perf_counter: Better align code
[safe/jmp/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
125 /*
126  * Bits that can be set in attr.read_format to request that
127  * reads on the counter should return the indicated quantities,
128  * in increasing order of bit value, after the counter value.
129  */
130 enum perf_counter_read_format {
131         PERF_FORMAT_TOTAL_TIME_ENABLED          = 1U << 0,
132         PERF_FORMAT_TOTAL_TIME_RUNNING          = 1U << 1,
133         PERF_FORMAT_ID                          = 1U << 2,
134 };
135
136 /*
137  * Hardware event to monitor via a performance monitoring counter:
138  */
139 struct perf_counter_attr {
140         /*
141          * Major type: hardware/software/tracepoint/etc.
142          */
143         __u32                   type;
144         __u32                   __reserved_1;
145
146         /*
147          * Type specific configuration information.
148          */
149         __u64                   config;
150
151         union {
152                 __u64           sample_period;
153                 __u64           sample_freq;
154         };
155
156         __u64                   sample_type;
157         __u64                   read_format;
158
159         __u64                   disabled       :  1, /* off by default        */
160                                 inherit        :  1, /* children inherit it   */
161                                 pinned         :  1, /* must always be on PMU */
162                                 exclusive      :  1, /* only group on PMU     */
163                                 exclude_user   :  1, /* don't count user      */
164                                 exclude_kernel :  1, /* ditto kernel          */
165                                 exclude_hv     :  1, /* ditto hypervisor      */
166                                 exclude_idle   :  1, /* don't count when idle */
167                                 mmap           :  1, /* include mmap data     */
168                                 comm           :  1, /* include comm data     */
169                                 freq           :  1, /* use freq, not period  */
170
171                                 __reserved_2   : 53;
172
173         __u32                   wakeup_events;  /* wakeup every n events */
174         __u32                   __reserved_3;
175
176         __u64                   __reserved_4;
177 };
178
179 /*
180  * Ioctls that can be done on a perf counter fd:
181  */
182 #define PERF_COUNTER_IOC_ENABLE         _IO ('$', 0)
183 #define PERF_COUNTER_IOC_DISABLE        _IO ('$', 1)
184 #define PERF_COUNTER_IOC_REFRESH        _IO ('$', 2)
185 #define PERF_COUNTER_IOC_RESET          _IO ('$', 3)
186 #define PERF_COUNTER_IOC_PERIOD         _IOW('$', 4, u64)
187
188 enum perf_counter_ioc_flags {
189         PERF_IOC_FLAG_GROUP             = 1U << 0,
190 };
191
192 /*
193  * Structure of the page that can be mapped via mmap
194  */
195 struct perf_counter_mmap_page {
196         __u32   version;                /* version number of this structure */
197         __u32   compat_version;         /* lowest version this is compat with */
198
199         /*
200          * Bits needed to read the hw counters in user-space.
201          *
202          *   u32 seq;
203          *   s64 count;
204          *
205          *   do {
206          *     seq = pc->lock;
207          *
208          *     barrier()
209          *     if (pc->index) {
210          *       count = pmc_read(pc->index - 1);
211          *       count += pc->offset;
212          *     } else
213          *       goto regular_read;
214          *
215          *     barrier();
216          *   } while (pc->lock != seq);
217          *
218          * NOTE: for obvious reason this only works on self-monitoring
219          *       processes.
220          */
221         __u32   lock;                   /* seqlock for synchronization */
222         __u32   index;                  /* hardware counter identifier */
223         __s64   offset;                 /* add to hardware counter value */
224
225         /*
226          * Control data for the mmap() data buffer.
227          *
228          * User-space reading this value should issue an rmb(), on SMP capable
229          * platforms, after reading this value -- see perf_counter_wakeup().
