perf_counter: Dynamically allocate tasks' perf_counter_context struct
[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         PERF_RECORD_CONFIG      = 1U << 6,
108         PERF_RECORD_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  =  1,
118         PERF_FORMAT_TOTAL_TIME_RUNNING  =  2,
119 };
120
121 /*
122  * Hardware event to monitor via a performance monitoring counter:
123  */
124 struct perf_counter_hw_event {
125         /*
126          * The MSB of the config word signifies if the rest contains cpu
127          * specific (raw) counter configuration data, if unset, the next
128          * 7 bits are an event type and the rest of the bits are the event
129          * identifier.
130          */
131         __u64                   config;
132
133         union {
134                 __u64           irq_period;
135                 __u64           irq_freq;
136         };
137
138         __u32                   record_type;
139         __u32                   read_format;
140
141         __u64                   disabled       :  1, /* off by default        */
142                                 nmi            :  1, /* NMI sampling          */
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   : 51;
156
157         __u32                   extra_config_len;
158         __u32                   wakeup_events;  /* wakeup every n events */
159
160         __u64                   __reserved_2;
161         __u64                   __reserved_3;
162 };
163
164 /*
165  * Ioctls that can be done on a perf counter fd:
166  */
167 #define PERF_COUNTER_IOC_ENABLE         _IOW('$', 0, u32)
168 #define PERF_COUNTER_IOC_DISABLE        _IOW('$', 1, u32)
169 #define PERF_COUNTER_IOC_REFRESH        _IOW('$', 2, u32)
170 #define PERF_COUNTER_IOC_RESET          _IOW('$', 3, u32)
171
172 enum perf_counter_ioc_flags {
173         PERF_IOC_FLAG_GROUP             = 1U << 0,
174 };
175
176 /*
177  * Structure of the page that can be mapped via mmap
178  */
179 struct perf_counter_mmap_page {
180         __u32   version;                /* version number of this structure */
181         __u32   compat_version;         /* lowest version this is compat with */
182
183         /*
184          * Bits needed to read the hw counters in user-space.
185          *
186          *   u32 seq;
187          *   s64 count;
188          *
189          *   do {
190          *     seq = pc->lock;
191          *
192          *     barrier()
193          *     if (pc->index) {
194          *       count = pmc_read(pc->index - 1);
195          *       count += pc->offset;
196          *     } else
197          *       goto regular_read;
198          *
199          *     barrier();
200          *   } while (pc->lock != seq);
201          *
202          * NOTE: for obvious reason this only works on self-monitoring
203          *       processes.
204          */
205         __u32   lock;                   /* seqlock for synchronization */
206         __u32   index;                  /* hardware counter identifier */
207         __s64   offset;                 /* add to hardware counter value */
208
209         /*
210          * Control data for the mmap() data buffer.
211          *
212          * User-space reading this value should issue an rmb(), on SMP capable
213          * platforms, after reading this value -- see perf_counter_wakeup().
214          */
215         __u32   data_head;              /* head in the data section */
216 };
217
218 #define PERF_EVENT_MISC_CPUMODE_MASK    (3 << 0)
219 #define PERF_EVENT_MISC_CPUMODE_UNKNOWN (0 << 0)
220 #define PERF_EVENT_MISC_KERNEL          (1 << 0)
221 #define PERF_EVENT_MISC_USER            (2 << 0)
222 #define PERF_EVENT_MISC_HYPERVISOR      (3 << 0)
223 #define PERF_EVENT_MISC_OVERFLOW        (1 << 2)
224
225 struct perf_event_header {
226         __u32   type;
227         __u16   misc;
228         __u16   size;
229 };
230
231 enum perf_event_type {
232
233         /*
234          * The MMAP events record the PROT_EXEC mappings so that we can
235          * correlate userspace IPs to code. They have the following structure:
236          *
237          * struct {
238          *      struct perf_event_header        header;
239          *
240          *      u32                             pid, tid;
241          *      u64                             addr;
242          *      u64                             len;
243          *      u64                             pgoff;
244          *      char                            filename[];
245          * };
246          */
247         PERF_EVENT_MMAP                 = 1,
248         PERF_EVENT_MUNMAP               = 2,
249
250         /*
251          * struct {
252          *      struct perf_event_header        header;
253          *
254          *      u32                             pid, tid;
255          *      char                            comm[];
256          * };
257          */
258         PERF_EVENT_COMM                 = 3,
259
260         /*
261          * struct {
262          *      struct perf_event_header        header;
263          *      u64                             irq_period;
264          * };
265          */
266         PERF_EVENT_PERIOD               = 4,
267
268         /*
269          * When header.misc & PERF_EVENT_MISC_OVERFLOW the event_type field
270          * will be PERF_RECORD_*
271          *
272          * struct {
273          *      struct perf_event_header        header;
274          *
275          *      { u64                   ip;       } && PERF_RECORD_IP
276          *      { u32                   pid, tid; } && PERF_RECORD_TID
