perf: Rework the PMU methods
[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                         struct hrtimer  hrtimer;
533                 };
534 #ifdef CONFIG_HAVE_HW_BREAKPOINT
535                 struct { /* breakpoint */
536                         struct arch_hw_breakpoint       info;
537                         struct list_head                bp_list;
538                 };
539 #endif
540         };
541         int                             state;
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 /*
554  * hw_perf_event::state flags
555  */
556 #define PERF_HES_STOPPED        0x01 /* the counter is stopped */
557 #define PERF_HES_UPTODATE       0x02 /* event->count up-to-date */
558 #define PERF_HES_ARCH           0x04
559
560 struct perf_event;
561
562 /*
563  * Common implementation detail of pmu::{start,commit,cancel}_txn
564  */
565 #define PERF_EVENT_TXN 0x1
566
567 /**
568  * struct pmu - generic performance monitoring unit
569  */
570 struct pmu {
571         struct list_head                entry;
572
573         int                             *pmu_disable_count;
574
575         /*
576          * Fully disable/enable this PMU, can be used to protect from the PMI
577          * as well as for lazy/batch writing of the MSRs.
578          */
579         void (*pmu_enable)              (struct pmu *pmu); /* optional */
580         void (*pmu_disable)             (struct pmu *pmu); /* optional */
581
582         /*
583          * Try and initialize the event for this PMU.
584          * Should return -ENOENT when the @event doesn't match this PMU.
585          */
586         int (*event_init)               (struct perf_event *event);
587
588 #define PERF_EF_START   0x01            /* start the counter when adding    */
589 #define PERF_EF_RELOAD  0x02            /* reload the counter when starting */
590 #define PERF_EF_UPDATE  0x04            /* update the counter when stopping */
591
592         /*
593          * Adds/Removes a counter to/from the PMU, can be done inside
594          * a transaction, see the ->*_txn() methods.
595          */
596         int  (*add)                     (struct perf_event *event, int flags);
597         void (*del)                     (struct perf_event *event, int flags);
598
599         /*
600          * Starts/Stops a counter present on the PMU. The PMI handler
601          * should stop the counter when perf_event_overflow() returns
602          * !0. ->start() will be used to continue.
603          */
604         void (*start)                   (struct perf_event *event, int flags);
605         void (*stop)                    (struct perf_event *event, int flags);
606
607         /*
608          * Updates the counter value of the event.
609          */
610         void (*read)                    (struct perf_event *event);
611
612         /*
613          * Group events scheduling is treated as a transaction, add
614          * group events as a whole and perform one schedulability test.
615          * If the test fails, roll back the whole group
616          *
617          * Start the transaction, after this ->add() doesn't need to
618          * do schedulability tests.
619          */
620         void (*start_txn)       (struct pmu *pmu); /* optional */
621         /*
622          * If ->start_txn() disabled the ->add() schedulability test
623          * then ->commit_txn() is required to perform one. On success
624          * the transaction is closed. On error the transaction is kept
625          * open until ->cancel_txn() is called.
626          */
627         int  (*commit_txn)      (struct pmu *pmu); /* optional */
628         /*
629          * Will cancel the transaction, assumes ->del() is called
630          * for each successfull ->add() during the transaction.