230          */
231         __u64   data_head;              /* head in the data section */
232 };
233
234 #define PERF_EVENT_MISC_CPUMODE_MASK            (3 << 0)
235 #define PERF_EVENT_MISC_CPUMODE_UNKNOWN         (0 << 0)
236 #define PERF_EVENT_MISC_KERNEL                  (1 << 0)
237 #define PERF_EVENT_MISC_USER                    (2 << 0)
238 #define PERF_EVENT_MISC_HYPERVISOR              (3 << 0)
239 #define PERF_EVENT_MISC_OVERFLOW                (1 << 2)
240
241 struct perf_event_header {
242         __u32   type;
243         __u16   misc;
244         __u16   size;
245 };
246
247 enum perf_event_type {
248
249         /*
250          * The MMAP events record the PROT_EXEC mappings so that we can
251          * correlate userspace IPs to code. They have the following structure:
252          *
253          * struct {
254          *      struct perf_event_header        header;
255          *
256          *      u32                             pid, tid;
257          *      u64                             addr;
258          *      u64                             len;
259          *      u64                             pgoff;
260          *      char                            filename[];
261          * };
262          */
263         PERF_EVENT_MMAP                 = 1,
264
265         /*
266          * struct {
267          *      struct perf_event_header        header;
268          *
269          *      u32                             pid, tid;
270          *      char                            comm[];
271          * };
272          */
273         PERF_EVENT_COMM                 = 3,
274
275         /*
276          * struct {
277          *      struct perf_event_header        header;
278          *      u64                             time;
279          *      u64                             id;
280          *      u64                             sample_period;
281          * };
282          */
283         PERF_EVENT_PERIOD               = 4,
284
285         /*
286          * struct {
287          *      struct perf_event_header        header;
288          *      u64                             time;
289          * };
290          */
291         PERF_EVENT_THROTTLE             = 5,
292         PERF_EVENT_UNTHROTTLE           = 6,
293
294         /*
295          * struct {
296          *      struct perf_event_header        header;
297          *      u32                             pid, ppid;
298          * };
299          */
300         PERF_EVENT_FORK                 = 7,
301
302         /*
303          * When header.misc & PERF_EVENT_MISC_OVERFLOW the event_type field
304          * will be PERF_RECORD_*
305          *
306          * struct {
307          *      struct perf_event_header        header;
308          *
309          *      { u64                   ip;       } && PERF_RECORD_IP
310          *      { u32                   pid, tid; } && PERF_RECORD_TID
311          *      { u64                   time;     } && PERF_RECORD_TIME
312          *      { u64                   addr;     } && PERF_RECORD_ADDR
313          *      { u64                   config;   } && PERF_RECORD_CONFIG
314          *      { u32                   cpu, res; } && PERF_RECORD_CPU
315          *
316          *      { u64                   nr;
317          *        { u64 id, val; }      cnt[nr];  } && PERF_RECORD_GROUP
318          *
319          *      { u16                   nr,
320          *                              hv,
321          *                              kernel,
322          *                              user;
323          *        u64                   ips[nr];  } && PERF_RECORD_CALLCHAIN
324          * };
325          */
326 };
327
328 #ifdef __KERNEL__
329 /*
330  * Kernel-internal data types and definitions:
331  */
332
333 #ifdef CONFIG_PERF_COUNTERS
334 # include <asm/perf_counter.h>
335 #endif
336
337 #include <linux/list.h>
338 #include <linux/mutex.h>
339 #include <linux/rculist.h>
340 #include <linux/rcupdate.h>
341 #include <linux/spinlock.h>
342 #include <linux/hrtimer.h>
343 #include <linux/fs.h>
344 #include <linux/pid_namespace.h>
345 #include <asm/atomic.h>
346
347 struct task_struct;
348
349 /**
350  * struct hw_perf_counter - performance counter hardware details:
351  */
352 struct hw_perf_counter {
353 #ifdef CONFIG_PERF_COUNTERS
354         union {
355                 struct { /* hardware */
356                         u64             config;
357                         unsigned long   config_base;
358                         unsigned long   counter_base;
359                         int             idx;
360                 };
361                 union { /* software */
362                         atomic64_t      count;
363                         struct hrtimer  hrtimer;
364                 };
365         };
366         atomic64_t                      prev_count;
367         u64                             sample_period;
368         u64                             last_period;
369         atomic64_t                      period_left;
370         u64                             interrupts;
371
372         u64                             freq_count;
373         u64                             freq_interrupts;
374         u64                             freq_stamp;
375 #endif
376 };
377
378 struct perf_counter;
379
380 /**
381  * struct pmu - generic performance monitoring unit
382  */
383 struct pmu {
384         int (*enable)                   (struct perf_counter *counter);
385         void (*disable)                 (struct perf_counter *counter);
386         void (*read)                    (struct perf_counter *counter);
387         void (*unthrottle)              (struct perf_counter *counter);
388 };
389
390 /**
391  * enum perf_counter_active_state - the states of a counter
392  */
393 enum perf_counter_active_state {
394         PERF_COUNTER_STATE_ERROR        = -2,
395         PERF_COUNTER_STATE_OFF          = -1,
396         PERF_COUNTER_STATE_INACTIVE     =  0,
397         PERF_COUNTER_STATE_ACTIVE       =  1,
398 };
399
400 struct file;
401
402 struct perf_mmap_data {
403         struct rcu_head                 rcu_head;
404         int                             nr_pages;       /* nr of data pages  */
405         int                             nr_locked;      /* nr pages mlocked  */
406
407         atomic_t                        poll;           /* POLL_ for wakeups */
408         atomic_t                        events;         /* event limit       */
409
410         atomic_long_t                   head;           /* write position    */
411         atomic_long_t                   done_head;      /* completed head    */
412
413         atomic_t                        lock;           /* concurrent writes */
414
415         atomic_t                        wakeup;         /* needs a wakeup    */
416
417         struct perf_counter_mmap_page   *user_page;
418         void                            *data_pages[0];
419 };
420
421 struct perf_pending_entry {
422         struct perf_pending_entry *next;
423         void (*func)(struct perf_pending_entry *);
424 };
425
426 /**
427  * struct perf_counter - performance counter kernel representation:
428  */
429 struct perf_counter {
430 #ifdef CONFIG_PERF_COUNTERS
431         struct list_head                list_entry;
432         struct list_head                event_entry;
433         struct list_head                sibling_list;
434         int                             nr_siblings;
435         struct perf_counter             *group_leader;
436         const struct pmu                *pmu;
437
438         enum perf_counter_active_state  state;
439         atomic64_t                      count;
440
441         /*
442          * These are the total time in nanoseconds that the counter
443          * has been enabled (i.e. eligible to run, and the task has
444          * been scheduled in, if this is a per-task counter)
445          * and running (scheduled onto the CPU), respectively.
