perf_counter: Implement more accurate per task statistics
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / include / linux / perf_counter.h
blobde70a10b5ec8128ffca3cc92e338938df40f99c1
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,
124 PERF_SAMPLE_MAX = 1U << 9, /* non-ABI */
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.
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,
137 PERF_FORMAT_MAX = 1U << 3, /* non-ABI */
140 #define PERF_ATTR_SIZE_VER0 64 /* sizeof first published struct */
143 * Hardware event to monitor via a performance monitoring counter:
145 struct perf_counter_attr {
148 * Major type: hardware/software/tracepoint/etc.
150 __u32 type;
153 * Size of the attr structure, for fwd/bwd compat.
155 __u32 size;
158 * Type specific configuration information.
160 __u64 config;
162 union {
163 __u64 sample_period;
164 __u64 sample_freq;
167 __u64 sample_type;
168 __u64 read_format;
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 inherit_stat : 1, /* per task counts */
183 __reserved_1 : 52;
185 __u32 wakeup_events; /* wakeup every n events */
186 __u32 __reserved_2;
188 __u64 __reserved_3;
192 * Ioctls that can be done on a perf counter fd:
194 #define PERF_COUNTER_IOC_ENABLE _IO ('$', 0)
195 #define PERF_COUNTER_IOC_DISABLE _IO ('$', 1)
196 #define PERF_COUNTER_IOC_REFRESH _IO ('$', 2)
197 #define PERF_COUNTER_IOC_RESET _IO ('$', 3)
198 #define PERF_COUNTER_IOC_PERIOD _IOW('$', 4, u64)
200 enum perf_counter_ioc_flags {
201 PERF_IOC_FLAG_GROUP = 1U << 0,
205 * Structure of the page that can be mapped via mmap
207 struct perf_counter_mmap_page {
208 __u32 version; /* version number of this structure */
209 __u32 compat_version; /* lowest version this is compat with */
212 * Bits needed to read the hw counters in user-space.
214 * u32 seq;
215 * s64 count;
217 * do {
218 * seq = pc->lock;
220 * barrier()
221 * if (pc->index) {
222 * count = pmc_read(pc->index - 1);
223 * count += pc->offset;
224 * } else
225 * goto regular_read;
227 * barrier();
228 * } while (pc->lock != seq);
230 * NOTE: for obvious reason this only works on self-monitoring
231 * processes.
233 __u32 lock; /* seqlock for synchronization */
234 __u32 index; /* hardware counter identifier */
235 __s64 offset; /* add to hardware counter value */
236 __u64 time_enabled; /* time counter active */
237 __u64 time_running; /* time counter on cpu */
240 * Hole for extension of the self monitor capabilities
243 __u64 __reserved[123]; /* align to 1k */
246 * Control data for the mmap() data buffer.
248 * User-space reading the @data_head value should issue an rmb(), on
249 * SMP capable platforms, after reading this value -- see
250 * perf_counter_wakeup().
252 * When the mapping is PROT_WRITE the @data_tail value should be
253 * written by userspace to reflect the last read data. In this case
254 * the kernel will not over-write unread data.
256 __u64 data_head; /* head in the data section */
257 __u64 data_tail; /* user-space written tail */
260 #define PERF_EVENT_MISC_CPUMODE_MASK (3 << 0)
261 #define PERF_EVENT_MISC_CPUMODE_UNKNOWN (0 << 0)
262 #define PERF_EVENT_MISC_KERNEL (1 << 0)
263 #define PERF_EVENT_MISC_USER (2 << 0)
264 #define PERF_EVENT_MISC_HYPERVISOR (3 << 0)
265 #define PERF_EVENT_MISC_OVERFLOW (1 << 2)
267 struct perf_event_header {
268 __u32 type;
269 __u16 misc;
270 __u16 size;
273 enum perf_event_type {
276 * The MMAP events record the PROT_EXEC mappings so that we can
277 * correlate userspace IPs to code. They have the following structure:
279 * struct {
280 * struct perf_event_header header;
282 * u32 pid, tid;
283 * u64 addr;
284 * u64 len;
285 * u64 pgoff;
286 * char filename[];
287 * };
289 PERF_EVENT_MMAP = 1,
292 * struct {
293 * struct perf_event_header header;
294 * u64 id;
295 * u64 lost;
296 * };
298 PERF_EVENT_LOST = 2,
301 * struct {
302 * struct perf_event_header header;
304 * u32 pid, tid;
305 * char comm[];
306 * };
308 PERF_EVENT_COMM = 3,
311 * struct {
312 * struct perf_event_header header;
313 * u64 time;
314 * u64 id;
315 * u64 sample_period;
316 * };
318 PERF_EVENT_PERIOD = 4,
321 * struct {
322 * struct perf_event_header header;
323 * u64 time;
324 * u64 id;
325 * };
327 PERF_EVENT_THROTTLE = 5,
