perf_counter: Fix/complete ftrace event records sampling
[linux-2.6/mini2440.git] / include / linux / perf_counter.h
bloba67dd5c5b6d3b72db134d4d81ed9a0c0aa6d9be9
1 /*
2 * Performance counters:
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
8 * Data type definitions, declarations, prototypes.
10 * Started by: Thomas Gleixner and Ingo Molnar
12 * For licencing details see kernel-base/COPYING
14 #ifndef _LINUX_PERF_COUNTER_H
15 #define _LINUX_PERF_COUNTER_H
17 #include <linux/types.h>
18 #include <linux/ioctl.h>
19 #include <asm/byteorder.h>
22 * User-space ABI bits:
26 * attr.type
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,
35 PERF_TYPE_MAX, /* non-ABI */
39 * Generalized performance counter event types, used by the
40 * attr.event_id parameter of the sys_perf_counter_open()
41 * syscall:
43 enum perf_hw_id {
45 * Common hardware events, generalized by the kernel:
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,
55 PERF_COUNT_HW_MAX, /* non-ABI */
59 * Generalized hardware cache counters:
61 * { L1-D, L1-I, LLC, ITLB, DTLB, BPU } x
62 * { read, write, prefetch } x
63 * { accesses, misses }
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,
73 PERF_COUNT_HW_CACHE_MAX, /* non-ABI */
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,
81 PERF_COUNT_HW_CACHE_OP_MAX, /* non-ABI */
84 enum perf_hw_cache_op_result_id {
85 PERF_COUNT_HW_CACHE_RESULT_ACCESS = 0,
86 PERF_COUNT_HW_CACHE_RESULT_MISS = 1,
88 PERF_COUNT_HW_CACHE_RESULT_MAX, /* non-ABI */
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):
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,
106 PERF_COUNT_SW_MAX, /* non-ABI */
110 * Bits that can be set in attr.sample_type to request information
111 * in the overflow packets.
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 PERF_SAMPLE_STREAM_ID = 1U << 9,
124 PERF_SAMPLE_TP_RECORD = 1U << 10,
126 PERF_SAMPLE_MAX = 1U << 11, /* non-ABI */
130 * Bits that can be set in attr.read_format to request that
131 * reads on the counter should return the indicated quantities,
132 * in increasing order of bit value, after the counter value.
134 enum perf_counter_read_format {
135 PERF_FORMAT_TOTAL_TIME_ENABLED = 1U << 0,
136 PERF_FORMAT_TOTAL_TIME_RUNNING = 1U << 1,
137 PERF_FORMAT_ID = 1U << 2,
139 PERF_FORMAT_MAX = 1U << 3, /* non-ABI */
142 #define PERF_ATTR_SIZE_VER0 64 /* sizeof first published struct */
145 * Hardware event to monitor via a performance monitoring counter:
147 struct perf_counter_attr {
150 * Major type: hardware/software/tracepoint/etc.
152 __u32 type;
155 * Size of the attr structure, for fwd/bwd compat.
157 __u32 size;
160 * Type specific configuration information.
162 __u64 config;
164 union {
165 __u64 sample_period;
166 __u64 sample_freq;
169 __u64 sample_type;
170 __u64 read_format;
172 __u64 disabled : 1, /* off by default */
173 inherit : 1, /* children inherit it */
174 pinned : 1, /* must always be on PMU */
175 exclusive : 1, /* only group on PMU */
176 exclude_user : 1, /* don't count user */
177 exclude_kernel : 1, /* ditto kernel */
178 exclude_hv : 1, /* ditto hypervisor */
179 exclude_idle : 1, /* don't count when idle */
180 mmap : 1, /* include mmap data */
181 comm : 1, /* include comm data */
182 freq : 1, /* use freq, not period */
183 inherit_stat : 1, /* per task counts */
184 enable_on_exec : 1, /* next exec enables */
185 task : 1, /* trace fork/exit */
187 __reserved_1 : 50;
189 __u32 wakeup_events; /* wakeup every n events */
190 __u32 __reserved_2;
192 __u64 __reserved_3;
196 * Ioctls that can be done on a perf counter fd:
198 #define PERF_COUNTER_IOC_ENABLE _IO ('$', 0)
199 #define PERF_COUNTER_IOC_DISABLE _IO ('$', 1)
200 #define PERF_COUNTER_IOC_REFRESH _IO ('$', 2)
201 #define PERF_COUNTER_IOC_RESET _IO ('$', 3)
202 #define PERF_COUNTER_IOC_PERIOD _IOW('$', 4, u64)
204 enum perf_counter_ioc_flags {
205 PERF_IOC_FLAG_GROUP = 1U << 0,
209 * Structure of the page that can be mapped via mmap
211 struct perf_counter_mmap_page {
212 __u32 version; /* version number of this structure */
213 __u32 compat_version; /* lowest version this is compat with */
216 * Bits needed to read the hw counters in user-space.
