573          */
574         int (*event_init)               (struct perf_event *event);
575
576         int  (*enable)                  (struct perf_event *event);
577         void (*disable)                 (struct perf_event *event);
578         int  (*start)                   (struct perf_event *event);
579         void (*stop)                    (struct perf_event *event);
580         void (*read)                    (struct perf_event *event);
581         void (*unthrottle)              (struct perf_event *event);
582
583         /*
584          * Group events scheduling is treated as a transaction, add
585          * group events as a whole and perform one schedulability test.
586          * If the test fails, roll back the whole group
587          */
588
589         /*
590          * Start the transaction, after this ->enable() doesn't need to
591          * do schedulability tests.
592          */
593         void (*start_txn)       (struct pmu *pmu);
594         /*
595          * If ->start_txn() disabled the ->enable() schedulability test
596          * then ->commit_txn() is required to perform one. On success
597          * the transaction is closed. On error the transaction is kept
598          * open until ->cancel_txn() is called.
599          */
600         int  (*commit_txn)      (struct pmu *pmu);
601         /*
602          * Will cancel the transaction, assumes ->disable() is called
603          * for each successfull ->enable() during the transaction.
604          */
605         void (*cancel_txn)      (struct pmu *pmu);
606 };
607
608 /**
609  * enum perf_event_active_state - the states of a event
610  */
611 enum perf_event_active_state {
612         PERF_EVENT_STATE_ERROR          = -2,
613         PERF_EVENT_STATE_OFF            = -1,
614         PERF_EVENT_STATE_INACTIVE       =  0,
615         PERF_EVENT_STATE_ACTIVE         =  1,
616 };
617
618 struct file;
619
620 #define PERF_BUFFER_WRITABLE            0x01
621
622 struct perf_buffer {
623         atomic_t                        refcount;
624         struct rcu_head                 rcu_head;
625 #ifdef CONFIG_PERF_USE_VMALLOC
626         struct work_struct              work;
627         int                             page_order;     /* allocation order  */
628 #endif
629         int                             nr_pages;       /* nr of data pages  */
630         int                             writable;       /* are we writable   */
631
632         atomic_t                        poll;           /* POLL_ for wakeups */
633
634         local_t                         head;           /* write position    */
635         local_t                         nest;           /* nested writers    */
636         local_t                         events;         /* event limit       */
637         local_t                         wakeup;         /* wakeup stamp      */
638         local_t                         lost;           /* nr records lost   */
639
640         long                            watermark;      /* wakeup watermark  */
641
642         struct perf_event_mmap_page     *user_page;
643         void                            *data_pages[0];
644 };
645
646 struct perf_pending_entry {
647         struct perf_pending_entry *next;
648         void (*func)(struct perf_pending_entry *);
649 };
650
651 struct perf_sample_data;
652
653 typedef void (*perf_overflow_handler_t)(struct perf_event *, int,
654                                         struct perf_sample_data *,
655                                         struct pt_regs *regs);
656
657 enum perf_group_flag {
658         PERF_GROUP_SOFTWARE = 0x1,
659 };
660
661 #define SWEVENT_HLIST_BITS      8
662 #define SWEVENT_HLIST_SIZE      (1 << SWEVENT_HLIST_BITS)
663
664 struct swevent_hlist {
665         struct hlist_head       heads[SWEVENT_HLIST_SIZE];
666         struct rcu_head         rcu_head;
667 };
668
669 #define PERF_ATTACH_CONTEXT     0x01
670 #define PERF_ATTACH_GROUP       0x02
671
672 /**
673  * struct perf_event - performance event kernel representation:
674  */
675 struct perf_event {
676 #ifdef CONFIG_PERF_EVENTS
677         struct list_head                group_entry;
678         struct list_head                event_entry;
679         struct list_head                sibling_list;
680         struct hlist_node               hlist_entry;
681         int                             nr_siblings;
682         int                             group_flags;
683         struct perf_event               *group_leader;
684         struct pmu              *pmu;
685
686         enum perf_event_active_state    state;
687         unsigned int                    attach_state;
688         local64_t                       count;
689         atomic64_t                      child_count;
690
691         /*
692          * These are the total time in nanoseconds that the event
693          * has been enabled (i.e. eligible to run, and the task has
694          * been scheduled in, if this is a per-task event)
695          * and running (scheduled onto the CPU), respectively.
696          *
697          * They are computed from tstamp_enabled, tstamp_running and
698          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
699          */
700         u64                             total_time_enabled;
701         u64                             total_time_running;
702
703         /*
704          * These are timestamps used for computing total_time_enabled
705          * and total_time_running when the event is in INACTIVE or
706          * ACTIVE state, measured in nanoseconds from an arbitrary point
707          * in time.
