perf events: Fix slow and broken cgroup context switch code
[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) 2008-2011, Red Hat, Inc., Ingo Molnar
6  *    Copyright (C) 2008-2011, 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         PERF_COUNT_HW_STALLED_CYCLES_FRONTEND   = 7,
56         PERF_COUNT_HW_STALLED_CYCLES_BACKEND    = 8,
57
58         PERF_COUNT_HW_MAX,                      /* non-ABI */
59 };
60
61 /*
62  * Generalized hardware cache events:
63  *
64  *       { L1-D, L1-I, LLC, ITLB, DTLB, BPU, NODE } x
65  *       { read, write, prefetch } x
66  *       { accesses, misses }
67  */
68 enum perf_hw_cache_id {
69         PERF_COUNT_HW_CACHE_L1D                 = 0,
70         PERF_COUNT_HW_CACHE_L1I                 = 1,
71         PERF_COUNT_HW_CACHE_LL                  = 2,
72         PERF_COUNT_HW_CACHE_DTLB                = 3,
73         PERF_COUNT_HW_CACHE_ITLB                = 4,
74         PERF_COUNT_HW_CACHE_BPU                 = 5,
75         PERF_COUNT_HW_CACHE_NODE                = 6,
76
77         PERF_COUNT_HW_CACHE_MAX,                /* non-ABI */
78 };
79
80 enum perf_hw_cache_op_id {
81         PERF_COUNT_HW_CACHE_OP_READ             = 0,
82         PERF_COUNT_HW_CACHE_OP_WRITE            = 1,
83         PERF_COUNT_HW_CACHE_OP_PREFETCH         = 2,
84
85         PERF_COUNT_HW_CACHE_OP_MAX,             /* non-ABI */
86 };
87
88 enum perf_hw_cache_op_result_id {
89         PERF_COUNT_HW_CACHE_RESULT_ACCESS       = 0,
90         PERF_COUNT_HW_CACHE_RESULT_MISS         = 1,
91
92         PERF_COUNT_HW_CACHE_RESULT_MAX,         /* non-ABI */
93 };
94
95 /*
96  * Special "software" events provided by the kernel, even if the hardware
97  * does not support performance events. These events measure various
98  * physical and sw events of the kernel (and allow the profiling of them as
99  * well):
100  */
101 enum perf_sw_ids {
102         PERF_COUNT_SW_CPU_CLOCK                 = 0,
103         PERF_COUNT_SW_TASK_CLOCK                = 1,
104         PERF_COUNT_SW_PAGE_FAULTS               = 2,
105         PERF_COUNT_SW_CONTEXT_SWITCHES          = 3,
106         PERF_COUNT_SW_CPU_MIGRATIONS            = 4,
107         PERF_COUNT_SW_PAGE_FAULTS_MIN           = 5,
108         PERF_COUNT_SW_PAGE_FAULTS_MAJ           = 6,
109         PERF_COUNT_SW_ALIGNMENT_FAULTS          = 7,
110         PERF_COUNT_SW_EMULATION_FAULTS          = 8,
111
112         PERF_COUNT_SW_MAX,                      /* non-ABI */
113 };
114
115 /*
116  * Bits that can be set in attr.sample_type to request information
117  * in the overflow packets.
118  */
119 enum perf_event_sample_format {
120         PERF_SAMPLE_IP                          = 1U << 0,
121         PERF_SAMPLE_TID                         = 1U << 1,
122         PERF_SAMPLE_TIME                        = 1U << 2,
123         PERF_SAMPLE_ADDR                        = 1U << 3,
124         PERF_SAMPLE_READ                        = 1U << 4,
125         PERF_SAMPLE_CALLCHAIN                   = 1U << 5,
126         PERF_SAMPLE_ID                          = 1U << 6,
127         PERF_SAMPLE_CPU                         = 1U << 7,
128         PERF_SAMPLE_PERIOD                      = 1U << 8,
129         PERF_SAMPLE_STREAM_ID                   = 1U << 9,
130         PERF_SAMPLE_RAW                         = 1U << 10,
131
132         PERF_SAMPLE_MAX = 1U << 11,             /* non-ABI */
133 };
134
135 /*
136  * The format of the data returned by read() on a perf event fd,
137  * as specified by attr.read_format:
138  *
139  * struct read_format {
140  *      { u64           value;
141  *        { u64         time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
142  *        { u64         time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
143  *        { u64         id;           } && PERF_FORMAT_ID
144  *      } && !PERF_FORMAT_GROUP
145  *
146  *      { u64           nr;
147  *        { u64         time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
148  *        { u64         time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
149  *        { u64         value;
150  *          { u64       id;           } && PERF_FORMAT_ID
151  *        }             cntr[nr];
152  *      } && PERF_FORMAT_GROUP
153  * };
154  */
155 enum perf_event_read_format {
156         PERF_FORMAT_TOTAL_TIME_ENABLED          = 1U << 0,
157         PERF_FORMAT_TOTAL_TIME_RUNNING          = 1U << 1,
158         PERF_FORMAT_ID                          = 1U << 2,
159         PERF_FORMAT_GROUP                       = 1U << 3,
160
161         PERF_FORMAT_MAX = 1U << 4,              /* non-ABI */
162 };
163
164 #define PERF_ATTR_SIZE_VER0     64      /* sizeof first published struct */
165
166 /*
167  * Hardware event_id to monitor via a performance monitoring event:
168  */
169 struct perf_event_attr {
170
171         /*
172          * Major type: hardware/software/tracepoint/etc.