218 * u32 seq;
219 * s64 count;
221 * do {
222 * seq = pc->lock;
224 * barrier()
225 * if (pc->index) {
226 * count = pmc_read(pc->index - 1);
227 * count += pc->offset;
228 * } else
229 * goto regular_read;
231 * barrier();
232 * } while (pc->lock != seq);
234 * NOTE: for obvious reason this only works on self-monitoring
235 * processes.
237 __u32 lock; /* seqlock for synchronization */
238 __u32 index; /* hardware counter identifier */
239 __s64 offset; /* add to hardware counter value */
240 __u64 time_enabled; /* time counter active */
241 __u64 time_running; /* time counter on cpu */
244 * Hole for extension of the self monitor capabilities
247 __u64 __reserved[123]; /* align to 1k */
250 * Control data for the mmap() data buffer.
252 * User-space reading the @data_head value should issue an rmb(), on
253 * SMP capable platforms, after reading this value -- see
254 * perf_counter_wakeup().
256 * When the mapping is PROT_WRITE the @data_tail value should be
257 * written by userspace to reflect the last read data. In this case
258 * the kernel will not over-write unread data.
260 __u64 data_head; /* head in the data section */
261 __u64 data_tail; /* user-space written tail */
264 #define PERF_EVENT_MISC_CPUMODE_MASK (3 << 0)
265 #define PERF_EVENT_MISC_CPUMODE_UNKNOWN (0 << 0)
266 #define PERF_EVENT_MISC_KERNEL (1 << 0)
267 #define PERF_EVENT_MISC_USER (2 << 0)
268 #define PERF_EVENT_MISC_HYPERVISOR (3 << 0)
270 struct perf_event_header {
271 __u32 type;
272 __u16 misc;
273 __u16 size;
276 enum perf_event_type {
279 * The MMAP events record the PROT_EXEC mappings so that we can
280 * correlate userspace IPs to code. They have the following structure:
282 * struct {
283 * struct perf_event_header header;
285 * u32 pid, tid;
286 * u64 addr;
287 * u64 len;
288 * u64 pgoff;
289 * char filename[];
290 * };
292 PERF_EVENT_MMAP = 1,
295 * struct {
296 * struct perf_event_header header;
297 * u64 id;
298 * u64 lost;
299 * };
301 PERF_EVENT_LOST = 2,
304 * struct {
305 * struct perf_event_header header;
307 * u32 pid, tid;
308 * char comm[];
309 * };
311 PERF_EVENT_COMM = 3,
314 * struct {
315 * struct perf_event_header header;
316 * u32 pid, ppid;
317 * u32 tid, ptid;
318 * };
320 PERF_EVENT_EXIT = 4,
323 * struct {
324 * struct perf_event_header header;
325 * u64 time;
326 * u64 id;
327 * u64 stream_id;
328 * };
330 PERF_EVENT_THROTTLE = 5,
331 PERF_EVENT_UNTHROTTLE = 6,
334 * struct {
335 * struct perf_event_header header;
336 * u32 pid, ppid;
337 * u32 tid, ptid;
338 * };
340 PERF_EVENT_FORK = 7,
343 * struct {
344 * struct perf_event_header header;
345 * u32 pid, tid;
346 * u64 value;
347 * { u64 time_enabled; } && PERF_FORMAT_ENABLED
348 * { u64 time_running; } && PERF_FORMAT_RUNNING
349 * { u64 parent_id; } && PERF_FORMAT_ID
350 * };
352 PERF_EVENT_READ = 8,
355 * struct {
356 * struct perf_event_header header;
358 * { u64 ip; } && PERF_SAMPLE_IP
359 * { u32 pid, tid; } && PERF_SAMPLE_TID
360 * { u64 time; } && PERF_SAMPLE_TIME
361 * { u64 addr; } && PERF_SAMPLE_ADDR
362 * { u64 id; } && PERF_SAMPLE_ID
363 * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID
364 * { u32 cpu, res; } && PERF_SAMPLE_CPU
365 * { u64 period; } && PERF_SAMPLE_PERIOD
367 * { u64 nr;
368 * { u64 id, val; } cnt[nr]; } && PERF_SAMPLE_GROUP
370 * { u64 nr,
371 * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN
372 * };
374 PERF_EVENT_SAMPLE = 9,