631          */
632         void (*cancel_txn)      (struct pmu *pmu); /* optional */
633 };
634
635 /**
636  * enum perf_event_active_state - the states of a event
637  */
638 enum perf_event_active_state {
639         PERF_EVENT_STATE_ERROR          = -2,
640         PERF_EVENT_STATE_OFF            = -1,
641         PERF_EVENT_STATE_INACTIVE       =  0,
642         PERF_EVENT_STATE_ACTIVE         =  1,
643 };
644
645 struct file;
646
647 #define PERF_BUFFER_WRITABLE            0x01
648
649 struct perf_buffer {
650         atomic_t                        refcount;
651         struct rcu_head                 rcu_head;
652 #ifdef CONFIG_PERF_USE_VMALLOC
653         struct work_struct              work;
654         int                             page_order;     /* allocation order  */
655 #endif
656         int                             nr_pages;       /* nr of data pages  */
657         int                             writable;       /* are we writable   */
658
659         atomic_t                        poll;           /* POLL_ for wakeups */
660
661         local_t                         head;           /* write position    */
662         local_t                         nest;           /* nested writers    */
663         local_t                         events;         /* event limit       */
664         local_t                         wakeup;         /* wakeup stamp      */
665         local_t                         lost;           /* nr records lost   */
666
667         long                            watermark;      /* wakeup watermark  */
668
669         struct perf_event_mmap_page     *user_page;
670         void                            *data_pages[0];
671 };
672
673 struct perf_pending_entry {
674         struct perf_pending_entry *next;
675         void (*func)(struct perf_pending_entry *);
676 };
677
678 struct perf_sample_data;
679
680 typedef void (*perf_overflow_handler_t)(struct perf_event *, int,
681                                         struct perf_sample_data *,
682                                         struct pt_regs *regs);
683
684 enum perf_group_flag {
685         PERF_GROUP_SOFTWARE = 0x1,
686 };
687
688 #define SWEVENT_HLIST_BITS      8
689 #define SWEVENT_HLIST_SIZE      (1 << SWEVENT_HLIST_BITS)
690
691 struct swevent_hlist {
692         struct hlist_head       heads[SWEVENT_HLIST_SIZE];
693         struct rcu_head         rcu_head;
694 };
695
696 #define PERF_ATTACH_CONTEXT     0x01
697 #define PERF_ATTACH_GROUP       0x02
698
699 /**
700  * struct perf_event - performance event kernel representation:
701  */
702 struct perf_event {
703 #ifdef CONFIG_PERF_EVENTS
704         struct list_head                group_entry;
705         struct list_head                event_entry;
706         struct list_head                sibling_list;
707         struct hlist_node               hlist_entry;
708         int                             nr_siblings;
709         int                             group_flags;
710         struct perf_event               *group_leader;
711         struct pmu                      *pmu;
712
713         enum perf_event_active_state    state;
714         unsigned int                    attach_state;
715         local64_t                       count;
716         atomic64_t                      child_count;
717
718         /*
719          * These are the total time in nanoseconds that the event
720          * has been enabled (i.e. eligible to run, and the task has
721          * been scheduled in, if this is a per-task event)
722          * and running (scheduled onto the CPU), respectively.
723          *
724          * They are computed from tstamp_enabled, tstamp_running and
725          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
726          */
727         u64                             total_time_enabled;
728         u64                             total_time_running;
729
730         /*
731          * These are timestamps used for computing total_time_enabled
732          * and total_time_running when the event is in INACTIVE or
733          * ACTIVE state, measured in nanoseconds from an arbitrary point
734          * in time.
735          * tstamp_enabled: the notional time when the event was enabled
736          * tstamp_running: the notional time when the event was scheduled on
737          * tstamp_stopped: in INACTIVE state, the notional time when the
738          *      event was scheduled off.
739          */
740         u64                             tstamp_enabled;
741         u64                             tstamp_running;
742         u64                             tstamp_stopped;
743
744         struct perf_event_attr          attr;
745         struct hw_perf_event            hw;
746
747         struct perf_event_context       *ctx;
748         struct file                     *filp;
749
750         /*
751          * These accumulate total time (in nanoseconds) that children
752          * events have been enabled and running, respectively.
