perf_counter: Full task tracing
[linux-2.6/mini2440.git] / include / linux / perf_counter.h
blobe604e6ef72dd5af13b6265d688451cef47fd56fa
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,
125 PERF_SAMPLE_MAX = 1U << 10, /* non-ABI */
129 * Bits that can be set in attr.read_format to request that
130 * reads on the counter should return the indicated quantities,
131 * in increasing order of bit value, after the counter value.
133 enum perf_counter_read_format {
134 PERF_FORMAT_TOTAL_TIME_ENABLED = 1U << 0,
135 PERF_FORMAT_TOTAL_TIME_RUNNING = 1U << 1,
136 PERF_FORMAT_ID = 1U << 2,
138 PERF_FORMAT_MAX = 1U << 3, /* non-ABI */
141 #define PERF_ATTR_SIZE_VER0 64 /* sizeof first published struct */
144 * Hardware event to monitor via a performance monitoring counter:
146 struct perf_counter_attr {
149 * Major type: hardware/software/tracepoint/etc.
151 __u32 type;
154 * Size of the attr structure, for fwd/bwd compat.
156 __u32 size;
159 * Type specific configuration information.
161 __u64 config;
163 union {
164 __u64 sample_period;
165 __u64 sample_freq;
168 __u64 sample_type;
169 __u64 read_format;
171 __u64 disabled : 1, /* off by default */
172 inherit : 1, /* children inherit it */
173 pinned : 1, /* must always be on PMU */
174 exclusive : 1, /* only group on PMU */
175 exclude_user : 1, /* don't count user */
176 exclude_kernel : 1, /* ditto kernel */
177 exclude_hv : 1, /* ditto hypervisor */
178 exclude_idle : 1, /* don't count when idle */
179 mmap : 1, /* include mmap data */
180 comm : 1, /* include comm data */
181 freq : 1, /* use freq, not period */
182 inherit_stat : 1, /* per task counts */
183 enable_on_exec : 1, /* next exec enables */
184 task : 1, /* trace fork/exit */
186 __reserved_1 : 50;
188 __u32 wakeup_events; /* wakeup every n events */
189 __u32 __reserved_2;
191 __u64 __reserved_3;
195 * Ioctls that can be done on a perf counter fd:
197 #define PERF_COUNTER_IOC_ENABLE _IO ('$', 0)
198 #define PERF_COUNTER_IOC_DISABLE _IO ('$', 1)
199 #define PERF_COUNTER_IOC_REFRESH _IO ('$', 2)
200 #define PERF_COUNTER_IOC_RESET _IO ('$', 3)
201 #define PERF_COUNTER_IOC_PERIOD _IOW('$', 4, u64)
203 enum perf_counter_ioc_flags {
204 PERF_IOC_FLAG_GROUP = 1U << 0,
208 * Structure of the page that can be mapped via mmap
210 struct perf_counter_mmap_page {
211 __u32 version; /* version number of this structure */
212 __u32 compat_version; /* lowest version this is compat with */
215 * Bits needed to read the hw counters in user-space.
217 * u32 seq;
218 * s64 count;
220 * do {
221 * seq = pc->lock;
223 * barrier()
224 * if (pc->index) {
225 * count = pmc_read(pc->index - 1);
226 * count += pc->offset;
227 * } else
228 * goto regular_read;
230 * barrier();
231 * } while (pc->lock != seq);
233 * NOTE: for obvious reason this only works on self-monitoring
234 * processes.
236 __u32 lock; /* seqlock for synchronization */
237 __u32 index; /* hardware counter identifier */
238 __s64 offset; /* add to hardware counter value */
239 __u64 time_enabled; /* time counter active */
240 __u64 time_running; /* time counter on cpu */
243 * Hole for extension of the self monitor capabilities
246 __u64 __reserved[123]; /* align to 1k */
249 * Control data for the mmap() data buffer.
251 * User-space reading the @data_head value should issue an rmb(), on
252 * SMP capable platforms, after reading this value -- see
253 * perf_counter_wakeup().
255 * When the mapping is PROT_WRITE the @data_tail value should be
256 * written by userspace to reflect the last read data. In this case
257 * the kernel will not over-write unread data.
