2 * Performance counters:
4 * Copyright(C) 2008, Thomas Gleixner <tglx@linutronix.de>
5 * Copyright(C) 2008, Red Hat, Inc., Ingo Molnar
7 * Data type definitions, declarations, prototypes.
9 * Started by: Thomas Gleixner and Ingo Molnar
11 * For licencing details see kernel-base/COPYING
13 #ifndef _LINUX_PERF_COUNTER_H
14 #define _LINUX_PERF_COUNTER_H
16 #include <linux/types.h>
17 #include <linux/ioctl.h>
18 #include <asm/byteorder.h>
21 * User-space ABI bits:
27 enum perf_event_types
{
28 PERF_TYPE_HARDWARE
= 0,
29 PERF_TYPE_SOFTWARE
= 1,
30 PERF_TYPE_TRACEPOINT
= 2,
33 * available TYPE space, raw is the max value.
40 * Generalized performance counter event types, used by the hw_event.event_id
41 * parameter of the sys_perf_counter_open() syscall:
45 * Common hardware events, generalized by the kernel:
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,
55 PERF_HW_EVENTS_MAX
= 7,
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
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,
73 PERF_SW_EVENTS_MAX
= 7,
76 #define __PERF_COUNTER_MASK(name) \
77 (((1ULL << PERF_COUNTER_##name##_BITS) - 1) << \
78 PERF_COUNTER_##name##_SHIFT)
80 #define PERF_COUNTER_RAW_BITS 1
81 #define PERF_COUNTER_RAW_SHIFT 63
82 #define PERF_COUNTER_RAW_MASK __PERF_COUNTER_MASK(RAW)
84 #define PERF_COUNTER_CONFIG_BITS 63
85 #define PERF_COUNTER_CONFIG_SHIFT 0
86 #define PERF_COUNTER_CONFIG_MASK __PERF_COUNTER_MASK(CONFIG)
88 #define PERF_COUNTER_TYPE_BITS 7
89 #define PERF_COUNTER_TYPE_SHIFT 56
90 #define PERF_COUNTER_TYPE_MASK __PERF_COUNTER_MASK(TYPE)
92 #define PERF_COUNTER_EVENT_BITS 56
93 #define PERF_COUNTER_EVENT_SHIFT 0
94 #define PERF_COUNTER_EVENT_MASK __PERF_COUNTER_MASK(EVENT)
97 * Bits that can be set in hw_event.record_type to request information
98 * in the overflow packets.
100 enum perf_counter_record_format
{
101 PERF_RECORD_IP
= 1U << 0,
102 PERF_RECORD_TID
= 1U << 1,
103 PERF_RECORD_GROUP
= 1U << 2,
104 PERF_RECORD_CALLCHAIN
= 1U << 3,
108 * Bits that can be set in hw_event.read_format to request that
109 * reads on the counter should return the indicated quantities,
110 * in increasing order of bit value, after the counter value.
112 enum perf_counter_read_format
{
113 PERF_FORMAT_TOTAL_TIME_ENABLED
= 1,
114 PERF_FORMAT_TOTAL_TIME_RUNNING
= 2,
118 * Hardware event to monitor via a performance monitoring counter:
120 struct perf_counter_hw_event
{
122 * The MSB of the config word signifies if the rest contains cpu
123 * specific (raw) counter configuration data, if unset, the next
124 * 7 bits are an event type and the rest of the bits are the event
133 __u64 disabled
: 1, /* off by default */
134 nmi
: 1, /* NMI sampling */
135 inherit
: 1, /* children inherit it */
136 pinned
: 1, /* must always be on PMU */
137 exclusive
: 1, /* only group on PMU */
138 exclude_user
: 1, /* don't count user */
139 exclude_kernel
: 1, /* ditto kernel */
140 exclude_hv
: 1, /* ditto hypervisor */
141 exclude_idle
: 1, /* don't count when idle */
142 mmap
: 1, /* include mmap data */
143 munmap
: 1, /* include munmap data */
147 __u32 extra_config_len
;
148 __u32 wakeup_events
; /* wakeup every n events */
155 * Ioctls that can be done on a perf counter fd:
157 #define PERF_COUNTER_IOC_ENABLE _IO('$', 0)
158 #define PERF_COUNTER_IOC_DISABLE _IO('$', 1)
161 * Structure of the page that can be mapped via mmap
163 struct perf_counter_mmap_page
{
164 __u32 version
; /* version number of this structure */
165 __u32 compat_version
; /* lowest version this is compat with */
168 * Bits needed to read the hw counters in user-space.
