]> git.openfabrics.org - ~shefty/rdma-dev.git/blob - include/linux/perf_event.h
perf, x86: Implement user-space RDPMC support, to allow fast, user-space access to...
[~shefty/rdma-dev.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         PERF_COUNT_HW_REF_CPU_CYCLES            = 9,
58
59         PERF_COUNT_HW_MAX,                      /* non-ABI */
60 };
61
62 /*
63  * Generalized hardware cache events:
64  *
65  *       { L1-D, L1-I, LLC, ITLB, DTLB, BPU, NODE } x
66  *       { read, write, prefetch } x
67  *       { accesses, misses }
68  */
69 enum perf_hw_cache_id {
70         PERF_COUNT_HW_CACHE_L1D                 = 0,
71         PERF_COUNT_HW_CACHE_L1I                 = 1,
72         PERF_COUNT_HW_CACHE_LL                  = 2,
73         PERF_COUNT_HW_CACHE_DTLB                = 3,
74         PERF_COUNT_HW_CACHE_ITLB                = 4,
75         PERF_COUNT_HW_CACHE_BPU                 = 5,
76         PERF_COUNT_HW_CACHE_NODE                = 6,
77
78         PERF_COUNT_HW_CACHE_MAX,                /* non-ABI */
79 };
80
81 enum perf_hw_cache_op_id {
82         PERF_COUNT_HW_CACHE_OP_READ             = 0,
83         PERF_COUNT_HW_CACHE_OP_WRITE            = 1,
84         PERF_COUNT_HW_CACHE_OP_PREFETCH         = 2,
85
86         PERF_COUNT_HW_CACHE_OP_MAX,             /* non-ABI */
87 };
88
89 enum perf_hw_cache_op_result_id {
90         PERF_COUNT_HW_CACHE_RESULT_ACCESS       = 0,
91         PERF_COUNT_HW_CACHE_RESULT_MISS         = 1,
92
93         PERF_COUNT_HW_CACHE_RESULT_MAX,         /* non-ABI */
94 };
95
96 /*
97  * Special "software" events provided by the kernel, even if the hardware
98  * does not support performance events. These events measure various
99  * physical and sw events of the kernel (and allow the profiling of them as
100  * well):
101  */
102 enum perf_sw_ids {
103         PERF_COUNT_SW_CPU_CLOCK                 = 0,
104         PERF_COUNT_SW_TASK_CLOCK                = 1,
105         PERF_COUNT_SW_PAGE_FAULTS               = 2,
106         PERF_COUNT_SW_CONTEXT_SWITCHES          = 3,
107         PERF_COUNT_SW_CPU_MIGRATIONS            = 4,
108         PERF_COUNT_SW_PAGE_FAULTS_MIN           = 5,
109         PERF_COUNT_SW_PAGE_FAULTS_MAJ           = 6,
110         PERF_COUNT_SW_ALIGNMENT_FAULTS          = 7,
111         PERF_COUNT_SW_EMULATION_FAULTS          = 8,
112
113         PERF_COUNT_SW_MAX,                      /* non-ABI */
114 };
115
116 /*
117  * Bits that can be set in attr.sample_type to request information
118  * in the overflow packets.
119  */
120 enum perf_event_sample_format {
121         PERF_SAMPLE_IP                          = 1U << 0,
122         PERF_SAMPLE_TID                         = 1U << 1,
123         PERF_SAMPLE_TIME                        = 1U << 2,
124         PERF_SAMPLE_ADDR                        = 1U << 3,
125         PERF_SAMPLE_READ                        = 1U << 4,
126         PERF_SAMPLE_CALLCHAIN                   = 1U << 5,
127         PERF_SAMPLE_ID                          = 1U << 6,
128         PERF_SAMPLE_CPU                         = 1U << 7,
129         PERF_SAMPLE_PERIOD                      = 1U << 8,
130         PERF_SAMPLE_STREAM_ID                   = 1U << 9,
131         PERF_SAMPLE_RAW                         = 1U << 10,
132
133         PERF_SAMPLE_MAX = 1U << 11,             /* non-ABI */
134 };
135
136 /*
137  * The format of the data returned by read() on a perf event fd,
138  * as specified by attr.read_format:
139  *
140  * struct read_format {
141  *      { u64           value;
142  *        { u64         time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
143  *        { u64         time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
144  *        { u64         id;           } && PERF_FORMAT_ID
145  *      } && !PERF_FORMAT_GROUP
146  *
147  *      { u64           nr;
148  *        { u64         time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
149  *        { u64         time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
150  *        { u64         value;
151  *          { u64       id;           } && PERF_FORMAT_ID
152  *        }             cntr[nr];
153  *      } && PERF_FORMAT_GROUP
154  * };
155  */
156 enum perf_event_read_format {
157         PERF_FORMAT_TOTAL_TIME_ENABLED          = 1U << 0,
158         PERF_FORMAT_TOTAL_TIME_RUNNING          = 1U << 1,
159         PERF_FORMAT_ID                          = 1U << 2,
160         PERF_FORMAT_GROUP                       = 1U << 3,
161
162         PERF_FORMAT_MAX = 1U << 4,              /* non-ABI */
163 };
164
165 #define PERF_ATTR_SIZE_VER0     64      /* sizeof first published struct */
166
167 /*
168  * Hardware event_id to monitor via a performance monitoring event:
169  */
170 struct perf_event_attr {
171
172         /*
173          * Major type: hardware/software/tracepoint/etc.