173          */
174         __u32                   type;
175
176         /*
177          * Size of the attr structure, for fwd/bwd compat.
178          */
179         __u32                   size;
180
181         /*
182          * Type specific configuration information.
183          */
184         __u64                   config;
185
186         union {
187                 __u64           sample_period;
188                 __u64           sample_freq;
189         };
190
191         __u64                   sample_type;
192         __u64                   read_format;
193
194         __u64                   disabled       :  1, /* off by default        */
195                                 inherit        :  1, /* children inherit it   */
196                                 pinned         :  1, /* must always be on PMU */
197                                 exclusive      :  1, /* only group on PMU     */
198                                 exclude_user   :  1, /* don't count user      */
199                                 exclude_kernel :  1, /* ditto kernel          */
200                                 exclude_hv     :  1, /* ditto hypervisor      */
201                                 exclude_idle   :  1, /* don't count when idle */
202                                 mmap           :  1, /* include mmap data     */
203                                 comm           :  1, /* include comm data     */
204                                 freq           :  1, /* use freq, not period  */
205                                 inherit_stat   :  1, /* per task counts       */
206                                 enable_on_exec :  1, /* next exec enables     */
207                                 task           :  1, /* trace fork/exit       */
208                                 watermark      :  1, /* wakeup_watermark      */
209                                 /*
210                                  * precise_ip:
211                                  *
212                                  *  0 - SAMPLE_IP can have arbitrary skid
213                                  *  1 - SAMPLE_IP must have constant skid
214                                  *  2 - SAMPLE_IP requested to have 0 skid
215                                  *  3 - SAMPLE_IP must have 0 skid
216                                  *
217                                  *  See also PERF_RECORD_MISC_EXACT_IP
218                                  */
219                                 precise_ip     :  2, /* skid constraint       */
220                                 mmap_data      :  1, /* non-exec mmap data    */
221                                 sample_id_all  :  1, /* sample_type all events */
222
223                                 __reserved_1   : 45;
224
225         union {
226                 __u32           wakeup_events;    /* wakeup every n events */
227                 __u32           wakeup_watermark; /* bytes before wakeup   */
228         };
229
230         __u32                   bp_type;
231         union {
232                 __u64           bp_addr;
233                 __u64           config1; /* extension of config */
234         };
235         union {
236                 __u64           bp_len;
237                 __u64           config2; /* extension of config1 */
238         };
239 };
240
241 /*
242  * Ioctls that can be done on a perf event fd:
243  */
244 #define PERF_EVENT_IOC_ENABLE           _IO ('$', 0)
245 #define PERF_EVENT_IOC_DISABLE          _IO ('$', 1)
246 #define PERF_EVENT_IOC_REFRESH          _IO ('$', 2)
247 #define PERF_EVENT_IOC_RESET            _IO ('$', 3)
248 #define PERF_EVENT_IOC_PERIOD           _IOW('$', 4, __u64)
249 #define PERF_EVENT_IOC_SET_OUTPUT       _IO ('$', 5)
250 #define PERF_EVENT_IOC_SET_FILTER       _IOW('$', 6, char *)
251
252 enum perf_event_ioc_flags {
253         PERF_IOC_FLAG_GROUP             = 1U << 0,
254 };
255
256 /*
257  * Structure of the page that can be mapped via mmap
258  */
259 struct perf_event_mmap_page {
260         __u32   version;                /* version number of this structure */
261         __u32   compat_version;         /* lowest version this is compat with */
262
263         /*
264          * Bits needed to read the hw events in user-space.
265          *
266          *   u32 seq;
267          *   s64 count;
268          *
269          *   do {
270          *     seq = pc->lock;
271          *
272          *     barrier()
273          *     if (pc->index) {
274          *       count = pmc_read(pc->index - 1);
275          *       count += pc->offset;
276          *     } else
277          *       goto regular_read;
278          *
279          *     barrier();
280          *   } while (pc->lock != seq);
281          *
282          * NOTE: for obvious reason this only works on self-monitoring
283          *       processes.
284          */
285         __u32   lock;                   /* seqlock for synchronization */
286         __u32   index;                  /* hardware event identifier */
287         __s64   offset;                 /* add to hardware event value */
288         __u64   time_enabled;           /* time event active */
289         __u64   time_running;           /* time event on cpu */
290
291                 /*
292                  * Hole for extension of the self monitor capabilities
293                  */
294
295         __u64   __reserved[123];        /* align to 1k */
296
297         /*
298          * Control data for the mmap() data buffer.
299          *
300          * User-space reading the @data_head value should issue an rmb(), on
301          * SMP capable platforms, after reading this value -- see
302          * perf_event_wakeup().