259 __u64 data_head; /* head in the data section */
260 __u64 data_tail; /* user-space written tail */
263 #define PERF_EVENT_MISC_CPUMODE_MASK (3 << 0)
264 #define PERF_EVENT_MISC_CPUMODE_UNKNOWN (0 << 0)
265 #define PERF_EVENT_MISC_KERNEL (1 << 0)
266 #define PERF_EVENT_MISC_USER (2 << 0)
267 #define PERF_EVENT_MISC_HYPERVISOR (3 << 0)
269 struct perf_event_header {
270 __u32 type;
271 __u16 misc;
272 __u16 size;
275 enum perf_event_type {
278 * The MMAP events record the PROT_EXEC mappings so that we can
279 * correlate userspace IPs to code. They have the following structure:
281 * struct {
282 * struct perf_event_header header;
284 * u32 pid, tid;
285 * u64 addr;
286 * u64 len;
287 * u64 pgoff;
288 * char filename[];
289 * };
291 PERF_EVENT_MMAP = 1,
294 * struct {
295 * struct perf_event_header header;
296 * u64 id;
297 * u64 lost;
298 * };
300 PERF_EVENT_LOST = 2,
303 * struct {
304 * struct perf_event_header header;
306 * u32 pid, tid;
307 * char comm[];
308 * };
310 PERF_EVENT_COMM = 3,
313 * struct {
314 * struct perf_event_header header;
315 * u32 pid, ppid;
316 * u32 tid, ptid;
317 * };
319 PERF_EVENT_EXIT = 4,
322 * struct {
323 * struct perf_event_header header;
324 * u64 time;
325 * u64 id;
326 * u64 stream_id;
327 * };
329 PERF_EVENT_THROTTLE = 5,
330 PERF_EVENT_UNTHROTTLE = 6,
333 * struct {
334 * struct perf_event_header header;
335 * u32 pid, ppid;
336 * u32 tid, ptid;
337 * };
339 PERF_EVENT_FORK = 7,
342 * struct {
343 * struct perf_event_header header;
344 * u32 pid, tid;
345 * u64 value;
346 * { u64 time_enabled; } && PERF_FORMAT_ENABLED
347 * { u64 time_running; } && PERF_FORMAT_RUNNING
348 * { u64 parent_id; } && PERF_FORMAT_ID
349 * };
351 PERF_EVENT_READ = 8,
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 id; } && PERF_SAMPLE_ID
362 * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID
363 * { u32 cpu, res; } && PERF_SAMPLE_CPU
364 * { u64 period; } && PERF_SAMPLE_PERIOD
366 * { u64 nr;
367 * { u64 id, val; } cnt[nr]; } && PERF_SAMPLE_GROUP
369 * { u64 nr,
370 * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN
371 * };
373 PERF_EVENT_SAMPLE = 9,
375 PERF_EVENT_MAX, /* non-ABI */
378 enum perf_callchain_context {
379 PERF_CONTEXT_HV = (__u64)-32,
380 PERF_CONTEXT_KERNEL = (__u64)-128,
381 PERF_CONTEXT_USER = (__u64)-512,
383 PERF_CONTEXT_GUEST = (__u64)-2048,
384 PERF_CONTEXT_GUEST_KERNEL = (__u64)-2176,
385 PERF_CONTEXT_GUEST_USER = (__u64)-2560,
387 PERF_CONTEXT_MAX = (__u64)-4095,
390 #ifdef __KERNEL__
392 * Kernel-internal data types and definitions:
395 #ifdef CONFIG_PERF_COUNTERS
396 # include <asm/perf_counter.h>
397 #endif
399 #include <linux/list.h>
400 #include <linux/mutex.h>
401 #include <linux/rculist.h>
402 #include <linux/rcupdate.h>
403 #include <linux/spinlock.h>
404 #include <linux/hrtimer.h>
405 #include <linux/fs.h>
406 #include <linux/pid_namespace.h>
407 #include <asm/atomic.h>
409 #define PERF_MAX_STACK_DEPTH 255
411 struct perf_callchain_entry {
412 __u64 nr;
413 __u64 ip[PERF_MAX_STACK_DEPTH];
416 struct task_struct;
419 * struct hw_perf_counter - performance counter hardware details:
421 struct hw_perf_counter {
422 #ifdef CONFIG_PERF_COUNTERS
423 union {
424 struct { /* hardware */
425 u64 config;
426 unsigned long config_base;
427 unsigned long counter_base;
428 int idx;
430 union { /* software */
431 atomic64_t count;
432 struct hrtimer hrtimer;
435 atomic64_t prev_count;
436 u64 sample_period;
437 u64 last_period;
438 atomic64_t period_left;
439 u64 interrupts;
441 u64 freq_count;
442 u64 freq_interrupts;
443 u64 freq_stamp;
444 #endif
447 struct perf_counter;
450 * struct pmu - generic performance monitoring unit
452 struct pmu {
453 int (*enable) (struct perf_counter *counter);
454 void (*disable) (struct perf_counter *counter);
455 void (*read) (struct perf_counter *counter);
456 void (*unthrottle) (struct perf_counter *counter);
460 * enum perf_counter_active_state - the states of a counter
462 enum perf_counter_active_state {
463 PERF_COUNTER_STATE_ERROR = -2,
464 PERF_COUNTER_STATE_OFF = -1,
