perf_counter: Remove the last nmi/irq bits
[deliverable/linux.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 * hw_event.type
26 */
27 enum perf_event_types {
28 PERF_TYPE_HARDWARE = 0,
29 PERF_TYPE_SOFTWARE = 1,
30 PERF_TYPE_TRACEPOINT = 2,
31
32 /*
33 * available TYPE space, raw is the max value.
34 */
35
36 PERF_TYPE_RAW = 128,
37 };
38
39 /*
40 * Generalized performance counter event types, used by the hw_event.event_id
41 * parameter of the sys_perf_counter_open() syscall:
42 */
43 enum hw_event_ids {
44 /*
45 * Common hardware events, generalized by the kernel:
46 */
47 PERF_COUNT_CPU_CYCLES = 0,
48 PERF_COUNT_INSTRUCTIONS = 1,
49 PERF_COUNT_CACHE_REFERENCES = 2,
50 PERF_COUNT_CACHE_MISSES = 3,
51 PERF_COUNT_BRANCH_INSTRUCTIONS = 4,
52 PERF_COUNT_BRANCH_MISSES = 5,
53 PERF_COUNT_BUS_CYCLES = 6,
54
55 PERF_HW_EVENTS_MAX = 7,
56 };
57
58 /*
59 * Special "software" counters provided by the kernel, even if the hardware
60 * does not support performance counters. These counters measure various
61 * physical and sw events of the kernel (and allow the profiling of them as
62 * well):
63 */
64 enum sw_event_ids {
65 PERF_COUNT_CPU_CLOCK = 0,
66 PERF_COUNT_TASK_CLOCK = 1,
67 PERF_COUNT_PAGE_FAULTS = 2,
68 PERF_COUNT_CONTEXT_SWITCHES = 3,
69 PERF_COUNT_CPU_MIGRATIONS = 4,
70 PERF_COUNT_PAGE_FAULTS_MIN = 5,
71 PERF_COUNT_PAGE_FAULTS_MAJ = 6,
72
73 PERF_SW_EVENTS_MAX = 7,
74 };
75
76 #define __PERF_COUNTER_MASK(name) \
77 (((1ULL << PERF_COUNTER_##name##_BITS) - 1) << \
78 PERF_COUNTER_##name##_SHIFT)
79
80 #define PERF_COUNTER_RAW_BITS 1
81 #define PERF_COUNTER_RAW_SHIFT 63
82 #define PERF_COUNTER_RAW_MASK __PERF_COUNTER_MASK(RAW)
83
84 #define PERF_COUNTER_CONFIG_BITS 63
85 #define PERF_COUNTER_CONFIG_SHIFT 0
86 #define PERF_COUNTER_CONFIG_MASK __PERF_COUNTER_MASK(CONFIG)
87
88 #define PERF_COUNTER_TYPE_BITS 7
89 #define PERF_COUNTER_TYPE_SHIFT 56
90 #define PERF_COUNTER_TYPE_MASK __PERF_COUNTER_MASK(TYPE)
91
92 #define PERF_COUNTER_EVENT_BITS 56
93 #define PERF_COUNTER_EVENT_SHIFT 0
94 #define PERF_COUNTER_EVENT_MASK __PERF_COUNTER_MASK(EVENT)
95
96 /*
97 * Bits that can be set in hw_event.sample_type to request information
98 * in the overflow packets.
99 */
100 enum perf_counter_sample_format {
101 PERF_SAMPLE_IP = 1U << 0,
102 PERF_SAMPLE_TID = 1U << 1,
103 PERF_SAMPLE_TIME = 1U << 2,
104 PERF_SAMPLE_ADDR = 1U << 3,
105 PERF_SAMPLE_GROUP = 1U << 4,
106 PERF_SAMPLE_CALLCHAIN = 1U << 5,
107 PERF_SAMPLE_CONFIG = 1U << 6,
108 PERF_SAMPLE_CPU = 1U << 7,
109 };
110
111 /*
112 * Bits that can be set in hw_event.read_format to request that
113 * reads on the counter should return the indicated quantities,
114 * in increasing order of bit value, after the counter value.
115 */
116 enum perf_counter_read_format {
117 PERF_FORMAT_TOTAL_TIME_ENABLED = 1U << 0,
118 PERF_FORMAT_TOTAL_TIME_RUNNING = 1U << 1,
119 PERF_FORMAT_ID = 1U << 2,
120 };
121
122 /*
123 * Hardware event to monitor via a performance monitoring counter:
124 */
125 struct perf_counter_hw_event {
126 /*
127 * The MSB of the config word signifies if the rest contains cpu
128 * specific (raw) counter configuration data, if unset, the next
129 * 7 bits are an event type and the rest of the bits are the event
130 * identifier.