446          *
447          * They are computed from tstamp_enabled, tstamp_running and
448          * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
449          */
450         u64                             total_time_enabled;
451         u64                             total_time_running;
452
453         /*
454          * These are timestamps used for computing total_time_enabled
455          * and total_time_running when the counter is in INACTIVE or
456          * ACTIVE state, measured in nanoseconds from an arbitrary point
457          * in time.
458          * tstamp_enabled: the notional time when the counter was enabled
459          * tstamp_running: the notional time when the counter was scheduled on
460          * tstamp_stopped: in INACTIVE state, the notional time when the
461          *      counter was scheduled off.
462          */
463         u64                             tstamp_enabled;
464         u64                             tstamp_running;
465         u64                             tstamp_stopped;
466
467         struct perf_counter_attr        attr;
468         struct hw_perf_counter          hw;
469
470         struct perf_counter_context     *ctx;
471         struct file                     *filp;
472
473         /*
474          * These accumulate total time (in nanoseconds) that children
475          * counters have been enabled and running, respectively.
476          */
477         atomic64_t                      child_total_time_enabled;
478         atomic64_t                      child_total_time_running;
479
480         /*
481          * Protect attach/detach and child_list:
482          */
483         struct mutex                    child_mutex;
484         struct list_head                child_list;
485         struct perf_counter             *parent;
486
487         int                             oncpu;
488         int                             cpu;
489
490         struct list_head                owner_entry;
491         struct task_struct              *owner;
492
493         /* mmap bits */
494         struct mutex                    mmap_mutex;
495         atomic_t                        mmap_count;
496         struct perf_mmap_data           *data;
497
498         /* poll related */
499         wait_queue_head_t               waitq;
500         struct fasync_struct            *fasync;
501
502         /* delayed work for NMIs and such */
503         int                             pending_wakeup;
504         int                             pending_kill;
505         int                             pending_disable;
506         struct perf_pending_entry       pending;
507
508         atomic_t                        event_limit;
509
510         void (*destroy)(struct perf_counter *);
511         struct rcu_head                 rcu_head;
512
513         struct pid_namespace            *ns;
514         u64                             id;
515 #endif
516 };
517
518 /**
519  * struct perf_counter_context - counter context structure
520  *
521  * Used as a container for task counters and CPU counters as well:
522  */
523 struct perf_counter_context {
524         /*
525          * Protect the states of the counters in the list,
526          * nr_active, and the list:
527          */
528         spinlock_t                      lock;
529         /*
530          * Protect the list of counters.  Locking either mutex or lock
531          * is sufficient to ensure the list doesn't change; to change
532          * the list you need to lock both the mutex and the spinlock.
533          */
534         struct mutex                    mutex;
535
536         struct list_head                counter_list;
537         struct list_head                event_list;
538         int                             nr_counters;
539         int                             nr_active;
540         int                             is_active;
541         atomic_t                        refcount;
542         struct task_struct              *task;
543
544         /*
545          * Context clock, runs when context enabled.
546          */
547         u64                             time;
548         u64                             timestamp;
549
550         /*
551          * These fields let us detect when two contexts have both
552          * been cloned (inherited) from a common ancestor.