174          */
175         __u32                   type;
176
177         /*
178          * Size of the attr structure, for fwd/bwd compat.
179          */
180         __u32                   size;
181
182         /*
183          * Type specific configuration information.
184          */
185         __u64                   config;
186
187         union {
188                 __u64           sample_period;
189                 __u64           sample_freq;
190         };
191
192         __u64                   sample_type;
193         __u64                   read_format;
194
195         __u64                   disabled       :  1, /* off by default        */
196                                 inherit        :  1, /* children inherit it   */
197                                 pinned         :  1, /* must always be on PMU */
198                                 exclusive      :  1, /* only group on PMU     */
199                                 exclude_user   :  1, /* don't count user      */
200                                 exclude_kernel :  1, /* ditto kernel          */
201                                 exclude_hv     :  1, /* ditto hypervisor      */
202                                 exclude_idle   :  1, /* don't count when idle */
203                                 mmap           :  1, /* include mmap data     */
204                                 comm           :  1, /* include comm data     */
205                                 freq           :  1, /* use freq, not period  */
206                                 inherit_stat   :  1, /* per task counts       */
207                                 enable_on_exec :  1, /* next exec enables     */
208                                 task           :  1, /* trace fork/exit       */
209                                 watermark      :  1, /* wakeup_watermark      */
210                                 /*
211                                  * precise_ip:
212                                  *
213                                  *  0 - SAMPLE_IP can have arbitrary skid
214                                  *  1 - SAMPLE_IP must have constant skid
215                                  *  2 - SAMPLE_IP requested to have 0 skid
216                                  *  3 - SAMPLE_IP must have 0 skid
217                                  *
218                                  *  See also PERF_RECORD_MISC_EXACT_IP
219                                  */
220                                 precise_ip     :  2, /* skid constraint       */
221                                 mmap_data      :  1, /* non-exec mmap data    */
222                                 sample_id_all  :  1, /* sample_type all events */
223
224                                 exclude_host   :  1, /* don't count in host   */
225                                 exclude_guest  :  1, /* don't count in guest  */
226
227                                 __reserved_1   : 43;
228
229         union {
230                 __u32           wakeup_events;    /* wakeup every n events */
231                 __u32           wakeup_watermark; /* bytes before wakeup   */
232         };
233
234         __u32                   bp_type;
235         union {
236                 __u64           bp_addr;
237                 __u64           config1; /* extension of config */
238         };
239         union {
240                 __u64           bp_len;
241                 __u64           config2; /* extension of config1 */
242         };
243 };
244
245 /*
246  * Ioctls that can be done on a perf event fd:
247  */
248 #define PERF_EVENT_IOC_ENABLE           _IO ('$', 0)
249 #define PERF_EVENT_IOC_DISABLE          _IO ('$', 1)
250 #define PERF_EVENT_IOC_REFRESH          _IO ('$', 2)
251 #define PERF_EVENT_IOC_RESET            _IO ('$', 3)
252 #define PERF_EVENT_IOC_PERIOD           _IOW('$', 4, __u64)
253 #define PERF_EVENT_IOC_SET_OUTPUT       _IO ('$', 5)
254 #define PERF_EVENT_IOC_SET_FILTER       _IOW('$', 6, char *)
255
256 enum perf_event_ioc_flags {
257         PERF_IOC_FLAG_GROUP             = 1U << 0,
258 };
259
260 /*
261  * Structure of the page that can be mapped via mmap
262  */
263 struct perf_event_mmap_page {
264         __u32   version;                /* version number of this structure */
265         __u32   compat_version;         /* lowest version this is compat with */
266
267         /*
268          * Bits needed to read the hw events in user-space.
269          *
270          *   u32 seq;
271          *   s64 count;
272          *
273          *   do {
274          *     seq = pc->lock;
275          *
276          *     barrier()
277          *     if (pc->index) {
278          *       count = pmc_read(pc->index - 1);
279          *       count += pc->offset;
280          *     } else
281          *       goto regular_read;
282          *
283          *     barrier();
284          *   } while (pc->lock != seq);
285          *
286          * NOTE: for obvious reason this only works on self-monitoring
287          *       processes.
288          */
289         __u32   lock;                   /* seqlock for synchronization */
290         __u32   index;                  /* hardware event identifier */
291         __s64   offset;                 /* add to hardware event value */
292         __u64   time_enabled;           /* time event active */
293         __u64   time_running;           /* time event on cpu */
294
295                 /*
296                  * Hole for extension of the self monitor capabilities
297                  */
298
299         __u64   __reserved[123];        /* align to 1k */
300
301         /*
302          * Control data for the mmap() data buffer.
303          *
304          * User-space reading the @data_head value should issue an rmb(), on
305          * SMP capable platforms, after reading this value -- see
306          * perf_event_wakeup().