277          *      { u64                   time;     } && PERF_RECORD_TIME
278          *      { u64                   addr;     } && PERF_RECORD_ADDR
279          *      { u64                   config;   } && PERF_RECORD_CONFIG
280          *      { u32                   cpu, res; } && PERF_RECORD_CPU
281          *
282          *      { u64                   nr;
283          *        { u64 event, val; }   cnt[nr];  } && PERF_RECORD_GROUP
284          *
285          *      { u16                   nr,
286          *                              hv,
287          *                              kernel,
288          *                              user;
289          *        u64                   ips[nr];  } && PERF_RECORD_CALLCHAIN
290          * };
291          */
292 };
293
294 #ifdef __KERNEL__
295 /*
296  * Kernel-internal data types and definitions:
297  */
298
299 #ifdef CONFIG_PERF_COUNTERS
300 # include <asm/perf_counter.h>
301 #endif
302
303 #include <linux/list.h>
304 #include <linux/mutex.h>
305 #include <linux/rculist.h>
306 #include <linux/rcupdate.h>
307 #include <linux/spinlock.h>
308 #include <linux/hrtimer.h>
309 #include <linux/fs.h>
310 #include <asm/atomic.h>
311
312 struct task_struct;
313
314 static inline u64 perf_event_raw(struct perf_counter_hw_event *hw_event)
315 {
316         return hw_event->config & PERF_COUNTER_RAW_MASK;
317 }
318
319 static inline u64 perf_event_config(struct perf_counter_hw_event *hw_event)
320 {
321         return hw_event->config & PERF_COUNTER_CONFIG_MASK;
322 }
323
324 static inline u64 perf_event_type(struct perf_counter_hw_event *hw_event)
325 {
326         return (hw_event->config & PERF_COUNTER_TYPE_MASK) >>
327                 PERF_COUNTER_TYPE_SHIFT;
328 }
329
330 static inline u64 perf_event_id(struct perf_counter_hw_event *hw_event)
331 {
332         return hw_event->config & PERF_COUNTER_EVENT_MASK;
333 }
334
335 /**
336  * struct hw_perf_counter - performance counter hardware details:
337  */
338 struct hw_perf_counter {
339 #ifdef CONFIG_PERF_COUNTERS
340         union {
341                 struct { /* hardware */
342                         u64                             config;
343                         unsigned long                   config_base;
344                         unsigned long                   counter_base;
345                         int                             nmi;
346                         int                             idx;
347                 };
348                 union { /* software */
349                         atomic64_t                      count;
350                         struct hrtimer                  hrtimer;
351                 };
352         };
353         atomic64_t                      prev_count;
354         u64                             irq_period;
355         atomic64_t                      period_left;
356         u64                             interrupts;
357 #endif
358 };
359
360 struct perf_counter;
361
362 /**
363  * struct pmu - generic performance monitoring unit
364  */
365 struct pmu {
366         int (*enable)                   (struct perf_counter *counter);
367         void (*disable)                 (struct perf_counter *counter);
368         void (*read)                    (struct perf_counter *counter);
369 };
370
371 /**
372  * enum perf_counter_active_state - the states of a counter
373  */
374 enum perf_counter_active_state {
375         PERF_COUNTER_STATE_ERROR        = -2,
376         PERF_COUNTER_STATE_OFF          = -1,
377         PERF_COUNTER_STATE_INACTIVE     =  0,
378         PERF_COUNTER_STATE_ACTIVE       =  1,
379 };
380
381 struct file;
382
383 struct perf_mmap_data {
384         struct rcu_head                 rcu_head;
385         int                             nr_pages;       /* nr of data pages  */
386         int                             nr_locked;      /* nr pages mlocked  */
387
388         atomic_t                        poll;           /* POLL_ for wakeups */
389         atomic_t                        head;           /* write position    */
390         atomic_t                        events;         /* event limit       */
391
392         atomic_t                        done_head;      /* completed head    */
393         atomic_t                        lock;           /* concurrent writes */
394
395         atomic_t                        wakeup;         /* needs a wakeup    */
396
397         struct perf_counter_mmap_page   *user_page;
398         void                            *data_pages[0];
399 };
400
401 struct perf_pending_entry {
402         struct perf_pending_entry *next;
403         void (*func)(struct perf_pending_entry *);
404 };
405
406 /**
407  * struct perf_counter - performance counter kernel representation:
408  */
409 struct perf_counter {
410 #ifdef CONFIG_PERF_COUNTERS
411         struct list_head                list_entry;
412         struct list_head                event_entry;
413         struct list_head                sibling_list;
414         int                             nr_siblings;
415         struct perf_counter             *group_leader;
416         const struct pmu                *pmu;
417
418         enum perf_counter_active_state  state;
419         enum perf_counter_active_state  prev_state;
420         atomic64_t                      count;
421
422         /*
423          * These are the total time in nanoseconds that the counter
424          * has been enabled (i.e. eligible to run, and the task has
425          * been scheduled in, if this is a per-task counter)
426          * and running (scheduled onto the CPU), respectively.