753          */
754         atomic64_t                      child_total_time_enabled;
755         atomic64_t                      child_total_time_running;
756
757         /*
758          * Protect attach/detach and child_list:
759          */
760         struct mutex                    child_mutex;
761         struct list_head                child_list;
762         struct perf_event               *parent;
763
764         int                             oncpu;
765         int                             cpu;
766
767         struct list_head                owner_entry;
768         struct task_struct              *owner;
769
770         /* mmap bits */
771         struct mutex                    mmap_mutex;
772         atomic_t                        mmap_count;
773         int                             mmap_locked;
774         struct user_struct              *mmap_user;
775         struct perf_buffer              *buffer;
776
777         /* poll related */
778         wait_queue_head_t               waitq;
779         struct fasync_struct            *fasync;
780
781         /* delayed work for NMIs and such */
782         int                             pending_wakeup;
783         int                             pending_kill;
784         int                             pending_disable;
785         struct perf_pending_entry       pending;
786
787         atomic_t                        event_limit;
788
789         void (*destroy)(struct perf_event *);
790         struct rcu_head                 rcu_head;
791
792         struct pid_namespace            *ns;
793         u64                             id;
794
795         perf_overflow_handler_t         overflow_handler;
796
797 #ifdef CONFIG_EVENT_TRACING
798         struct ftrace_event_call        *tp_event;
799         struct event_filter             *filter;
800 #endif
801
802 #endif /* CONFIG_PERF_EVENTS */
803 };
804
805 /**
806  * struct perf_event_context - event context structure
807  *
808  * Used as a container for task events and CPU events as well:
809  */
810 struct perf_event_context {
811         /*
812          * Protect the states of the events in the list,
813          * nr_active, and the list:
814          */
815         raw_spinlock_t                  lock;
816         /*
817          * Protect the list of events.  Locking either mutex or lock
818          * is sufficient to ensure the list doesn't change; to change
819          * the list you need to lock both the mutex and the spinlock.
820          */
821         struct mutex                    mutex;
822
823         struct list_head                pinned_groups;
824         struct list_head                flexible_groups;
825         struct list_head                event_list;
826         int                             nr_events;
827         int                             nr_active;
828         int                             is_active;
829         int                             nr_stat;
830         atomic_t                        refcount;
831         struct task_struct              *task;
832
833         /*
834          * Context clock, runs when context enabled.
835          */
836         u64                             time;
837         u64                             timestamp;
838
839         /*
840          * These fields let us detect when two contexts have both
841          * been cloned (inherited) from a common ancestor.
842          */
843         struct perf_event_context       *parent_ctx;
844         u64                             parent_gen;
845         u64                             generation;
846         int                             pin_count;
847         struct rcu_head                 rcu_head;
848 };
849
850 /*
851  * Number of contexts where an event can trigger:
852  *      task, softirq, hardirq, nmi.
853  */
854 #define PERF_NR_CONTEXTS        4
855
856 /**
857  * struct perf_event_cpu_context - per cpu event context structure
858  */
859 struct perf_cpu_context {
860         struct perf_event_context       ctx;
861         struct perf_event_context       *task_ctx;
862         int                             active_oncpu;
863         int                             max_pertask;
864         int                             exclusive;
865         struct swevent_hlist            *swevent_hlist;
866         struct mutex                    hlist_mutex;
867         int                             hlist_refcount;
868
869         /* Recursion avoidance in each contexts */
870         int                             recursion[PERF_NR_CONTEXTS];
871 };
872
873 struct perf_output_handle {
874         struct perf_event               *event;
875         struct perf_buffer              *buffer;
876         unsigned long                   wakeup;
877         unsigned long                   size;
878         void                            *addr;
879         int                             page;
880         int                             nmi;
881         int                             sample;
882 };
883
884 #ifdef CONFIG_PERF_EVENTS
885
886 /*
887  * Set by architecture code:
888  */
889 extern int perf_max_events;
890
891 extern int perf_pmu_register(struct pmu *pmu);
892 extern void perf_pmu_unregister(struct pmu *pmu);