131 */
132 __u64 config;
133
134 union {
135 __u64 sample_period;
136 __u64 sample_freq;
137 };
138
139 __u64 sample_type;
140 __u64 read_format;
141
142 __u64 disabled : 1, /* off by default */
143 inherit : 1, /* children inherit it */
144 pinned : 1, /* must always be on PMU */
145 exclusive : 1, /* only group on PMU */
146 exclude_user : 1, /* don't count user */
147 exclude_kernel : 1, /* ditto kernel */
148 exclude_hv : 1, /* ditto hypervisor */
149 exclude_idle : 1, /* don't count when idle */
150 mmap : 1, /* include mmap data */
151 munmap : 1, /* include munmap data */
152 comm : 1, /* include comm data */
153 freq : 1, /* use freq, not period */
154
155 __reserved_1 : 52;
156
157 __u32 wakeup_events; /* wakeup every n events */
158 __u32 __reserved_2;
159
160 __u64 __reserved_3;
161 __u64 __reserved_4;
162 };
163
164 /*
165 * Ioctls that can be done on a perf counter fd:
166 */
167 #define PERF_COUNTER_IOC_ENABLE _IOW('$', 0, u32)
168 #define PERF_COUNTER_IOC_DISABLE _IOW('$', 1, u32)
169 #define PERF_COUNTER_IOC_REFRESH _IOW('$', 2, u32)
170 #define PERF_COUNTER_IOC_RESET _IOW('$', 3, u32)
171
172 enum perf_counter_ioc_flags {
173 PERF_IOC_FLAG_GROUP = 1U << 0,
174 };
175
176 /*
177 * Structure of the page that can be mapped via mmap
178 */
179 struct perf_counter_mmap_page {
180 __u32 version; /* version number of this structure */
181 __u32 compat_version; /* lowest version this is compat with */
182
183 /*
184 * Bits needed to read the hw counters in user-space.
185 *
186 * u32 seq;
187 * s64 count;
188 *
189 * do {
190 * seq = pc->lock;
191 *
192 * barrier()
193 * if (pc->index) {
194 * count = pmc_read(pc->index - 1);
195 * count += pc->offset;
196 * } else
197 * goto regular_read;
198 *
199 * barrier();
200 * } while (pc->lock != seq);
201 *
202 * NOTE: for obvious reason this only works on self-monitoring
203 * processes.
204 */
205 __u32 lock; /* seqlock for synchronization */
206 __u32 index; /* hardware counter identifier */
207 __s64 offset; /* add to hardware counter value */
208
209 /*
210 * Control data for the mmap() data buffer.
211 *
212 * User-space reading this value should issue an rmb(), on SMP capable
213 * platforms, after reading this value -- see perf_counter_wakeup().
214 */
215 __u32 data_head; /* head in the data section */
216 };
217
218 #define PERF_EVENT_MISC_CPUMODE_MASK (3 << 0)
219 #define PERF_EVENT_MISC_CPUMODE_UNKNOWN (0 << 0)
220 #define PERF_EVENT_MISC_KERNEL (1 << 0)
221 #define PERF_EVENT_MISC_USER (2 << 0)
222 #define PERF_EVENT_MISC_HYPERVISOR (3 << 0)
223 #define PERF_EVENT_MISC_OVERFLOW (1 << 2)
224
225 struct perf_event_header {
226 __u32 type;
227 __u16 misc;
228 __u16 size;
229 };
230
231 enum perf_event_type {
232
233 /*
234 * The MMAP events record the PROT_EXEC mappings so that we can
235 * correlate userspace IPs to code. They have the following structure:
236 *
237 * struct {
238 * struct perf_event_header header;
239 *
240 * u32 pid, tid;
241 * u64 addr;
242 * u64 len;
243 * u64 pgoff;
244 * char filename[];
245 * };
246 */
247 PERF_EVENT_MMAP = 1,
248 PERF_EVENT_MUNMAP = 2,
249
250 /*
251 * struct {
252 * struct perf_event_header header;
253 *
254 * u32 pid, tid;
255 * char comm[];
256 * };
257 */
258 PERF_EVENT_COMM = 3,
259
260 /*
261 * struct {
262 * struct perf_event_header header;
263 * u64 time;
264 * u64 sample_period;
265 * };
266 */
267 PERF_EVENT_PERIOD = 4,
268
269 /*
270 * struct {
271 * struct perf_event_header header;
272 * u64 time;
273 * };
274 */
275 PERF_EVENT_THROTTLE = 5,
276 PERF_EVENT_UNTHROTTLE = 6,
277
278 /*
279 * When header.misc & PERF_EVENT_MISC_OVERFLOW the event_type field
280 * will be PERF_RECORD_*
281 *
282 * struct {
283 * struct perf_event_header header;
284 *
285 * { u64 ip; } && PERF_RECORD_IP
286 * { u32 pid, tid; } && PERF_RECORD_TID
287 * { u64 time; } && PERF_RECORD_TIME
