perf_counter: Add fork event
[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  * attr.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 attr.event_id
41  * parameter of the sys_perf_counter_open() syscall:
42  */
43 enum attr_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 attr.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 attr.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_attr {
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          * struct {
281          *      struct perf_event_header        header;
282          *      u32                             pid, ppid;
283          * };
284          */
285         PERF_EVENT_FORK                 = 7,
286
287         /*
288          * When header.misc & PERF_EVENT_MISC_OVERFLOW the event_type field
289          * will be PERF_RECORD_*
290          *
291          * struct {
292          *      struct perf_event_header        header;
293          *
294          *      { u64                   ip;       } && PERF_RECORD_IP
295          *      { u32                   pid, tid; } && PERF_RECORD_TID
296          *      { u64                   time;     } && PERF_RECORD_TIME
297          *      { u64                   addr;     } && PERF_RECORD_ADDR
298          *      { u64                   config;   } && PERF_RECORD_CONFIG
299          *      { u32                   cpu, res; } && PERF_RECORD_CPU
300          *
301          *      { u64                   nr;
302          *        { u64 id, val; }      cnt[nr];  } && PERF_RECORD_GROUP
303          *
304          *      { u16                   nr,
305          *                              hv,
306          *                              kernel,
307          *                              user;
308          *        u64                   ips[nr];  } && PERF_RECORD_CALLCHAIN
309          * };
310          */
311 };
312
313 #ifdef __KERNEL__
314 /*
315  * Kernel-internal data types and definitions:
316  */
317
318 #ifdef CONFIG_PERF_COUNTERS
319 # include <asm/perf_counter.h>
320 #endif
321
322 #include <linux/list.h>
323 #include <linux/mutex.h>
324 #include <linux/rculist.h>
325 #include <linux/rcupdate.h>
326 #include <linux/spinlock.h>
327 #include <linux/hrtimer.h>
328 #include <linux/fs.h>
329 #include <linux/pid_namespace.h>
330 #include <asm/atomic.h>
331
332 struct task_struct;
333
334 static inline u64 perf_event_raw(struct perf_counter_attr *attr)
335 {
336         return attr->config & PERF_COUNTER_RAW_MASK;
337 }
338
339 static inline u64 perf_event_config(struct perf_counter_attr *attr)
340 {
341         return attr->config & PERF_COUNTER_CONFIG_MASK;
342 }
343
344 static inline u64 perf_event_type(struct perf_counter_attr *attr)
345 {
346         return (attr->config & PERF_COUNTER_TYPE_MASK) >>
347                 PERF_COUNTER_TYPE_SHIFT;
348 }
349
350 static inline u64 perf_event_id(struct perf_counter_attr *attr)
351 {
352         return attr->config & PERF_COUNTER_EVENT_MASK;
353 }
354
355 /**
356  * struct hw_perf_counter - performance counter hardware details:
357  */
358 struct hw_perf_counter {
359 #ifdef CONFIG_PERF_COUNTERS
360         union {
361                 struct { /* hardware */
362                         u64                             config;
363                         unsigned long                   config_base;
364                         unsigned long                   counter_base;
365                         int                             idx;
366                 };
367                 union { /* software */
368                         atomic64_t                      count;
369                         struct hrtimer                  hrtimer;
370                 };
371         };
372         atomic64_t                      prev_count;
373         u64                             sample_period;
374         atomic64_t                      period_left;
375         u64                             interrupts;
376 #endif
377 };
378
379 struct perf_counter;
380
381 /**
382  * struct pmu - generic performance monitoring unit
383  */
384 struct pmu {
385         int (*enable)                   (struct perf_counter *counter);
386         void (*disable)                 (struct perf_counter *counter);
387         void (*read)                    (struct perf_counter *counter);
388         void (*unthrottle)              (struct perf_counter *counter);
389 };
390
391 /**
392  * enum perf_counter_active_state - the states of a counter
393  */
394 enum perf_counter_active_state {
395         PERF_COUNTER_STATE_ERROR        = -2,
396         PERF_COUNTER_STATE_OFF          = -1,
397         PERF_COUNTER_STATE_INACTIVE     =  0,
398         PERF_COUNTER_STATE_ACTIVE       =  1,
399 };
400
401 struct file;
402
403 struct perf_mmap_data {
404         struct rcu_head                 rcu_head;
405         int                             nr_pages;       /* nr of data pages  */
406         int                             nr_locked;      /* nr pages mlocked  */
407
408         atomic_t                        poll;           /* POLL_ for wakeups */
409         atomic_t                        events;         /* event limit       */
410
411         atomic_long_t                   head;           /* write position    */
412         atomic_long_t                   done_head;      /* completed head    */
413
414         atomic_t                        lock;           /* concurrent writes */
415
416         atomic_t                        wakeup;         /* needs a wakeup    */
417
418         struct perf_counter_mmap_page   *user_page;
419         void                            *data_pages[0];
420 };
421
422 struct perf_pending_entry {
423         struct perf_pending_entry *next;
424         void (*func)(struct perf_pending_entry *);
425 };
426
427 /**
428  * struct perf_counter - performance counter kernel representation:
429  */
430 struct perf_counter {
431 #ifdef CONFIG_PERF_COUNTERS
432         struct list_head                list_entry;
433         struct list_head                event_entry;
434         struct list_head                sibling_list;
435         int                             nr_siblings;
436         struct perf_counter             *group_leader;
437         const struct pmu                *pmu;
438
439         enum perf_counter_active_state  state;
440         atomic64_t                      count;
441
442         /*
443          * These are the total time in nanoseconds that the counter
444          * has been enabled (i.e. eligible to run, and the task has
445          * been scheduled in, if this is a per-task counter)
446          * and running (scheduled onto the CPU), respectively.
