0fcbf34a4f73b047cd5c7b1473ae05ff69c71446
[linux-2.6.git] / include / linux / perf_counter.h
1 /*
2  *  Performance counters:
3  *
4  *   Copyright(C) 2008, Thomas Gleixner <tglx@linutronix.de>
5  *   Copyright(C) 2008, Red Hat, Inc., Ingo Molnar
6  *
7  *  Data type definitions, declarations, prototypes.
8  *
9  *  Started by: Thomas Gleixner and Ingo Molnar
10  *
11  *  For licencing details see kernel-base/COPYING
12  */
13 #ifndef _LINUX_PERF_COUNTER_H
14 #define _LINUX_PERF_COUNTER_H
15
16 #include <linux/types.h>
17 #include <linux/ioctl.h>
18 #include <asm/byteorder.h>
19
20 /*
21  * User-space ABI bits:
22  */
23
24 /*
25  * hw_event.type
26  */
27 enum perf_event_types {
28         PERF_TYPE_HARDWARE              = 0,
29         PERF_TYPE_SOFTWARE              = 1,
30         PERF_TYPE_TRACEPOINT            = 2,
31
32         /*
33          * available TYPE space, raw is the max value.
34          */
35
36         PERF_TYPE_RAW                   = 128,
37 };
38
39 /*
40  * Generalized performance counter event types, used by the hw_event.event_id
41  * parameter of the sys_perf_counter_open() syscall:
42  */
43 enum hw_event_ids {
44         /*
45          * Common hardware events, generalized by the kernel:
46          */
47         PERF_COUNT_CPU_CYCLES           = 0,
48         PERF_COUNT_INSTRUCTIONS         = 1,
49         PERF_COUNT_CACHE_REFERENCES     = 2,
50         PERF_COUNT_CACHE_MISSES         = 3,
51         PERF_COUNT_BRANCH_INSTRUCTIONS  = 4,
52         PERF_COUNT_BRANCH_MISSES        = 5,
53         PERF_COUNT_BUS_CYCLES           = 6,
54
55         PERF_HW_EVENTS_MAX              = 7,
56 };
57
58 /*
59  * Special "software" counters provided by the kernel, even if the hardware
60  * does not support performance counters. These counters measure various
61  * physical and sw events of the kernel (and allow the profiling of them as
62  * well):
63  */
64 enum sw_event_ids {
65         PERF_COUNT_CPU_CLOCK            = 0,
66         PERF_COUNT_TASK_CLOCK           = 1,
67         PERF_COUNT_PAGE_FAULTS          = 2,
68         PERF_COUNT_CONTEXT_SWITCHES     = 3,
69         PERF_COUNT_CPU_MIGRATIONS       = 4,
70         PERF_COUNT_PAGE_FAULTS_MIN      = 5,
71         PERF_COUNT_PAGE_FAULTS_MAJ      = 6,
72
73         PERF_SW_EVENTS_MAX              = 7,
74 };
75
76 #define __PERF_COUNTER_MASK(name)                       \
77         (((1ULL << PERF_COUNTER_##name##_BITS) - 1) <<  \
78          PERF_COUNTER_##name##_SHIFT)
79
80 #define PERF_COUNTER_RAW_BITS           1
81 #define PERF_COUNTER_RAW_SHIFT          63
82 #define PERF_COUNTER_RAW_MASK           __PERF_COUNTER_MASK(RAW)
83
84 #define PERF_COUNTER_CONFIG_BITS        63
85 #define PERF_COUNTER_CONFIG_SHIFT       0
86 #define PERF_COUNTER_CONFIG_MASK        __PERF_COUNTER_MASK(CONFIG)
87
88 #define PERF_COUNTER_TYPE_BITS          7
89 #define PERF_COUNTER_TYPE_SHIFT         56
90 #define PERF_COUNTER_TYPE_MASK          __PERF_COUNTER_MASK(TYPE)
91
92 #define PERF_COUNTER_EVENT_BITS         56
93 #define PERF_COUNTER_EVENT_SHIFT        0
94 #define PERF_COUNTER_EVENT_MASK         __PERF_COUNTER_MASK(EVENT)
95
96 /*
97  * Bits that can be set in hw_event.record_type to request information
98  * in the overflow packets.