465 PERF_COUNTER_STATE_INACTIVE = 0,
466 PERF_COUNTER_STATE_ACTIVE = 1,
469 struct file;
471 struct perf_mmap_data {
472 struct rcu_head rcu_head;
473 int nr_pages; /* nr of data pages */
474 int writable; /* are we writable */
475 int nr_locked; /* nr pages mlocked */
477 atomic_t poll; /* POLL_ for wakeups */
478 atomic_t events; /* event limit */
480 atomic_long_t head; /* write position */
481 atomic_long_t done_head; /* completed head */
483 atomic_t lock; /* concurrent writes */
484 atomic_t wakeup; /* needs a wakeup */
485 atomic_t lost; /* nr records lost */
487 struct perf_counter_mmap_page *user_page;
488 void *data_pages[0];
491 struct perf_pending_entry {
492 struct perf_pending_entry *next;
493 void (*func)(struct perf_pending_entry *);
497 * struct perf_counter - performance counter kernel representation:
499 struct perf_counter {
500 #ifdef CONFIG_PERF_COUNTERS
501 struct list_head list_entry;
502 struct list_head event_entry;
503 struct list_head sibling_list;
504 int nr_siblings;
505 struct perf_counter *group_leader;
506 const struct pmu *pmu;
508 enum perf_counter_active_state state;
509 atomic64_t count;
512 * These are the total time in nanoseconds that the counter
513 * has been enabled (i.e. eligible to run, and the task has
514 * been scheduled in, if this is a per-task counter)
515 * and running (scheduled onto the CPU), respectively.
517 * They are computed from tstamp_enabled, tstamp_running and
518 * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
520 u64 total_time_enabled;
521 u64 total_time_running;
524 * These are timestamps used for computing total_time_enabled
525 * and total_time_running when the counter is in INACTIVE or
526 * ACTIVE state, measured in nanoseconds from an arbitrary point
527 * in time.
528 * tstamp_enabled: the notional time when the counter was enabled
529 * tstamp_running: the notional time when the counter was scheduled on
530 * tstamp_stopped: in INACTIVE state, the notional time when the
531 * counter was scheduled off.
533 u64 tstamp_enabled;
534 u64 tstamp_running;
535 u64 tstamp_stopped;
537 struct perf_counter_attr attr;
538 struct hw_perf_counter hw;
540 struct perf_counter_context *ctx;
541 struct file *filp;
544 * These accumulate total time (in nanoseconds) that children
545 * counters have been enabled and running, respectively.
547 atomic64_t child_total_time_enabled;
548 atomic64_t child_total_time_running;
551 * Protect attach/detach and child_list:
553 struct mutex child_mutex;
554 struct list_head child_list;
555 struct perf_counter *parent;
557 int oncpu;
558 int cpu;
560 struct list_head owner_entry;
561 struct task_struct *owner;
563 /* mmap bits */
564 struct mutex mmap_mutex;
565 atomic_t mmap_count;
566 struct perf_mmap_data *data;
568 /* poll related */
569 wait_queue_head_t waitq;
570 struct fasync_struct *fasync;
572 /* delayed work for NMIs and such */
573 int pending_wakeup;
574 int pending_kill;
575 int pending_disable;
576 struct perf_pending_entry pending;
578 atomic_t event_limit;
580 void (*destroy)(struct perf_counter *);
581 struct rcu_head rcu_head;
583 struct pid_namespace *ns;
584 u64 id;
585 #endif
589 * struct perf_counter_context - counter context structure
591 * Used as a container for task counters and CPU counters as well:
593 struct perf_counter_context {
595 * Protect the states of the counters in the list,
596 * nr_active, and the list:
598 spinlock_t lock;
600 * Protect the list of counters. Locking either mutex or lock
601 * is sufficient to ensure the list doesn't change; to change
602 * the list you need to lock both the mutex and the spinlock.
604 struct mutex mutex;
606 struct list_head counter_list;
607 struct list_head event_list;
608 int nr_counters;
609 int nr_active;
610 int is_active;
611 int nr_stat;
612 atomic_t refcount;
613 struct task_struct *task;
616 * Context clock, runs when context enabled.
618 u64 time;
619 u64 timestamp;
622 * These fields let us detect when two contexts have both
623 * been cloned (inherited) from a common ancestor.