708          * tstamp_enabled: the notional time when the event was enabled
709          * tstamp_running: the notional time when the event was scheduled on
710          * tstamp_stopped: in INACTIVE state, the notional time when the
711          *      event was scheduled off.
712          */
713         u64                             tstamp_enabled;
714         u64                             tstamp_running;
715         u64                             tstamp_stopped;
716
717         struct perf_event_attr          attr;
718         struct hw_perf_event            hw;
719
720         struct perf_event_context       *ctx;
721         struct file                     *filp;
722
723         /*
724          * These accumulate total time (in nanoseconds) that children
725          * events have been enabled and running, respectively.
726          */
727         atomic64_t                      child_total_time_enabled;
728         atomic64_t                      child_total_time_running;
729
730         /*
731          * Protect attach/detach and child_list:
732          */
733         struct mutex                    child_mutex;
734         struct list_head                child_list;
735         struct perf_event               *parent;
736
737         int                             oncpu;
738         int                             cpu;
739
740         struct list_head                owner_entry;
741         struct task_struct              *owner;
742
743         /* mmap bits */
744         struct mutex                    mmap_mutex;
745         atomic_t                        mmap_count;
746         int                             mmap_locked;
747         struct user_struct              *mmap_user;
748         struct perf_buffer              *buffer;
749
750         /* poll related */
751         wait_queue_head_t               waitq;
752         struct fasync_struct            *fasync;
753
754         /* delayed work for NMIs and such */
755         int                             pending_wakeup;
756         int                             pending_kill;
757         int                             pending_disable;
758         struct perf_pending_entry       pending;
759
760         atomic_t                        event_limit;
761
762         void (*destroy)(struct perf_event *);
763         struct rcu_head                 rcu_head;
764
765         struct pid_namespace            *ns;
766         u64                             id;
767
768         perf_overflow_handler_t         overflow_handler;
769
770 #ifdef CONFIG_EVENT_TRACING
771         struct ftrace_event_call        *tp_event;
772         struct event_filter             *filter;
773 #endif
774
775 #endif /* CONFIG_PERF_EVENTS */
776 };
777
778 /**
779  * struct perf_event_context - event context structure
780  *
781  * Used as a container for task events and CPU events as well:
782  */
783 struct perf_event_context {
784         /*
785          * Protect the states of the events in the list,
786          * nr_active, and the list:
787          */
788         raw_spinlock_t                  lock;
789         /*
790          * Protect the list of events.  Locking either mutex or lock
791          * is sufficient to ensure the list doesn't change; to change
792          * the list you need to lock both the mutex and the spinlock.
793          */
794         struct mutex                    mutex;
795
796         struct list_head                pinned_groups;
797         struct list_head                flexible_groups;
798         struct list_head                event_list;
799         int                             nr_events;
800         int                             nr_active;
801         int                             is_active;
802         int                             nr_stat;
803         atomic_t                        refcount;
804         struct task_struct              *task;
805
806         /*
807          * Context clock, runs when context enabled.
808          */
809         u64                             time;
810         u64                             timestamp;
811
812         /*
813          * These fields let us detect when two contexts have both
814          * been cloned (inherited) from a common ancestor.
815          */
816         struct perf_event_context       *parent_ctx;
817         u64                             parent_gen;
818         u64                             generation;
819         int                             pin_count;
820         struct rcu_head                 rcu_head;
821 };
822
823 /*
824  * Number of contexts where an event can trigger:
825  *      task, softirq, hardirq, nmi.