303          *
304          * When the mapping is PROT_WRITE the @data_tail value should be
305          * written by userspace to reflect the last read data. In this case
306          * the kernel will not over-write unread data.
307          */
308         __u64   data_head;              /* head in the data section */
309         __u64   data_tail;              /* user-space written tail */
310 };
311
312 #define PERF_RECORD_MISC_CPUMODE_MASK           (7 << 0)
313 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN        (0 << 0)
314 #define PERF_RECORD_MISC_KERNEL                 (1 << 0)
315 #define PERF_RECORD_MISC_USER                   (2 << 0)
316 #define PERF_RECORD_MISC_HYPERVISOR             (3 << 0)
317 #define PERF_RECORD_MISC_GUEST_KERNEL           (4 << 0)
318 #define PERF_RECORD_MISC_GUEST_USER             (5 << 0)
319
320 /*
321  * Indicates that the content of PERF_SAMPLE_IP points to
322  * the actual instruction that triggered the event. See also
323  * perf_event_attr::precise_ip.
324  */
325 #define PERF_RECORD_MISC_EXACT_IP               (1 << 14)
326 /*
327  * Reserve the last bit to indicate some extended misc field
328  */
329 #define PERF_RECORD_MISC_EXT_RESERVED           (1 << 15)
330
331 struct perf_event_header {
332         __u32   type;
333         __u16   misc;
334         __u16   size;
335 };
336
337 enum perf_event_type {
338
339         /*
340          * If perf_event_attr.sample_id_all is set then all event types will
341          * have the sample_type selected fields related to where/when
342          * (identity) an event took place (TID, TIME, ID, CPU, STREAM_ID)
343          * described in PERF_RECORD_SAMPLE below, it will be stashed just after
344          * the perf_event_header and the fields already present for the existing
345          * fields, i.e. at the end of the payload. That way a newer perf.data
346          * file will be supported by older perf tools, with these new optional
347          * fields being ignored.
348          *
349          * The MMAP events record the PROT_EXEC mappings so that we can
350          * correlate userspace IPs to code. They have the following structure:
351          *
352          * struct {
353          *      struct perf_event_header        header;
354          *
355          *      u32                             pid, tid;
356          *      u64                             addr;
357          *      u64                             len;
358          *      u64                             pgoff;
359          *      char                            filename[];
360          * };
361          */
362         PERF_RECORD_MMAP                        = 1,
363
364         /*
365          * struct {
366          *      struct perf_event_header        header;
367          *      u64                             id;
368          *      u64                             lost;
369          * };
370          */
371         PERF_RECORD_LOST                        = 2,
372
373         /*
374          * struct {
375          *      struct perf_event_header        header;
376          *
377          *      u32                             pid, tid;
378          *      char                            comm[];
379          * };
380          */
381         PERF_RECORD_COMM                        = 3,
382
383         /*
384          * struct {
385          *      struct perf_event_header        header;
386          *      u32                             pid, ppid;
387          *      u32                             tid, ptid;
388          *      u64                             time;
389          * };
390          */
391         PERF_RECORD_EXIT                        = 4,
392
393         /*
394          * struct {
395          *      struct perf_event_header        header;
396          *      u64                             time;
397          *      u64                             id;
398          *      u64                             stream_id;
399          * };
400          */
401         PERF_RECORD_THROTTLE                    = 5,
402         PERF_RECORD_UNTHROTTLE                  = 6,
403
404         /*
405          * struct {
406          *      struct perf_event_header        header;
407          *      u32                             pid, ppid;
408          *      u32                             tid, ptid;
409          *      u64                             time;
410          * };
411          */
412         PERF_RECORD_FORK                        = 7,
413
414         /*
415          * struct {
416          *      struct perf_event_header        header;
417          *      u32                             pid, tid;
418          *
419          *      struct read_format              values;
420          * };
421          */
422         PERF_RECORD_READ                        = 8,
423
424         /*
425          * struct {
426          *      struct perf_event_header        header;
427          *
428          *      { u64                   ip;       } && PERF_SAMPLE_IP
429          *      { u32                   pid, tid; } && PERF_SAMPLE_TID
430          *      { u64                   time;     } && PERF_SAMPLE_TIME
431          *      { u64                   addr;     } && PERF_SAMPLE_ADDR
432          *      { u64                   id;       } && PERF_SAMPLE_ID
433          *      { u64                   stream_id;} && PERF_SAMPLE_STREAM_ID
434          *      { u32                   cpu, res; } && PERF_SAMPLE_CPU
435          *      { u64                   period;   } && PERF_SAMPLE_PERIOD
436          *
437          *      { struct read_format    values;   } && PERF_SAMPLE_READ
438          *
439          *      { u64                   nr,
440          *        u64                   ips[nr];  } && PERF_SAMPLE_CALLCHAIN
441          *
442          *      #
443          *      # The RAW record below is opaque data wrt the ABI
444          *      #
445          *      # That is, the ABI doesn't make any promises wrt to
446          *      # the stability of its content, it may vary depending
447          *      # on event, hardware, kernel version and phase of
448          *      # the moon.
