perf_counter: Generate mmap events for install_special_mapping()
[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 * attr.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 attr.event_id
41 * parameter of the sys_perf_counter_open() syscall:
42 */
43 enum attr_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 attr.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 attr.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_attr {
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 comm : 1, /* include comm data */
152 freq : 1, /* use freq, not period */
153
154 __reserved_1 : 53;
155
156 __u32 wakeup_events; /* wakeup every n events */
157 __u32 __reserved_2;
158
159 __u64 __reserved_3;
160 __u64 __reserved_4;
161 };
162
163 /*
164 * Ioctls that can be done on a perf counter fd:
165 */
166 #define PERF_COUNTER_IOC_ENABLE _IO ('$', 0)
167 #define PERF_COUNTER_IOC_DISABLE _IO ('$', 1)
168 #define PERF_COUNTER_IOC_REFRESH _IO ('$', 2)
169 #define PERF_COUNTER_IOC_RESET _IO ('$', 3)
170 #define PERF_COUNTER_IOC_PERIOD _IOW('$', 4, u64)
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 __u64 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
249 /*
250 * struct {
251 * struct perf_event_header header;
252 *
253 * u32 pid, tid;
254 * char comm[];
255 * };
256 */
257 PERF_EVENT_COMM = 3,
258
259 /*
260 * struct {
261 * struct perf_event_header header;
262 * u64 time;
263 * u64 sample_period;
264 * };
265 */
266 PERF_EVENT_PERIOD = 4,
267
268 /*
269 * struct {
270 * struct perf_event_header header;
271 * u64 time;
272 * };
273 */
274 PERF_EVENT_THROTTLE = 5,
275 PERF_EVENT_UNTHROTTLE = 6,
276
277 /*
278 * struct {
279 * struct perf_event_header header;
280 * u32 pid, ppid;
281 * };
282 */
283 PERF_EVENT_FORK = 7,
284
285 /*
286 * When header.misc & PERF_EVENT_MISC_OVERFLOW the event_type field
287 * will be PERF_RECORD_*
288 *
289 * struct {
290 * struct perf_event_header header;
291 *
292 * { u64 ip; } && PERF_RECORD_IP
293 * { u32 pid, tid; } && PERF_RECORD_TID
294 * { u64 time; } && PERF_RECORD_TIME
295 * { u64 addr; } && PERF_RECORD_ADDR
296 * { u64 config; } && PERF_RECORD_CONFIG
297 * { u32 cpu, res; } && PERF_RECORD_CPU
298 *
299 * { u64 nr;
300 * { u64 id, val; } cnt[nr]; } && PERF_RECORD_GROUP
301 *
302 * { u16 nr,
303 * hv,
304 * kernel,
305 * user;
306 * u64 ips[nr]; } && PERF_RECORD_CALLCHAIN
307 * };
308 */
309 };
310
311 #ifdef __KERNEL__
312 /*
313 * Kernel-internal data types and definitions:
314 */
315
316 #ifdef CONFIG_PERF_COUNTERS
317 # include <asm/perf_counter.h>
318 #endif
319
320 #include <linux/list.h>
321 #include <linux/mutex.h>
322 #include <linux/rculist.h>
323 #include <linux/rcupdate.h>
324 #include <linux/spinlock.h>
325 #include <linux/hrtimer.h>
326 #include <linux/fs.h>
327 #include <linux/pid_namespace.h>
328 #include <asm/atomic.h>
329
330 struct task_struct;
331
332 static inline u64 perf_event_raw(struct perf_counter_attr *attr)
333 {
334 return attr->config & PERF_COUNTER_RAW_MASK;
335 }
336
337 static inline u64 perf_event_config(struct perf_counter_attr *attr)
338 {
339 return attr->config & PERF_COUNTER_CONFIG_MASK;
340 }
341
342 static inline u64 perf_event_type(struct perf_counter_attr *attr)
343 {
344 return (attr->config & PERF_COUNTER_TYPE_MASK) >>
345 PERF_COUNTER_TYPE_SHIFT;
346 }
347
348 static inline u64 perf_event_id(struct perf_counter_attr *attr)
349 {
350 return attr->config & PERF_COUNTER_EVENT_MASK;
351 }
352
353 /**
354 * struct hw_perf_counter - performance counter hardware details:
355 */
356 struct hw_perf_counter {
357 #ifdef CONFIG_PERF_COUNTERS
358 union {
359 struct { /* hardware */
360 u64 config;
361 unsigned long config_base;
362 unsigned long counter_base;
363 int idx;
364 };
365 union { /* software */
366 atomic64_t count;
367 struct hrtimer hrtimer;
368 };
369 };
370 atomic64_t prev_count;
371 u64 sample_period;
372 atomic64_t period_left;
373 u64 interrupts;
374 #endif
375 };
376
377 struct perf_counter;
378
379 /**
380 * struct pmu - generic performance monitoring unit
381 */
382 struct pmu {
383 int (*enable) (struct perf_counter *counter);
384 void (*disable) (struct perf_counter *counter);
385 void (*read) (struct perf_counter *counter);
386 void (*unthrottle) (struct perf_counter *counter);
387 };
388
389 /**
390 * enum perf_counter_active_state - the states of a counter