99  */
100 enum perf_counter_record_format {
101         PERF_RECORD_IP          = 1U << 0,
102         PERF_RECORD_TID         = 1U << 1,
103         PERF_RECORD_TIME        = 1U << 2,
104         PERF_RECORD_ADDR        = 1U << 3,
105         PERF_RECORD_GROUP       = 1U << 4,
106         PERF_RECORD_CALLCHAIN   = 1U << 5,
107 };
108
109 /*
110  * Bits that can be set in hw_event.read_format to request that
111  * reads on the counter should return the indicated quantities,
112  * in increasing order of bit value, after the counter value.
113  */
114 enum perf_counter_read_format {
115         PERF_FORMAT_TOTAL_TIME_ENABLED  =  1,
116         PERF_FORMAT_TOTAL_TIME_RUNNING  =  2,
117 };
118
119 /*
120  * Hardware event to monitor via a performance monitoring counter:
121  */
122 struct perf_counter_hw_event {
123         /*
124          * The MSB of the config word signifies if the rest contains cpu
125          * specific (raw) counter configuration data, if unset, the next
126          * 7 bits are an event type and the rest of the bits are the event
127          * identifier.
128          */
129         __u64                   config;
130
131         __u64                   irq_period;
132         __u32                   record_type;
133         __u32                   read_format;
134
135         __u64                   disabled       :  1, /* off by default        */
136                                 nmi            :  1, /* NMI sampling          */
137                                 inherit        :  1, /* children inherit it   */
138                                 pinned         :  1, /* must always be on PMU */
139                                 exclusive      :  1, /* only group on PMU     */
140                                 exclude_user   :  1, /* don't count user      */
141                                 exclude_kernel :  1, /* ditto kernel          */
142                                 exclude_hv     :  1, /* ditto hypervisor      */
143                                 exclude_idle   :  1, /* don't count when idle */
144                                 mmap           :  1, /* include mmap data     */
145                                 munmap         :  1, /* include munmap data   */
146                                 comm           :  1, /* include comm data     */
147
148                                 __reserved_1   : 52;
149
150         __u32                   extra_config_len;
151         __u32                   wakeup_events;  /* wakeup every n events */
152
153         __u64                   __reserved_2;
154         __u64                   __reserved_3;
155 };
156
157 /*
158  * Ioctls that can be done on a perf counter fd:
159  */
160 #define PERF_COUNTER_IOC_ENABLE         _IO ('$', 0)
161 #define PERF_COUNTER_IOC_DISABLE        _IO ('$', 1)
162 #define PERF_COUNTER_IOC_REFRESH        _IOW('$', 2, u32)
163 #define PERF_COUNTER_IOC_RESET          _IO ('$', 3)
164
165 /*
166  * Structure of the page that can be mapped via mmap
167  */
168 struct perf_counter_mmap_page {
169         __u32   version;                /* version number of this structure */
170         __u32   compat_version;         /* lowest version this is compat with */
171
172         /*
173          * Bits needed to read the hw counters in user-space.
174          *
175          *   u32 seq;
176          *   s64 count;
177          *
178          *   do {
179          *     seq = pc->lock;
180          *
181          *     barrier()
182          *     if (pc->index) {
183          *       count = pmc_read(pc->index - 1);
184          *       count += pc->offset;
185          *     } else
186          *       goto regular_read;
187          *
188          *     barrier();
189          *   } while (pc->lock != seq);
190          *
191          * NOTE: for obvious reason this only works on self-monitoring
192          *       processes.
193          */
194         __u32   lock;                   /* seqlock for synchronization */
195         __u32   index;                  /* hardware counter identifier */
196         __s64   offset;                 /* add to hardware counter value */
197
198         /*
199          * Control data for the mmap() data buffer.
200          *
201          * User-space reading this value should issue an rmb(), on SMP capable
202          * platforms, after reading this value -- see perf_counter_wakeup().