449          *      #
450          *      # In other words, PERF_SAMPLE_RAW contents are not an ABI.
451          *      #
452          *
453          *      { u32                   size;
454          *        char                  data[size];}&& PERF_SAMPLE_RAW
455          * };
456          */
457         PERF_RECORD_SAMPLE                      = 9,
458
459         PERF_RECORD_MAX,                        /* non-ABI */
460 };
461
462 enum perf_callchain_context {
463         PERF_CONTEXT_HV                 = (__u64)-32,
464         PERF_CONTEXT_KERNEL             = (__u64)-128,
465         PERF_CONTEXT_USER               = (__u64)-512,
466
467         PERF_CONTEXT_GUEST              = (__u64)-2048,
468         PERF_CONTEXT_GUEST_KERNEL       = (__u64)-2176,
469         PERF_CONTEXT_GUEST_USER         = (__u64)-2560,
470
471         PERF_CONTEXT_MAX                = (__u64)-4095,
472 };
473
474 #define PERF_FLAG_FD_NO_GROUP           (1U << 0)
475 #define PERF_FLAG_FD_OUTPUT             (1U << 1)
476 #define PERF_FLAG_PID_CGROUP            (1U << 2) /* pid=cgroup id, per-cpu mode only */
477
478 #ifdef __KERNEL__
479 /*
480  * Kernel-internal data types and definitions:
481  */
482
483 #ifdef CONFIG_PERF_EVENTS
484 # include <linux/cgroup.h>
485 # include <asm/perf_event.h>
486 # include <asm/local64.h>
487 #endif
488
489 struct perf_guest_info_callbacks {
490         int                             (*is_in_guest)(void);
491         int                             (*is_user_mode)(void);
492         unsigned long                   (*get_guest_ip)(void);
493 };
494
495 #ifdef CONFIG_HAVE_HW_BREAKPOINT
496 #include <asm/hw_breakpoint.h>
497 #endif
498
499 #include <linux/list.h>
500 #include <linux/mutex.h>
501 #include <linux/rculist.h>
502 #include <linux/rcupdate.h>
503 #include <linux/spinlock.h>
504 #include <linux/hrtimer.h>
505 #include <linux/fs.h>
506 #include <linux/pid_namespace.h>
507 #include <linux/workqueue.h>
508 #include <linux/ftrace.h>
509 #include <linux/cpu.h>
510 #include <linux/irq_work.h>
511 #include <linux/jump_label.h>
512 #include <linux/atomic.h>
513 #include <asm/local.h>
514
515 #define PERF_MAX_STACK_DEPTH            255
516
517 struct perf_callchain_entry {
518         __u64                           nr;
519         __u64                           ip[PERF_MAX_STACK_DEPTH];
520 };
521
522 struct perf_raw_record {
523         u32                             size;
524         void                            *data;
525 };
526
527 struct perf_branch_entry {
528         __u64                           from;
529         __u64                           to;
530         __u64                           flags;
531 };
532
533 struct perf_branch_stack {
534         __u64                           nr;
535         struct perf_branch_entry        entries[0];
536 };
537
538 struct task_struct;
539
540 /*
541  * extra PMU register associated with an event
542  */
543 struct hw_perf_event_extra {
544         u64             config; /* register value */
545         unsigned int    reg;    /* register address or index */
546         int             alloc;  /* extra register already allocated */
547         int             idx;    /* index in shared_regs->regs[] */
548 };
549
550 /**
551  * struct hw_perf_event - performance event hardware details:
552  */
553 struct hw_perf_event {
554 #ifdef CONFIG_PERF_EVENTS
555         union {
556                 struct { /* hardware */
557                         u64             config;
558                         u64             last_tag;
559                         unsigned long   config_base;
560                         unsigned long   event_base;
561                         int             idx;
562                         int             last_cpu;
563                         struct hw_perf_event_extra extra_reg;
564                 };
565                 struct { /* software */
566                         struct hrtimer  hrtimer;
567                 };
568 #ifdef CONFIG_HAVE_HW_BREAKPOINT
569                 struct { /* breakpoint */
570                         struct arch_hw_breakpoint       info;
571                         struct list_head                bp_list;
572                         /*
573                          * Crufty hack to avoid the chicken and egg
574                          * problem hw_breakpoint has with context
575                          * creation and event initalization.
576                          */
577                         struct task_struct              *bp_target;
578                 };
579 #endif
580         };
581         int                             state;
582         local64_t                       prev_count;
583         u64                             sample_period;
584         u64                             last_period;
585         local64_t                       period_left;
586         u64                             interrupts;
587
588         u64                             freq_time_stamp;
589         u64                             freq_count_stamp;
590 #endif
591 };
592
593 /*
594  * hw_perf_event::state flags
595  */
596 #define PERF_HES_STOPPED        0x01 /* the counter is stopped */
597 #define PERF_HES_UPTODATE       0x02 /* event->count up-to-date */
598 #define PERF_HES_ARCH           0x04
599
600 struct perf_event;
601
602 /*
603  * Common implementation detail of pmu::{start,commit,cancel}_txn
604  */
605 #define PERF_EVENT_TXN 0x1
606
607 /**
608  * struct pmu - generic performance monitoring unit
609  */
610 struct pmu {
611         struct list_head                entry;
612
613         struct device                   *dev;
614         char                            *name;
615         int                             type;
616
617         int * __percpu                  pmu_disable_count;
618         struct perf_cpu_context * __percpu pmu_cpu_context;
619         int                             task_ctx_nr;
620
621         /*
622          * Fully disable/enable this PMU, can be used to protect from the PMI
623          * as well as for lazy/batch writing of the MSRs.
