Merge branch 'fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/davej/cpufreq
[linux-2.6] / include / linux / perf_counter.h
1 /*
2  *  Performance counters:
3  *
4  *    Copyright (C) 2008-2009, Thomas Gleixner <tglx@linutronix.de>
5  *    Copyright (C) 2008-2009, Red Hat, Inc., Ingo Molnar
6  *    Copyright (C) 2008-2009, Red Hat, Inc., Peter Zijlstra
7  *
8  *  Data type definitions, declarations, prototypes.
9  *
10  *    Started by: Thomas Gleixner and Ingo Molnar
11  *
12  *  For licencing details see kernel-base/COPYING
13  */
14 #ifndef _LINUX_PERF_COUNTER_H
15 #define _LINUX_PERF_COUNTER_H
16
17 #include <linux/types.h>
18 #include <linux/ioctl.h>
19 #include <asm/byteorder.h>
20
21 /*
22  * User-space ABI bits:
23  */
24
25 /*
26  * attr.type
27  */
28 enum perf_type_id {
29         PERF_TYPE_HARDWARE                      = 0,
30         PERF_TYPE_SOFTWARE                      = 1,
31         PERF_TYPE_TRACEPOINT                    = 2,
32         PERF_TYPE_HW_CACHE                      = 3,
33         PERF_TYPE_RAW                           = 4,
34
35         PERF_TYPE_MAX,                          /* non-ABI */
36 };
37
38 /*
39  * Generalized performance counter event types, used by the
40  * attr.event_id parameter of the sys_perf_counter_open()
41  * syscall:
42  */
43 enum perf_hw_id {
44         /*
45          * Common hardware events, generalized by the kernel:
46          */
47         PERF_COUNT_HW_CPU_CYCLES                = 0,
48         PERF_COUNT_HW_INSTRUCTIONS              = 1,
49         PERF_COUNT_HW_CACHE_REFERENCES          = 2,
50         PERF_COUNT_HW_CACHE_MISSES              = 3,
51         PERF_COUNT_HW_BRANCH_INSTRUCTIONS       = 4,
52         PERF_COUNT_HW_BRANCH_MISSES             = 5,
53         PERF_COUNT_HW_BUS_CYCLES                = 6,
54
55         PERF_COUNT_HW_MAX,                      /* non-ABI */
56 };
57
58 /*
59  * Generalized hardware cache counters:
60  *
61  *       { L1-D, L1-I, LLC, ITLB, DTLB, BPU } x
62  *       { read, write, prefetch } x
63  *       { accesses, misses }
64  */
65 enum perf_hw_cache_id {
66         PERF_COUNT_HW_CACHE_L1D                 = 0,
67         PERF_COUNT_HW_CACHE_L1I                 = 1,
68         PERF_COUNT_HW_CACHE_LL                  = 2,
69         PERF_COUNT_HW_CACHE_DTLB                = 3,
70         PERF_COUNT_HW_CACHE_ITLB                = 4,
71         PERF_COUNT_HW_CACHE_BPU                 = 5,
72
73         PERF_COUNT_HW_CACHE_MAX,                /* non-ABI */
74 };
75
76 enum perf_hw_cache_op_id {
77         PERF_COUNT_HW_CACHE_OP_READ             = 0,
78         PERF_COUNT_HW_CACHE_OP_WRITE            = 1,
79         PERF_COUNT_HW_CACHE_OP_PREFETCH         = 2,
80
81         PERF_COUNT_HW_CACHE_OP_MAX,             /* non-ABI */
82 };
83
84 enum perf_hw_cache_op_result_id {
85         PERF_COUNT_HW_CACHE_RESULT_ACCESS       = 0,
86         PERF_COUNT_HW_CACHE_RESULT_MISS         = 1,
87
88         PERF_COUNT_HW_CACHE_RESULT_MAX,         /* non-ABI */
89 };
90
91 /*
92  * Special "software" counters provided by the kernel, even if the hardware
93  * does not support performance counters. These counters measure various
94  * physical and sw events of the kernel (and allow the profiling of them as
95  * well):
96  */
97 enum perf_sw_ids {
98         PERF_COUNT_SW_CPU_CLOCK                 = 0,
99         PERF_COUNT_SW_TASK_CLOCK                = 1,
100         PERF_COUNT_SW_PAGE_FAULTS               = 2,
101         PERF_COUNT_SW_CONTEXT_SWITCHES          = 3,
102         PERF_COUNT_SW_CPU_MIGRATIONS            = 4,
103         PERF_COUNT_SW_PAGE_FAULTS_MIN           = 5,
104         PERF_COUNT_SW_PAGE_FAULTS_MAJ           = 6,
105
106         PERF_COUNT_SW_MAX,                      /* non-ABI */
107 };
108
109 /*
110  * Bits that can be set in attr.sample_type to request information
111  * in the overflow packets.