307          *
308          * When the mapping is PROT_WRITE the @data_tail value should be
309          * written by userspace to reflect the last read data. In this case
310          * the kernel will not over-write unread data.
311          */
312         __u64   data_head;              /* head in the data section */
313         __u64   data_tail;              /* user-space written tail */
314 };
315
316 #define PERF_RECORD_MISC_CPUMODE_MASK           (7 << 0)
317 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN        (0 << 0)
318 #define PERF_RECORD_MISC_KERNEL                 (1 << 0)
319 #define PERF_RECORD_MISC_USER                   (2 << 0)
320 #define PERF_RECORD_MISC_HYPERVISOR             (3 << 0)
321 #define PERF_RECORD_MISC_GUEST_KERNEL           (4 << 0)
322 #define PERF_RECORD_MISC_GUEST_USER             (5 << 0)
323
324 /*
325  * Indicates that the content of PERF_SAMPLE_IP points to
326  * the actual instruction that triggered the event. See also
327  * perf_event_attr::precise_ip.
328  */
329 #define PERF_RECORD_MISC_EXACT_IP               (1 << 14)
330 /*
331  * Reserve the last bit to indicate some extended misc field
332  */
333 #define PERF_RECORD_MISC_EXT_RESERVED           (1 << 15)
334
335 struct perf_event_header {
336         __u32   type;
337         __u16   misc;
338         __u16   size;
339 };
340
341 enum perf_event_type {
342
343         /*
344          * If perf_event_attr.sample_id_all is set then all event types will
345          * have the sample_type selected fields related to where/when
346          * (identity) an event took place (TID, TIME, ID, CPU, STREAM_ID)
347          * described in PERF_RECORD_SAMPLE below, it will be stashed just after
348          * the perf_event_header and the fields already present for the existing
349          * fields, i.e. at the end of the payload. That way a newer perf.data
350          * file will be supported by older perf tools, with these new optional
351          * fields being ignored.
352          *
353          * The MMAP events record the PROT_EXEC mappings so that we can
354          * correlate userspace IPs to code. They have the following structure:
355          *
356          * struct {
357          *      struct perf_event_header        header;
358          *
359          *      u32                             pid, tid;
360          *      u64                             addr;
361          *      u64                             len;
362          *      u64                             pgoff;
363          *      char                            filename[];
364          * };
365          */
366         PERF_RECORD_MMAP                        = 1,
367
368         /*
369          * struct {
370          *      struct perf_event_header        header;
371          *      u64                             id;
372          *      u64                             lost;
373          * };
374          */
375         PERF_RECORD_LOST                        = 2,
376
377         /*
378          * struct {
379          *      struct perf_event_header        header;
380          *
381          *      u32                             pid, tid;
382          *      char                            comm[];
383          * };
384          */
385         PERF_RECORD_COMM                        = 3,
386
387         /*
388          * struct {
389          *      struct perf_event_header        header;
390          *      u32                             pid, ppid;
391          *      u32                             tid, ptid;
392          *      u64                             time;
393          * };
394          */
395         PERF_RECORD_EXIT                        = 4,
396
397         /*
398          * struct {
399          *      struct perf_event_header        header;
400          *      u64                             time;
401          *      u64                             id;
402          *      u64                             stream_id;
403          * };
404          */
405         PERF_RECORD_THROTTLE                    = 5,
406         PERF_RECORD_UNTHROTTLE                  = 6,
407
408         /*
409          * struct {
410          *      struct perf_event_header        header;
411          *      u32                             pid, ppid;
412          *      u32                             tid, ptid;
413          *      u64                             time;
414          * };
415          */
416         PERF_RECORD_FORK                        = 7,
417
418         /*
419          * struct {
420          *      struct perf_event_header        header;
421          *      u32                             pid, tid;
422          *
423          *      struct read_format              values;
424          * };
425          */
426         PERF_RECORD_READ                        = 8,
427
428         /*
429          * struct {
430          *      struct perf_event_header        header;
431          *
432          *      { u64                   ip;       } && PERF_SAMPLE_IP
433          *      { u32                   pid, tid; } && PERF_SAMPLE_TID
434          *      { u64                   time;     } && PERF_SAMPLE_TIME
435          *      { u64                   addr;     } && PERF_SAMPLE_ADDR
436          *      { u64                   id;       } && PERF_SAMPLE_ID
437          *      { u64                   stream_id;} && PERF_SAMPLE_STREAM_ID
438          *      { u32                   cpu, res; } && PERF_SAMPLE_CPU
439          *      { u64                   period;   } && PERF_SAMPLE_PERIOD
440          *
441          *      { struct read_format    values;   } && PERF_SAMPLE_READ
442          *
443          *      { u64                   nr,
444          *        u64                   ips[nr];  } && PERF_SAMPLE_CALLCHAIN
445          *
446          *      #
447          *      # The RAW record below is opaque data wrt the ABI
448          *      #
449          *      # That is, the ABI doesn't make any promises wrt to
450          *      # the stability of its content, it may vary depending
451          *      # on event, hardware, kernel version and phase of
452          *      # the moon.
453          *      #
454          *      # In other words, PERF_SAMPLE_RAW contents are not an ABI.