376 PERF_EVENT_MAX, /* non-ABI */
379 enum perf_callchain_context {
380 PERF_CONTEXT_HV = (__u64)-32,
381 PERF_CONTEXT_KERNEL = (__u64)-128,
382 PERF_CONTEXT_USER = (__u64)-512,
384 PERF_CONTEXT_GUEST = (__u64)-2048,
385 PERF_CONTEXT_GUEST_KERNEL = (__u64)-2176,
386 PERF_CONTEXT_GUEST_USER = (__u64)-2560,
388 PERF_CONTEXT_MAX = (__u64)-4095,
391 #ifdef __KERNEL__
393 * Kernel-internal data types and definitions:
396 #ifdef CONFIG_PERF_COUNTERS
397 # include <asm/perf_counter.h>
398 #endif
400 #include <linux/list.h>
401 #include <linux/mutex.h>
402 #include <linux/rculist.h>
403 #include <linux/rcupdate.h>
404 #include <linux/spinlock.h>
405 #include <linux/hrtimer.h>
406 #include <linux/fs.h>
407 #include <linux/pid_namespace.h>
408 #include <asm/atomic.h>
410 #define PERF_MAX_STACK_DEPTH 255
412 struct perf_callchain_entry {
413 __u64 nr;
414 __u64 ip[PERF_MAX_STACK_DEPTH];
417 struct perf_tracepoint_record {
418 int size;
419 char *record;
422 struct task_struct;
425 * struct hw_perf_counter - performance counter hardware details:
427 struct hw_perf_counter {
428 #ifdef CONFIG_PERF_COUNTERS
429 union {
430 struct { /* hardware */
431 u64 config;
432 unsigned long config_base;
433 unsigned long counter_base;
434 int idx;
436 union { /* software */
437 atomic64_t count;
438 struct hrtimer hrtimer;
441 atomic64_t prev_count;
442 u64 sample_period;
443 u64 last_period;
444 atomic64_t period_left;
445 u64 interrupts;
447 u64 freq_count;
448 u64 freq_interrupts;
449 u64 freq_stamp;
450 #endif
453 struct perf_counter;
456 * struct pmu - generic performance monitoring unit
458 struct pmu {
459 int (*enable) (struct perf_counter *counter);
460 void (*disable) (struct perf_counter *counter);
461 void (*read) (struct perf_counter *counter);
462 void (*unthrottle) (struct perf_counter *counter);
466 * enum perf_counter_active_state - the states of a counter
468 enum perf_counter_active_state {
469 PERF_COUNTER_STATE_ERROR = -2,
470 PERF_COUNTER_STATE_OFF = -1,
471 PERF_COUNTER_STATE_INACTIVE = 0,
472 PERF_COUNTER_STATE_ACTIVE = 1,
475 struct file;
477 struct perf_mmap_data {
478 struct rcu_head rcu_head;
479 int nr_pages; /* nr of data pages */
480 int writable; /* are we writable */
481 int nr_locked; /* nr pages mlocked */
483 atomic_t poll; /* POLL_ for wakeups */
484 atomic_t events; /* event limit */
486 atomic_long_t head; /* write position */
487 atomic_long_t done_head; /* completed head */
489 atomic_t lock; /* concurrent writes */
490 atomic_t wakeup; /* needs a wakeup */
491 atomic_t lost; /* nr records lost */
493 struct perf_counter_mmap_page *user_page;
494 void *data_pages[0];
497 struct perf_pending_entry {
498 struct perf_pending_entry *next;
499 void (*func)(struct perf_pending_entry *);
503 * struct perf_counter - performance counter kernel representation:
505 struct perf_counter {
506 #ifdef CONFIG_PERF_COUNTERS
507 struct list_head list_entry;
508 struct list_head event_entry;
509 struct list_head sibling_list;
510 int nr_siblings;
511 struct perf_counter *group_leader;
512 const struct pmu *pmu;
514 enum perf_counter_active_state state;
515 atomic64_t count;
518 * These are the total time in nanoseconds that the counter
519 * has been enabled (i.e. eligible to run, and the task has
520 * been scheduled in, if this is a per-task counter)
521 * and running (scheduled onto the CPU), respectively.