447          *
448          * They are computed from tstamp_enabled, tstamp_running and
449          * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
450          */
451         u64                             total_time_enabled;
452         u64                             total_time_running;
453
454         /*
455          * These are timestamps used for computing total_time_enabled
456          * and total_time_running when the counter is in INACTIVE or
457          * ACTIVE state, measured in nanoseconds from an arbitrary point
458          * in time.
459          * tstamp_enabled: the notional time when the counter was enabled
460          * tstamp_running: the notional time when the counter was scheduled on
461          * tstamp_stopped: in INACTIVE state, the notional time when the
462          *      counter was scheduled off.
463          */
464         u64                             tstamp_enabled;
465         u64                             tstamp_running;
466         u64                             tstamp_stopped;
467
468         struct perf_counter_attr        attr;
469         struct hw_perf_counter          hw;
470
471         struct perf_counter_context     *ctx;
472         struct file                     *filp;
473
474         /*
475          * These accumulate total time (in nanoseconds) that children
476          * counters have been enabled and running, respectively.
477          */
478         atomic64_t                      child_total_time_enabled;
479         atomic64_t                      child_total_time_running;
480
481         /*
482          * Protect attach/detach and child_list:
483          */
484         struct mutex                    child_mutex;
485         struct list_head                child_list;
486         struct perf_counter             *parent;
487
488         int                             oncpu;
489         int                             cpu;
490
491         struct list_head                owner_entry;
492         struct task_struct              *owner;
493
494         /* mmap bits */
495         struct mutex                    mmap_mutex;
496         atomic_t                        mmap_count;
497         struct perf_mmap_data           *data;
498
499         /* poll related */
500         wait_queue_head_t               waitq;
501         struct fasync_struct            *fasync;
502
503         /* delayed work for NMIs and such */
504         int                             pending_wakeup;
505         int                             pending_kill;
506         int                             pending_disable;
507         struct perf_pending_entry       pending;
508
509         atomic_t                        event_limit;
510
511         void (*destroy)(struct perf_counter *);
512         struct rcu_head                 rcu_head;
513
514         struct pid_namespace            *ns;
515         u64                             id;
516 #endif
517 };
518
519 /**
520  * struct perf_counter_context - counter context structure
521  *
522  * Used as a container for task counters and CPU counters as well:
523  */
524 struct perf_counter_context {
525         /*
526          * Protect the states of the counters in the list,
527          * nr_active, and the list:
528          */
529         spinlock_t              lock;
530         /*
531          * Protect the list of counters.  Locking either mutex or lock
532          * is sufficient to ensure the list doesn't change; to change
533          * the list you need to lock both the mutex and the spinlock.
534          */
535         struct mutex            mutex;
536
537         struct list_head        counter_list;
538         struct list_head        event_list;
539         int                     nr_counters;
540         int                     nr_active;
541         int                     is_active;
542         atomic_t                refcount;
543         struct task_struct      *task;
544
545         /*
546          * Context clock, runs when context enabled.
547          */
548         u64                     time;
549         u64                     timestamp;
550
551         /*
552          * These fields let us detect when two contexts have both
553          * been cloned (inherited) from a common ancestor.