625 struct perf_counter_context *parent_ctx;
626 u64 parent_gen;
627 u64 generation;
628 int pin_count;
629 struct rcu_head rcu_head;
633 * struct perf_counter_cpu_context - per cpu counter context structure
635 struct perf_cpu_context {
636 struct perf_counter_context ctx;
637 struct perf_counter_context *task_ctx;
638 int active_oncpu;
639 int max_pertask;
640 int exclusive;
643 * Recursion avoidance:
645 * task, softirq, irq, nmi context
647 int recursion[4];
650 #ifdef CONFIG_PERF_COUNTERS
653 * Set by architecture code:
655 extern int perf_max_counters;
657 extern const struct pmu *hw_perf_counter_init(struct perf_counter *counter);
659 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
660 extern void perf_counter_task_sched_out(struct task_struct *task,
661 struct task_struct *next, int cpu);
662 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
663 extern int perf_counter_init_task(struct task_struct *child);
664 extern void perf_counter_exit_task(struct task_struct *child);
665 extern void perf_counter_free_task(struct task_struct *task);
666 extern void set_perf_counter_pending(void);
667 extern void perf_counter_do_pending(void);
668 extern void perf_counter_print_debug(void);
669 extern void __perf_disable(void);
670 extern bool __perf_enable(void);
671 extern void perf_disable(void);
672 extern void perf_enable(void);
673 extern int perf_counter_task_disable(void);
674 extern int perf_counter_task_enable(void);
675 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
676 struct perf_cpu_context *cpuctx,
677 struct perf_counter_context *ctx, int cpu);
678 extern void perf_counter_update_userpage(struct perf_counter *counter);
680 struct perf_sample_data {
681 struct pt_regs *regs;
682 u64 addr;
683 u64 period;
686 extern int perf_counter_overflow(struct perf_counter *counter, int nmi,
687 struct perf_sample_data *data);
690 * Return 1 for a software counter, 0 for a hardware counter
692 static inline int is_software_counter(struct perf_counter *counter)
694 return (counter->attr.type != PERF_TYPE_RAW) &&
695 (counter->attr.type != PERF_TYPE_HARDWARE) &&
696 (counter->attr.type != PERF_TYPE_HW_CACHE);
699 extern atomic_t perf_swcounter_enabled[PERF_COUNT_SW_MAX];
701 extern void __perf_swcounter_event(u32, u64, int, struct pt_regs *, u64);
703 static inline void
704 perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
706 if (atomic_read(&perf_swcounter_enabled[event]))
707 __perf_swcounter_event(event, nr, nmi, regs, addr);
710 extern void __perf_counter_mmap(struct vm_area_struct *vma);
712 static inline void perf_counter_mmap(struct vm_area_struct *vma)
714 if (vma->vm_flags & VM_EXEC)
715 __perf_counter_mmap(vma);
718 extern void perf_counter_comm(struct task_struct *tsk);
719 extern void perf_counter_fork(struct task_struct *tsk);
721 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
723 extern int sysctl_perf_counter_paranoid;
724 extern int sysctl_perf_counter_mlock;
725 extern int sysctl_perf_counter_sample_rate;
727 extern void perf_counter_init(void);
729 #ifndef perf_misc_flags
730 #define perf_misc_flags(regs) (user_mode(regs) ? PERF_EVENT_MISC_USER : \
731 PERF_EVENT_MISC_KERNEL)
732 #define perf_instruction_pointer(regs) instruction_pointer(regs)
733 #endif
735 #else
736 static inline void
737 perf_counter_task_sched_in(struct task_struct *task, int cpu) { }
738 static inline void
739 perf_counter_task_sched_out(struct task_struct *task,
740 struct task_struct *next, int cpu) { }
741 static inline void
742 perf_counter_task_tick(struct task_struct *task, int cpu) { }
743 static inline int perf_counter_init_task(struct task_struct *child) { return 0; }
744 static inline void perf_counter_exit_task(struct task_struct *child) { }
745 static inline void perf_counter_free_task(struct task_struct *task) { }
746 static inline void perf_counter_do_pending(void) { }
747 static inline void perf_counter_print_debug(void) { }
748 static inline void perf_disable(void) { }
749 static inline void perf_enable(void) { }
750 static inline int perf_counter_task_disable(void) { return -EINVAL; }
751 static inline int perf_counter_task_enable(void) { return -EINVAL; }
753 static inline void
754 perf_swcounter_event(u32 event, u64 nr, int nmi,
755 struct pt_regs *regs, u64 addr) { }
757 static inline void perf_counter_mmap(struct vm_area_struct *vma) { }
758 static inline void perf_counter_comm(struct task_struct *tsk) { }
759 static inline void perf_counter_fork(struct task_struct *tsk) { }
760 static inline void perf_counter_init(void) { }
761 #endif
763 #endif /* __KERNEL__ */
764 #endif /* _LINUX_PERF_COUNTER_H */