624          */
625         void (*pmu_enable)              (struct pmu *pmu); /* optional */
626         void (*pmu_disable)             (struct pmu *pmu); /* optional */
627
628         /*
629          * Try and initialize the event for this PMU.
630          * Should return -ENOENT when the @event doesn't match this PMU.
631          */
632         int (*event_init)               (struct perf_event *event);
633
634 #define PERF_EF_START   0x01            /* start the counter when adding    */
635 #define PERF_EF_RELOAD  0x02            /* reload the counter when starting */
636 #define PERF_EF_UPDATE  0x04            /* update the counter when stopping */
637
638         /*
639          * Adds/Removes a counter to/from the PMU, can be done inside
640          * a transaction, see the ->*_txn() methods.
641          */
642         int  (*add)                     (struct perf_event *event, int flags);
643         void (*del)                     (struct perf_event *event, int flags);
644
645         /*
646          * Starts/Stops a counter present on the PMU. The PMI handler
647          * should stop the counter when perf_event_overflow() returns
648          * !0. ->start() will be used to continue.
649          */
650         void (*start)                   (struct perf_event *event, int flags);
651         void (*stop)                    (struct perf_event *event, int flags);
652
653         /*
654          * Updates the counter value of the event.
655          */
656         void (*read)                    (struct perf_event *event);
657
658         /*
659          * Group events scheduling is treated as a transaction, add
660          * group events as a whole and perform one schedulability test.
661          * If the test fails, roll back the whole group
662          *
663          * Start the transaction, after this ->add() doesn't need to
664          * do schedulability tests.
665          */
666         void (*start_txn)               (struct pmu *pmu); /* optional */
667         /*
668          * If ->start_txn() disabled the ->add() schedulability test
669          * then ->commit_txn() is required to perform one. On success
670          * the transaction is closed. On error the transaction is kept
671          * open until ->cancel_txn() is called.
672          */
673         int  (*commit_txn)              (struct pmu *pmu); /* optional */
674         /*
675          * Will cancel the transaction, assumes ->del() is called
676          * for each successful ->add() during the transaction.
677          */
678         void (*cancel_txn)              (struct pmu *pmu); /* optional */
679 };
680
681 /**
682  * enum perf_event_active_state - the states of a event
683  */
684 enum perf_event_active_state {
685         PERF_EVENT_STATE_ERROR          = -2,
686         PERF_EVENT_STATE_OFF            = -1,
687         PERF_EVENT_STATE_INACTIVE       =  0,
688         PERF_EVENT_STATE_ACTIVE         =  1,
689 };
690
691 struct file;
692 struct perf_sample_data;
693
694 typedef void (*perf_overflow_handler_t)(struct perf_event *,
695                                         struct perf_sample_data *,
696                                         struct pt_regs *regs);
697
698 enum perf_group_flag {
699         PERF_GROUP_SOFTWARE             = 0x1,
700 };
701
702 #define SWEVENT_HLIST_BITS              8
703 #define SWEVENT_HLIST_SIZE              (1 << SWEVENT_HLIST_BITS)
704
705 struct swevent_hlist {
706         struct hlist_head               heads[SWEVENT_HLIST_SIZE];
707         struct rcu_head                 rcu_head;
708 };
709
710 #define PERF_ATTACH_CONTEXT     0x01
711 #define PERF_ATTACH_GROUP       0x02
712 #define PERF_ATTACH_TASK        0x04
713
714 #ifdef CONFIG_CGROUP_PERF
715 /*
716  * perf_cgroup_info keeps track of time_enabled for a cgroup.
717  * This is a per-cpu dynamically allocated data structure.
718  */
719 struct perf_cgroup_info {
720         u64                             time;
721         u64                             timestamp;
722 };
723
724 struct perf_cgroup {
725         struct                          cgroup_subsys_state css;
726         struct                          perf_cgroup_info *info; /* timing info, one per cpu */
727 };
728 #endif
729
730 struct ring_buffer;
731
732 /**
733  * struct perf_event - performance event kernel representation:
734  */
735 struct perf_event {
736 #ifdef CONFIG_PERF_EVENTS
737         struct list_head                group_entry;
738         struct list_head                event_entry;
739         struct list_head                sibling_list;
740         struct hlist_node               hlist_entry;
741         int                             nr_siblings;
742         int                             group_flags;
743         struct perf_event               *group_leader;
744         struct pmu                      *pmu;
745
746         enum perf_event_active_state    state;
747         unsigned int                    attach_state;
748         local64_t                       count;
749         atomic64_t                      child_count;
750
751         /*
752          * These are the total time in nanoseconds that the event
753          * has been enabled (i.e. eligible to run, and the task has
754          * been scheduled in, if this is a per-task event)
755          * and running (scheduled onto the CPU), respectively.
