perf_counter: Add ioctl for changing the sample period/frequency
[safe/jmp/linux-2.6] / 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.sample_type to request information
98  * in the overflow packets.
99  */
100 enum perf_counter_sample_format {
101         PERF_SAMPLE_IP                  = 1U << 0,
102         PERF_SAMPLE_TID                 = 1U << 1,
103         PERF_SAMPLE_TIME                = 1U << 2,
104         PERF_SAMPLE_ADDR                = 1U << 3,
105         PERF_SAMPLE_GROUP               = 1U << 4,
106         PERF_SAMPLE_CALLCHAIN           = 1U << 5,
107         PERF_SAMPLE_CONFIG              = 1U << 6,
108         PERF_SAMPLE_CPU                 = 1U << 7,
109 };
110
111 /*
112  * Bits that can be set in hw_event.read_format to request that
113  * reads on the counter should return the indicated quantities,
114  * in increasing order of bit value, after the counter value.
115  */
116 enum perf_counter_read_format {
117         PERF_FORMAT_TOTAL_TIME_ENABLED  =  1U << 0,
118         PERF_FORMAT_TOTAL_TIME_RUNNING  =  1U << 1,
119         PERF_FORMAT_ID                  =  1U << 2,
120 };
121
122 /*
123  * Hardware event to monitor via a performance monitoring counter:
124  */
125 struct perf_counter_hw_event {
126         /*
127          * The MSB of the config word signifies if the rest contains cpu
128          * specific (raw) counter configuration data, if unset, the next
129          * 7 bits are an event type and the rest of the bits are the event
130          * identifier.
131          */
132         __u64                   config;
133
134         union {
135                 __u64           sample_period;
136                 __u64           sample_freq;
137         };
138
139         __u64                   sample_type;
140         __u64                   read_format;
141
142         __u64                   disabled       :  1, /* off by default        */
143                                 inherit        :  1, /* children inherit it   */
144                                 pinned         :  1, /* must always be on PMU */
145                                 exclusive      :  1, /* only group on PMU     */
146                                 exclude_user   :  1, /* don't count user      */
147                                 exclude_kernel :  1, /* ditto kernel          */
148                                 exclude_hv     :  1, /* ditto hypervisor      */
149                                 exclude_idle   :  1, /* don't count when idle */
150                                 mmap           :  1, /* include mmap data     */
151                                 munmap         :  1, /* include munmap data   */
152                                 comm           :  1, /* include comm data     */
153                                 freq           :  1, /* use freq, not period  */
154
155                                 __reserved_1   : 52;
156
157         __u32                   wakeup_events;  /* wakeup every n events */
158         __u32                   __reserved_2;
159
160         __u64                   __reserved_3;
161         __u64                   __reserved_4;
162 };
163
164 /*
165  * Ioctls that can be done on a perf counter fd:
166  */
167 #define PERF_COUNTER_IOC_ENABLE         _IO ('$', 0)
168 #define PERF_COUNTER_IOC_DISABLE        _IO ('$', 1)
169 #define PERF_COUNTER_IOC_REFRESH        _IO ('$', 2)
170 #define PERF_COUNTER_IOC_RESET          _IO ('$', 3)
171 #define PERF_COUNTER_IOC_PERIOD         _IOW('$', 4, u64)
172
173 enum perf_counter_ioc_flags {
174         PERF_IOC_FLAG_GROUP             = 1U << 0,
175 };
176
177 /*
178  * Structure of the page that can be mapped via mmap
179  */
180 struct perf_counter_mmap_page {
181         __u32   version;                /* version number of this structure */
182         __u32   compat_version;         /* lowest version this is compat with */
183
184         /*
185          * Bits needed to read the hw counters in user-space.
186          *
187          *   u32 seq;
188          *   s64 count;
189          *
190          *   do {
191          *     seq = pc->lock;
192          *
193          *     barrier()
194          *     if (pc->index) {
195          *       count = pmc_read(pc->index - 1);
196          *       count += pc->offset;
197          *     } else
198          *       goto regular_read;
199          *
200          *     barrier();
201          *   } while (pc->lock != seq);
202          *
203          * NOTE: for obvious reason this only works on self-monitoring
204          *       processes.