391 */
392 enum perf_counter_active_state {
393 PERF_COUNTER_STATE_ERROR = -2,
394 PERF_COUNTER_STATE_OFF = -1,
395 PERF_COUNTER_STATE_INACTIVE = 0,
396 PERF_COUNTER_STATE_ACTIVE = 1,
397 };
398
399 struct file;
400
401 struct perf_mmap_data {
402 struct rcu_head rcu_head;
403 int nr_pages; /* nr of data pages */
404 int nr_locked; /* nr pages mlocked */
405
406 atomic_t poll; /* POLL_ for wakeups */
407 atomic_t events; /* event limit */
408
409 atomic_long_t head; /* write position */
410 atomic_long_t done_head; /* completed head */
411
412 atomic_t lock; /* concurrent writes */
413
414 atomic_t wakeup; /* needs a wakeup */
415
416 struct perf_counter_mmap_page *user_page;
417 void *data_pages[0];
418 };
419
420 struct perf_pending_entry {
421 struct perf_pending_entry *next;
422 void (*func)(struct perf_pending_entry *);
423 };
424
425 /**
426 * struct perf_counter - performance counter kernel representation:
427 */
428 struct perf_counter {
429 #ifdef CONFIG_PERF_COUNTERS
430 struct list_head list_entry;
431 struct list_head event_entry;
432 struct list_head sibling_list;
433 int nr_siblings;
434 struct perf_counter *group_leader;
435 const struct pmu *pmu;
436
437 enum perf_counter_active_state state;
438 atomic64_t count;
439
440 /*
441 * These are the total time in nanoseconds that the counter
442 * has been enabled (i.e. eligible to run, and the task has
443 * been scheduled in, if this is a per-task counter)
444 * and running (scheduled onto the CPU), respectively.
445 *
446 * They are computed from tstamp_enabled, tstamp_running and
447 * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
448 */
449 u64 total_time_enabled;
450 u64 total_time_running;
451
452 /*
453 * These are timestamps used for computing total_time_enabled
454 * and total_time_running when the counter is in INACTIVE or
455 * ACTIVE state, measured in nanoseconds from an arbitrary point
456 * in time.
457 * tstamp_enabled: the notional time when the counter was enabled
458 * tstamp_running: the notional time when the counter was scheduled on
459 * tstamp_stopped: in INACTIVE state, the notional time when the
460 * counter was scheduled off.
461 */
462 u64 tstamp_enabled;
463 u64 tstamp_running;
464 u64 tstamp_stopped;
465
466 struct perf_counter_attr attr;
467 struct hw_perf_counter hw;
468
469 struct perf_counter_context *ctx;
470 struct file *filp;
471
472 /*
473 * These accumulate total time (in nanoseconds) that children
474 * counters have been enabled and running, respectively.
475 */
476 atomic64_t child_total_time_enabled;
477 atomic64_t child_total_time_running;
478
479 /*
480 * Protect attach/detach and child_list:
481 */
482 struct mutex child_mutex;
483 struct list_head child_list;
484 struct perf_counter *parent;
485
486 int oncpu;
487 int cpu;
488
489 struct list_head owner_entry;
490 struct task_struct *owner;
491
492 /* mmap bits */
493 struct mutex mmap_mutex;
494 atomic_t mmap_count;
495 struct perf_mmap_data *data;
496
497 /* poll related */
498 wait_queue_head_t waitq;
499 struct fasync_struct *fasync;
500
501 /* delayed work for NMIs and such */
502 int pending_wakeup;
503 int pending_kill;
504 int pending_disable;
505 struct perf_pending_entry pending;
506
507 atomic_t event_limit;
508
509 void (*destroy)(struct perf_counter *);
510 struct rcu_head rcu_head;
511
512 struct pid_namespace *ns;
513 u64 id;
514 #endif
515 };
516
517 /**
518 * struct perf_counter_context - counter context structure
519 *
520 * Used as a container for task counters and CPU counters as well:
521 */
522 struct perf_counter_context {
523 /*
524 * Protect the states of the counters in the list,
525 * nr_active, and the list:
526 */
527 spinlock_t lock;
528 /*
529 * Protect the list of counters. Locking either mutex or lock
530 * is sufficient to ensure the list doesn't change; to change
531 * the list you need to lock both the mutex and the spinlock.
532 */
533 struct mutex mutex;
534
535 struct list_head counter_list;
536 struct list_head event_list;
537 int nr_counters;
538 int nr_active;
539 int is_active;
540 atomic_t refcount;
541 struct task_struct *task;
542
543 /*
544 * Context clock, runs when context enabled.
545 */
546 u64 time;
547 u64 timestamp;
548
549 /*
550 * These fields let us detect when two contexts have both
551 * been cloned (inherited) from a common ancestor.