328 PERF_EVENT_UNTHROTTLE = 6,
331 * struct {
332 * struct perf_event_header header;
333 * u32 pid, ppid;
334 * };
336 PERF_EVENT_FORK = 7,
339 * struct {
340 * struct perf_event_header header;
341 * u32 pid, tid;
342 * u64 value;
343 * { u64 time_enabled; } && PERF_FORMAT_ENABLED
344 * { u64 time_running; } && PERF_FORMAT_RUNNING
345 * { u64 parent_id; } && PERF_FORMAT_ID
346 * };
348 PERF_EVENT_READ = 8,
351 * When header.misc & PERF_EVENT_MISC_OVERFLOW the event_type field
352 * will be PERF_SAMPLE_*
354 * struct {
355 * struct perf_event_header header;
357 * { u64 ip; } && PERF_SAMPLE_IP
358 * { u32 pid, tid; } && PERF_SAMPLE_TID
359 * { u64 time; } && PERF_SAMPLE_TIME
360 * { u64 addr; } && PERF_SAMPLE_ADDR
361 * { u64 config; } && PERF_SAMPLE_CONFIG
362 * { u32 cpu, res; } && PERF_SAMPLE_CPU
364 * { u64 nr;
365 * { u64 id, val; } cnt[nr]; } && PERF_SAMPLE_GROUP
367 * { u64 nr,
368 * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN
369 * };
373 enum perf_callchain_context {
374 PERF_CONTEXT_HV = (__u64)-32,
375 PERF_CONTEXT_KERNEL = (__u64)-128,
376 PERF_CONTEXT_USER = (__u64)-512,
378 PERF_CONTEXT_GUEST = (__u64)-2048,
379 PERF_CONTEXT_GUEST_KERNEL = (__u64)-2176,
380 PERF_CONTEXT_GUEST_USER = (__u64)-2560,
382 PERF_CONTEXT_MAX = (__u64)-4095,
385 #ifdef __KERNEL__
387 * Kernel-internal data types and definitions:
390 #ifdef CONFIG_PERF_COUNTERS
391 # include <asm/perf_counter.h>
392 #endif
394 #include <linux/list.h>
395 #include <linux/mutex.h>
396 #include <linux/rculist.h>
397 #include <linux/rcupdate.h>
398 #include <linux/spinlock.h>
399 #include <linux/hrtimer.h>
400 #include <linux/fs.h>
401 #include <linux/pid_namespace.h>
402 #include <asm/atomic.h>
404 #define PERF_MAX_STACK_DEPTH 255
406 struct perf_callchain_entry {
407 __u64 nr;
408 __u64 ip[PERF_MAX_STACK_DEPTH];
411 struct task_struct;
414 * struct hw_perf_counter - performance counter hardware details:
416 struct hw_perf_counter {
417 #ifdef CONFIG_PERF_COUNTERS
418 union {
419 struct { /* hardware */
420 u64 config;
421 unsigned long config_base;
422 unsigned long counter_base;
423 int idx;
425 union { /* software */
426 atomic64_t count;
427 struct hrtimer hrtimer;
430 atomic64_t prev_count;
431 u64 sample_period;
432 u64 last_period;
433 atomic64_t period_left;
434 u64 interrupts;
436 u64 freq_count;
437 u64 freq_interrupts;
438 u64 freq_stamp;
439 #endif
442 struct perf_counter;
445 * struct pmu - generic performance monitoring unit
447 struct pmu {
448 int (*enable) (struct perf_counter *counter);
449 void (*disable) (struct perf_counter *counter);
450 void (*read) (struct perf_counter *counter);
451 void (*unthrottle) (struct perf_counter *counter);
455 * enum perf_counter_active_state - the states of a counter
457 enum perf_counter_active_state {
458 PERF_COUNTER_STATE_ERROR = -2,
459 PERF_COUNTER_STATE_OFF = -1,
460 PERF_COUNTER_STATE_INACTIVE = 0,
461 PERF_COUNTER_STATE_ACTIVE = 1,
464 struct file;
466 struct perf_mmap_data {
467 struct rcu_head rcu_head;
468 int nr_pages; /* nr of data pages */
469 int writable; /* are we writable */
470 int nr_locked; /* nr pages mlocked */
472 atomic_t poll; /* POLL_ for wakeups */
473 atomic_t events; /* event limit */
475 atomic_long_t head; /* write position */
476 atomic_long_t done_head; /* completed head */
478 atomic_t lock; /* concurrent writes */
479 atomic_t wakeup; /* needs a wakeup */
480 atomic_t lost; /* nr records lost */
482 struct perf_counter_mmap_page *user_page;
483 void *data_pages[0];
486 struct perf_pending_entry {
487 struct perf_pending_entry *next;
488 void (*func)(struct perf_pending_entry *);
492 * struct perf_counter - performance counter kernel representation:
494 struct perf_counter {
495 #ifdef CONFIG_PERF_COUNTERS
496 struct list_head list_entry;
497 struct list_head event_entry;
498 struct list_head sibling_list;
499 int nr_siblings;
500 struct perf_counter *group_leader;
501 const struct pmu *pmu;
503 enum perf_counter_active_state state;
504 atomic64_t count;
507 * These are the total time in nanoseconds that the counter
508 * has been enabled (i.e. eligible to run, and the task has
509 * been scheduled in, if this is a per-task counter)
510 * and running (scheduled onto the CPU), respectively.