205          */
206         __u32   lock;                   /* seqlock for synchronization */
207         __u32   index;                  /* hardware counter identifier */
208         __s64   offset;                 /* add to hardware counter value */
209
210         /*
211          * Control data for the mmap() data buffer.
212          *
213          * User-space reading this value should issue an rmb(), on SMP capable
214          * platforms, after reading this value -- see perf_counter_wakeup().
215          */
216         __u64   data_head;              /* head in the data section */
217 };
218
219 #define PERF_EVENT_MISC_CPUMODE_MASK    (3 << 0)
220 #define PERF_EVENT_MISC_CPUMODE_UNKNOWN (0 << 0)
221 #define PERF_EVENT_MISC_KERNEL          (1 << 0)
222 #define PERF_EVENT_MISC_USER            (2 << 0)
223 #define PERF_EVENT_MISC_HYPERVISOR      (3 << 0)
224 #define PERF_EVENT_MISC_OVERFLOW        (1 << 2)
225
226 struct perf_event_header {
227         __u32   type;
228         __u16   misc;
229         __u16   size;
230 };
231
232 enum perf_event_type {
233
234         /*
235          * The MMAP events record the PROT_EXEC mappings so that we can
236          * correlate userspace IPs to code. They have the following structure:
237          *
238          * struct {
239          *      struct perf_event_header        header;
240          *
241          *      u32                             pid, tid;
242          *      u64                             addr;
243          *      u64                             len;
244          *      u64                             pgoff;
245          *      char                            filename[];
246          * };
247          */
248         PERF_EVENT_MMAP                 = 1,
249         PERF_EVENT_MUNMAP               = 2,
250
251         /*
252          * struct {
253          *      struct perf_event_header        header;
254          *
255          *      u32                             pid, tid;
256          *      char                            comm[];
257          * };
258          */
259         PERF_EVENT_COMM                 = 3,
260
261         /*
262          * struct {
263          *      struct perf_event_header        header;
264          *      u64                             time;
265          *      u64                             sample_period;
266          * };
267          */
268         PERF_EVENT_PERIOD               = 4,
269
270         /*
271          * struct {
272          *      struct perf_event_header        header;
273          *      u64                             time;
274          * };
275          */
276         PERF_EVENT_THROTTLE             = 5,
277         PERF_EVENT_UNTHROTTLE           = 6,
278
279         /*
280          * When header.misc & PERF_EVENT_MISC_OVERFLOW the event_type field
281          * will be PERF_RECORD_*
282          *
283          * struct {
284          *      struct perf_event_header        header;
285          *
286          *      { u64                   ip;       } && PERF_RECORD_IP
287          *      { u32                   pid, tid; } && PERF_RECORD_TID
288          *      { u64                   time;     } && PERF_RECORD_TIME
289          *      { u64                   addr;     } && PERF_RECORD_ADDR
290          *      { u64                   config;   } && PERF_RECORD_CONFIG
291          *      { u32                   cpu, res; } && PERF_RECORD_CPU
292          *
293          *      { u64                   nr;
294          *        { u64 id, val; }      cnt[nr];  } && PERF_RECORD_GROUP
295          *
296          *      { u16                   nr,
297          *                              hv,
298          *                              kernel,
299          *                              user;
300          *        u64                   ips[nr];  } && PERF_RECORD_CALLCHAIN
301          * };
302          */
303 };
304
305 #ifdef __KERNEL__
306 /*
307  * Kernel-internal data types and definitions:
308  */
309
310 #ifdef CONFIG_PERF_COUNTERS
311 # include <asm/perf_counter.h>
312 #endif
313
314 #include <linux/list.h>
315 #include <linux/mutex.h>
316 #include <linux/rculist.h>
317 #include <linux/rcupdate.h>
318 #include <linux/spinlock.h>
319 #include <linux/hrtimer.h>
320 #include <linux/fs.h>
321 #include <linux/pid_namespace.h>
322 #include <asm/atomic.h>
323
324 struct task_struct;
325
326 static inline u64 perf_event_raw(struct perf_counter_hw_event *hw_event)
327 {
328         return hw_event->config & PERF_COUNTER_RAW_MASK;
329 }
330
331 static inline u64 perf_event_config(struct perf_counter_hw_event *hw_event)
332 {
