perf_counter: Simplify and fix task migration counting
[deliverable/linux.git] / 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 the @data_head value should issue an rmb(), on
240 * SMP capable platforms, after reading this value -- see
241 * perf_counter_wakeup().
242 *
243 * When the mapping is PROT_WRITE the @data_tail value should be
244 * written by userspace to reflect the last read data. In this case
245 * the kernel will not over-write unread data.
246 */
247 __u64 data_head; /* head in the data section */
248 __u64 data_tail; /* user-space written tail */
249 };
250
251 #define PERF_EVENT_MISC_CPUMODE_MASK (3 << 0)
252 #define PERF_EVENT_MISC_CPUMODE_UNKNOWN (0 << 0)
253 #define PERF_EVENT_MISC_KERNEL (1 << 0)
254 #define PERF_EVENT_MISC_USER (2 << 0)
255 #define PERF_EVENT_MISC_HYPERVISOR (3 << 0)
256 #define PERF_EVENT_MISC_OVERFLOW (1 << 2)
257
258 struct perf_event_header {
259 __u32 type;
260 __u16 misc;
261 __u16 size;
262 };
263
264 enum perf_event_type {
265
266 /*
267 * The MMAP events record the PROT_EXEC mappings so that we can
268 * correlate userspace IPs to code. They have the following structure:
269 *
270 * struct {
271 * struct perf_event_header header;
272 *
273 * u32 pid, tid;
274 * u64 addr;
275 * u64 len;
276 * u64 pgoff;
277 * char filename[];
278 * };
279 */
280 PERF_EVENT_MMAP = 1,
281
282 /*
283 * struct {
284 * struct perf_event_header header;
285 * u64 id;
286 * u64 lost;
287 * };
288 */
289 PERF_EVENT_LOST = 2,
290
291 /*
292 * struct {
293 * struct perf_event_header header;
294 *
295 * u32 pid, tid;
296 * char comm[];
297 * };
298 */
299 PERF_EVENT_COMM = 3,
300
301 /*
302 * struct {
303 * struct perf_event_header header;
304 * u64 time;
305 * u64 id;
306 * u64 sample_period;
307 * };
308 */
309 PERF_EVENT_PERIOD = 4,
310
311 /*
312 * struct {
313 * struct perf_event_header header;
314 * u64 time;
315 * u64 id;
316 * };
317 */
318 PERF_EVENT_THROTTLE = 5,
319 PERF_EVENT_UNTHROTTLE = 6,
320
321 /*
322 * struct {
323 * struct perf_event_header header;
324 * u32 pid, ppid;
325 * };
326 */
327 PERF_EVENT_FORK = 7,
328
329 /*
330 * When header.misc & PERF_EVENT_MISC_OVERFLOW the event_type field
331 * will be PERF_SAMPLE_*
332 *
333 * struct {
334 * struct perf_event_header header;
335 *
336 * { u64 ip; } && PERF_SAMPLE_IP
337 * { u32 pid, tid; } && PERF_SAMPLE_TID
338 * { u64 time; } && PERF_SAMPLE_TIME
339 * { u64 addr; } && PERF_SAMPLE_ADDR
340 * { u64 config; } && PERF_SAMPLE_CONFIG
341 * { u32 cpu, res; } && PERF_SAMPLE_CPU
342 *
343 * { u64 nr;
344 * { u64 id, val; } cnt[nr]; } && PERF_SAMPLE_GROUP
345 *
346 * { u64 nr,
347 * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN
348 * };
349 */
350 };
351
352 enum perf_callchain_context {
353 PERF_CONTEXT_HV = (__u64)-32,
354 PERF_CONTEXT_KERNEL = (__u64)-128,
355 PERF_CONTEXT_USER = (__u64)-512,
356
357 PERF_CONTEXT_GUEST = (__u64)-2048,
358 PERF_CONTEXT_GUEST_KERNEL = (__u64)-2176,
359 PERF_CONTEXT_GUEST_USER = (__u64)-2560,
360
361 PERF_CONTEXT_MAX = (__u64)-4095,
362 };
363
364 #ifdef __KERNEL__
365 /*
366 * Kernel-internal data types and definitions:
367 */
368
369 #ifdef CONFIG_PERF_COUNTERS
370 # include <asm/perf_counter.h>
371 #endif
372
373 #include <linux/list.h>
374 #include <linux/mutex.h>
375 #include <linux/rculist.h>
376 #include <linux/rcupdate.h>
377 #include <linux/spinlock.h>
378 #include <linux/hrtimer.h>
379 #include <linux/fs.h>
380 #include <linux/pid_namespace.h>
381 #include <asm/atomic.h>
382
383 #define PERF_MAX_STACK_DEPTH 255
384
385 struct perf_callchain_entry {
386 __u64 nr;
387 __u64 ip[PERF_MAX_STACK_DEPTH];
388 };
389
390 struct task_struct;
391
392 /**
393 * struct hw_perf_counter - performance counter hardware details:
394 */
395 struct hw_perf_counter {
396 #ifdef CONFIG_PERF_COUNTERS
397 union {
398 struct { /* hardware */
399 u64 config;
400 unsigned long config_base;
401 unsigned long counter_base;
402 int idx;
403 };
404 union { /* software */
405 atomic64_t count;
406 struct hrtimer hrtimer;
407 };
408 };
409 atomic64_t prev_count;
410 u64 sample_period;
411 u64 last_period;
412 atomic64_t period_left;
413 u64 interrupts;
414
415 u64 freq_count;
416 u64 freq_interrupts;
417 u64 freq_stamp;
418 #endif
419 };
420
421 struct perf_counter;
422
423 /**
424 * struct pmu - generic performance monitoring unit
425 */
426 struct pmu {
427 int (*enable) (struct perf_counter *counter);
428 void (*disable) (struct perf_counter *counter);
429 void (*read) (struct perf_counter *counter);
430 void (*unthrottle) (struct perf_counter *counter);
431 };
432
433 /**
434 * enum perf_counter_active_state - the states of a counter
435 */
436 enum perf_counter_active_state {
437 PERF_COUNTER_STATE_ERROR = -2,
438 PERF_COUNTER_STATE_OFF = -1,
439 PERF_COUNTER_STATE_INACTIVE = 0,
440 PERF_COUNTER_STATE_ACTIVE = 1,
441 };
442
443 struct file;
444
445 struct perf_mmap_data {
446 struct rcu_head rcu_head;
447 int nr_pages; /* nr of data pages */
448 int writable; /* are we writable */
449 int nr_locked; /* nr pages mlocked */
450
451 atomic_t poll; /* POLL_ for wakeups */
452 atomic_t events; /* event limit */
453
454 atomic_long_t head; /* write position */
455 atomic_long_t done_head; /* completed head */
456
457 atomic_t lock; /* concurrent writes */
458 atomic_t wakeup; /* needs a wakeup */
459 atomic_t lost; /* nr records lost */
460
461 struct perf_counter_mmap_page *user_page;
462 void *data_pages[0];
463 };
464
465 struct perf_pending_entry {
466 struct perf_pending_entry *next;
467 void (*func)(struct perf_pending_entry *);
468 };
469
470 /**
471 * struct perf_counter - performance counter kernel representation:
472 */
473 struct perf_counter {
474 #ifdef CONFIG_PERF_COUNTERS
475 struct list_head list_entry;
476 struct list_head event_entry;
477 struct list_head sibling_list;
478 int nr_siblings;
479 struct perf_counter *group_leader;
480 const struct pmu *pmu;
481
482 enum perf_counter_active_state state;
483 atomic64_t count;
484
485 /*
486 * These are the total time in nanoseconds that the counter
487 * has been enabled (i.e. eligible to run, and the task has
488 * been scheduled in, if this is a per-task counter)
489 * and running (scheduled onto the CPU), respectively.