512 * They are computed from tstamp_enabled, tstamp_running and
513 * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
515 u64 total_time_enabled;
516 u64 total_time_running;
519 * These are timestamps used for computing total_time_enabled
520 * and total_time_running when the counter is in INACTIVE or
521 * ACTIVE state, measured in nanoseconds from an arbitrary point
522 * in time.
523 * tstamp_enabled: the notional time when the counter was enabled
524 * tstamp_running: the notional time when the counter was scheduled on
525 * tstamp_stopped: in INACTIVE state, the notional time when the
526 * counter was scheduled off.
528 u64 tstamp_enabled;
529 u64 tstamp_running;
530 u64 tstamp_stopped;
532 struct perf_counter_attr attr;
533 struct hw_perf_counter hw;
535 struct perf_counter_context *ctx;
536 struct file *filp;
539 * These accumulate total time (in nanoseconds) that children
540 * counters have been enabled and running, respectively.
542 atomic64_t child_total_time_enabled;
543 atomic64_t child_total_time_running;
546 * Protect attach/detach and child_list:
548 struct mutex child_mutex;
549 struct list_head child_list;
550 struct perf_counter *parent;
552 int oncpu;
553 int cpu;
555 struct list_head owner_entry;
556 struct task_struct *owner;
558 /* mmap bits */
559 struct mutex mmap_mutex;
560 atomic_t mmap_count;
561 struct perf_mmap_data *data;
563 /* poll related */
564 wait_queue_head_t waitq;
565 struct fasync_struct *fasync;
567 /* delayed work for NMIs and such */
568 int pending_wakeup;
569 int pending_kill;
570 int pending_disable;
571 struct perf_pending_entry pending;
573 atomic_t event_limit;
575 void (*destroy)(struct perf_counter *);
576 struct rcu_head rcu_head;
578 struct pid_namespace *ns;
579 u64 id;
580 #endif
584 * struct perf_counter_context - counter context structure
586 * Used as a container for task counters and CPU counters as well:
588 struct perf_counter_context {
590 * Protect the states of the counters in the list,
591 * nr_active, and the list:
593 spinlock_t lock;
595 * Protect the list of counters. Locking either mutex or lock
596 * is sufficient to ensure the list doesn't change; to change
597 * the list you need to lock both the mutex and the spinlock.
599 struct mutex mutex;
601 struct list_head counter_list;
602 struct list_head event_list;
603 int nr_counters;
604 int nr_active;
605 int is_active;
606 int nr_stat;
607 atomic_t refcount;
608 struct task_struct *task;
611 * Context clock, runs when context enabled.
613 u64 time;
614 u64 timestamp;
617 * These fields let us detect when two contexts have both
618 * been cloned (inherited) from a common ancestor.