333         return hw_event->config & PERF_COUNTER_CONFIG_MASK;
334 }
335
336 static inline u64 perf_event_type(struct perf_counter_hw_event *hw_event)
337 {
338         return (hw_event->config & PERF_COUNTER_TYPE_MASK) >>
339                 PERF_COUNTER_TYPE_SHIFT;
340 }
341
342 static inline u64 perf_event_id(struct perf_counter_hw_event *hw_event)
343 {
344         return hw_event->config & PERF_COUNTER_EVENT_MASK;
345 }
346
347 /**
348  * struct hw_perf_counter - performance counter hardware details:
349  */
350 struct hw_perf_counter {
351 #ifdef CONFIG_PERF_COUNTERS
352         union {
353                 struct { /* hardware */
354                         u64                             config;
355                         unsigned long                   config_base;
356                         unsigned long                   counter_base;
357                         int                             idx;
358                 };
359                 union { /* software */
360                         atomic64_t                      count;
361                         struct hrtimer                  hrtimer;
362                 };
363         };
364         atomic64_t                      prev_count;
365         u64                             sample_period;
366         atomic64_t                      period_left;
367         u64                             interrupts;
368 #endif
369 };
370
371 struct perf_counter;
372
373 /**
374  * struct pmu - generic performance monitoring unit
375  */
376 struct pmu {
377         int (*enable)                   (struct perf_counter *counter);
378         void (*disable)                 (struct perf_counter *counter);
379         void (*read)                    (struct perf_counter *counter);
380         void (*unthrottle)              (struct perf_counter *counter);
381 };
382
383 /**
384  * enum perf_counter_active_state - the states of a counter
385  */
386 enum perf_counter_active_state {
387         PERF_COUNTER_STATE_ERROR        = -2,
388         PERF_COUNTER_STATE_OFF          = -1,
389         PERF_COUNTER_STATE_INACTIVE     =  0,
390         PERF_COUNTER_STATE_ACTIVE       =  1,
391 };
392
393 struct file;
394
395 struct perf_mmap_data {
396         struct rcu_head                 rcu_head;
397         int                             nr_pages;       /* nr of data pages  */
398         int                             nr_locked;      /* nr pages mlocked  */
399
400         atomic_t                        poll;           /* POLL_ for wakeups */
401         atomic_t                        events;         /* event limit       */
402
403         atomic_long_t                   head;           /* write position    */
404         atomic_long_t                   done_head;      /* completed head    */
405
406         atomic_t                        lock;           /* concurrent writes */
407
408         atomic_t                        wakeup;         /* needs a wakeup    */
409
410         struct perf_counter_mmap_page   *user_page;
411         void                            *data_pages[0];
412 };
413
414 struct perf_pending_entry {
415         struct perf_pending_entry *next;
416         void (*func)(struct perf_pending_entry *);
417 };
418
419 /**
420  * struct perf_counter - performance counter kernel representation:
421  */
422 struct perf_counter {
423 #ifdef CONFIG_PERF_COUNTERS
424         struct list_head                list_entry;
425         struct list_head                event_entry;
426         struct list_head                sibling_list;
427         int                             nr_siblings;
428         struct perf_counter             *group_leader;
429         const struct pmu                *pmu;
430
431         enum perf_counter_active_state  state;
432         atomic64_t                      count;
433
434         /*
435          * These are the total time in nanoseconds that the counter
436          * has been enabled (i.e. eligible to run, and the task has
437          * been scheduled in, if this is a per-task counter)
438          * and running (scheduled onto the CPU), respectively.
439          *
440          * They are computed from tstamp_enabled, tstamp_running and
441          * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
442          */
443         u64                             total_time_enabled;
444         u64                             total_time_running;
445
446         /*
447          * These are timestamps used for computing total_time_enabled
448          * and total_time_running when the counter is in INACTIVE or
449          * ACTIVE state, measured in nanoseconds from an arbitrary point
450          * in time.