427          *
428          * They are computed from tstamp_enabled, tstamp_running and
429          * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
430          */
431         u64                             total_time_enabled;
432         u64                             total_time_running;
433
434         /*
435          * These are timestamps used for computing total_time_enabled
436          * and total_time_running when the counter is in INACTIVE or
437          * ACTIVE state, measured in nanoseconds from an arbitrary point
438          * in time.
439          * tstamp_enabled: the notional time when the counter was enabled
440          * tstamp_running: the notional time when the counter was scheduled on
441          * tstamp_stopped: in INACTIVE state, the notional time when the
442          *      counter was scheduled off.
443          */
444         u64                             tstamp_enabled;
445         u64                             tstamp_running;
446         u64                             tstamp_stopped;
447
448         struct perf_counter_hw_event    hw_event;
449         struct hw_perf_counter          hw;
450
451         struct perf_counter_context     *ctx;
452         struct file                     *filp;
453
454         struct perf_counter             *parent;
455         struct list_head                child_list;
456
457         /*
458          * These accumulate total time (in nanoseconds) that children
459          * counters have been enabled and running, respectively.
460          */
461         atomic64_t                      child_total_time_enabled;
462         atomic64_t                      child_total_time_running;
463
464         /*
465          * Protect attach/detach and child_list:
466          */
467         struct mutex                    mutex;
468
469         int                             oncpu;
470         int                             cpu;
471
472         /* mmap bits */
473         struct mutex                    mmap_mutex;
474         atomic_t                        mmap_count;
475         struct perf_mmap_data           *data;
476
477         /* poll related */
478         wait_queue_head_t               waitq;
479         struct fasync_struct            *fasync;
480
481         /* delayed work for NMIs and such */
482         int                             pending_wakeup;
483         int                             pending_kill;
484         int                             pending_disable;
485         struct perf_pending_entry       pending;
486
487         atomic_t                        event_limit;
488
489         void (*destroy)(struct perf_counter *);
490         struct rcu_head                 rcu_head;
491 #endif
492 };
493
494 /**
495  * struct perf_counter_context - counter context structure
496  *
497  * Used as a container for task counters and CPU counters as well:
498  */
499 struct perf_counter_context {
500         /*
501          * Protect the states of the counters in the list,
502          * nr_active, and the list:
503          */
504         spinlock_t              lock;
505         /*
506          * Protect the list of counters.  Locking either mutex or lock
507          * is sufficient to ensure the list doesn't change; to change
508          * the list you need to lock both the mutex and the spinlock.
509          */
510         struct mutex            mutex;
511
512         struct list_head        counter_list;
513         struct list_head        event_list;
514         int                     nr_counters;
515         int                     nr_active;
516         int                     is_active;
517         atomic_t                refcount;
518         struct task_struct      *task;
519
520         /*
521          * Context clock, runs when context enabled.