288 * { u64 addr; } && PERF_RECORD_ADDR
289 * { u64 config; } && PERF_RECORD_CONFIG
290 * { u32 cpu, res; } && PERF_RECORD_CPU
291 *
292 * { u64 nr;
293 * { u64 id, val; } cnt[nr]; } && PERF_RECORD_GROUP
294 *
295 * { u16 nr,
296 * hv,
297 * kernel,
298 * user;
299 * u64 ips[nr]; } && PERF_RECORD_CALLCHAIN
300 * };
301 */
302 };
303
304 #ifdef __KERNEL__
305 /*
306 * Kernel-internal data types and definitions:
307 */
308
309 #ifdef CONFIG_PERF_COUNTERS
310 # include <asm/perf_counter.h>
311 #endif
312
313 #include <linux/list.h>
314 #include <linux/mutex.h>
315 #include <linux/rculist.h>
316 #include <linux/rcupdate.h>
317 #include <linux/spinlock.h>
318 #include <linux/hrtimer.h>
319 #include <linux/fs.h>
320 #include <linux/pid_namespace.h>
321 #include <asm/atomic.h>
322
323 struct task_struct;
324
325 static inline u64 perf_event_raw(struct perf_counter_hw_event *hw_event)
326 {
327 return hw_event->config & PERF_COUNTER_RAW_MASK;
328 }
329
330 static inline u64 perf_event_config(struct perf_counter_hw_event *hw_event)
331 {
332 return hw_event->config & PERF_COUNTER_CONFIG_MASK;
333 }
334
335 static inline u64 perf_event_type(struct perf_counter_hw_event *hw_event)
336 {
337 return (hw_event->config & PERF_COUNTER_TYPE_MASK) >>
338 PERF_COUNTER_TYPE_SHIFT;
339 }
340
341 static inline u64 perf_event_id(struct perf_counter_hw_event *hw_event)
342 {
343 return hw_event->config & PERF_COUNTER_EVENT_MASK;
344 }
345
346 /**
347 * struct hw_perf_counter - performance counter hardware details:
348 */
349 struct hw_perf_counter {
350 #ifdef CONFIG_PERF_COUNTERS
351 union {
352 struct { /* hardware */
353 u64 config;
354 unsigned long config_base;
355 unsigned long counter_base;
356 int idx;
357 };
358 union { /* software */
359 atomic64_t count;
360 struct hrtimer hrtimer;
361 };
362 };
363 atomic64_t prev_count;
364 u64 sample_period;
365 atomic64_t period_left;
366 u64 interrupts;
367 #endif
368 };
369
370 struct perf_counter;
371
372 /**
373 * struct pmu - generic performance monitoring unit
374 */
375 struct pmu {
376 int (*enable) (struct perf_counter *counter);
377 void (*disable) (struct perf_counter *counter);
378 void (*read) (struct perf_counter *counter);
379 void (*unthrottle) (struct perf_counter *counter);
380 };
381
382 /**
383 * enum perf_counter_active_state - the states of a counter
384 */
385 enum perf_counter_active_state {
386 PERF_COUNTER_STATE_ERROR = -2,
387 PERF_COUNTER_STATE_OFF = -1,
388 PERF_COUNTER_STATE_INACTIVE = 0,
389 PERF_COUNTER_STATE_ACTIVE = 1,
390 };
391
392 struct file;
393
394 struct perf_mmap_data {
395 struct rcu_head rcu_head;
396 int nr_pages; /* nr of data pages */
397 int nr_locked; /* nr pages mlocked */
398
399 atomic_t poll; /* POLL_ for wakeups */
400 atomic_t head; /* write position */
401 atomic_t events; /* event limit */
402
403 atomic_t done_head; /* completed head */
404 atomic_t lock; /* concurrent writes */
405
406 atomic_t wakeup; /* needs a wakeup */
407
408 struct perf_counter_mmap_page *user_page;
409 void *data_pages[0];
410 };
411
412 struct perf_pending_entry {
413 struct perf_pending_entry *next;
414 void (*func)(struct perf_pending_entry *);
415 };
416
417 /**
418 * struct perf_counter - performance counter kernel representation:
419 */
420 struct perf_counter {
421 #ifdef CONFIG_PERF_COUNTERS
422 struct list_head list_entry;
423 struct list_head event_entry;
424 struct list_head sibling_list;
425 int nr_siblings;
426 struct perf_counter *group_leader;
427 const struct pmu *pmu;
428
429 enum perf_counter_active_state state;
430 atomic64_t count;
431
432 /*
433 * These are the total time in nanoseconds that the counter
434 * has been enabled (i.e. eligible to run, and the task has
435 * been scheduled in, if this is a per-task counter)
436 * and running (scheduled onto the CPU), respectively.