451          * tstamp_enabled: the notional time when the counter was enabled
452          * tstamp_running: the notional time when the counter was scheduled on
453          * tstamp_stopped: in INACTIVE state, the notional time when the
454          *      counter was scheduled off.
455          */
456         u64                             tstamp_enabled;
457         u64                             tstamp_running;
458         u64                             tstamp_stopped;
459
460         struct perf_counter_hw_event    hw_event;
461         struct hw_perf_counter          hw;
462
463         struct perf_counter_context     *ctx;
464         struct file                     *filp;
465
466         /*
467          * These accumulate total time (in nanoseconds) that children
468          * counters have been enabled and running, respectively.
469          */
470         atomic64_t                      child_total_time_enabled;
471         atomic64_t                      child_total_time_running;
472
473         /*
474          * Protect attach/detach and child_list:
475          */
476         struct mutex                    child_mutex;
477         struct list_head                child_list;
478         struct perf_counter             *parent;
479
480         int                             oncpu;
481         int                             cpu;
482
483         struct list_head                owner_entry;
484         struct task_struct              *owner;
485
486         /* mmap bits */
487         struct mutex                    mmap_mutex;
488         atomic_t                        mmap_count;
489         struct perf_mmap_data           *data;
490
491         /* poll related */
492         wait_queue_head_t               waitq;
493         struct fasync_struct            *fasync;
494
495         /* delayed work for NMIs and such */
496         int                             pending_wakeup;
497         int                             pending_kill;
498         int                             pending_disable;
499         struct perf_pending_entry       pending;
500
501         atomic_t                        event_limit;
502
503         void (*destroy)(struct perf_counter *);
504         struct rcu_head                 rcu_head;
505
506         struct pid_namespace            *ns;
507         u64                             id;
508 #endif
509 };
510
511 /**
512  * struct perf_counter_context - counter context structure
513  *
514  * Used as a container for task counters and CPU counters as well:
515  */
516 struct perf_counter_context {
517         /*
518          * Protect the states of the counters in the list,
519          * nr_active, and the list:
520          */
521         spinlock_t              lock;
522         /*
523          * Protect the list of counters.  Locking either mutex or lock
524          * is sufficient to ensure the list doesn't change; to change
525          * the list you need to lock both the mutex and the spinlock.
526          */
527         struct mutex            mutex;
528
529         struct list_head        counter_list;
530         struct list_head        event_list;
531         int                     nr_counters;
532         int                     nr_active;
533         int                     is_active;
534         atomic_t                refcount;
535         struct task_struct      *task;
536
537         /*
538          * Context clock, runs when context enabled.
539          */
540         u64                     time;
541         u64                     timestamp;
542
543         /*
544          * These fields let us detect when two contexts have both
545          * been cloned (inherited) from a common ancestor.