893
894 extern void perf_event_task_sched_in(struct task_struct *task);
895 extern void perf_event_task_sched_out(struct task_struct *task, struct task_struct *next);
896 extern void perf_event_task_tick(struct task_struct *task);
897 extern int perf_event_init_task(struct task_struct *child);
898 extern void perf_event_exit_task(struct task_struct *child);
899 extern void perf_event_free_task(struct task_struct *task);
900 extern void set_perf_event_pending(void);
901 extern void perf_event_do_pending(void);
902 extern void perf_event_print_debug(void);
903 extern void perf_pmu_disable(struct pmu *pmu);
904 extern void perf_pmu_enable(struct pmu *pmu);
905 extern int perf_event_task_disable(void);
906 extern int perf_event_task_enable(void);
907 extern void perf_event_update_userpage(struct perf_event *event);
908 extern int perf_event_release_kernel(struct perf_event *event);
909 extern struct perf_event *
910 perf_event_create_kernel_counter(struct perf_event_attr *attr,
911                                 int cpu,
912                                 pid_t pid,
913                                 perf_overflow_handler_t callback);
914 extern u64 perf_event_read_value(struct perf_event *event,
915                                  u64 *enabled, u64 *running);
916
917 struct perf_sample_data {
918         u64                             type;
919
920         u64                             ip;
921         struct {
922                 u32     pid;
923                 u32     tid;
924         }                               tid_entry;
925         u64                             time;
926         u64                             addr;
927         u64                             id;
928         u64                             stream_id;
929         struct {
930                 u32     cpu;
931                 u32     reserved;
932         }                               cpu_entry;
933         u64                             period;
934         struct perf_callchain_entry     *callchain;
935         struct perf_raw_record          *raw;
936 };
937
938 static inline
939 void perf_sample_data_init(struct perf_sample_data *data, u64 addr)
940 {
941         data->addr = addr;
942         data->raw  = NULL;
943 }
944
945 extern void perf_output_sample(struct perf_output_handle *handle,
946                                struct perf_event_header *header,
947                                struct perf_sample_data *data,
948                                struct perf_event *event);
949 extern void perf_prepare_sample(struct perf_event_header *header,
950                                 struct perf_sample_data *data,
951                                 struct perf_event *event,
952                                 struct pt_regs *regs);
953
954 extern int perf_event_overflow(struct perf_event *event, int nmi,
955                                  struct perf_sample_data *data,
956                                  struct pt_regs *regs);
957
958 /*
959  * Return 1 for a software event, 0 for a hardware event
960  */
961 static inline int is_software_event(struct perf_event *event)
962 {
963         switch (event->attr.type) {
964         case PERF_TYPE_SOFTWARE:
965         case PERF_TYPE_TRACEPOINT:
966         /* for now the breakpoint stuff also works as software event */
967         case PERF_TYPE_BREAKPOINT:
968                 return 1;
969         }
970         return 0;
971 }
972
973 extern atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
974
975 extern void __perf_sw_event(u32, u64, int, struct pt_regs *, u64);
976
977 #ifndef perf_arch_fetch_caller_regs
978 static inline void
979 perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
980 #endif
981
982 /*
983  * Take a snapshot of the regs. Skip ip and frame pointer to
984  * the nth caller. We only need a few of the regs:
985  * - ip for PERF_SAMPLE_IP
986  * - cs for user_mode() tests
987  * - bp for callchains
988  * - eflags, for future purposes, just in case
989  */
990 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
991 {
992         memset(regs, 0, sizeof(*regs));
993
994         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
995 }
996
997 static inline void
998 perf_sw_event(u32 event_id, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
999 {
1000         if (atomic_read(&perf_swevent_enabled[event_id])) {
1001                 struct pt_regs hot_regs;
1002
1003                 if (!regs) {
1004                         perf_fetch_caller_regs(&hot_regs);
1005                         regs = &hot_regs;
1006                 }
1007                 __perf_sw_event(event_id, nr, nmi, regs, addr);
1008         }
1009 }
1010
1011 extern void perf_event_mmap(struct vm_area_struct *vma);
1012 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1013 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1014 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1015
1016 extern void perf_event_comm(struct task_struct *tsk);
1017 extern void perf_event_fork(struct task_struct *tsk);
1018
1019 /* Callchains */
1020 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1021
1022 extern void perf_callchain_user(struct perf_callchain_entry *entry,
