perf_counter: Rename perf_counter_limit sysctl
[sfrench/cifs-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  * attr.type
26  */
27 enum perf_event_types {
28         PERF_TYPE_HARDWARE              = 0,
29         PERF_TYPE_SOFTWARE              = 1,
30         PERF_TYPE_TRACEPOINT            = 2,
31         PERF_TYPE_HW_CACHE              = 3,
32
33         /*
34          * available TYPE space, raw is the max value.
35          */
36
37         PERF_TYPE_RAW                   = 128,
38 };
39
40 /*
41  * Generalized performance counter event types, used by the attr.event_id
42  * parameter of the sys_perf_counter_open() syscall:
43  */
44 enum attr_ids {
45         /*
46          * Common hardware events, generalized by the kernel:
47          */
48         PERF_COUNT_CPU_CYCLES           = 0,
49         PERF_COUNT_INSTRUCTIONS         = 1,
50         PERF_COUNT_CACHE_REFERENCES     = 2,
51         PERF_COUNT_CACHE_MISSES         = 3,
52         PERF_COUNT_BRANCH_INSTRUCTIONS  = 4,
53         PERF_COUNT_BRANCH_MISSES        = 5,
54         PERF_COUNT_BUS_CYCLES           = 6,
55
56         PERF_HW_EVENTS_MAX              = 7,
57 };
58
59 /*
60  * Generalized hardware cache counters:
61  *
62  *       { L1-D, L1-I, L2, LLC, ITLB, DTLB, BPU } x
63  *       { read, write, prefetch } x
64  *       { accesses, misses }
65  */
66 enum hw_cache_id {
67         PERF_COUNT_HW_CACHE_L1D,
68         PERF_COUNT_HW_CACHE_L1I,
69         PERF_COUNT_HW_CACHE_L2,
70         PERF_COUNT_HW_CACHE_DTLB,
71         PERF_COUNT_HW_CACHE_ITLB,
72         PERF_COUNT_HW_CACHE_BPU,
73
74         PERF_COUNT_HW_CACHE_MAX,
75 };
76
77 enum hw_cache_op_id {
78         PERF_COUNT_HW_CACHE_OP_READ,
79         PERF_COUNT_HW_CACHE_OP_WRITE,
80         PERF_COUNT_HW_CACHE_OP_PREFETCH,
81
82         PERF_COUNT_HW_CACHE_OP_MAX,
83 };
84
85 enum hw_cache_op_result_id {
86         PERF_COUNT_HW_CACHE_RESULT_ACCESS,
87         PERF_COUNT_HW_CACHE_RESULT_MISS,
88
89         PERF_COUNT_HW_CACHE_RESULT_MAX,
90 };
91
92 /*
93  * Special "software" counters provided by the kernel, even if the hardware
94  * does not support performance counters. These counters measure various
95  * physical and sw events of the kernel (and allow the profiling of them as
96  * well):
97  */
98 enum sw_event_ids {
99         PERF_COUNT_CPU_CLOCK            = 0,
100         PERF_COUNT_TASK_CLOCK           = 1,
101         PERF_COUNT_PAGE_FAULTS          = 2,
102         PERF_COUNT_CONTEXT_SWITCHES     = 3,
103         PERF_COUNT_CPU_MIGRATIONS       = 4,
104         PERF_COUNT_PAGE_FAULTS_MIN      = 5,
105         PERF_COUNT_PAGE_FAULTS_MAJ      = 6,
106
107         PERF_SW_EVENTS_MAX              = 7,
108 };
109
110 /*
111  * Bits that can be set in attr.sample_type to request information
112  * in the overflow packets.