756          *
757          * They are computed from tstamp_enabled, tstamp_running and
758          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
759          */
760         u64                             total_time_enabled;
761         u64                             total_time_running;
762
763         /*
764          * These are timestamps used for computing total_time_enabled
765          * and total_time_running when the event is in INACTIVE or
766          * ACTIVE state, measured in nanoseconds from an arbitrary point
767          * in time.
768          * tstamp_enabled: the notional time when the event was enabled
769          * tstamp_running: the notional time when the event was scheduled on
770          * tstamp_stopped: in INACTIVE state, the notional time when the
771          *      event was scheduled off.
772          */
773         u64                             tstamp_enabled;
774         u64                             tstamp_running;
775         u64                             tstamp_stopped;
776
777         /*
778          * timestamp shadows the actual context timing but it can
779          * be safely used in NMI interrupt context. It reflects the
780          * context time as it was when the event was last scheduled in.
781          *
782          * ctx_time already accounts for ctx->timestamp. Therefore to
783          * compute ctx_time for a sample, simply add perf_clock().
784          */
785         u64                             shadow_ctx_time;
786
787         struct perf_event_attr          attr;
788         u16                             header_size;
789         u16                             id_header_size;
790         u16                             read_size;
791         struct hw_perf_event            hw;
792
793         struct perf_event_context       *ctx;
794         struct file                     *filp;
795
796         /*
797          * These accumulate total time (in nanoseconds) that children
798          * events have been enabled and running, respectively.
799          */
800         atomic64_t                      child_total_time_enabled;
801         atomic64_t                      child_total_time_running;
802
803         /*
804          * Protect attach/detach and child_list:
805          */
806         struct mutex                    child_mutex;
807         struct list_head                child_list;
808         struct perf_event               *parent;
809
810         int                             oncpu;
811         int                             cpu;
812
813         struct list_head                owner_entry;
814         struct task_struct              *owner;
815
816         /* mmap bits */
817         struct mutex                    mmap_mutex;
818         atomic_t                        mmap_count;
819         int                             mmap_locked;
820         struct user_struct              *mmap_user;
821         struct ring_buffer              *rb;
822
823         /* poll related */
824         wait_queue_head_t               waitq;
825         struct fasync_struct            *fasync;
826
827         /* delayed work for NMIs and such */
828         int                             pending_wakeup;
829         int                             pending_kill;
830         int                             pending_disable;
831         struct irq_work                 pending;
832
833         atomic_t                        event_limit;
834
835         void (*destroy)(struct perf_event *);
836         struct rcu_head                 rcu_head;
837
838         struct pid_namespace            *ns;
839         u64                             id;
840
841         perf_overflow_handler_t         overflow_handler;
842         void                            *overflow_handler_context;
843
844 #ifdef CONFIG_EVENT_TRACING
845         struct ftrace_event_call        *tp_event;
846         struct event_filter             *filter;
847 #endif
848
849 #ifdef CONFIG_CGROUP_PERF
850         struct perf_cgroup              *cgrp; /* cgroup event is attach to */
851         int                             cgrp_defer_enabled;
852 #endif
853
854 #endif /* CONFIG_PERF_EVENTS */
855 };
856
857 enum perf_event_context_type {
858         task_context,
859         cpu_context,
860 };
861
862 /**
863  * struct perf_event_context - event context structure
864  *
865  * Used as a container for task events and CPU events as well:
866  */
867 struct perf_event_context {
868         struct pmu                      *pmu;
869         enum perf_event_context_type    type;
870         /*
871          * Protect the states of the events in the list,
872          * nr_active, and the list:
873          */
874         raw_spinlock_t                  lock;
875         /*
876          * Protect the list of events.  Locking either mutex or lock
877          * is sufficient to ensure the list doesn't change; to change
878          * the list you need to lock both the mutex and the spinlock.
879          */
880         struct mutex                    mutex;
881
882         struct list_head                pinned_groups;
883         struct list_head                flexible_groups;
884         struct list_head                event_list;
885         int                             nr_events;
886         int                             nr_active;
887         int                             is_active;
888         int                             nr_stat;
889         int                             rotate_disable;
890         atomic_t                        refcount;
891         struct task_struct              *task;
892
893         /*
894          * Context clock, runs when context enabled.
895          */
896         u64                             time;
897         u64                             timestamp;
898
899         /*
900          * These fields let us detect when two contexts have both
901          * been cloned (inherited) from a common ancestor.
902          */
903         struct perf_event_context       *parent_ctx;
904         u64                             parent_gen;
905         u64                             generation;
906         int                             pin_count;
907         int                             nr_cgroups; /* cgroup events present */
908         struct rcu_head                 rcu_head;
909 };
910
911 /*
912  * Number of contexts where an event can trigger:
913  *      task, softirq, hardirq, nmi.