554          */
555         struct perf_counter_context *parent_ctx;
556         u64                     parent_gen;
557         u64                     generation;
558         int                     pin_count;
559         struct rcu_head         rcu_head;
560 };
561
562 /**
563  * struct perf_counter_cpu_context - per cpu counter context structure
564  */
565 struct perf_cpu_context {
566         struct perf_counter_context     ctx;
567         struct perf_counter_context     *task_ctx;
568         int                             active_oncpu;
569         int                             max_pertask;
570         int                             exclusive;
571
572         /*
573          * Recursion avoidance:
574          *
575          * task, softirq, irq, nmi context
576          */
577         int                             recursion[4];
578 };
579
580 #ifdef CONFIG_PERF_COUNTERS
581
582 /*
583  * Set by architecture code:
584  */
585 extern int perf_max_counters;
586
587 extern const struct pmu *hw_perf_counter_init(struct perf_counter *counter);
588
589 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
590 extern void perf_counter_task_sched_out(struct task_struct *task,
591                                         struct task_struct *next, int cpu);
592 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
593 extern int perf_counter_init_task(struct task_struct *child);
594 extern void perf_counter_exit_task(struct task_struct *child);
595 extern void perf_counter_free_task(struct task_struct *task);
596 extern void perf_counter_do_pending(void);
597 extern void perf_counter_print_debug(void);
598 extern void __perf_disable(void);
599 extern bool __perf_enable(void);
600 extern void perf_disable(void);
601 extern void perf_enable(void);
602 extern int perf_counter_task_disable(void);
603 extern int perf_counter_task_enable(void);
604 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
605                struct perf_cpu_context *cpuctx,
606                struct perf_counter_context *ctx, int cpu);
607 extern void perf_counter_update_userpage(struct perf_counter *counter);
608
609 extern int perf_counter_overflow(struct perf_counter *counter,
610                                  int nmi, struct pt_regs *regs, u64 addr);
611 /*
612  * Return 1 for a software counter, 0 for a hardware counter
613  */
614 static inline int is_software_counter(struct perf_counter *counter)
615 {
616         return !perf_event_raw(&counter->attr) &&
617                 perf_event_type(&counter->attr) != PERF_TYPE_HARDWARE;
618 }
619
620 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *, u64);
621
622 extern void perf_counter_mmap(unsigned long addr, unsigned long len,
623                               unsigned long pgoff, struct file *file);
624
625 extern void perf_counter_munmap(unsigned long addr, unsigned long len,
626                                 unsigned long pgoff, struct file *file);
627
628 extern void perf_counter_comm(struct task_struct *tsk);
629 extern void perf_counter_fork(struct task_struct *tsk);
630
631 extern void perf_counter_task_migration(struct task_struct *task, int cpu);
632
633 #define MAX_STACK_DEPTH         255
634
635 struct perf_callchain_entry {
636         u16     nr, hv, kernel, user;
637         u64     ip[MAX_STACK_DEPTH];
638 };
639
640 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
641
642 extern int sysctl_perf_counter_priv;
643 extern int sysctl_perf_counter_mlock;
644 extern int sysctl_perf_counter_limit;
645
646 extern void perf_counter_init(void);
647
648 #ifndef perf_misc_flags
649 #define perf_misc_flags(regs)   (user_mode(regs) ? PERF_EVENT_MISC_USER : \
650                                  PERF_EVENT_MISC_KERNEL)
651 #define perf_instruction_pointer(regs)  instruction_pointer(regs)
652 #endif
653
654 #else
655 static inline void
656 perf_counter_task_sched_in(struct task_struct *task, int cpu)           { }
657 static inline void
658 perf_counter_task_sched_out(struct task_struct *task,
659                             struct task_struct *next, int cpu)          { }
660 static inline void
661 perf_counter_task_tick(struct task_struct *task, int cpu)               { }
662 static inline int perf_counter_init_task(struct task_struct *child)     { return 0; }
663 static inline void perf_counter_exit_task(struct task_struct *child)    { }
664 static inline void perf_counter_free_task(struct task_struct *task)     { }
665 static inline void perf_counter_do_pending(void)                        { }
666 static inline void perf_counter_print_debug(void)                       { }
667 static inline void perf_disable(void)                                   { }
668 static inline void perf_enable(void)                                    { }
669 static inline int perf_counter_task_disable(void)       { return -EINVAL; }
670 static inline int perf_counter_task_enable(void)        { return -EINVAL; }
671
672 static inline void
673 perf_swcounter_event(u32 event, u64 nr, int nmi,
674                      struct pt_regs *regs, u64 addr)                    { }
675
676 static inline void
677 perf_counter_mmap(unsigned long addr, unsigned long len,
678                   unsigned long pgoff, struct file *file)               { }
679
680 static inline void
681 perf_counter_munmap(unsigned long addr, unsigned long len,
682                     unsigned long pgoff, struct file *file)             { }
683
684 static inline void perf_counter_comm(struct task_struct *tsk)           { }
685 static inline void perf_counter_fork(struct task_struct *tsk)           { }
686 static inline void perf_counter_init(void)                              { }
687 static inline void perf_counter_task_migration(struct task_struct *task,
688                                                int cpu)                 { }
689 #endif
690
691 #endif /* __KERNEL__ */
692 #endif /* _LINUX_PERF_COUNTER_H */