620 struct perf_counter_context *parent_ctx;
621 u64 parent_gen;
622 u64 generation;
623 int pin_count;
624 struct rcu_head rcu_head;
628 * struct perf_counter_cpu_context - per cpu counter context structure
630 struct perf_cpu_context {
631 struct perf_counter_context ctx;
632 struct perf_counter_context *task_ctx;
633 int active_oncpu;
634 int max_pertask;
635 int exclusive;
638 * Recursion avoidance:
640 * task, softirq, irq, nmi context
642 int recursion[4];
645 #ifdef CONFIG_PERF_COUNTERS
648 * Set by architecture code:
650 extern int perf_max_counters;
652 extern const struct pmu *hw_perf_counter_init(struct perf_counter *counter);
654 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
655 extern void perf_counter_task_sched_out(struct task_struct *task,
656 struct task_struct *next, int cpu);
657 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
658 extern int perf_counter_init_task(struct task_struct *child);
659 extern void perf_counter_exit_task(struct task_struct *child);
660 extern void perf_counter_free_task(struct task_struct *task);
661 extern void set_perf_counter_pending(void);
662 extern void perf_counter_do_pending(void);
663 extern void perf_counter_print_debug(void);
664 extern void __perf_disable(void);
665 extern bool __perf_enable(void);
666 extern void perf_disable(void);
667 extern void perf_enable(void);
668 extern int perf_counter_task_disable(void);
669 extern int perf_counter_task_enable(void);
670 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
671 struct perf_cpu_context *cpuctx,
672 struct perf_counter_context *ctx, int cpu);
673 extern void perf_counter_update_userpage(struct perf_counter *counter);
675 struct perf_sample_data {
676 struct pt_regs *regs;
677 u64 addr;
678 u64 period;
681 extern int perf_counter_overflow(struct perf_counter *counter, int nmi,
682 struct perf_sample_data *data);
685 * Return 1 for a software counter, 0 for a hardware counter
687 static inline int is_software_counter(struct perf_counter *counter)
689 return (counter->attr.type != PERF_TYPE_RAW) &&
690 (counter->attr.type != PERF_TYPE_HARDWARE) &&
691 (counter->attr.type != PERF_TYPE_HW_CACHE);
694 extern atomic_t perf_swcounter_enabled[PERF_COUNT_SW_MAX];
696 extern void __perf_swcounter_event(u32, u64, int, struct pt_regs *, u64);
698 static inline void
699 perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
701 if (atomic_read(&perf_swcounter_enabled[event]))
702 __perf_swcounter_event(event, nr, nmi, regs, addr);
705 extern void __perf_counter_mmap(struct vm_area_struct *vma);
707 static inline void perf_counter_mmap(struct vm_area_struct *vma)
709 if (vma->vm_flags & VM_EXEC)
710 __perf_counter_mmap(vma);
713 extern void perf_counter_comm(struct task_struct *tsk);
714 extern void perf_counter_fork(struct task_struct *tsk);
716 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
718 extern int sysctl_perf_counter_paranoid;
719 extern int sysctl_perf_counter_mlock;
720 extern int sysctl_perf_counter_sample_rate;
722 extern void perf_counter_init(void);
724 #ifndef perf_misc_flags
725 #define perf_misc_flags(regs) (user_mode(regs) ? PERF_EVENT_MISC_USER : \
726 PERF_EVENT_MISC_KERNEL)
727 #define perf_instruction_pointer(regs) instruction_pointer(regs)
728 #endif
730 #else
731 static inline void
732 perf_counter_task_sched_in(struct task_struct *task, int cpu) { }
733 static inline void
734 perf_counter_task_sched_out(struct task_struct *task,
735 struct task_struct *next, int cpu) { }
736 static inline void
737 perf_counter_task_tick(struct task_struct *task, int cpu) { }
738 static inline int perf_counter_init_task(struct task_struct *child) { return 0; }
739 static inline void perf_counter_exit_task(struct task_struct *child) { }
740 static inline void perf_counter_free_task(struct task_struct *task) { }
741 static inline void perf_counter_do_pending(void) { }
742 static inline void perf_counter_print_debug(void) { }
743 static inline void perf_disable(void) { }
744 static inline void perf_enable(void) { }
745 static inline int perf_counter_task_disable(void) { return -EINVAL; }
746 static inline int perf_counter_task_enable(void) { return -EINVAL; }
748 static inline void
749 perf_swcounter_event(u32 event, u64 nr, int nmi,
750 struct pt_regs *regs, u64 addr) { }
752 static inline void perf_counter_mmap(struct vm_area_struct *vma) { }
753 static inline void perf_counter_comm(struct task_struct *tsk) { }
754 static inline void perf_counter_fork(struct task_struct *tsk) { }
755 static inline void perf_counter_init(void) { }
756 #endif
758 #endif /* __KERNEL__ */
759 #endif /* _LINUX_PERF_COUNTER_H */