490 *
491 * They are computed from tstamp_enabled, tstamp_running and
492 * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
493 */
494 u64 total_time_enabled;
495 u64 total_time_running;
496
497 /*
498 * These are timestamps used for computing total_time_enabled
499 * and total_time_running when the counter is in INACTIVE or
500 * ACTIVE state, measured in nanoseconds from an arbitrary point
501 * in time.
502 * tstamp_enabled: the notional time when the counter was enabled
503 * tstamp_running: the notional time when the counter was scheduled on
504 * tstamp_stopped: in INACTIVE state, the notional time when the
505 * counter was scheduled off.
506 */
507 u64 tstamp_enabled;
508 u64 tstamp_running;
509 u64 tstamp_stopped;
510
511 struct perf_counter_attr attr;
512 struct hw_perf_counter hw;
513
514 struct perf_counter_context *ctx;
515 struct file *filp;
516
517 /*
518 * These accumulate total time (in nanoseconds) that children
519 * counters have been enabled and running, respectively.
520 */
521 atomic64_t child_total_time_enabled;
522 atomic64_t child_total_time_running;
523
524 /*
525 * Protect attach/detach and child_list:
526 */
527 struct mutex child_mutex;
528 struct list_head child_list;
529 struct perf_counter *parent;
530
531 int oncpu;
532 int cpu;
533
534 struct list_head owner_entry;
535 struct task_struct *owner;
536
537 /* mmap bits */
538 struct mutex mmap_mutex;
539 atomic_t mmap_count;
540 struct perf_mmap_data *data;
541
542 /* poll related */
543 wait_queue_head_t waitq;
544 struct fasync_struct *fasync;
545
546 /* delayed work for NMIs and such */
547 int pending_wakeup;
548 int pending_kill;
549 int pending_disable;
550 struct perf_pending_entry pending;
551
552 atomic_t event_limit;
553
554 void (*destroy)(struct perf_counter *);
555 struct rcu_head rcu_head;
556
557 struct pid_namespace *ns;
558 u64 id;
559 #endif
560 };
561
562 /**
563 * struct perf_counter_context - counter context structure
564 *
565 * Used as a container for task counters and CPU counters as well:
566 */
567 struct perf_counter_context {
568 /*
569 * Protect the states of the counters in the list,
570 * nr_active, and the list:
571 */
572 spinlock_t lock;
573 /*
574 * Protect the list of counters. Locking either mutex or lock
575 * is sufficient to ensure the list doesn't change; to change
576 * the list you need to lock both the mutex and the spinlock.
577 */
578 struct mutex mutex;
579
580 struct list_head counter_list;
581 struct list_head event_list;
582 int nr_counters;
583 int nr_active;
584 int is_active;
585 atomic_t refcount;
586 struct task_struct *task;
587
588 /*
589 * Context clock, runs when context enabled.
590 */
591 u64 time;
592 u64 timestamp;
593
594 /*
595 * These fields let us detect when two contexts have both
596 * been cloned (inherited) from a common ancestor.