455          *      #
456          *
457          *      { u32                   size;
458          *        char                  data[size];}&& PERF_SAMPLE_RAW
459          * };
460          */
461         PERF_RECORD_SAMPLE                      = 9,
462
463         PERF_RECORD_MAX,                        /* non-ABI */
464 };
465
466 enum perf_callchain_context {
467         PERF_CONTEXT_HV                 = (__u64)-32,
468         PERF_CONTEXT_KERNEL             = (__u64)-128,
469         PERF_CONTEXT_USER               = (__u64)-512,
470
471         PERF_CONTEXT_GUEST              = (__u64)-2048,
472         PERF_CONTEXT_GUEST_KERNEL       = (__u64)-2176,
473         PERF_CONTEXT_GUEST_USER         = (__u64)-2560,
474
475         PERF_CONTEXT_MAX                = (__u64)-4095,
476 };
477
478 #define PERF_FLAG_FD_NO_GROUP           (1U << 0)
479 #define PERF_FLAG_FD_OUTPUT             (1U << 1)
480 #define PERF_FLAG_PID_CGROUP            (1U << 2) /* pid=cgroup id, per-cpu mode only */
481
482 #ifdef __KERNEL__
483 /*
484  * Kernel-internal data types and definitions:
485  */
486
487 #ifdef CONFIG_PERF_EVENTS
488 # include <linux/cgroup.h>
489 # include <asm/perf_event.h>
490 # include <asm/local64.h>
491 #endif
492
493 struct perf_guest_info_callbacks {
494         int                             (*is_in_guest)(void);
495         int                             (*is_user_mode)(void);
496         unsigned long                   (*get_guest_ip)(void);
497 };
498
499 #ifdef CONFIG_HAVE_HW_BREAKPOINT
500 #include <asm/hw_breakpoint.h>
501 #endif
502
503 #include <linux/list.h>
504 #include <linux/mutex.h>
505 #include <linux/rculist.h>
506 #include <linux/rcupdate.h>
507 #include <linux/spinlock.h>
508 #include <linux/hrtimer.h>
509 #include <linux/fs.h>
510 #include <linux/pid_namespace.h>
511 #include <linux/workqueue.h>
512 #include <linux/ftrace.h>
513 #include <linux/cpu.h>
514 #include <linux/irq_work.h>
515 #include <linux/jump_label.h>
516 #include <linux/atomic.h>
517 #include <asm/local.h>
518
519 #define PERF_MAX_STACK_DEPTH            255
520
521 struct perf_callchain_entry {
522         __u64                           nr;
523         __u64                           ip[PERF_MAX_STACK_DEPTH];
524 };
525
526 struct perf_raw_record {
527         u32                             size;
528         void                            *data;
529 };
530
531 struct perf_branch_entry {
532         __u64                           from;
533         __u64                           to;
534         __u64                           flags;
535 };
536
537 struct perf_branch_stack {
538         __u64                           nr;
539         struct perf_branch_entry        entries[0];
540 };
541
542 struct task_struct;
543
544 /*
545  * extra PMU register associated with an event
546  */
547 struct hw_perf_event_extra {
548         u64             config; /* register value */
549         unsigned int    reg;    /* register address or index */
550         int             alloc;  /* extra register already allocated */
551         int             idx;    /* index in shared_regs->regs[] */
552 };
553
554 /**
555  * struct hw_perf_event - performance event hardware details:
556  */
557 struct hw_perf_event {
558 #ifdef CONFIG_PERF_EVENTS
559         union {
560                 struct { /* hardware */
561                         u64             config;
562                         u64             last_tag;
563                         unsigned long   config_base;
564                         unsigned long   event_base;
565                         int             idx;
566                         int             last_cpu;
567                         struct hw_perf_event_extra extra_reg;
568                 };
569                 struct { /* software */
570                         struct hrtimer  hrtimer;
571                 };
572 #ifdef CONFIG_HAVE_HW_BREAKPOINT
573                 struct { /* breakpoint */
574                         struct arch_hw_breakpoint       info;
575                         struct list_head                bp_list;
576                         /*
577                          * Crufty hack to avoid the chicken and egg
578                          * problem hw_breakpoint has with context
579                          * creation and event initalization.
580                          */
581                         struct task_struct              *bp_target;
582                 };
583 #endif
584         };
585         int                             state;
586         local64_t                       prev_count;
587         u64                             sample_period;
588         u64                             last_period;
589         local64_t                       period_left;
590         u64                             interrupts;
591
592         u64                             freq_time_stamp;
593         u64                             freq_count_stamp;
594 #endif
595 };
596
597 /*
598  * hw_perf_event::state flags
599  */
600 #define PERF_HES_STOPPED        0x01 /* the counter is stopped */
601 #define PERF_HES_UPTODATE       0x02 /* event->count up-to-date */
602 #define PERF_HES_ARCH           0x04
603
604 struct perf_event;
605
606 /*
607  * Common implementation detail of pmu::{start,commit,cancel}_txn
608  */
609 #define PERF_EVENT_TXN 0x1
610
611 /**
612  * struct pmu - generic performance monitoring unit
613  */
614 struct pmu {
615         struct list_head                entry;
616
617         struct device                   *dev;
618         char                            *name;
619         int                             type;
620
621         int * __percpu                  pmu_disable_count;
622         struct perf_cpu_context * __percpu pmu_cpu_context;
623         int                             task_ctx_nr;
624
625         /*
626          * Fully disable/enable this PMU, can be used to protect from the PMI
627          * as well as for lazy/batch writing of the MSRs.