523 * They are computed from tstamp_enabled, tstamp_running and
524 * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
526 u64 total_time_enabled;
527 u64 total_time_running;
530 * These are timestamps used for computing total_time_enabled
531 * and total_time_running when the counter is in INACTIVE or
532 * ACTIVE state, measured in nanoseconds from an arbitrary point
533 * in time.
534 * tstamp_enabled: the notional time when the counter was enabled
535 * tstamp_running: the notional time when the counter was scheduled on
536 * tstamp_stopped: in INACTIVE state, the notional time when the
537 * counter was scheduled off.
539 u64 tstamp_enabled;
540 u64 tstamp_running;
541 u64 tstamp_stopped;
543 struct perf_counter_attr attr;
544 struct hw_perf_counter hw;
546 struct perf_counter_context *ctx;
547 struct file *filp;
550 * These accumulate total time (in nanoseconds) that children
551 * counters have been enabled and running, respectively.
553 atomic64_t child_total_time_enabled;
554 atomic64_t child_total_time_running;
557 * Protect attach/detach and child_list:
559 struct mutex child_mutex;
560 struct list_head child_list;
561 struct perf_counter *parent;
563 int oncpu;
564 int cpu;
566 struct list_head owner_entry;
567 struct task_struct *owner;
569 /* mmap bits */
570 struct mutex mmap_mutex;
571 atomic_t mmap_count;
572 struct perf_mmap_data *data;
574 /* poll related */
575 wait_queue_head_t waitq;
576 struct fasync_struct *fasync;
578 /* delayed work for NMIs and such */
579 int pending_wakeup;
580 int pending_kill;
581 int pending_disable;
582 struct perf_pending_entry pending;
584 atomic_t event_limit;
586 void (*destroy)(struct perf_counter *);
587 struct rcu_head rcu_head;
589 struct pid_namespace *ns;
590 u64 id;
591 #endif
595 * struct perf_counter_context - counter context structure
597 * Used as a container for task counters and CPU counters as well:
599 struct perf_counter_context {
601 * Protect the states of the counters in the list,
602 * nr_active, and the list:
604 spinlock_t lock;
606 * Protect the list of counters. Locking either mutex or lock
607 * is sufficient to ensure the list doesn't change; to change
608 * the list you need to lock both the mutex and the spinlock.
610 struct mutex mutex;
612 struct list_head counter_list;
613 struct list_head event_list;
614 int nr_counters;
615 int nr_active;
616 int is_active;
617 int nr_stat;
618 atomic_t refcount;
619 struct task_struct *task;
622 * Context clock, runs when context enabled.
624 u64 time;
625 u64 timestamp;
628 * These fields let us detect when two contexts have both
629 * been cloned (inherited) from a common ancestor.