203          */
204         __u32   data_head;              /* head in the data section */
205 };
206
207 #define PERF_EVENT_MISC_KERNEL          (1 << 0)
208 #define PERF_EVENT_MISC_USER            (1 << 1)
209 #define PERF_EVENT_MISC_OVERFLOW        (1 << 2)
210
211 struct perf_event_header {
212         __u32   type;
213         __u16   misc;
214         __u16   size;
215 };
216
217 enum perf_event_type {
218
219         /*
220          * The MMAP events record the PROT_EXEC mappings so that we can
221          * correlate userspace IPs to code. They have the following structure:
222          *
223          * struct {
224          *      struct perf_event_header        header;
225          *
226          *      u32                             pid, tid;
227          *      u64                             addr;
228          *      u64                             len;
229          *      u64                             pgoff;
230          *      char                            filename[];
231          * };
232          */
233         PERF_EVENT_MMAP                 = 1,
234         PERF_EVENT_MUNMAP               = 2,
235
236         /*
237          * struct {
238          *      struct perf_event_header        header;
239          *
240          *      u32                             pid, tid;
241          *      char                            comm[];
242          * };
243          */
244         PERF_EVENT_COMM                 = 3,
245
246         /*
247          * When header.misc & PERF_EVENT_MISC_OVERFLOW the event_type field
248          * will be PERF_RECORD_*
249          *
250          * struct {
251          *      struct perf_event_header        header;
252          *
253          *      { u64                   ip;       } && PERF_RECORD_IP
254          *      { u32                   pid, tid; } && PERF_RECORD_TID
255          *      { u64                   time;     } && PERF_RECORD_TIME
256          *      { u64                   addr;     } && PERF_RECORD_ADDR
257          *
258          *      { u64                   nr;
259          *        { u64 event, val; }   cnt[nr];  } && PERF_RECORD_GROUP
260          *
261          *      { u16                   nr,
262          *                              hv,
263          *                              kernel,
264          *                              user;
265          *        u64                   ips[nr];  } && PERF_RECORD_CALLCHAIN
266          * };
267          */
268 };
269
270 #ifdef __KERNEL__
271 /*
272  * Kernel-internal data types and definitions:
273  */
274
275 #ifdef CONFIG_PERF_COUNTERS
276 # include <asm/perf_counter.h>
277 #endif
278
279 #include <linux/list.h>
280 #include <linux/mutex.h>
281 #include <linux/rculist.h>
282 #include <linux/rcupdate.h>
283 #include <linux/spinlock.h>
284 #include <linux/hrtimer.h>
285 #include <linux/fs.h>
286 #include <asm/atomic.h>
287
288 struct task_struct;
289
290 static inline u64 perf_event_raw(struct perf_counter_hw_event *hw_event)
291 {
292         return hw_event->config & PERF_COUNTER_RAW_MASK;
293 }
294
295 static inline u64 perf_event_config(struct perf_counter_hw_event *hw_event)
296 {
297         return hw_event->config & PERF_COUNTER_CONFIG_MASK;
298 }
299
300 static inline u64 perf_event_type(struct perf_counter_hw_event *hw_event)
301 {
302         return (hw_event->config & PERF_COUNTER_TYPE_MASK) >>
303                 PERF_COUNTER_TYPE_SHIFT;
304 }
305
306 static inline u64 perf_event_id(struct perf_counter_hw_event *hw_event)
307 {
308         return hw_event->config & PERF_COUNTER_EVENT_MASK;
309 }
310
311 /**
312  * struct hw_perf_counter - performance counter hardware details:
313  */
314 struct hw_perf_counter {
315 #ifdef CONFIG_PERF_COUNTERS
316         union {
317                 struct { /* hardware */
318                         u64                             config;
319                         unsigned long                   config_base;
320                         unsigned long                   counter_base;
321                         int                             nmi;
322                         int                             idx;
323                 };
324                 union { /* software */
325                         atomic64_t                      count;
326                         struct hrtimer                  hrtimer;
327                 };
328         };
329         atomic64_t                      prev_count;
330         u64                             irq_period;
331         atomic64_t                      period_left;
332 #endif
333 };
334
335 struct perf_counter;
336
337 /**
338  * struct pmu - generic performance monitoring unit
339  */
340 struct pmu {
341         int (*enable)                   (struct perf_counter *counter);
342         void (*disable)                 (struct perf_counter *counter);
343         void (*read)                    (struct perf_counter *counter);
344 };
345
346 /**
347  * enum perf_counter_active_state - the states of a counter
348  */
349 enum perf_counter_active_state {
350         PERF_COUNTER_STATE_ERROR        = -2,
351         PERF_COUNTER_STATE_OFF          = -1,
352         PERF_COUNTER_STATE_INACTIVE     =  0,
353         PERF_COUNTER_STATE_ACTIVE       =  1,
354 };
355
356 struct file;
357
358 struct perf_mmap_data {
359         struct rcu_head                 rcu_head;
360         int                             nr_pages;       /* nr of data pages  */
361
362         atomic_t                        poll;           /* POLL_ for wakeups */