914  */
915 #define PERF_NR_CONTEXTS        4
916
917 /**
918  * struct perf_event_cpu_context - per cpu event context structure
919  */
920 struct perf_cpu_context {
921         struct perf_event_context       ctx;
922         struct perf_event_context       *task_ctx;
923         int                             active_oncpu;
924         int                             exclusive;
925         struct list_head                rotation_list;
926         int                             jiffies_interval;
927         struct pmu                      *active_pmu;
928         struct perf_cgroup              *cgrp;
929 };
930
931 struct perf_output_handle {
932         struct perf_event               *event;
933         struct ring_buffer              *rb;
934         unsigned long                   wakeup;
935         unsigned long                   size;
936         void                            *addr;
937         int                             page;
938 };
939
940 #ifdef CONFIG_PERF_EVENTS
941
942 extern int perf_pmu_register(struct pmu *pmu, char *name, int type);
943 extern void perf_pmu_unregister(struct pmu *pmu);
944
945 extern int perf_num_counters(void);
946 extern const char *perf_pmu_name(void);
947 extern void __perf_event_task_sched_in(struct task_struct *prev,
948                                        struct task_struct *task);
949 extern void __perf_event_task_sched_out(struct task_struct *prev,
950                                         struct task_struct *next);
951 extern int perf_event_init_task(struct task_struct *child);
952 extern void perf_event_exit_task(struct task_struct *child);
953 extern void perf_event_free_task(struct task_struct *task);
954 extern void perf_event_delayed_put(struct task_struct *task);
955 extern void perf_event_print_debug(void);
956 extern void perf_pmu_disable(struct pmu *pmu);
957 extern void perf_pmu_enable(struct pmu *pmu);
958 extern int perf_event_task_disable(void);
959 extern int perf_event_task_enable(void);
960 extern int perf_event_refresh(struct perf_event *event, int refresh);
961 extern void perf_event_update_userpage(struct perf_event *event);
962 extern int perf_event_release_kernel(struct perf_event *event);
963 extern struct perf_event *
964 perf_event_create_kernel_counter(struct perf_event_attr *attr,
965                                 int cpu,
966                                 struct task_struct *task,
967                                 perf_overflow_handler_t callback,
968                                 void *context);
969 extern u64 perf_event_read_value(struct perf_event *event,
970                                  u64 *enabled, u64 *running);
971
972 struct perf_sample_data {
973         u64                             type;
974
975         u64                             ip;
976         struct {
977                 u32     pid;
978                 u32     tid;
979         }                               tid_entry;
980         u64                             time;
981         u64                             addr;
982         u64                             id;
983         u64                             stream_id;
984         struct {
985                 u32     cpu;
986                 u32     reserved;
987         }                               cpu_entry;
988         u64                             period;
989         struct perf_callchain_entry     *callchain;
990         struct perf_raw_record          *raw;
991 };
992
993 static inline void perf_sample_data_init(struct perf_sample_data *data, u64 addr)
994 {
995         data->addr = addr;
996         data->raw  = NULL;
997 }
998
999 extern void perf_output_sample(struct perf_output_handle *handle,
1000                                struct perf_event_header *header,
1001                                struct perf_sample_data *data,
1002                                struct perf_event *event);
1003 extern void perf_prepare_sample(struct perf_event_header *header,
1004                                 struct perf_sample_data *data,
1005                                 struct perf_event *event,
1006                                 struct pt_regs *regs);
1007
1008 extern int perf_event_overflow(struct perf_event *event,
1009                                  struct perf_sample_data *data,
1010                                  struct pt_regs *regs);
1011
1012 static inline bool is_sampling_event(struct perf_event *event)
1013 {
1014         return event->attr.sample_period != 0;
1015 }
1016
1017 /*
1018  * Return 1 for a software event, 0 for a hardware event
1019  */
1020 static inline int is_software_event(struct perf_event *event)
1021 {
1022         return event->pmu->task_ctx_nr == perf_sw_context;
1023 }
1024
1025 extern struct jump_label_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
1026
1027 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
1028
1029 #ifndef perf_arch_fetch_caller_regs
1030 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
1031 #endif
1032
1033 /*
1034  * Take a snapshot of the regs. Skip ip and frame pointer to
1035  * the nth caller. We only need a few of the regs:
1036  * - ip for PERF_SAMPLE_IP
1037  * - cs for user_mode() tests
1038  * - bp for callchains
1039  * - eflags, for future purposes, just in case
1040  */
1041 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1042 {
1043         memset(regs, 0, sizeof(*regs));
1044
1045         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1046 }
1047
1048 static __always_inline void
1049 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1050 {
1051         struct pt_regs hot_regs;
1052
1053         if (static_branch(&perf_swevent_enabled[event_id])) {
1054                 if (!regs) {
1055                         perf_fetch_caller_regs(&hot_regs);
1056                         regs = &hot_regs;
1057                 }
1058                 __perf_sw_event(event_id, nr, regs, addr);
1059         }
1060 }
1061
1062 extern struct jump_label_key perf_sched_events;
1063
1064 static inline void perf_event_task_sched_in(struct task_struct *prev,
1065                                             struct task_struct *task)
1066 {
1067         if (static_branch(&perf_sched_events))
1068                 __perf_event_task_sched_in(prev, task);