631 struct perf_counter_context *parent_ctx;
632 u64 parent_gen;
633 u64 generation;
634 int pin_count;
635 struct rcu_head rcu_head;
639 * struct perf_counter_cpu_context - per cpu counter context structure
641 struct perf_cpu_context {
642 struct perf_counter_context ctx;
643 struct perf_counter_context *task_ctx;
644 int active_oncpu;
645 int max_pertask;
646 int exclusive;
649 * Recursion avoidance:
651 * task, softirq, irq, nmi context
653 int recursion[4];
656 #ifdef CONFIG_PERF_COUNTERS
659 * Set by architecture code:
661 extern int perf_max_counters;
663 extern const struct pmu *hw_perf_counter_init(struct perf_counter *counter);
665 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
666 extern void perf_counter_task_sched_out(struct task_struct *task,
667 struct task_struct *next, int cpu);
668 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
669 extern int perf_counter_init_task(struct task_struct *child);
670 extern void perf_counter_exit_task(struct task_struct *child);
671 extern void perf_counter_free_task(struct task_struct *task);
672 extern void set_perf_counter_pending(void);
673 extern void perf_counter_do_pending(void);
674 extern void perf_counter_print_debug(void);
675 extern void __perf_disable(void);
676 extern bool __perf_enable(void);
677 extern void perf_disable(void);
678 extern void perf_enable(void);
679 extern int perf_counter_task_disable(void);
680 extern int perf_counter_task_enable(void);
681 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
682 struct perf_cpu_context *cpuctx,
683 struct perf_counter_context *ctx, int cpu);
684 extern void perf_counter_update_userpage(struct perf_counter *counter);
686 struct perf_sample_data {
687 struct pt_regs *regs;
688 u64 addr;
689 u64 period;
690 void *private;
693 extern int perf_counter_overflow(struct perf_counter *counter, int nmi,
694 struct perf_sample_data *data);
697 * Return 1 for a software counter, 0 for a hardware counter
699 static inline int is_software_counter(struct perf_counter *counter)
701 return (counter->attr.type != PERF_TYPE_RAW) &&
702 (counter->attr.type != PERF_TYPE_HARDWARE) &&
703 (counter->attr.type != PERF_TYPE_HW_CACHE);
706 extern atomic_t perf_swcounter_enabled[PERF_COUNT_SW_MAX];
708 extern void __perf_swcounter_event(u32, u64, int, struct pt_regs *, u64);
710 static inline void
711 perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
713 if (atomic_read(&perf_swcounter_enabled[event]))
714 __perf_swcounter_event(event, nr, nmi, regs, addr);
717 extern void __perf_counter_mmap(struct vm_area_struct *vma);
719 static inline void perf_counter_mmap(struct vm_area_struct *vma)
721 if (vma->vm_flags & VM_EXEC)
722 __perf_counter_mmap(vma);
725 extern void perf_counter_comm(struct task_struct *tsk);
726 extern void perf_counter_fork(struct task_struct *tsk);
728 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
730 extern int sysctl_perf_counter_paranoid;
731 extern int sysctl_perf_counter_mlock;
732 extern int sysctl_perf_counter_sample_rate;
734 extern void perf_counter_init(void);
736 #ifndef perf_misc_flags
737 #define perf_misc_flags(regs) (user_mode(regs) ? PERF_EVENT_MISC_USER : \
738 PERF_EVENT_MISC_KERNEL)
739 #define perf_instruction_pointer(regs) instruction_pointer(regs)
740 #endif
742 #else
743 static inline void
744 perf_counter_task_sched_in(struct task_struct *task, int cpu) { }
745 static inline void
746 perf_counter_task_sched_out(struct task_struct *task,
747 struct task_struct *next, int cpu) { }
748 static inline void
749 perf_counter_task_tick(struct task_struct *task, int cpu) { }
750 static inline int perf_counter_init_task(struct task_struct *child) { return 0; }
751 static inline void perf_counter_exit_task(struct task_struct *child) { }
752 static inline void perf_counter_free_task(struct task_struct *task) { }
753 static inline void perf_counter_do_pending(void) { }
754 static inline void perf_counter_print_debug(void) { }
755 static inline void perf_disable(void) { }
756 static inline void perf_enable(void) { }
757 static inline int perf_counter_task_disable(void) { return -EINVAL; }
758 static inline int perf_counter_task_enable(void) { return -EINVAL; }
760 static inline void
761 perf_swcounter_event(u32 event, u64 nr, int nmi,
762 struct pt_regs *regs, u64 addr) { }
764 static inline void perf_counter_mmap(struct vm_area_struct *vma) { }
765 static inline void perf_counter_comm(struct task_struct *tsk) { }
766 static inline void perf_counter_fork(struct task_struct *tsk) { }
767 static inline void perf_counter_init(void) { }
768 #endif
770 #endif /* __KERNEL__ */
771 #endif /* _LINUX_PERF_COUNTER_H */