826  */
827 #define PERF_NR_CONTEXTS        4
828
829 /**
830  * struct perf_event_cpu_context - per cpu event context structure
831  */
832 struct perf_cpu_context {
833         struct perf_event_context       ctx;
834         struct perf_event_context       *task_ctx;
835         int                             active_oncpu;
836         int                             max_pertask;
837         int                             exclusive;
838         struct swevent_hlist            *swevent_hlist;
839         struct mutex                    hlist_mutex;
840         int                             hlist_refcount;
841
842         /* Recursion avoidance in each contexts */
843         int                             recursion[PERF_NR_CONTEXTS];
844 };
845
846 struct perf_output_handle {
847         struct perf_event               *event;
848         struct perf_buffer              *buffer;
849         unsigned long                   wakeup;
850         unsigned long                   size;
851         void                            *addr;
852         int                             page;
853         int                             nmi;
854         int                             sample;
855 };
856
857 #ifdef CONFIG_PERF_EVENTS
858
859 /*
860  * Set by architecture code:
861  */
862 extern int perf_max_events;
863
864 extern int perf_pmu_register(struct pmu *pmu);
865 extern void perf_pmu_unregister(struct pmu *pmu);
866
867 extern void perf_event_task_sched_in(struct task_struct *task);
868 extern void perf_event_task_sched_out(struct task_struct *task, struct task_struct *next);
869 extern void perf_event_task_tick(struct task_struct *task);
870 extern int perf_event_init_task(struct task_struct *child);
871 extern void perf_event_exit_task(struct task_struct *child);
872 extern void perf_event_free_task(struct task_struct *task);
873 extern void set_perf_event_pending(void);
874 extern void perf_event_do_pending(void);
875 extern void perf_event_print_debug(void);
876 extern void perf_pmu_disable(struct pmu *pmu);
877 extern void perf_pmu_enable(struct pmu *pmu);
878 extern int perf_event_task_disable(void);
879 extern int perf_event_task_enable(void);
880 extern void perf_event_update_userpage(struct perf_event *event);
881 extern int perf_event_release_kernel(struct perf_event *event);
882 extern struct perf_event *
883 perf_event_create_kernel_counter(struct perf_event_attr *attr,
884                                 int cpu,
885                                 pid_t pid,
886                                 perf_overflow_handler_t callback);
887 extern u64 perf_event_read_value(struct perf_event *event,
888                                  u64 *enabled, u64 *running);
889
890 struct perf_sample_data {
891         u64                             type;
892
893         u64                             ip;
894         struct {
895                 u32     pid;
896                 u32     tid;
897         }                               tid_entry;
898         u64                             time;
899         u64                             addr;
900         u64                             id;
901         u64                             stream_id;
902         struct {
903                 u32     cpu;
904                 u32     reserved;
905         }                               cpu_entry;
906         u64                             period;
907         struct perf_callchain_entry     *callchain;
908         struct perf_raw_record          *raw;
909 };
910
911 static inline
912 void perf_sample_data_init(struct perf_sample_data *data, u64 addr)
913 {
914         data->addr = addr;
915         data->raw  = NULL;
916 }
917
918 extern void perf_output_sample(struct perf_output_handle *handle,
919                                struct perf_event_header *header,
920                                struct perf_sample_data *data,
921                                struct perf_event *event);
922 extern void perf_prepare_sample(struct perf_event_header *header,
923                                 struct perf_sample_data *data,
924                                 struct perf_event *event,
925                                 struct pt_regs *regs);
926
927 extern int perf_event_overflow(struct perf_event *event, int nmi,
928                                  struct perf_sample_data *data,
929                                  struct pt_regs *regs);
930
931 /*
932  * Return 1 for a software event, 0 for a hardware event
933  */
934 static inline int is_software_event(struct perf_event *event)
935 {
936         switch (event->attr.type) {
937         case PERF_TYPE_SOFTWARE:
938         case PERF_TYPE_TRACEPOINT:
939         /* for now the breakpoint stuff also works as software event */
940         case PERF_TYPE_BREAKPOINT:
941                 return 1;
942         }
943         return 0;
944 }
945
946 extern atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
947
948 extern void __perf_sw_event(u32, u64, int, struct pt_regs *, u64);
949
950 #ifndef perf_arch_fetch_caller_regs
951 static inline void
952 perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
953 #endif
954
955 /*
956  * Take a snapshot of the regs. Skip ip and frame pointer to
957  * the nth caller. We only need a few of the regs:
958  * - ip for PERF_SAMPLE_IP
959  * - cs for user_mode() tests
960  * - bp for callchains
961  * - eflags, for future purposes, just in case
962  */
963 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
964 {
965         memset(regs, 0, sizeof(*regs));
966
967         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
968 }
969
970 static inline void
971 perf_sw_event(u32 event_id, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
972 {
973         if (atomic_read(&perf_swevent_enabled[event_id])) {
974                 struct pt_regs hot_regs;
975
976                 if (!regs) {
977                         perf_fetch_caller_regs(&hot_regs);