553          */
554         struct perf_counter_context     *parent_ctx;
555         u64                             parent_gen;
556         u64                             generation;
557         int                             pin_count;
558         struct rcu_head                 rcu_head;
559 };
560
561 /**
562  * struct perf_counter_cpu_context - per cpu counter context structure
563  */
564 struct perf_cpu_context {
565         struct perf_counter_context     ctx;
566         struct perf_counter_context     *task_ctx;
567         int                             active_oncpu;
568         int                             max_pertask;
569         int                             exclusive;
570
571         /*
572          * Recursion avoidance:
573          *
574          * task, softirq, irq, nmi context
575          */
576         int                             recursion[4];
577 };
578
579 #ifdef CONFIG_PERF_COUNTERS
580
581 /*
582  * Set by architecture code:
583  */
584 extern int perf_max_counters;
585
586 extern const struct pmu *hw_perf_counter_init(struct perf_counter *counter);
587
588 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
589 extern void perf_counter_task_sched_out(struct task_struct *task,
590                                         struct task_struct *next, int cpu);
591 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
592 extern int perf_counter_init_task(struct task_struct *child);
593 extern void perf_counter_exit_task(struct task_struct *child);
594 extern void perf_counter_free_task(struct task_struct *task);
595 extern void perf_counter_do_pending(void);
596 extern void perf_counter_print_debug(void);
597 extern void __perf_disable(void);
598 extern bool __perf_enable(void);
599 extern void perf_disable(void);
600 extern void perf_enable(void);
601 extern int perf_counter_task_disable(void);
602 extern int perf_counter_task_enable(void);
603 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
604                struct perf_cpu_context *cpuctx,
605                struct perf_counter_context *ctx, int cpu);
606 extern void perf_counter_update_userpage(struct perf_counter *counter);
607
608 struct perf_sample_data {
609         struct pt_regs                  *regs;
610         u64                             addr;
611         u64                             period;
612 };
613
614 extern int perf_counter_overflow(struct perf_counter *counter, int nmi,
615                                  struct perf_sample_data *data);
616
617 /*
618  * Return 1 for a software counter, 0 for a hardware counter
619  */
620 static inline int is_software_counter(struct perf_counter *counter)
621 {
622         return (counter->attr.type != PERF_TYPE_RAW) &&
623                 (counter->attr.type != PERF_TYPE_HARDWARE);
624 }
625
626 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *, u64);
627
628 extern void __perf_counter_mmap(struct vm_area_struct *vma);
629
630 static inline void perf_counter_mmap(struct vm_area_struct *vma)
631 {
632         if (vma->vm_flags & VM_EXEC)
633                 __perf_counter_mmap(vma);
634 }
635
636 extern void perf_counter_comm(struct task_struct *tsk);
637 extern void perf_counter_fork(struct task_struct *tsk);
638
639 extern void perf_counter_task_migration(struct task_struct *task, int cpu);
640
641 #define MAX_STACK_DEPTH                 255
642
643 struct perf_callchain_entry {
644         u16                             nr;
645         u16                             hv;
646         u16                             kernel;
647         u16                             user;
648         u64                             ip[MAX_STACK_DEPTH];
649 };
650
651 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
652
653 extern int sysctl_perf_counter_paranoid;
654 extern int sysctl_perf_counter_mlock;
655 extern int sysctl_perf_counter_sample_rate;
656
657 extern void perf_counter_init(void);
658
659 #ifndef perf_misc_flags
660 #define perf_misc_flags(regs)   (user_mode(regs) ? PERF_EVENT_MISC_USER : \
661                                  PERF_EVENT_MISC_KERNEL)
662 #define perf_instruction_pointer(regs)  instruction_pointer(regs)
663 #endif
664
665 #else
666 static inline void
667 perf_counter_task_sched_in(struct task_struct *task, int cpu)           { }
668 static inline void
669 perf_counter_task_sched_out(struct task_struct *task,
670                             struct task_struct *next, int cpu)          { }
671 static inline void
672 perf_counter_task_tick(struct task_struct *task, int cpu)               { }
673 static inline int perf_counter_init_task(struct task_struct *child)     { return 0; }
674 static inline void perf_counter_exit_task(struct task_struct *child)    { }
675 static inline void perf_counter_free_task(struct task_struct *task)     { }
676 static inline void perf_counter_do_pending(void)                        { }
677 static inline void perf_counter_print_debug(void)                       { }
678 static inline void perf_disable(void)                                   { }
679 static inline void perf_enable(void)                                    { }
680 static inline int perf_counter_task_disable(void)       { return -EINVAL; }
681 static inline int perf_counter_task_enable(void)        { return -EINVAL; }
682
683 static inline void
684 perf_swcounter_event(u32 event, u64 nr, int nmi,
685                      struct pt_regs *regs, u64 addr)                    { }
686
687 static inline void perf_counter_mmap(struct vm_area_struct *vma)        { }
688 static inline void perf_counter_comm(struct task_struct *tsk)           { }
689 static inline void perf_counter_fork(struct task_struct *tsk)           { }
690 static inline void perf_counter_init(void)                              { }
691 static inline void perf_counter_task_migration(struct task_struct *task,
692                                                int cpu)                 { }
693 #endif
694
695 #endif /* __KERNEL__ */
696 #endif /* _LINUX_PERF_COUNTER_H */