113  */
114 enum perf_counter_sample_format {
115         PERF_SAMPLE_IP                  = 1U << 0,
116         PERF_SAMPLE_TID                 = 1U << 1,
117         PERF_SAMPLE_TIME                = 1U << 2,
118         PERF_SAMPLE_ADDR                = 1U << 3,
119         PERF_SAMPLE_GROUP               = 1U << 4,
120         PERF_SAMPLE_CALLCHAIN           = 1U << 5,
121         PERF_SAMPLE_ID                  = 1U << 6,
122         PERF_SAMPLE_CPU                 = 1U << 7,
123         PERF_SAMPLE_PERIOD              = 1U << 8,
124 };
125
126 /*
127  * Bits that can be set in attr.read_format to request that
128  * reads on the counter should return the indicated quantities,
129  * in increasing order of bit value, after the counter value.
130  */
131 enum perf_counter_read_format {
132         PERF_FORMAT_TOTAL_TIME_ENABLED  =  1U << 0,
133         PERF_FORMAT_TOTAL_TIME_RUNNING  =  1U << 1,
134         PERF_FORMAT_ID                  =  1U << 2,
135 };
136
137 /*
138  * Hardware event to monitor via a performance monitoring counter:
139  */
140 struct perf_counter_attr {
141         /*
142          * Major type: hardware/software/tracepoint/etc.
143          */
144         __u32                   type;
145         __u32                   __reserved_1;
146
147         /*
148          * Type specific configuration information.
149          */
150         __u64                   config;
151
152         union {
153                 __u64           sample_period;
154                 __u64           sample_freq;
155         };
156
157         __u64                   sample_type;
158         __u64                   read_format;
159
160         __u64                   disabled       :  1, /* off by default        */
161                                 inherit        :  1, /* children inherit it   */
162                                 pinned         :  1, /* must always be on PMU */
163                                 exclusive      :  1, /* only group on PMU     */
164                                 exclude_user   :  1, /* don't count user      */
165                                 exclude_kernel :  1, /* ditto kernel          */
166                                 exclude_hv     :  1, /* ditto hypervisor      */
167                                 exclude_idle   :  1, /* don't count when idle */
168                                 mmap           :  1, /* include mmap data     */
169                                 comm           :  1, /* include comm data     */
170                                 freq           :  1, /* use freq, not period  */
171
172                                 __reserved_2   : 53;
173
174         __u32                   wakeup_events;  /* wakeup every n events */
175         __u32                   __reserved_3;
176
177         __u64                   __reserved_4;
178 };
179
180 /*
181  * Ioctls that can be done on a perf counter fd:
182  */
183 #define PERF_COUNTER_IOC_ENABLE         _IO ('$', 0)
184 #define PERF_COUNTER_IOC_DISABLE        _IO ('$', 1)
185 #define PERF_COUNTER_IOC_REFRESH        _IO ('$', 2)
186 #define PERF_COUNTER_IOC_RESET          _IO ('$', 3)
187 #define PERF_COUNTER_IOC_PERIOD         _IOW('$', 4, u64)
188
189 enum perf_counter_ioc_flags {
190         PERF_IOC_FLAG_GROUP             = 1U << 0,
191 };
192
193 /*
194  * Structure of the page that can be mapped via mmap
195  */
196 struct perf_counter_mmap_page {
197         __u32   version;                /* version number of this structure */
198         __u32   compat_version;         /* lowest version this is compat with */
199
200         /*
201          * Bits needed to read the hw counters in user-space.
202          *
203          *   u32 seq;
204          *   s64 count;
205          *
206          *   do {
207          *     seq = pc->lock;
208          *
209          *     barrier()
210          *     if (pc->index) {
211          *       count = pmc_read(pc->index - 1);
212          *       count += pc->offset;
213          *     } else
214          *       goto regular_read;
215          *
216          *     barrier();
217          *   } while (pc->lock != seq);
218          *
219          * NOTE: for obvious reason this only works on self-monitoring
220          *       processes.
221          */
222         __u32   lock;                   /* seqlock for synchronization */
223         __u32   index;                  /* hardware counter identifier */
224         __s64   offset;                 /* add to hardware counter value */
225
226         /*
227          * Control data for the mmap() data buffer.
228          *
229          * User-space reading this value should issue an rmb(), on SMP capable
230          * platforms, after reading this value -- see perf_counter_wakeup().