628          */
629         void (*pmu_enable)              (struct pmu *pmu); /* optional */
630         void (*pmu_disable)             (struct pmu *pmu); /* optional */
631
632         /*
633          * Try and initialize the event for this PMU.
634          * Should return -ENOENT when the @event doesn't match this PMU.
635          */
636         int (*event_init)               (struct perf_event *event);
637
638 #define PERF_EF_START   0x01            /* start the counter when adding    */
639 #define PERF_EF_RELOAD  0x02            /* reload the counter when starting */
640 #define PERF_EF_UPDATE  0x04            /* update the counter when stopping */
641
642         /*
643          * Adds/Removes a counter to/from the PMU, can be done inside
644          * a transaction, see the ->*_txn() methods.
645          */
646         int  (*add)                     (struct perf_event *event, int flags);
647         void (*del)                     (struct perf_event *event, int flags);
648
649         /*
650          * Starts/Stops a counter present on the PMU. The PMI handler
651          * should stop the counter when perf_event_overflow() returns
652          * !0. ->start() will be used to continue.
653          */
654         void (*start)                   (struct perf_event *event, int flags);
655         void (*stop)                    (struct perf_event *event, int flags);
656
657         /*
658          * Updates the counter value of the event.
659          */
660         void (*read)                    (struct perf_event *event);
661
662         /*
663          * Group events scheduling is treated as a transaction, add
664          * group events as a whole and perform one schedulability test.
665          * If the test fails, roll back the whole group
666          *
667          * Start the transaction, after this ->add() doesn't need to
668          * do schedulability tests.
669          */
670         void (*start_txn)               (struct pmu *pmu); /* optional */
671         /*
672          * If ->start_txn() disabled the ->add() schedulability test
673          * then ->commit_txn() is required to perform one. On success
674          * the transaction is closed. On error the transaction is kept
675          * open until ->cancel_txn() is called.
676          */
677         int  (*commit_txn)              (struct pmu *pmu); /* optional */
678         /*
679          * Will cancel the transaction, assumes ->del() is called
680          * for each successful ->add() during the transaction.
681          */
682         void (*cancel_txn)              (struct pmu *pmu); /* optional */
683
684         /*
685          * Will return the value for perf_event_mmap_page::index for this event,
686          * if no implementation is provided it will default to: event->hw.idx + 1.
687          */
688         int (*event_idx)                (struct perf_event *event); /*optional */
689 };
690
691 /**
692  * enum perf_event_active_state - the states of a event
693  */
694 enum perf_event_active_state {
695         PERF_EVENT_STATE_ERROR          = -2,
696         PERF_EVENT_STATE_OFF            = -1,
697         PERF_EVENT_STATE_INACTIVE       =  0,
698         PERF_EVENT_STATE_ACTIVE         =  1,
699 };
700
701 struct file;
702 struct perf_sample_data;
703
704 typedef void (*perf_overflow_handler_t)(struct perf_event *,
705                                         struct perf_sample_data *,
706                                         struct pt_regs *regs);
707
708 enum perf_group_flag {
709         PERF_GROUP_SOFTWARE             = 0x1,
710 };
711
712 #define SWEVENT_HLIST_BITS              8
713 #define SWEVENT_HLIST_SIZE              (1 << SWEVENT_HLIST_BITS)
714
715 struct swevent_hlist {
716         struct hlist_head               heads[SWEVENT_HLIST_SIZE];
717         struct rcu_head                 rcu_head;
718 };
719
720 #define PERF_ATTACH_CONTEXT     0x01
721 #define PERF_ATTACH_GROUP       0x02
722 #define PERF_ATTACH_TASK        0x04
723
724 #ifdef CONFIG_CGROUP_PERF
725 /*
726  * perf_cgroup_info keeps track of time_enabled for a cgroup.
727  * This is a per-cpu dynamically allocated data structure.
728  */
729 struct perf_cgroup_info {
730         u64                             time;
731         u64                             timestamp;
732 };
733
734 struct perf_cgroup {
735         struct                          cgroup_subsys_state css;
736         struct                          perf_cgroup_info *info; /* timing info, one per cpu */
737 };
738 #endif
739
740 struct ring_buffer;
741
742 /**
743  * struct perf_event - performance event kernel representation:
744  */
745 struct perf_event {
746 #ifdef CONFIG_PERF_EVENTS
747         struct list_head                group_entry;
748         struct list_head                event_entry;
749         struct list_head                sibling_list;
750         struct hlist_node               hlist_entry;
751         int                             nr_siblings;
752         int                             group_flags;
753         struct perf_event               *group_leader;
754         struct pmu                      *pmu;
755
756         enum perf_event_active_state    state;
757         unsigned int                    attach_state;
758         local64_t                       count;
759         atomic64_t                      child_count;
760
761         /*
762          * These are the total time in nanoseconds that the event
763          * has been enabled (i.e. eligible to run, and the task has
764          * been scheduled in, if this is a per-task event)
765          * and running (scheduled onto the CPU), respectively.