552 */
553 struct perf_counter_context *parent_ctx;
554 u64 parent_gen;
555 u64 generation;
556 int pin_count;
557 struct rcu_head rcu_head;
558 };
559
560 /**
561 * struct perf_counter_cpu_context - per cpu counter context structure
562 */
563 struct perf_cpu_context {
564 struct perf_counter_context ctx;
565 struct perf_counter_context *task_ctx;
566 int active_oncpu;
567 int max_pertask;
568 int exclusive;
569
570 /*
571 * Recursion avoidance:
572 *
573 * task, softirq, irq, nmi context
574 */
575 int recursion[4];
576 };
577
578 #ifdef CONFIG_PERF_COUNTERS
579
580 /*
581 * Set by architecture code:
582 */
583 extern int perf_max_counters;
584
585 extern const struct pmu *hw_perf_counter_init(struct perf_counter *counter);
586
587 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
588 extern void perf_counter_task_sched_out(struct task_struct *task,
589 struct task_struct *next, int cpu);
590 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
591 extern int perf_counter_init_task(struct task_struct *child);
592 extern void perf_counter_exit_task(struct task_struct *child);
593 extern void perf_counter_free_task(struct task_struct *task);
594 extern void perf_counter_do_pending(void);
595 extern void perf_counter_print_debug(void);
596 extern void __perf_disable(void);
597 extern bool __perf_enable(void);
598 extern void perf_disable(void);
599 extern void perf_enable(void);
600 extern int perf_counter_task_disable(void);
601 extern int perf_counter_task_enable(void);
602 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
603 struct perf_cpu_context *cpuctx,
604 struct perf_counter_context *ctx, int cpu);
605 extern void perf_counter_update_userpage(struct perf_counter *counter);
606
607 extern int perf_counter_overflow(struct perf_counter *counter,
608 int nmi, struct pt_regs *regs, u64 addr);
609 /*
610 * Return 1 for a software counter, 0 for a hardware counter
611 */
612 static inline int is_software_counter(struct perf_counter *counter)
613 {
614 return !perf_event_raw(&counter->attr) &&
615 perf_event_type(&counter->attr) != PERF_TYPE_HARDWARE;
616 }
617
618 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *, u64);
619
620 extern void __perf_counter_mmap(struct vm_area_struct *vma);
621
622 static inline void perf_counter_mmap(struct vm_area_struct *vma)
623 {
624 if (vma->vm_flags & VM_EXEC)
625 __perf_counter_mmap(vma);
626 }
627
628 extern void perf_counter_comm(struct task_struct *tsk);
629 extern void perf_counter_fork(struct task_struct *tsk);
630
631 extern void perf_counter_task_migration(struct task_struct *task, int cpu);
632
633 #define MAX_STACK_DEPTH 255
634
635 struct perf_callchain_entry {
636 u16 nr, hv, kernel, user;
637 u64 ip[MAX_STACK_DEPTH];
638 };
639
640 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
641
642 extern int sysctl_perf_counter_priv;
643 extern int sysctl_perf_counter_mlock;
644 extern int sysctl_perf_counter_limit;
645
646 extern void perf_counter_init(void);
647
648 #ifndef perf_misc_flags
649 #define perf_misc_flags(regs) (user_mode(regs) ? PERF_EVENT_MISC_USER : \
650 PERF_EVENT_MISC_KERNEL)
651 #define perf_instruction_pointer(regs) instruction_pointer(regs)
652 #endif
653
654 #else
655 static inline void
656 perf_counter_task_sched_in(struct task_struct *task, int cpu) { }
657 static inline void
658 perf_counter_task_sched_out(struct task_struct *task,
659 struct task_struct *next, int cpu) { }
660 static inline void
661 perf_counter_task_tick(struct task_struct *task, int cpu) { }
662 static inline int perf_counter_init_task(struct task_struct *child) { return 0; }
663 static inline void perf_counter_exit_task(struct task_struct *child) { }
664 static inline void perf_counter_free_task(struct task_struct *task) { }
665 static inline void perf_counter_do_pending(void) { }
666 static inline void perf_counter_print_debug(void) { }
667 static inline void perf_disable(void) { }
668 static inline void perf_enable(void) { }
669 static inline int perf_counter_task_disable(void) { return -EINVAL; }
670 static inline int perf_counter_task_enable(void) { return -EINVAL; }
671
672 static inline void
673 perf_swcounter_event(u32 event, u64 nr, int nmi,
674 struct pt_regs *regs, u64 addr) { }
675
676 static inline void perf_counter_mmap(struct vm_area_struct *vma) { }
677 static inline void perf_counter_comm(struct task_struct *tsk) { }
678 static inline void perf_counter_fork(struct task_struct *tsk) { }
679 static inline void perf_counter_init(void) { }
680 static inline void perf_counter_task_migration(struct task_struct *task,
681 int cpu) { }
682 #endif
683
684 #endif /* __KERNEL__ */
685 #endif /* _LINUX_PERF_COUNTER_H */
This page took 0.048307 seconds and 6 git commands to generate.