231          */
232         __u64   data_head;              /* head in the data section */
233 };
234
235 #define PERF_EVENT_MISC_CPUMODE_MASK    (3 << 0)
236 #define PERF_EVENT_MISC_CPUMODE_UNKNOWN (0 << 0)
237 #define PERF_EVENT_MISC_KERNEL          (1 << 0)
238 #define PERF_EVENT_MISC_USER            (2 << 0)
239 #define PERF_EVENT_MISC_HYPERVISOR      (3 << 0)
240 #define PERF_EVENT_MISC_OVERFLOW        (1 << 2)
241
242 struct perf_event_header {
243         __u32   type;
244         __u16   misc;
245         __u16   size;
246 };
247
248 enum perf_event_type {
249
250         /*
251          * The MMAP events record the PROT_EXEC mappings so that we can
252          * correlate userspace IPs to code. They have the following structure:
253          *
254          * struct {
255          *      struct perf_event_header        header;
256          *
257          *      u32                             pid, tid;
258          *      u64                             addr;
259          *      u64                             len;
260          *      u64                             pgoff;
261          *      char                            filename[];
262          * };
263          */
264         PERF_EVENT_MMAP                 = 1,
265
266         /*
267          * struct {
268          *      struct perf_event_header        header;
269          *
270          *      u32                             pid, tid;
271          *      char                            comm[];
272          * };
273          */
274         PERF_EVENT_COMM                 = 3,
275
276         /*
277          * struct {
278          *      struct perf_event_header        header;
279          *      u64                             time;
280          *      u64                             id;
281          *      u64                             sample_period;
282          * };
283          */
284         PERF_EVENT_PERIOD               = 4,
285
286         /*
287          * struct {
288          *      struct perf_event_header        header;
289          *      u64                             time;
290          * };
291          */
292         PERF_EVENT_THROTTLE             = 5,
293         PERF_EVENT_UNTHROTTLE           = 6,
294
295         /*
296          * struct {
297          *      struct perf_event_header        header;
298          *      u32                             pid, ppid;
299          * };
300          */
301         PERF_EVENT_FORK                 = 7,
302
303         /*
304          * When header.misc & PERF_EVENT_MISC_OVERFLOW the event_type field
305          * will be PERF_RECORD_*
306          *
307          * struct {
308          *      struct perf_event_header        header;
309          *
310          *      { u64                   ip;       } && PERF_RECORD_IP
311          *      { u32                   pid, tid; } && PERF_RECORD_TID
312          *      { u64                   time;     } && PERF_RECORD_TIME
313          *      { u64                   addr;     } && PERF_RECORD_ADDR
314          *      { u64                   config;   } && PERF_RECORD_CONFIG
315          *      { u32                   cpu, res; } && PERF_RECORD_CPU
316          *
317          *      { u64                   nr;
318          *        { u64 id, val; }      cnt[nr];  } && PERF_RECORD_GROUP
319          *
320          *      { u16                   nr,
321          *                              hv,
322          *                              kernel,
323          *                              user;
324          *        u64                   ips[nr];  } && PERF_RECORD_CALLCHAIN
325          * };
326          */
327 };
328
329 #ifdef __KERNEL__
330 /*
331  * Kernel-internal data types and definitions:
332  */
333
334 #ifdef CONFIG_PERF_COUNTERS
335 # include <asm/perf_counter.h>
336 #endif
337
338 #include <linux/list.h>
339 #include <linux/mutex.h>
340 #include <linux/rculist.h>
341 #include <linux/rcupdate.h>
342 #include <linux/spinlock.h>
343 #include <linux/hrtimer.h>
344 #include <linux/fs.h>
345 #include <linux/pid_namespace.h>
346 #include <asm/atomic.h>
347
348 struct task_struct;
349
350 /**
351  * struct hw_perf_counter - performance counter hardware details:
352  */
353 struct hw_perf_counter {
354 #ifdef CONFIG_PERF_COUNTERS
355         union {
356                 struct { /* hardware */
357                         u64                             config;
358                         unsigned long                   config_base;