766          *
767          * They are computed from tstamp_enabled, tstamp_running and
768          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
769          */
770         u64                             total_time_enabled;
771         u64                             total_time_running;
772
773         /*
774          * These are timestamps used for computing total_time_enabled
775          * and total_time_running when the event is in INACTIVE or
776          * ACTIVE state, measured in nanoseconds from an arbitrary point
777          * in time.
778          * tstamp_enabled: the notional time when the event was enabled
779          * tstamp_running: the notional time when the event was scheduled on
780          * tstamp_stopped: in INACTIVE state, the notional time when the
781          *      event was scheduled off.
782          */
783         u64                             tstamp_enabled;
784         u64                             tstamp_running;
785         u64                             tstamp_stopped;
786
787         /*
788          * timestamp shadows the actual context timing but it can
789          * be safely used in NMI interrupt context. It reflects the
790          * context time as it was when the event was last scheduled in.
791          *
792          * ctx_time already accounts for ctx->timestamp. Therefore to
793          * compute ctx_time for a sample, simply add perf_clock().
794          */
795         u64                             shadow_ctx_time;
796
797         struct perf_event_attr          attr;
798         u16                             header_size;
799         u16                             id_header_size;
800         u16                             read_size;
801         struct hw_perf_event            hw;
802
803         struct perf_event_context       *ctx;
804         struct file                     *filp;
805
806         /*
807          * These accumulate total time (in nanoseconds) that children
808          * events have been enabled and running, respectively.
809          */
810         atomic64_t                      child_total_time_enabled;
811         atomic64_t                      child_total_time_running;
812
813         /*
814          * Protect attach/detach and child_list:
815          */
816         struct mutex                    child_mutex;
817         struct list_head                child_list;
818         struct perf_event               *parent;
819
820         int                             oncpu;
821         int                             cpu;
822
823         struct list_head                owner_entry;
824         struct task_struct              *owner;
825
826         /* mmap bits */
827         struct mutex                    mmap_mutex;
828         atomic_t                        mmap_count;
829         int                             mmap_locked;
830         struct user_struct              *mmap_user;
831         struct ring_buffer              *rb;
832         struct list_head                rb_entry;
833
834         /* poll related */
835         wait_queue_head_t               waitq;
836         struct fasync_struct            *fasync;
837
838         /* delayed work for NMIs and such */
839         int                             pending_wakeup;
840         int                             pending_kill;
841         int                             pending_disable;
842         struct irq_work                 pending;
843
844         atomic_t                        event_limit;
845
846         void (*destroy)(struct perf_event *);
847         struct rcu_head                 rcu_head;
848
849         struct pid_namespace            *ns;
850         u64                             id;
851
852         perf_overflow_handler_t         overflow_handler;
853         void                            *overflow_handler_context;
854
855 #ifdef CONFIG_EVENT_TRACING
856         struct ftrace_event_call        *tp_event;
857         struct event_filter             *filter;
858 #endif
859
860 #ifdef CONFIG_CGROUP_PERF
861         struct perf_cgroup              *cgrp; /* cgroup event is attach to */
862         int                             cgrp_defer_enabled;
863 #endif
864
865 #endif /* CONFIG_PERF_EVENTS */
866 };
867
868 enum perf_event_context_type {
869         task_context,
870         cpu_context,
871 };
872
873 /**
874  * struct perf_event_context - event context structure
875  *
876  * Used as a container for task events and CPU events as well:
877  */
878 struct perf_event_context {
879         struct pmu                      *pmu;
880         enum perf_event_context_type    type;
881         /*
882          * Protect the states of the events in the list,
883          * nr_active, and the list:
884          */
885         raw_spinlock_t                  lock;
886         /*
887          * Protect the list of events.  Locking either mutex or lock
888          * is sufficient to ensure the list doesn't change; to change
889          * the list you need to lock both the mutex and the spinlock.
890          */
891         struct mutex                    mutex;
892
893         struct list_head                pinned_groups;
894         struct list_head                flexible_groups;
895         struct list_head                event_list;
896         int                             nr_events;
897         int                             nr_active;
898         int                             is_active;
899         int                             nr_stat;
900         int                             nr_freq;
901         int                             rotate_disable;
902         atomic_t                        refcount;
903         struct task_struct              *task;
904
905         /*
906          * Context clock, runs when context enabled.
907          */
908         u64                             time;
909         u64                             timestamp;
910
911         /*
912          * These fields let us detect when two contexts have both
913          * been cloned (inherited) from a common ancestor.
914          */
915         struct perf_event_context       *parent_ctx;
916         u64                             parent_gen;
917         u64                             generation;
918         int                             pin_count;
919         int                             nr_cgroups; /* cgroup events present */
920         struct rcu_head                 rcu_head;
921 };
922
923 /*
924  * Number of contexts where an event can trigger:
925  *      task, softirq, hardirq, nmi.