112  */
113 enum perf_counter_sample_format {
114         PERF_SAMPLE_IP                          = 1U << 0,
115         PERF_SAMPLE_TID                         = 1U << 1,
116         PERF_SAMPLE_TIME                        = 1U << 2,
117         PERF_SAMPLE_ADDR                        = 1U << 3,
118         PERF_SAMPLE_GROUP                       = 1U << 4,
119         PERF_SAMPLE_CALLCHAIN                   = 1U << 5,
120         PERF_SAMPLE_ID                          = 1U << 6,
121         PERF_SAMPLE_CPU                         = 1U << 7,
122         PERF_SAMPLE_PERIOD                      = 1U << 8,
123
124         PERF_SAMPLE_MAX = 1U << 9,              /* non-ABI */
125 };
126
127 /*
128  * Bits that can be set in attr.read_format to request that
129  * reads on the counter should return the indicated quantities,
130  * in increasing order of bit value, after the counter value.
131  */
132 enum perf_counter_read_format {
133         PERF_FORMAT_TOTAL_TIME_ENABLED          = 1U << 0,
134         PERF_FORMAT_TOTAL_TIME_RUNNING          = 1U << 1,
135         PERF_FORMAT_ID                          = 1U << 2,
136
137         PERF_FORMAT_MAX = 1U << 3,              /* non-ABI */
138 };
139
140 #define PERF_ATTR_SIZE_VER0     64      /* sizeof first published struct */
141
142 /*
143  * Hardware event to monitor via a performance monitoring counter:
144  */
145 struct perf_counter_attr {
146
147         /*
148          * Major type: hardware/software/tracepoint/etc.
149          */
150         __u32                   type;
151
152         /*
153          * Size of the attr structure, for fwd/bwd compat.
154          */
155         __u32                   size;
156
157         /*
158          * Type specific configuration information.
159          */
160         __u64                   config;
161
162         union {
163                 __u64           sample_period;
164                 __u64           sample_freq;
165         };
166
167         __u64                   sample_type;
168         __u64                   read_format;
169
170         __u64                   disabled       :  1, /* off by default        */
171                                 inherit        :  1, /* children inherit it   */
172                                 pinned         :  1, /* must always be on PMU */
173                                 exclusive      :  1, /* only group on PMU     */
174                                 exclude_user   :  1, /* don't count user      */
175                                 exclude_kernel :  1, /* ditto kernel          */
176                                 exclude_hv     :  1, /* ditto hypervisor      */
177                                 exclude_idle   :  1, /* don't count when idle */
178                                 mmap           :  1, /* include mmap data     */
179                                 comm           :  1, /* include comm data     */
180                                 freq           :  1, /* use freq, not period  */
181
182                                 __reserved_1   : 53;
183
184         __u32                   wakeup_events;  /* wakeup every n events */
185         __u32                   __reserved_2;
186
187         __u64                   __reserved_3;
188 };
189
190 /*
191  * Ioctls that can be done on a perf counter fd:
192  */
193 #define PERF_COUNTER_IOC_ENABLE         _IO ('$', 0)
194 #define PERF_COUNTER_IOC_DISABLE        _IO ('$', 1)
195 #define PERF_COUNTER_IOC_REFRESH        _IO ('$', 2)
196 #define PERF_COUNTER_IOC_RESET          _IO ('$', 3)
197 #define PERF_COUNTER_IOC_PERIOD         _IOW('$', 4, u64)
198
199 enum perf_counter_ioc_flags {
200         PERF_IOC_FLAG_GROUP             = 1U << 0,
201 };
202
203 /*
204  * Structure of the page that can be mapped via mmap
205  */
206 struct perf_counter_mmap_page {
207         __u32   version;                /* version number of this structure */
208         __u32   compat_version;         /* lowest version this is compat with */
209
210         /*
211          * Bits needed to read the hw counters in user-space.