978                         regs = &hot_regs;
979                 }
980                 __perf_sw_event(event_id, nr, nmi, regs, addr);
981         }
982 }
983
984 extern void perf_event_mmap(struct vm_area_struct *vma);
985 extern struct perf_guest_info_callbacks *perf_guest_cbs;
986 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
987 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
988
989 extern void perf_event_comm(struct task_struct *tsk);
990 extern void perf_event_fork(struct task_struct *tsk);
991
992 /* Callchains */
993 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
994
995 extern void perf_callchain_user(struct perf_callchain_entry *entry,
996                                 struct pt_regs *regs);
997 extern void perf_callchain_kernel(struct perf_callchain_entry *entry,
998                                   struct pt_regs *regs);
999
1000
1001 static inline void
1002 perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
1003 {
1004         if (entry->nr < PERF_MAX_STACK_DEPTH)
1005                 entry->ip[entry->nr++] = ip;
1006 }
1007
1008 extern int sysctl_perf_event_paranoid;
1009 extern int sysctl_perf_event_mlock;
1010 extern int sysctl_perf_event_sample_rate;
1011
1012 static inline bool perf_paranoid_tracepoint_raw(void)
1013 {
1014         return sysctl_perf_event_paranoid > -1;
1015 }
1016
1017 static inline bool perf_paranoid_cpu(void)
1018 {
1019         return sysctl_perf_event_paranoid > 0;
1020 }
1021
1022 static inline bool perf_paranoid_kernel(void)
1023 {
1024         return sysctl_perf_event_paranoid > 1;
1025 }
1026
1027 extern void perf_event_init(void);
1028 extern void perf_tp_event(u64 addr, u64 count, void *record,
1029                           int entry_size, struct pt_regs *regs,
1030                           struct hlist_head *head, int rctx);
1031 extern void perf_bp_event(struct perf_event *event, void *data);
1032
1033 #ifndef perf_misc_flags
1034 #define perf_misc_flags(regs)   (user_mode(regs) ? PERF_RECORD_MISC_USER : \
1035                                  PERF_RECORD_MISC_KERNEL)
1036 #define perf_instruction_pointer(regs)  instruction_pointer(regs)
1037 #endif
1038
1039 extern int perf_output_begin(struct perf_output_handle *handle,
1040                              struct perf_event *event, unsigned int size,
1041                              int nmi, int sample);
1042 extern void perf_output_end(struct perf_output_handle *handle);
1043 extern void perf_output_copy(struct perf_output_handle *handle,
1044                              const void *buf, unsigned int len);
1045 extern int perf_swevent_get_recursion_context(void);
1046 extern void perf_swevent_put_recursion_context(int rctx);
1047 extern void perf_event_enable(struct perf_event *event);
1048 extern void perf_event_disable(struct perf_event *event);
1049 #else
1050 static inline void
1051 perf_event_task_sched_in(struct task_struct *task)                      { }
1052 static inline void
1053 perf_event_task_sched_out(struct task_struct *task,
1054                             struct task_struct *next)                   { }
1055 static inline void
1056 perf_event_task_tick(struct task_struct *task)                          { }
1057 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
1058 static inline void perf_event_exit_task(struct task_struct *child)      { }
1059 static inline void perf_event_free_task(struct task_struct *task)       { }
1060 static inline void perf_event_do_pending(void)                          { }
1061 static inline void perf_event_print_debug(void)                         { }
1062 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
1063 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
1064
1065 static inline void
1066 perf_sw_event(u32 event_id, u64 nr, int nmi,
1067                      struct pt_regs *regs, u64 addr)                    { }
1068 static inline void
1069 perf_bp_event(struct perf_event *event, void *data)                     { }
1070
1071 static inline int perf_register_guest_info_callbacks
1072 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1073 static inline int perf_unregister_guest_info_callbacks
1074 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1075
1076 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
1077 static inline void perf_event_comm(struct task_struct *tsk)             { }
1078 static inline void perf_event_fork(struct task_struct *tsk)             { }
1079 static inline void perf_event_init(void)                                { }
1080 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
1081 static inline void perf_swevent_put_recursion_context(int rctx)         { }
1082 static inline void perf_event_enable(struct perf_event *event)          { }
1083 static inline void perf_event_disable(struct perf_event *event)         { }
1084 #endif
1085
1086 #define perf_output_put(handle, x) \
1087         perf_output_copy((handle), &(x), sizeof(x))
1088
1089 /*
1090  * This has to have a higher priority than migration_notifier in sched.c.
1091  */
1092 #define perf_cpu_notifier(fn)                                   \
1093 do {                                                            \
1094         static struct notifier_block fn##_nb __cpuinitdata =    \
1095                 { .notifier_call = fn, .priority = CPU_PRI_PERF }; \
1096         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,             \
1097                 (void *)(unsigned long)smp_processor_id());     \
1098         fn(&fn##_nb, (unsigned long)CPU_STARTING,               \
1099                 (void *)(unsigned long)smp_processor_id());     \
1100         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                 \
1101                 (void *)(unsigned long)smp_processor_id());     \
1102         register_cpu_notifier(&fn##_nb);                        \
1103 } while (0)
1104
1105 #endif /* __KERNEL__ */
1106 #endif /* _LINUX_PERF_EVENT_H */