437 *
438 * They are computed from tstamp_enabled, tstamp_running and
439 * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
440 */
441 u64 total_time_enabled;
442 u64 total_time_running;
443
444 /*
445 * These are timestamps used for computing total_time_enabled
446 * and total_time_running when the counter is in INACTIVE or
447 * ACTIVE state, measured in nanoseconds from an arbitrary point
448 * in time.
449 * tstamp_enabled: the notional time when the counter was enabled
450 * tstamp_running: the notional time when the counter was scheduled on
451 * tstamp_stopped: in INACTIVE state, the notional time when the
452 * counter was scheduled off.
453 */
454 u64 tstamp_enabled;
455 u64 tstamp_running;
456 u64 tstamp_stopped;
457
458 struct perf_counter_hw_event hw_event;
459 struct hw_perf_counter hw;
460
461 struct perf_counter_context *ctx;
462 struct file *filp;
463
464 /*
465 * These accumulate total time (in nanoseconds) that children
466 * counters have been enabled and running, respectively.
467 */
468 atomic64_t child_total_time_enabled;
469 atomic64_t child_total_time_running;
470
471 /*
472 * Protect attach/detach and child_list:
473 */
474 struct mutex child_mutex;
475 struct list_head child_list;
476 struct perf_counter *parent;
477
478 int oncpu;
479 int cpu;
480
481 struct list_head owner_entry;
482 struct task_struct *owner;
483
484 /* mmap bits */
485 struct mutex mmap_mutex;
486 atomic_t mmap_count;
487 struct perf_mmap_data *data;
488
489 /* poll related */
490 wait_queue_head_t waitq;
491 struct fasync_struct *fasync;
492
493 /* delayed work for NMIs and such */
494 int pending_wakeup;
495 int pending_kill;
496 int pending_disable;
497 struct perf_pending_entry pending;
498
499 atomic_t event_limit;
500
501 void (*destroy)(struct perf_counter *);
502 struct rcu_head rcu_head;
503
504 struct pid_namespace *ns;
505 u64 id;
506 #endif
507 };
508
509 /**
510 * struct perf_counter_context - counter context structure
511 *
512 * Used as a container for task counters and CPU counters as well:
513 */
514 struct perf_counter_context {
515 /*
516 * Protect the states of the counters in the list,
517 * nr_active, and the list:
518 */
519 spinlock_t lock;
520 /*
521 * Protect the list of counters. Locking either mutex or lock
522 * is sufficient to ensure the list doesn't change; to change
523 * the list you need to lock both the mutex and the spinlock.
524 */
525 struct mutex mutex;
526
527 struct list_head counter_list;
528 struct list_head event_list;
529 int nr_counters;
530 int nr_active;
531 int is_active;
532 atomic_t refcount;
533 struct task_struct *task;
534
535 /*
536 * Context clock, runs when context enabled.
537 */
538 u64 time;
539 u64 timestamp;
540
541 /*
542 * These fields let us detect when two contexts have both
543 * been cloned (inherited) from a common ancestor.