212          *
213          *   u32 seq;
214          *   s64 count;
215          *
216          *   do {
217          *     seq = pc->lock;
218          *
219          *     barrier()
220          *     if (pc->index) {
221          *       count = pmc_read(pc->index - 1);
222          *       count += pc->offset;
223          *     } else
224          *       goto regular_read;
225          *
226          *     barrier();
227          *   } while (pc->lock != seq);
228          *
229          * NOTE: for obvious reason this only works on self-monitoring
230          *       processes.
231          */
232         __u32   lock;                   /* seqlock for synchronization */
233         __u32   index;                  /* hardware counter identifier */
234         __s64   offset;                 /* add to hardware counter value */
235
236         /*
237          * Control data for the mmap() data buffer.
238          *
239          * User-space reading this value should issue an rmb(), on SMP capable
240          * platforms, after reading this value -- see perf_counter_wakeup().
241          */
242         __u64   data_head;              /* head in the data section */
243 };
244
245 #define PERF_EVENT_MISC_CPUMODE_MASK            (3 << 0)
246 #define PERF_EVENT_MISC_CPUMODE_UNKNOWN         (0 << 0)
247 #define PERF_EVENT_MISC_KERNEL                  (1 << 0)
248 #define PERF_EVENT_MISC_USER                    (2 << 0)
249 #define PERF_EVENT_MISC_HYPERVISOR              (3 << 0)
250 #define PERF_EVENT_MISC_OVERFLOW                (1 << 2)
251
252 struct perf_event_header {
253         __u32   type;
254         __u16   misc;
255         __u16   size;
256 };
257
258 enum perf_event_type {
259
260         /*
261          * The MMAP events record the PROT_EXEC mappings so that we can
262          * correlate userspace IPs to code. They have the following structure:
263          *
264          * struct {
265          *      struct perf_event_header        header;
266          *
267          *      u32                             pid, tid;
268          *      u64                             addr;
269          *      u64                             len;
270          *      u64                             pgoff;
271          *      char                            filename[];
272          * };
273          */
274         PERF_EVENT_MMAP                 = 1,
275
276         /*
277          * struct {
278          *      struct perf_event_header        header;
279          *
280          *      u32                             pid, tid;
281          *      char                            comm[];
282          * };
283          */
284         PERF_EVENT_COMM                 = 3,
285
286         /*
287          * struct {
288          *      struct perf_event_header        header;
289          *      u64                             time;
290          *      u64                             id;
291          *      u64                             sample_period;
292          * };
293          */
294         PERF_EVENT_PERIOD               = 4,
295
296         /*
297          * struct {
298          *      struct perf_event_header        header;
299          *      u64                             time;
300          *      u64                             id;
301          * };
302          */
303         PERF_EVENT_THROTTLE             = 5,
304         PERF_EVENT_UNTHROTTLE           = 6,
305
306         /*
307          * struct {
308          *      struct perf_event_header        header;
309          *      u32                             pid, ppid;
310          * };
311          */
312         PERF_EVENT_FORK                 = 7,
313
314         /*
315          * When header.misc & PERF_EVENT_MISC_OVERFLOW the event_type field
316          * will be PERF_RECORD_*
317          *
318          * struct {
319          *      struct perf_event_header        header;
320          *
321          *      { u64                   ip;       } && PERF_RECORD_IP
322          *      { u32                   pid, tid; } && PERF_RECORD_TID
323          *      { u64                   time;     } && PERF_RECORD_TIME
324          *      { u64                   addr;     } && PERF_RECORD_ADDR
325          *      { u64                   config;   } && PERF_RECORD_CONFIG
326          *      { u32                   cpu, res; } && PERF_RECORD_CPU
327          *
328          *      { u64                   nr;
329          *        { u64 id, val; }      cnt[nr];  } && PERF_RECORD_GROUP
330          *
331          *      { u16                   nr,
332          *                              hv,
333          *                              kernel,
334          *                              user;
335          *        u64                   ips[nr];  } && PERF_RECORD_CALLCHAIN
336          * };
337          */
338 };
339