1069 }
1070
1071 static inline void perf_event_task_sched_out(struct task_struct *prev,
1072                                              struct task_struct *next)
1073 {
1074         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, NULL, 0);
1075
1076         if (static_branch(&perf_sched_events))
1077                 __perf_event_task_sched_out(prev, next);
1078 }
1079
1080 extern void perf_event_mmap(struct vm_area_struct *vma);
1081 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1082 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1083 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1084
1085 extern void perf_event_comm(struct task_struct *tsk);
1086 extern void perf_event_fork(struct task_struct *tsk);
1087
1088 /* Callchains */
1089 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1090
1091 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
1092 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
1093
1094 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
1095 {
1096         if (entry->nr < PERF_MAX_STACK_DEPTH)
1097                 entry->ip[entry->nr++] = ip;
1098 }
1099
1100 extern int sysctl_perf_event_paranoid;
1101 extern int sysctl_perf_event_mlock;
1102 extern int sysctl_perf_event_sample_rate;
1103
1104 extern int perf_proc_update_handler(struct ctl_table *table, int write,
1105                 void __user *buffer, size_t *lenp,
1106                 loff_t *ppos);
1107
1108 static inline bool perf_paranoid_tracepoint_raw(void)
1109 {
1110         return sysctl_perf_event_paranoid > -1;
1111 }
1112
1113 static inline bool perf_paranoid_cpu(void)
1114 {
1115         return sysctl_perf_event_paranoid > 0;
1116 }
1117
1118 static inline bool perf_paranoid_kernel(void)
1119 {
1120         return sysctl_perf_event_paranoid > 1;
1121 }
1122
1123 extern void perf_event_init(void);
1124 extern void perf_tp_event(u64 addr, u64 count, void *record,
1125                           int entry_size, struct pt_regs *regs,
1126                           struct hlist_head *head, int rctx);
1127 extern void perf_bp_event(struct perf_event *event, void *data);
1128
1129 #ifndef perf_misc_flags
1130 # define perf_misc_flags(regs) \
1131                 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1132 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1133 #endif
1134
1135 extern int perf_output_begin(struct perf_output_handle *handle,
1136                              struct perf_event *event, unsigned int size);
1137 extern void perf_output_end(struct perf_output_handle *handle);
1138 extern void perf_output_copy(struct perf_output_handle *handle,
1139                              const void *buf, unsigned int len);
1140 extern int perf_swevent_get_recursion_context(void);
1141 extern void perf_swevent_put_recursion_context(int rctx);
1142 extern void perf_event_enable(struct perf_event *event);
1143 extern void perf_event_disable(struct perf_event *event);
1144 extern void perf_event_task_tick(void);
1145 #else
1146 static inline void
1147 perf_event_task_sched_in(struct task_struct *prev,
1148                          struct task_struct *task)                      { }
1149 static inline void
1150 perf_event_task_sched_out(struct task_struct *prev,
1151                           struct task_struct *next)                     { }
1152 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
1153 static inline void perf_event_exit_task(struct task_struct *child)      { }
1154 static inline void perf_event_free_task(struct task_struct *task)       { }
1155 static inline void perf_event_delayed_put(struct task_struct *task)     { }
1156 static inline void perf_event_print_debug(void)                         { }
1157 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
1158 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
1159 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1160 {
1161         return -EINVAL;
1162 }
1163
1164 static inline void
1165 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)     { }
1166 static inline void
1167 perf_bp_event(struct perf_event *event, void *data)                     { }
1168
1169 static inline int perf_register_guest_info_callbacks
1170 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1171 static inline int perf_unregister_guest_info_callbacks
1172 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1173
1174 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
1175 static inline void perf_event_comm(struct task_struct *tsk)             { }
1176 static inline void perf_event_fork(struct task_struct *tsk)             { }
1177 static inline void perf_event_init(void)                                { }
1178 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
1179 static inline void perf_swevent_put_recursion_context(int rctx)         { }
1180 static inline void perf_event_enable(struct perf_event *event)          { }
1181 static inline void perf_event_disable(struct perf_event *event)         { }
1182 static inline void perf_event_task_tick(void)                           { }
1183 #endif
1184
1185 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1186
1187 /*
1188  * This has to have a higher priority than migration_notifier in sched.c.
1189  */
1190 #define perf_cpu_notifier(fn)                                           \
1191 do {                                                                    \
1192         static struct notifier_block fn##_nb __cpuinitdata =            \
1193                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
1194         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,                     \
1195                 (void *)(unsigned long)smp_processor_id());             \
1196         fn(&fn##_nb, (unsigned long)CPU_STARTING,                       \
1197                 (void *)(unsigned long)smp_processor_id());             \
1198         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                         \
1199                 (void *)(unsigned long)smp_processor_id());             \
1200         register_cpu_notifier(&fn##_nb);                                \
1201 } while (0)
1202
1203 #endif /* __KERNEL__ */
1204 #endif /* _LINUX_PERF_EVENT_H */