544 */
545 struct perf_counter_context *parent_ctx;
546 u64 parent_gen;
547 u64 generation;
548 int pin_count;
549 struct rcu_head rcu_head;
550 };
551
552 /**
553 * struct perf_counter_cpu_context - per cpu counter context structure
554 */
555 struct perf_cpu_context {
556 struct perf_counter_context ctx;
557 struct perf_counter_context *task_ctx;
558 int active_oncpu;
559 int max_pertask;
560 int exclusive;
561
562 /*
563 * Recursion avoidance:
564 *
565 * task, softirq, irq, nmi context
566 */
567 int recursion[4];
568 };
569
570 #ifdef CONFIG_PERF_COUNTERS
571
572 /*
573 * Set by architecture code:
574 */
575 extern int perf_max_counters;
576
577 extern const struct pmu *hw_perf_counter_init(struct perf_counter *counter);
578
579 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
580 extern void perf_counter_task_sched_out(struct task_struct *task,
581 struct task_struct *next, int cpu);
582 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
583 extern int perf_counter_init_task(struct task_struct *child);
584 extern void perf_counter_exit_task(struct task_struct *child);
585 extern void perf_counter_free_task(struct task_struct *task);
586 extern void perf_counter_do_pending(void);
587 extern void perf_counter_print_debug(void);
588 extern void __perf_disable(void);
589 extern bool __perf_enable(void);
590 extern void perf_disable(void);
591 extern void perf_enable(void);
592 extern int perf_counter_task_disable(void);
593 extern int perf_counter_task_enable(void);
594 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
595 struct perf_cpu_context *cpuctx,
596 struct perf_counter_context *ctx, int cpu);
597 extern void perf_counter_update_userpage(struct perf_counter *counter);
598
599 extern int perf_counter_overflow(struct perf_counter *counter,
600 int nmi, struct pt_regs *regs, u64 addr);
601 /*
602 * Return 1 for a software counter, 0 for a hardware counter
603 */
604 static inline int is_software_counter(struct perf_counter *counter)
605 {
606 return !perf_event_raw(&counter->hw_event) &&
607 perf_event_type(&counter->hw_event) != PERF_TYPE_HARDWARE;
608 }
609
610 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *, u64);
611
612 extern void perf_counter_mmap(unsigned long addr, unsigned long len,
613 unsigned long pgoff, struct file *file);
614
615 extern void perf_counter_munmap(unsigned long addr, unsigned long len,
616 unsigned long pgoff, struct file *file);
617
618 extern void perf_counter_comm(struct task_struct *tsk);
619
620 extern void perf_counter_task_migration(struct task_struct *task, int cpu);
621
622 #define MAX_STACK_DEPTH 255
623
624 struct perf_callchain_entry {
625 u16 nr, hv, kernel, user;
626 u64 ip[MAX_STACK_DEPTH];
627 };
628
629 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
630
631 extern int sysctl_perf_counter_priv;
632 extern int sysctl_perf_counter_mlock;
633 extern int sysctl_perf_counter_limit;
634
635 extern void perf_counter_init(void);
636
637 #ifndef perf_misc_flags
638 #define perf_misc_flags(regs) (user_mode(regs) ? PERF_EVENT_MISC_USER : \
639 PERF_EVENT_MISC_KERNEL)
640 #define perf_instruction_pointer(regs) instruction_pointer(regs)
641 #endif
642
643 #else
644 static inline void
645 perf_counter_task_sched_in(struct task_struct *task, int cpu) { }
646 static inline void
647 perf_counter_task_sched_out(struct task_struct *task,
648 struct task_struct *next, int cpu) { }
649 static inline void
650 perf_counter_task_tick(struct task_struct *task, int cpu) { }
651 static inline int perf_counter_init_task(struct task_struct *child) { return 0; }
652 static inline void perf_counter_exit_task(struct task_struct *child) { }
653 static inline void perf_counter_free_task(struct task_struct *task) { }
654 static inline void perf_counter_do_pending(void) { }
655 static inline void perf_counter_print_debug(void) { }
656 static inline void perf_disable(void) { }
657 static inline void perf_enable(void) { }
658 static inline int perf_counter_task_disable(void) { return -EINVAL; }
659 static inline int perf_counter_task_enable(void) { return -EINVAL; }
660
661 static inline void
662 perf_swcounter_event(u32 event, u64 nr, int nmi,
663 struct pt_regs *regs, u64 addr) { }
664
665 static inline void
666 perf_counter_mmap(unsigned long addr, unsigned long len,
667 unsigned long pgoff, struct file *file) { }
668
669 static inline void
670 perf_counter_munmap(unsigned long addr, unsigned long len,
671 unsigned long pgoff, struct file *file) { }
672
673 static inline void perf_counter_comm(struct task_struct *tsk) { }
674 static inline void perf_counter_init(void) { }
675 static inline void perf_counter_task_migration(struct task_struct *task,
676 int cpu) { }
677 #endif
678
679 #endif /* __KERNEL__ */
680 #endif /* _LINUX_PERF_COUNTER_H */
This page took 0.046558 seconds and 6 git commands to generate.