340 #ifdef __KERNEL__
341 /*
342  * Kernel-internal data types and definitions:
343  */
344
345 #ifdef CONFIG_PERF_COUNTERS
346 # include <asm/perf_counter.h>
347 #endif
348
349 #include <linux/list.h>
350 #include <linux/mutex.h>
351 #include <linux/rculist.h>
352 #include <linux/rcupdate.h>
353 #include <linux/spinlock.h>
354 #include <linux/hrtimer.h>
355 #include <linux/fs.h>
356 #include <linux/pid_namespace.h>
357 #include <asm/atomic.h>
358
359 struct task_struct;
360
361 /**
362  * struct hw_perf_counter - performance counter hardware details:
363  */
364 struct hw_perf_counter {
365 #ifdef CONFIG_PERF_COUNTERS
366         union {
367                 struct { /* hardware */
368                         u64             config;
369                         unsigned long   config_base;
370                         unsigned long   counter_base;
371                         int             idx;
372                 };
373                 union { /* software */
374                         atomic64_t      count;
375                         struct hrtimer  hrtimer;
376                 };
377         };
378         atomic64_t                      prev_count;
379         u64                             sample_period;
380         u64                             last_period;
381         atomic64_t                      period_left;
382         u64                             interrupts;
383
384         u64                             freq_count;
385         u64                             freq_interrupts;
386         u64                             freq_stamp;
387 #endif
388 };
389
390 struct perf_counter;
391
392 /**
393  * struct pmu - generic performance monitoring unit
394  */
395 struct pmu {
396         int (*enable)                   (struct perf_counter *counter);
397         void (*disable)                 (struct perf_counter *counter);
398         void (*read)                    (struct perf_counter *counter);
399         void (*unthrottle)              (struct perf_counter *counter);
400 };
401
402 /**
403  * enum perf_counter_active_state - the states of a counter
404  */
405 enum perf_counter_active_state {
406         PERF_COUNTER_STATE_ERROR        = -2,
407         PERF_COUNTER_STATE_OFF          = -1,
408         PERF_COUNTER_STATE_INACTIVE     =  0,
409         PERF_COUNTER_STATE_ACTIVE       =  1,
410 };
411
412 struct file;
413
414 struct perf_mmap_data {
415         struct rcu_head                 rcu_head;
416         int                             nr_pages;       /* nr of data pages  */
417         int                             nr_locked;      /* nr pages mlocked  */
418
419         atomic_t                        poll;           /* POLL_ for wakeups */
420         atomic_t                        events;         /* event limit       */
421
422         atomic_long_t                   head;           /* write position    */
423         atomic_long_t                   done_head;      /* completed head    */
424
425         atomic_t                        lock;           /* concurrent writes */
426
427         atomic_t                        wakeup;         /* needs a wakeup    */
428
429         struct perf_counter_mmap_page   *user_page;
430         void                            *data_pages[0];
431 };
432
433 struct perf_pending_entry {
434         struct perf_pending_entry *next;
435         void (*func)(struct perf_pending_entry *);
436 };
437
438 /**
439  * struct perf_counter - performance counter kernel representation:
440  */
441 struct perf_counter {
442 #ifdef CONFIG_PERF_COUNTERS
443         struct list_head                list_entry;
444         struct list_head                event_entry;
445         struct list_head                sibling_list;
446         int                             nr_siblings;
447         struct perf_counter             *group_leader;
448         const struct pmu                *pmu;
449
450         enum perf_counter_active_state  state;
451         atomic64_t                      count;
452
453         /*
454          * These are the total time in nanoseconds that the counter
455          * has been enabled (i.e. eligible to run, and the task has
456          * been scheduled in, if this is a per-task counter)
457          * and running (scheduled onto the CPU), respectively.
458          *
459          * They are computed from tstamp_enabled, tstamp_running and
460          * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
461          */
462         u64                             total_time_enabled;
463         u64                             total_time_running;
464
465         /*
466          * These are timestamps used for computing total_time_enabled
467          * and total_time_running when the counter is in INACTIVE or
468          * ACTIVE state, measured in nanoseconds from an arbitrary point
469          * in time.