363         atomic_t                        head;           /* write position    */
364         atomic_t                        events;         /* event limit       */
365
366         atomic_t                        done_head;      /* completed head    */
367         atomic_t                        lock;           /* concurrent writes */
368
369         atomic_t                        wakeup;         /* needs a wakeup    */
370
371         struct perf_counter_mmap_page   *user_page;
372         void                            *data_pages[0];
373 };
374
375 struct perf_pending_entry {
376         struct perf_pending_entry *next;
377         void (*func)(struct perf_pending_entry *);
378 };
379
380 /**
381  * struct perf_counter - performance counter kernel representation:
382  */
383 struct perf_counter {
384 #ifdef CONFIG_PERF_COUNTERS
385         struct list_head                list_entry;
386         struct list_head                event_entry;
387         struct list_head                sibling_list;
388         int                             nr_siblings;
389         struct perf_counter             *group_leader;
390         const struct pmu                *pmu;
391
392         enum perf_counter_active_state  state;
393         enum perf_counter_active_state  prev_state;
394         atomic64_t                      count;
395
396         /*
397          * These are the total time in nanoseconds that the counter
398          * has been enabled (i.e. eligible to run, and the task has
399          * been scheduled in, if this is a per-task counter)
400          * and running (scheduled onto the CPU), respectively.
401          *
402          * They are computed from tstamp_enabled, tstamp_running and
403          * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
404          */
405         u64                             total_time_enabled;
406         u64                             total_time_running;
407
408         /*
409          * These are timestamps used for computing total_time_enabled
410          * and total_time_running when the counter is in INACTIVE or
411          * ACTIVE state, measured in nanoseconds from an arbitrary point
412          * in time.
413          * tstamp_enabled: the notional time when the counter was enabled
414          * tstamp_running: the notional time when the counter was scheduled on
415          * tstamp_stopped: in INACTIVE state, the notional time when the
416          *      counter was scheduled off.
417          */
418         u64                             tstamp_enabled;
419         u64                             tstamp_running;
420         u64                             tstamp_stopped;
421
422         struct perf_counter_hw_event    hw_event;
423         struct hw_perf_counter          hw;
424
425         struct perf_counter_context     *ctx;
426         struct task_struct              *task;
427         struct file                     *filp;
428
429         struct perf_counter             *parent;
430         struct list_head                child_list;
431
432         /*
433          * These accumulate total time (in nanoseconds) that children
434          * counters have been enabled and running, respectively.
435          */
436         atomic64_t                      child_total_time_enabled;
437         atomic64_t                      child_total_time_running;
438
439         /*
440          * Protect attach/detach and child_list:
441          */
442         struct mutex                    mutex;
443
444         int                             oncpu;
445         int                             cpu;
446
447         /* mmap bits */
448         struct mutex                    mmap_mutex;
449         atomic_t                        mmap_count;
450         struct perf_mmap_data           *data;
451
452         /* poll related */
453         wait_queue_head_t               waitq;
454         struct fasync_struct            *fasync;
455
456         /* delayed work for NMIs and such */
457         int                             pending_wakeup;
458         int                             pending_kill;
459         int                             pending_disable;
460         struct perf_pending_entry       pending;
461
462         atomic_t                        event_limit;
463
464         void (*destroy)(struct perf_counter *);
465         struct rcu_head                 rcu_head;
466 #endif
467 };
468
469 /**
470  * struct perf_counter_context - counter context structure
471  *
472  * Used as a container for task counters and CPU counters as well:
473  */
474 struct perf_counter_context {
475 #ifdef CONFIG_PERF_COUNTERS
476         /*
477          * Protect the states of the counters in the list,
478          * nr_active, and the list:
479          */
480         spinlock_t              lock;
481         /*
482          * Protect the list of counters.  Locking either mutex or lock
483          * is sufficient to ensure the list doesn't change; to change
484          * the list you need to lock both the mutex and the spinlock.
485          */
486         struct mutex            mutex;
487
488         struct list_head        counter_list;
489         struct list_head        event_list;
490         int                     nr_counters;
491         int                     nr_active;
492         int                     is_active;
493         struct task_struct      *task;
494
495         /*
496          * Context clock, runs when context enabled.