522          */
523         u64                     time;
524         u64                     timestamp;
525 };
526
527 /**
528  * struct perf_counter_cpu_context - per cpu counter context structure
529  */
530 struct perf_cpu_context {
531         struct perf_counter_context     ctx;
532         struct perf_counter_context     *task_ctx;
533         int                             active_oncpu;
534         int                             max_pertask;
535         int                             exclusive;
536
537         /*
538          * Recursion avoidance:
539          *
540          * task, softirq, irq, nmi context
541          */
542         int                     recursion[4];
543 };
544
545 #ifdef CONFIG_PERF_COUNTERS
546
547 /*
548  * Set by architecture code:
549  */
550 extern int perf_max_counters;
551
552 extern const struct pmu *hw_perf_counter_init(struct perf_counter *counter);
553
554 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
555 extern void perf_counter_task_sched_out(struct task_struct *task, int cpu);
556 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
557 extern void perf_counter_init_task(struct task_struct *child);
558 extern void perf_counter_exit_task(struct task_struct *child);
559 extern void perf_counter_do_pending(void);
560 extern void perf_counter_print_debug(void);
561 extern void perf_counter_unthrottle(void);
562 extern void __perf_disable(void);
563 extern bool __perf_enable(void);
564 extern void perf_disable(void);
565 extern void perf_enable(void);
566 extern int perf_counter_task_disable(void);
567 extern int perf_counter_task_enable(void);
568 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
569                struct perf_cpu_context *cpuctx,
570                struct perf_counter_context *ctx, int cpu);
571 extern void perf_counter_update_userpage(struct perf_counter *counter);
572
573 extern int perf_counter_overflow(struct perf_counter *counter,
574                                  int nmi, struct pt_regs *regs, u64 addr);
575 /*
576  * Return 1 for a software counter, 0 for a hardware counter
577  */
578 static inline int is_software_counter(struct perf_counter *counter)
579 {
580         return !perf_event_raw(&counter->hw_event) &&
581                 perf_event_type(&counter->hw_event) != PERF_TYPE_HARDWARE;
582 }
583
584 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *, u64);
585
586 extern void perf_counter_mmap(unsigned long addr, unsigned long len,
587                               unsigned long pgoff, struct file *file);
588
589 extern void perf_counter_munmap(unsigned long addr, unsigned long len,
590                                 unsigned long pgoff, struct file *file);
591
592 extern void perf_counter_comm(struct task_struct *tsk);
593
594 #define MAX_STACK_DEPTH         255
595
596 struct perf_callchain_entry {
597         u16     nr, hv, kernel, user;
598         u64     ip[MAX_STACK_DEPTH];
599 };
600
601 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
602
603 extern int sysctl_perf_counter_priv;
604 extern int sysctl_perf_counter_mlock;
605
606 extern void perf_counter_init(void);
607
608 #ifndef perf_misc_flags
609 #define perf_misc_flags(regs)   (user_mode(regs) ? PERF_EVENT_MISC_USER : \
610                                  PERF_EVENT_MISC_KERNEL)
611 #define perf_instruction_pointer(regs)  instruction_pointer(regs)
612 #endif
613
614 #else
615 static inline void
616 perf_counter_task_sched_in(struct task_struct *task, int cpu)           { }
617 static inline void
618 perf_counter_task_sched_out(struct task_struct *task, int cpu)          { }
619 static inline void
620 perf_counter_task_tick(struct task_struct *task, int cpu)               { }
621 static inline void perf_counter_init_task(struct task_struct *child)    { }
622 static inline void perf_counter_exit_task(struct task_struct *child)    { }
623 static inline void perf_counter_do_pending(void)                        { }
624 static inline void perf_counter_print_debug(void)                       { }
625 static inline void perf_counter_unthrottle(void)                        { }
626 static inline void perf_disable(void)                                   { }
627 static inline void perf_enable(void)                                    { }
628 static inline int perf_counter_task_disable(void)       { return -EINVAL; }
629 static inline int perf_counter_task_enable(void)        { return -EINVAL; }
630
631 static inline void
632 perf_swcounter_event(u32 event, u64 nr, int nmi,
633                      struct pt_regs *regs, u64 addr)                    { }
634
635 static inline void
636 perf_counter_mmap(unsigned long addr, unsigned long len,
637                   unsigned long pgoff, struct file *file)               { }
638
639 static inline void
640 perf_counter_munmap(unsigned long addr, unsigned long len,
641                     unsigned long pgoff, struct file *file)             { }
642
643 static inline void perf_counter_comm(struct task_struct *tsk)           { }
644 static inline void perf_counter_init(void)                              { }
645 #endif
646
647 #endif /* __KERNEL__ */
648 #endif /* _LINUX_PERF_COUNTER_H */