470          * tstamp_enabled: the notional time when the counter was enabled
471          * tstamp_running: the notional time when the counter was scheduled on
472          * tstamp_stopped: in INACTIVE state, the notional time when the
473          *      counter was scheduled off.
474          */
475         u64                             tstamp_enabled;
476         u64                             tstamp_running;
477         u64                             tstamp_stopped;
478
479         struct perf_counter_attr        attr;
480         struct hw_perf_counter          hw;
481
482         struct perf_counter_context     *ctx;
483         struct file                     *filp;
484
485         /*
486          * These accumulate total time (in nanoseconds) that children
487          * counters have been enabled and running, respectively.
488          */
489         atomic64_t                      child_total_time_enabled;
490         atomic64_t                      child_total_time_running;
491
492         /*
493          * Protect attach/detach and child_list:
494          */
495         struct mutex                    child_mutex;
496         struct list_head                child_list;
497         struct perf_counter             *parent;
498
499         int                             oncpu;
500         int                             cpu;
501
502         struct list_head                owner_entry;
503         struct task_struct              *owner;
504
505         /* mmap bits */
506         struct mutex                    mmap_mutex;
507         atomic_t                        mmap_count;
508         struct perf_mmap_data           *data;
509
510         /* poll related */
511         wait_queue_head_t               waitq;
512         struct fasync_struct            *fasync;
513
514         /* delayed work for NMIs and such */
515         int                             pending_wakeup;
516         int                             pending_kill;
517         int                             pending_disable;
518         struct perf_pending_entry       pending;
519
520         atomic_t                        event_limit;
521
522         void (*destroy)(struct perf_counter *);
523         struct rcu_head                 rcu_head;
524
525         struct pid_namespace            *ns;
526         u64                             id;
527 #endif
528 };
529
530 /**
531  * struct perf_counter_context - counter context structure
532  *
533  * Used as a container for task counters and CPU counters as well:
534  */
535 struct perf_counter_context {
536         /*
537          * Protect the states of the counters in the list,
538          * nr_active, and the list:
539          */
540         spinlock_t                      lock;
541         /*
542          * Protect the list of counters.  Locking either mutex or lock
543          * is sufficient to ensure the list doesn't change; to change
544          * the list you need to lock both the mutex and the spinlock.
545          */
546         struct mutex                    mutex;
547
548         struct list_head                counter_list;
549         struct list_head                event_list;
550         int                             nr_counters;
551         int                             nr_active;
552         int                             is_active;
553         atomic_t                        refcount;
554         struct task_struct              *task;
555
556         /*
557          * Context clock, runs when context enabled.
558          */
559         u64                             time;
560         u64                             timestamp;
561
562         /*
563          * These fields let us detect when two contexts have both
564          * been cloned (inherited) from a common ancestor.