926  */
927 #define PERF_NR_CONTEXTS        4
928
929 /**
930  * struct perf_event_cpu_context - per cpu event context structure
931  */
932 struct perf_cpu_context {
933         struct perf_event_context       ctx;
934         struct perf_event_context       *task_ctx;
935         int                             active_oncpu;
936         int                             exclusive;
937         struct list_head                rotation_list;
938         int                             jiffies_interval;
939         struct pmu                      *active_pmu;
940         struct perf_cgroup              *cgrp;
941 };
942
943 struct perf_output_handle {
944         struct perf_event               *event;
945         struct ring_buffer              *rb;
946         unsigned long                   wakeup;
947         unsigned long                   size;
948         void                            *addr;
949         int                             page;
950 };
951
952 #ifdef CONFIG_PERF_EVENTS
953
954 extern int perf_pmu_register(struct pmu *pmu, char *name, int type);
955 extern void perf_pmu_unregister(struct pmu *pmu);
956
957 extern int perf_num_counters(void);
958 extern const char *perf_pmu_name(void);
959 extern void __perf_event_task_sched_in(struct task_struct *prev,
960                                        struct task_struct *task);
961 extern void __perf_event_task_sched_out(struct task_struct *prev,
962                                         struct task_struct *next);
963 extern int perf_event_init_task(struct task_struct *child);
964 extern void perf_event_exit_task(struct task_struct *child);
965 extern void perf_event_free_task(struct task_struct *task);
966 extern void perf_event_delayed_put(struct task_struct *task);
967 extern void perf_event_print_debug(void);
968 extern void perf_pmu_disable(struct pmu *pmu);
969 extern void perf_pmu_enable(struct pmu *pmu);
970 extern int perf_event_task_disable(void);
971 extern int perf_event_task_enable(void);
972 extern int perf_event_refresh(struct perf_event *event, int refresh);
973 extern void perf_event_update_userpage(struct perf_event *event);
974 extern int perf_event_release_kernel(struct perf_event *event);
975 extern struct perf_event *
976 perf_event_create_kernel_counter(struct perf_event_attr *attr,
977                                 int cpu,
978                                 struct task_struct *task,
979                                 perf_overflow_handler_t callback,
980                                 void *context);
981 extern u64 perf_event_read_value(struct perf_event *event,
982                                  u64 *enabled, u64 *running);
983
984 struct perf_sample_data {
985         u64                             type;
986
987         u64                             ip;
988         struct {
989                 u32     pid;
990                 u32     tid;
991         }                               tid_entry;
992         u64                             time;
993         u64                             addr;
994         u64                             id;
995         u64                             stream_id;
996         struct {
997                 u32     cpu;
998                 u32     reserved;
999         }                               cpu_entry;
1000         u64                             period;
1001         struct perf_callchain_entry     *callchain;
1002         struct perf_raw_record          *raw;
1003 };
1004
1005 static inline void perf_sample_data_init(struct perf_sample_data *data, u64 addr)
1006 {
1007         data->addr = addr;
1008         data->raw  = NULL;
1009 }
1010
1011 extern void perf_output_sample(struct perf_output_handle *handle,
1012                                struct perf_event_header *header,
1013                                struct perf_sample_data *data,
1014                                struct perf_event *event);
1015 extern void perf_prepare_sample(struct perf_event_header *header,
1016                                 struct perf_sample_data *data,
1017                                 struct perf_event *event,
1018                                 struct pt_regs *regs);
1019
1020 extern int perf_event_overflow(struct perf_event *event,
1021                                  struct perf_sample_data *data,
1022                                  struct pt_regs *regs);
1023
1024 static inline bool is_sampling_event(struct perf_event *event)
1025 {
1026         return event->attr.sample_period != 0;
1027 }
1028
1029 /*
1030  * Return 1 for a software event, 0 for a hardware event
1031  */
1032 static inline int is_software_event(struct perf_event *event)
1033 {
1034         return event->pmu->task_ctx_nr == perf_sw_context;
1035 }
1036
1037 extern struct jump_label_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
1038
1039 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
1040
1041 #ifndef perf_arch_fetch_caller_regs
1042 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
1043 #endif
1044
1045 /*
1046  * Take a snapshot of the regs. Skip ip and frame pointer to
1047  * the nth caller. We only need a few of the regs:
1048  * - ip for PERF_SAMPLE_IP
1049  * - cs for user_mode() tests
1050  * - bp for callchains
1051  * - eflags, for future purposes, just in case
1052  */
1053 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1054 {
1055         memset(regs, 0, sizeof(*regs));
1056
1057         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1058 }
1059
1060 static __always_inline void
1061 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1062 {
1063         struct pt_regs hot_regs;
1064
1065         if (static_branch(&perf_swevent_enabled[event_id])) {
1066                 if (!regs) {
1067                         perf_fetch_caller_regs(&hot_regs);
1068                         regs = &hot_regs;
1069                 }
1070                 __perf_sw_event(event_id, nr, regs, addr);
1071         }
1072 }
1073
1074 extern struct jump_label_key_deferred perf_sched_events;
1075
1076 static inline void perf_event_task_sched_in(struct task_struct *prev,
1077                                             struct task_struct *task)
1078 {
1079         if (static_branch(&perf_sched_events.key))