597 */
598 struct perf_counter_context *parent_ctx;
599 u64 parent_gen;
600 u64 generation;
601 int pin_count;
602 struct rcu_head rcu_head;
603 };
604
605 /**
606 * struct perf_counter_cpu_context - per cpu counter context structure
607 */
608 struct perf_cpu_context {
609 struct perf_counter_context ctx;
610 struct perf_counter_context *task_ctx;
611 int active_oncpu;
612 int max_pertask;
613 int exclusive;
614
615 /*
616 * Recursion avoidance:
617 *
618 * task, softirq, irq, nmi context
619 */
620 int recursion[4];
621 };
622
623 #ifdef CONFIG_PERF_COUNTERS
624
625 /*
626 * Set by architecture code:
627 */
628 extern int perf_max_counters;
629
630 extern const struct pmu *hw_perf_counter_init(struct perf_counter *counter);
631
632 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
633 extern void perf_counter_task_sched_out(struct task_struct *task,
634 struct task_struct *next, int cpu);
635 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
636 extern int perf_counter_init_task(struct task_struct *child);
637 extern void perf_counter_exit_task(struct task_struct *child);
638 extern void perf_counter_free_task(struct task_struct *task);
639 extern void set_perf_counter_pending(void);
640 extern void perf_counter_do_pending(void);
641 extern void perf_counter_print_debug(void);
642 extern void __perf_disable(void);
643 extern bool __perf_enable(void);
644 extern void perf_disable(void);
645 extern void perf_enable(void);
646 extern int perf_counter_task_disable(void);
647 extern int perf_counter_task_enable(void);
648 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
649 struct perf_cpu_context *cpuctx,
650 struct perf_counter_context *ctx, int cpu);
651 extern void perf_counter_update_userpage(struct perf_counter *counter);
652
653 struct perf_sample_data {
654 struct pt_regs *regs;
655 u64 addr;
656 u64 period;
657 };
658
659 extern int perf_counter_overflow(struct perf_counter *counter, int nmi,
660 struct perf_sample_data *data);
661
662 /*
663 * Return 1 for a software counter, 0 for a hardware counter
664 */
665 static inline int is_software_counter(struct perf_counter *counter)
666 {
667 return (counter->attr.type != PERF_TYPE_RAW) &&
668 (counter->attr.type != PERF_TYPE_HARDWARE) &&
669 (counter->attr.type != PERF_TYPE_HW_CACHE);
670 }
671
672 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *, u64);
673
674 extern void __perf_counter_mmap(struct vm_area_struct *vma);
675
676 static inline void perf_counter_mmap(struct vm_area_struct *vma)
677 {
678 if (vma->vm_flags & VM_EXEC)
679 __perf_counter_mmap(vma);
680 }
681
682 extern void perf_counter_comm(struct task_struct *tsk);
683 extern void perf_counter_fork(struct task_struct *tsk);
684
685 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
686
687 extern int sysctl_perf_counter_paranoid;
688 extern int sysctl_perf_counter_mlock;
689 extern int sysctl_perf_counter_sample_rate;
690
691 extern void perf_counter_init(void);
692
693 #ifndef perf_misc_flags
694 #define perf_misc_flags(regs) (user_mode(regs) ? PERF_EVENT_MISC_USER : \
695 PERF_EVENT_MISC_KERNEL)
696 #define perf_instruction_pointer(regs) instruction_pointer(regs)
697 #endif
698
699 #else
700 static inline void
701 perf_counter_task_sched_in(struct task_struct *task, int cpu) { }
702 static inline void
703 perf_counter_task_sched_out(struct task_struct *task,
704 struct task_struct *next, int cpu) { }
705 static inline void
706 perf_counter_task_tick(struct task_struct *task, int cpu) { }
707 static inline int perf_counter_init_task(struct task_struct *child) { return 0; }
708 static inline void perf_counter_exit_task(struct task_struct *child) { }
709 static inline void perf_counter_free_task(struct task_struct *task) { }
710 static inline void perf_counter_do_pending(void) { }
711 static inline void perf_counter_print_debug(void) { }
712 static inline void perf_disable(void) { }
713 static inline void perf_enable(void) { }
714 static inline int perf_counter_task_disable(void) { return -EINVAL; }
715 static inline int perf_counter_task_enable(void) { return -EINVAL; }
716
717 static inline void
718 perf_swcounter_event(u32 event, u64 nr, int nmi,
719 struct pt_regs *regs, u64 addr) { }
720
721 static inline void perf_counter_mmap(struct vm_area_struct *vma) { }
722 static inline void perf_counter_comm(struct task_struct *tsk) { }
723 static inline void perf_counter_fork(struct task_struct *tsk) { }
724 static inline void perf_counter_init(void) { }
725 #endif
726
727 #endif /* __KERNEL__ */
728 #endif /* _LINUX_PERF_COUNTER_H */
This page took 0.062846 seconds and 5 git commands to generate.