565          */
566         struct perf_counter_context     *parent_ctx;
567         u64                             parent_gen;
568         u64                             generation;
569         int                             pin_count;
570         struct rcu_head                 rcu_head;
571 };
572
573 /**
574  * struct perf_counter_cpu_context - per cpu counter context structure
575  */
576 struct perf_cpu_context {
577         struct perf_counter_context     ctx;
578         struct perf_counter_context     *task_ctx;
579         int                             active_oncpu;
580         int                             max_pertask;
581         int                             exclusive;
582
583         /*
584          * Recursion avoidance:
585          *
586          * task, softirq, irq, nmi context
587          */
588         int                             recursion[4];
589 };
590
591 #ifdef CONFIG_PERF_COUNTERS
592
593 /*
594  * Set by architecture code:
595  */
596 extern int perf_max_counters;
597
598 extern const struct pmu *hw_perf_counter_init(struct perf_counter *counter);
599
600 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
601 extern void perf_counter_task_sched_out(struct task_struct *task,
602                                         struct task_struct *next, int cpu);
603 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
604 extern int perf_counter_init_task(struct task_struct *child);
605 extern void perf_counter_exit_task(struct task_struct *child);
606 extern void perf_counter_free_task(struct task_struct *task);
607 extern void perf_counter_do_pending(void);
608 extern void perf_counter_print_debug(void);
609 extern void __perf_disable(void);
610 extern bool __perf_enable(void);
611 extern void perf_disable(void);
612 extern void perf_enable(void);
613 extern int perf_counter_task_disable(void);
614 extern int perf_counter_task_enable(void);
615 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
616                struct perf_cpu_context *cpuctx,
617                struct perf_counter_context *ctx, int cpu);
618 extern void perf_counter_update_userpage(struct perf_counter *counter);
619
620 struct perf_sample_data {
621         struct pt_regs                  *regs;
622         u64                             addr;
623         u64                             period;
624 };
625
626 extern int perf_counter_overflow(struct perf_counter *counter, int nmi,
627                                  struct perf_sample_data *data);
628
629 /*
630  * Return 1 for a software counter, 0 for a hardware counter
631  */
632 static inline int is_software_counter(struct perf_counter *counter)
633 {
634         return (counter->attr.type != PERF_TYPE_RAW) &&
635                 (counter->attr.type != PERF_TYPE_HARDWARE) &&
636                 (counter->attr.type != PERF_TYPE_HW_CACHE);
637 }
638
639 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *, u64);
640
641 extern void __perf_counter_mmap(struct vm_area_struct *vma);
642
643 static inline void perf_counter_mmap(struct vm_area_struct *vma)
644 {
645         if (vma->vm_flags & VM_EXEC)
646                 __perf_counter_mmap(vma);
647 }
648
649 extern void perf_counter_comm(struct task_struct *tsk);
650 extern void perf_counter_fork(struct task_struct *tsk);
651
652 extern void perf_counter_task_migration(struct task_struct *task, int cpu);
653
654 #define MAX_STACK_DEPTH                 255
655
656 struct perf_callchain_entry {
657         u16                             nr;
658         u16                             hv;
659         u16                             kernel;
660         u16                             user;
661         u64                             ip[MAX_STACK_DEPTH];
662 };
663
664 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
665
666 extern int sysctl_perf_counter_paranoid;
667 extern int sysctl_perf_counter_mlock;
668 extern int sysctl_perf_counter_sample_rate;
669
670 extern void perf_counter_init(void);
671
672 #ifndef perf_misc_flags
673 #define perf_misc_flags(regs)   (user_mode(regs) ? PERF_EVENT_MISC_USER : \
674                                  PERF_EVENT_MISC_KERNEL)
675 #define perf_instruction_pointer(regs)  instruction_pointer(regs)
676 #endif
677
678 #else
679 static inline void
680 perf_counter_task_sched_in(struct task_struct *task, int cpu)           { }
681 static inline void
682 perf_counter_task_sched_out(struct task_struct *task,
683                             struct task_struct *next, int cpu)          { }
684 static inline void
685 perf_counter_task_tick(struct task_struct *task, int cpu)               { }
686 static inline int perf_counter_init_task(struct task_struct *child)     { return 0; }
687 static inline void perf_counter_exit_task(struct task_struct *child)    { }
688 static inline void perf_counter_free_task(struct task_struct *task)     { }
689 static inline void perf_counter_do_pending(void)                        { }
690 static inline void perf_counter_print_debug(void)                       { }
691 static inline void perf_disable(void)                                   { }
692 static inline void perf_enable(void)                                    { }
693 static inline int perf_counter_task_disable(void)       { return -EINVAL; }
694 static inline int perf_counter_task_enable(void)        { return -EINVAL; }
695
696 static inline void
697 perf_swcounter_event(u32 event, u64 nr, int nmi,
698                      struct pt_regs *regs, u64 addr)                    { }
699
700 static inline void perf_counter_mmap(struct vm_area_struct *vma)        { }
701 static inline void perf_counter_comm(struct task_struct *tsk)           { }
702 static inline void perf_counter_fork(struct task_struct *tsk)           { }
703 static inline void perf_counter_init(void)                              { }
704 static inline void perf_counter_task_migration(struct task_struct *task,
705                                                int cpu)                 { }
706 #endif
707
708 #endif /* __KERNEL__ */
709 #endif /* _LINUX_PERF_COUNTER_H */