1023                                 struct pt_regs *regs);
1024 extern void perf_callchain_kernel(struct perf_callchain_entry *entry,
1025                                   struct pt_regs *regs);
1026
1027
1028 static inline void
1029 perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
1030 {
1031         if (entry->nr < PERF_MAX_STACK_DEPTH)
1032                 entry->ip[entry->nr++] = ip;
1033 }
1034
1035 extern int sysctl_perf_event_paranoid;
1036 extern int sysctl_perf_event_mlock;
1037 extern int sysctl_perf_event_sample_rate;
1038
1039 static inline bool perf_paranoid_tracepoint_raw(void)
1040 {
1041         return sysctl_perf_event_paranoid > -1;
1042 }
1043
1044 static inline bool perf_paranoid_cpu(void)
1045 {
1046         return sysctl_perf_event_paranoid > 0;
1047 }
1048
1049 static inline bool perf_paranoid_kernel(void)
1050 {
1051         return sysctl_perf_event_paranoid > 1;
1052 }
1053
1054 extern void perf_event_init(void);
1055 extern void perf_tp_event(u64 addr, u64 count, void *record,
1056                           int entry_size, struct pt_regs *regs,
1057                           struct hlist_head *head, int rctx);
1058 extern void perf_bp_event(struct perf_event *event, void *data);
1059
1060 #ifndef perf_misc_flags
1061 #define perf_misc_flags(regs)   (user_mode(regs) ? PERF_RECORD_MISC_USER : \
1062                                  PERF_RECORD_MISC_KERNEL)
1063 #define perf_instruction_pointer(regs)  instruction_pointer(regs)
1064 #endif
1065
1066 extern int perf_output_begin(struct perf_output_handle *handle,
1067                              struct perf_event *event, unsigned int size,
1068                              int nmi, int sample);
1069 extern void perf_output_end(struct perf_output_handle *handle);
1070 extern void perf_output_copy(struct perf_output_handle *handle,
1071                              const void *buf, unsigned int len);
1072 extern int perf_swevent_get_recursion_context(void);
1073 extern void perf_swevent_put_recursion_context(int rctx);
1074 extern void perf_event_enable(struct perf_event *event);
1075 extern void perf_event_disable(struct perf_event *event);
1076 #else
1077 static inline void
1078 perf_event_task_sched_in(struct task_struct *task)                      { }
1079 static inline void
1080 perf_event_task_sched_out(struct task_struct *task,
1081                             struct task_struct *next)                   { }
1082 static inline void
1083 perf_event_task_tick(struct task_struct *task)                          { }
1084 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
1085 static inline void perf_event_exit_task(struct task_struct *child)      { }
1086 static inline void perf_event_free_task(struct task_struct *task)       { }
1087 static inline void perf_event_do_pending(void)                          { }
1088 static inline void perf_event_print_debug(void)                         { }
1089 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
1090 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
1091
1092 static inline void
1093 perf_sw_event(u32 event_id, u64 nr, int nmi,
1094                      struct pt_regs *regs, u64 addr)                    { }
1095 static inline void
1096 perf_bp_event(struct perf_event *event, void *data)                     { }
1097
1098 static inline int perf_register_guest_info_callbacks
1099 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1100 static inline int perf_unregister_guest_info_callbacks
1101 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1102
1103 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
1104 static inline void perf_event_comm(struct task_struct *tsk)             { }
1105 static inline void perf_event_fork(struct task_struct *tsk)             { }
1106 static inline void perf_event_init(void)                                { }
1107 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
1108 static inline void perf_swevent_put_recursion_context(int rctx)         { }
1109 static inline void perf_event_enable(struct perf_event *event)          { }
1110 static inline void perf_event_disable(struct perf_event *event)         { }
1111 #endif
1112
1113 #define perf_output_put(handle, x) \
1114         perf_output_copy((handle), &(x), sizeof(x))
1115
1116 /*
1117  * This has to have a higher priority than migration_notifier in sched.c.
1118  */
1119 #define perf_cpu_notifier(fn)                                   \
1120 do {                                                            \
1121         static struct notifier_block fn##_nb __cpuinitdata =    \
1122                 { .notifier_call = fn, .priority = CPU_PRI_PERF }; \
1123         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,             \
1124                 (void *)(unsigned long)smp_processor_id());     \
1125         fn(&fn##_nb, (unsigned long)CPU_STARTING,               \
1126                 (void *)(unsigned long)smp_processor_id());     \
1127         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                 \
1128                 (void *)(unsigned long)smp_processor_id());     \
1129         register_cpu_notifier(&fn##_nb);                        \
1130 } while (0)
1131
1132 #endif /* __KERNEL__ */
1133 #endif /* _LINUX_PERF_EVENT_H */