1080                 __perf_event_task_sched_in(prev, task);
1081 }
1082
1083 static inline void perf_event_task_sched_out(struct task_struct *prev,
1084                                              struct task_struct *next)
1085 {
1086         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, NULL, 0);
1087
1088         if (static_branch(&perf_sched_events.key))
1089                 __perf_event_task_sched_out(prev, next);
1090 }
1091
1092 extern void perf_event_mmap(struct vm_area_struct *vma);
1093 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1094 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1095 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1096
1097 extern void perf_event_comm(struct task_struct *tsk);
1098 extern void perf_event_fork(struct task_struct *tsk);
1099
1100 /* Callchains */
1101 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1102
1103 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
1104 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
1105
1106 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
1107 {
1108         if (entry->nr < PERF_MAX_STACK_DEPTH)
1109                 entry->ip[entry->nr++] = ip;
1110 }
1111
1112 extern int sysctl_perf_event_paranoid;
1113 extern int sysctl_perf_event_mlock;
1114 extern int sysctl_perf_event_sample_rate;
1115
1116 extern int perf_proc_update_handler(struct ctl_table *table, int write,
1117                 void __user *buffer, size_t *lenp,
1118                 loff_t *ppos);
1119
1120 static inline bool perf_paranoid_tracepoint_raw(void)
1121 {
1122         return sysctl_perf_event_paranoid > -1;
1123 }
1124
1125 static inline bool perf_paranoid_cpu(void)
1126 {
1127         return sysctl_perf_event_paranoid > 0;
1128 }
1129
1130 static inline bool perf_paranoid_kernel(void)
1131 {
1132         return sysctl_perf_event_paranoid > 1;
1133 }
1134
1135 extern void perf_event_init(void);
1136 extern void perf_tp_event(u64 addr, u64 count, void *record,
1137                           int entry_size, struct pt_regs *regs,
1138                           struct hlist_head *head, int rctx);
1139 extern void perf_bp_event(struct perf_event *event, void *data);
1140
1141 #ifndef perf_misc_flags
1142 # define perf_misc_flags(regs) \
1143                 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1144 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1145 #endif
1146
1147 extern int perf_output_begin(struct perf_output_handle *handle,
1148                              struct perf_event *event, unsigned int size);
1149 extern void perf_output_end(struct perf_output_handle *handle);
1150 extern void perf_output_copy(struct perf_output_handle *handle,
1151                              const void *buf, unsigned int len);
1152 extern int perf_swevent_get_recursion_context(void);
1153 extern void perf_swevent_put_recursion_context(int rctx);
1154 extern void perf_event_enable(struct perf_event *event);
1155 extern void perf_event_disable(struct perf_event *event);
1156 extern void perf_event_task_tick(void);
1157 #else
1158 static inline void
1159 perf_event_task_sched_in(struct task_struct *prev,
1160                          struct task_struct *task)                      { }
1161 static inline void
1162 perf_event_task_sched_out(struct task_struct *prev,
1163                           struct task_struct *next)                     { }
1164 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
1165 static inline void perf_event_exit_task(struct task_struct *child)      { }
1166 static inline void perf_event_free_task(struct task_struct *task)       { }
1167 static inline void perf_event_delayed_put(struct task_struct *task)     { }
1168 static inline void perf_event_print_debug(void)                         { }
1169 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
1170 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
1171 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1172 {
1173         return -EINVAL;
1174 }
1175
1176 static inline void
1177 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)     { }
1178 static inline void
1179 perf_bp_event(struct perf_event *event, void *data)                     { }
1180
1181 static inline int perf_register_guest_info_callbacks
1182 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1183 static inline int perf_unregister_guest_info_callbacks
1184 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1185
1186 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
1187 static inline void perf_event_comm(struct task_struct *tsk)             { }
1188 static inline void perf_event_fork(struct task_struct *tsk)             { }
1189 static inline void perf_event_init(void)                                { }
1190 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
1191 static inline void perf_swevent_put_recursion_context(int rctx)         { }
1192 static inline void perf_event_enable(struct perf_event *event)          { }
1193 static inline void perf_event_disable(struct perf_event *event)         { }
1194 static inline void perf_event_task_tick(void)                           { }
1195 #endif
1196
1197 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1198
1199 /*
1200  * This has to have a higher priority than migration_notifier in sched.c.
1201  */
1202 #define perf_cpu_notifier(fn)                                           \
1203 do {                                                                    \
1204         static struct notifier_block fn##_nb __cpuinitdata =            \
1205                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
1206         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,                     \
1207                 (void *)(unsigned long)smp_processor_id());             \
1208         fn(&fn##_nb, (unsigned long)CPU_STARTING,                       \
1209                 (void *)(unsigned long)smp_processor_id());             \
1210         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                         \
1211                 (void *)(unsigned long)smp_processor_id());             \
1212         register_cpu_notifier(&fn##_nb);                                \
1213 } while (0)
1214
1215 #endif /* __KERNEL__ */
1216 #endif /* _LINUX_PERF_EVENT_H */