546          */
547         struct perf_counter_context *parent_ctx;
548         u64                     parent_gen;
549         u64                     generation;
550         int                     pin_count;
551         struct rcu_head         rcu_head;
552 };
553
554 /**
555  * struct perf_counter_cpu_context - per cpu counter context structure
556  */
557 struct perf_cpu_context {
558         struct perf_counter_context     ctx;
559         struct perf_counter_context     *task_ctx;
560         int                             active_oncpu;
561         int                             max_pertask;
562         int                             exclusive;
563
564         /*
565          * Recursion avoidance:
566          *
567          * task, softirq, irq, nmi context
568          */
569         int                             recursion[4];
570 };
571
572 #ifdef CONFIG_PERF_COUNTERS
573
574 /*
575  * Set by architecture code:
576  */
577 extern int perf_max_counters;
578
579 extern const struct pmu *hw_perf_counter_init(struct perf_counter *counter);
580
581 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
582 extern void perf_counter_task_sched_out(struct task_struct *task,
583                                         struct task_struct *next, int cpu);
584 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
585 extern int perf_counter_init_task(struct task_struct *child);
586 extern void perf_counter_exit_task(struct task_struct *child);
587 extern void perf_counter_free_task(struct task_struct *task);
588 extern void perf_counter_do_pending(void);
589 extern void perf_counter_print_debug(void);
590 extern void __perf_disable(void);
591 extern bool __perf_enable(void);
592 extern void perf_disable(void);
593 extern void perf_enable(void);
594 extern int perf_counter_task_disable(void);
595 extern int perf_counter_task_enable(void);
596 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
597                struct perf_cpu_context *cpuctx,
598                struct perf_counter_context *ctx, int cpu);
599 extern void perf_counter_update_userpage(struct perf_counter *counter);
600
601 extern int perf_counter_overflow(struct perf_counter *counter,
602                                  int nmi, struct pt_regs *regs, u64 addr);
603 /*
604  * Return 1 for a software counter, 0 for a hardware counter
605  */
606 static inline int is_software_counter(struct perf_counter *counter)
607 {
608         return !perf_event_raw(&counter->hw_event) &&
609                 perf_event_type(&counter->hw_event) != PERF_TYPE_HARDWARE;
610 }
611
612 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *, u64);
613
614 extern void perf_counter_mmap(unsigned long addr, unsigned long len,
615                               unsigned long pgoff, struct file *file);
616
617 extern void perf_counter_munmap(unsigned long addr, unsigned long len,
618                                 unsigned long pgoff, struct file *file);
619
620 extern void perf_counter_comm(struct task_struct *tsk);
621
622 extern void perf_counter_task_migration(struct task_struct *task, int cpu);
623
624 #define MAX_STACK_DEPTH         255
625
626 struct perf_callchain_entry {
627         u16     nr, hv, kernel, user;
628         u64     ip[MAX_STACK_DEPTH];
629 };
630
631 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
632
633 extern int sysctl_perf_counter_priv;
634 extern int sysctl_perf_counter_mlock;
635 extern int sysctl_perf_counter_limit;
636
637 extern void perf_counter_init(void);
638
639 #ifndef perf_misc_flags
640 #define perf_misc_flags(regs)   (user_mode(regs) ? PERF_EVENT_MISC_USER : \
641                                  PERF_EVENT_MISC_KERNEL)
642 #define perf_instruction_pointer(regs)  instruction_pointer(regs)
643 #endif
644
645 #else
646 static inline void
647 perf_counter_task_sched_in(struct task_struct *task, int cpu)           { }
648 static inline void
649 perf_counter_task_sched_out(struct task_struct *task,
650                             struct task_struct *next, int cpu)          { }
651 static inline void
652 perf_counter_task_tick(struct task_struct *task, int cpu)               { }
653 static inline int perf_counter_init_task(struct task_struct *child)     { return 0; }
654 static inline void perf_counter_exit_task(struct task_struct *child)    { }
655 static inline void perf_counter_free_task(struct task_struct *task)     { }
656 static inline void perf_counter_do_pending(void)                        { }
657 static inline void perf_counter_print_debug(void)                       { }
658 static inline void perf_disable(void)                                   { }
659 static inline void perf_enable(void)                                    { }
660 static inline int perf_counter_task_disable(void)       { return -EINVAL; }
661 static inline int perf_counter_task_enable(void)        { return -EINVAL; }
662
663 static inline void
664 perf_swcounter_event(u32 event, u64 nr, int nmi,
665                      struct pt_regs *regs, u64 addr)                    { }
666
667 static inline void
668 perf_counter_mmap(unsigned long addr, unsigned long len,
669                   unsigned long pgoff, struct file *file)               { }
670
671 static inline void
672 perf_counter_munmap(unsigned long addr, unsigned long len,
673                     unsigned long pgoff, struct file *file)             { }
674
675 static inline void perf_counter_comm(struct task_struct *tsk)           { }
676 static inline void perf_counter_init(void)                              { }
677 static inline void perf_counter_task_migration(struct task_struct *task,
678                                                int cpu)                 { }
679 #endif
680
681 #endif /* __KERNEL__ */
682 #endif /* _LINUX_PERF_COUNTER_H */