497          */
498         u64                     time;
499         u64                     timestamp;
500 #endif
501 };
502
503 /**
504  * struct perf_counter_cpu_context - per cpu counter context structure
505  */
506 struct perf_cpu_context {
507         struct perf_counter_context     ctx;
508         struct perf_counter_context     *task_ctx;
509         int                             active_oncpu;
510         int                             max_pertask;
511         int                             exclusive;
512
513         /*
514          * Recursion avoidance:
515          *
516          * task, softirq, irq, nmi context
517          */
518         int                     recursion[4];
519 };
520
521 #ifdef CONFIG_PERF_COUNTERS
522
523 /*
524  * Set by architecture code:
525  */
526 extern int perf_max_counters;
527
528 extern const struct pmu *hw_perf_counter_init(struct perf_counter *counter);
529
530 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
531 extern void perf_counter_task_sched_out(struct task_struct *task, int cpu);
532 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
533 extern void perf_counter_init_task(struct task_struct *child);
534 extern void perf_counter_exit_task(struct task_struct *child);
535 extern void perf_counter_do_pending(void);
536 extern void perf_counter_print_debug(void);
537 extern void perf_counter_unthrottle(void);
538 extern u64 hw_perf_save_disable(void);
539 extern void hw_perf_restore(u64 ctrl);
540 extern int perf_counter_task_disable(void);
541 extern int perf_counter_task_enable(void);
542 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
543                struct perf_cpu_context *cpuctx,
544                struct perf_counter_context *ctx, int cpu);
545 extern void perf_counter_update_userpage(struct perf_counter *counter);
546
547 extern int perf_counter_overflow(struct perf_counter *counter,
548                                  int nmi, struct pt_regs *regs, u64 addr);
549 /*
550  * Return 1 for a software counter, 0 for a hardware counter
551  */
552 static inline int is_software_counter(struct perf_counter *counter)
553 {
554         return !perf_event_raw(&counter->hw_event) &&
555                 perf_event_type(&counter->hw_event) != PERF_TYPE_HARDWARE;
556 }
557
558 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *, u64);
559
560 extern void perf_counter_mmap(unsigned long addr, unsigned long len,
561                               unsigned long pgoff, struct file *file);
562
563 extern void perf_counter_munmap(unsigned long addr, unsigned long len,
564                                 unsigned long pgoff, struct file *file);
565
566 extern void perf_counter_comm(struct task_struct *tsk);
567
568 #define MAX_STACK_DEPTH         255
569
570 struct perf_callchain_entry {
571         u16     nr, hv, kernel, user;
572         u64     ip[MAX_STACK_DEPTH];
573 };
574
575 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
576
577 extern int sysctl_perf_counter_priv;
578
579 extern void perf_counter_init(void);
580
581 #else
582 static inline void
583 perf_counter_task_sched_in(struct task_struct *task, int cpu)           { }
584 static inline void
585 perf_counter_task_sched_out(struct task_struct *task, int cpu)          { }
586 static inline void
587 perf_counter_task_tick(struct task_struct *task, int cpu)               { }
588 static inline void perf_counter_init_task(struct task_struct *child)    { }
589 static inline void perf_counter_exit_task(struct task_struct *child)    { }
590 static inline void perf_counter_do_pending(void)                        { }
591 static inline void perf_counter_print_debug(void)                       { }
592 static inline void perf_counter_unthrottle(void)                        { }
593 static inline void hw_perf_restore(u64 ctrl)                            { }
594 static inline u64 hw_perf_save_disable(void)                  { return 0; }
595 static inline int perf_counter_task_disable(void)       { return -EINVAL; }
596 static inline int perf_counter_task_enable(void)        { return -EINVAL; }
597
598 static inline void
599 perf_swcounter_event(u32 event, u64 nr, int nmi,
600                      struct pt_regs *regs, u64 addr)                    { }
601
602 static inline void
603 perf_counter_mmap(unsigned long addr, unsigned long len,
604                   unsigned long pgoff, struct file *file)               { }
605
606 static inline void
607 perf_counter_munmap(unsigned long addr, unsigned long len,
608                     unsigned long pgoff, struct file *file)             { }
609
610 static inline void perf_counter_comm(struct task_struct *tsk)           { }
611 static inline void perf_counter_init(void)                              { }
612 #endif
613
614 #endif /* __KERNEL__ */
615 #endif /* _LINUX_PERF_COUNTER_H */