359                         unsigned long                   counter_base;
360                         int                             idx;
361                 };
362                 union { /* software */
363                         atomic64_t                      count;
364                         struct hrtimer                  hrtimer;
365                 };
366         };
367         atomic64_t                      prev_count;
368         u64                             sample_period;
369         u64                             last_period;
370         atomic64_t                      period_left;
371         u64                             interrupts;
372
373         u64                             freq_count;
374         u64                             freq_interrupts;
375         u64                             freq_stamp;
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 struct perf_sample_data {
610         struct pt_regs          *regs;
611         u64                     addr;
612         u64                     period;
613 };
614
615 extern int perf_counter_overflow(struct perf_counter *counter, int nmi,
616                                  struct perf_sample_data *data);
617
618 /*
619  * Return 1 for a software counter, 0 for a hardware counter
620  */
621 static inline int is_software_counter(struct perf_counter *counter)
622 {
623         return (counter->attr.type != PERF_TYPE_RAW) &&
624                 (counter->attr.type != PERF_TYPE_HARDWARE);
625 }
626
627 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *, u64);
628
629 extern void __perf_counter_mmap(struct vm_area_struct *vma);
630
631 static inline void perf_counter_mmap(struct vm_area_struct *vma)
632 {
633         if (vma->vm_flags & VM_EXEC)
634                 __perf_counter_mmap(vma);
635 }
636
637 extern void perf_counter_comm(struct task_struct *tsk);
638 extern void perf_counter_fork(struct task_struct *tsk);
639
640 extern void perf_counter_task_migration(struct task_struct *task, int cpu);
641
642 #define MAX_STACK_DEPTH         255
643
644 struct perf_callchain_entry {
645         u16     nr, hv, kernel, user;
646         u64     ip[MAX_STACK_DEPTH];
647 };
648
649 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
650
651 extern int sysctl_perf_counter_paranoid;
652 extern int sysctl_perf_counter_mlock;
653 extern int sysctl_perf_counter_sample_rate;
654
655 extern void perf_counter_init(void);
656
657 #ifndef perf_misc_flags
658 #define perf_misc_flags(regs)   (user_mode(regs) ? PERF_EVENT_MISC_USER : \
659                                  PERF_EVENT_MISC_KERNEL)
660 #define perf_instruction_pointer(regs)  instruction_pointer(regs)
661 #endif
662
663 #else
664 static inline void
665 perf_counter_task_sched_in(struct task_struct *task, int cpu)           { }
666 static inline void
667 perf_counter_task_sched_out(struct task_struct *task,
668                             struct task_struct *next, int cpu)          { }
669 static inline void
670 perf_counter_task_tick(struct task_struct *task, int cpu)               { }
671 static inline int perf_counter_init_task(struct task_struct *child)     { return 0; }
672 static inline void perf_counter_exit_task(struct task_struct *child)    { }
673 static inline void perf_counter_free_task(struct task_struct *task)     { }
674 static inline void perf_counter_do_pending(void)                        { }
675 static inline void perf_counter_print_debug(void)                       { }
676 static inline void perf_disable(void)                                   { }
677 static inline void perf_enable(void)                                    { }
678 static inline int perf_counter_task_disable(void)       { return -EINVAL; }
679 static inline int perf_counter_task_enable(void)        { return -EINVAL; }
680
681 static inline void
682 perf_swcounter_event(u32 event, u64 nr, int nmi,
683                      struct pt_regs *regs, u64 addr)                    { }
684
685 static inline void perf_counter_mmap(struct vm_area_struct *vma)        { }
686 static inline void perf_counter_comm(struct task_struct *tsk)           { }
687 static inline void perf_counter_fork(struct task_struct *tsk)           { }
688 static inline void perf_counter_init(void)                              { }
689 static inline void perf_counter_task_migration(struct task_struct *task,
690                                                int cpu)                 { }
691 #endif
692
693 #endif /* __KERNEL__ */
694 #endif /* _LINUX_PERF_COUNTER_H */