178 * count = pmc_read(pc->index - 1);
179 * count += pc->offset;
184 * } while (pc->lock != seq);
186 * NOTE: for obvious reason this only works on self-monitoring
189 __u32 lock
; /* seqlock for synchronization */
190 __u32 index
; /* hardware counter identifier */
191 __s64 offset
; /* add to hardware counter value */
194 * Control data for the mmap() data buffer.
196 * User-space reading this value should issue an rmb(), on SMP capable
197 * platforms, after reading this value -- see perf_counter_wakeup().
199 __u32 data_head
; /* head in the data section */
202 struct perf_event_header
{
207 enum perf_event_type
{
210 PERF_EVENT_MUNMAP
= 2,
213 * Half the event type space is reserved for the counter overflow
214 * bitfields, as found in hw_event.record_type.
216 * These events will have types of the form:
217 * PERF_EVENT_COUNTER_OVERFLOW { | __PERF_EVENT_* } *
219 PERF_EVENT_COUNTER_OVERFLOW
= 1UL << 31,
220 __PERF_EVENT_IP
= PERF_RECORD_IP
,
221 __PERF_EVENT_TID
= PERF_RECORD_TID
,
222 __PERF_EVENT_GROUP
= PERF_RECORD_GROUP
,
223 __PERF_EVENT_CALLCHAIN
= PERF_RECORD_CALLCHAIN
,
228 * Kernel-internal data types and definitions:
231 #ifdef CONFIG_PERF_COUNTERS
232 # include <asm/perf_counter.h>
235 #include <linux/list.h>
236 #include <linux/mutex.h>
237 #include <linux/rculist.h>
238 #include <linux/rcupdate.h>
239 #include <linux/spinlock.h>
240 #include <linux/hrtimer.h>
241 #include <asm/atomic.h>
245 static inline u64
perf_event_raw(struct perf_counter_hw_event
*hw_event
)
247 return hw_event
->config
& PERF_COUNTER_RAW_MASK
;
250 static inline u64
perf_event_config(struct perf_counter_hw_event
*hw_event
)
252 return hw_event
->config
& PERF_COUNTER_CONFIG_MASK
;
255 static inline u64
perf_event_type(struct perf_counter_hw_event
*hw_event
)
257 return (hw_event
->config
& PERF_COUNTER_TYPE_MASK
) >>
258 PERF_COUNTER_TYPE_SHIFT
;
261 static inline u64
perf_event_id(struct perf_counter_hw_event
*hw_event
)
263 return hw_event
->config
& PERF_COUNTER_EVENT_MASK
;
267 * struct hw_perf_counter - performance counter hardware details:
269 struct hw_perf_counter
{
270 #ifdef CONFIG_PERF_COUNTERS
272 struct { /* hardware */
274 unsigned long config_base
;
275 unsigned long counter_base
;
279 union { /* software */
281 struct hrtimer hrtimer
;
284 atomic64_t prev_count
;
286 atomic64_t period_left
;
293 * struct hw_perf_counter_ops - performance counter hw ops
295 struct hw_perf_counter_ops
{
296 int (*enable
) (struct perf_counter
*counter
);
297 void (*disable
) (struct perf_counter
*counter
);
298 void (*read
) (struct perf_counter
*counter
);
302 * enum perf_counter_active_state - the states of a counter
304 enum perf_counter_active_state
{
305 PERF_COUNTER_STATE_ERROR
= -2,
306 PERF_COUNTER_STATE_OFF
= -1,
307 PERF_COUNTER_STATE_INACTIVE
= 0,
308 PERF_COUNTER_STATE_ACTIVE
= 1,
313 struct perf_mmap_data
{
314 struct rcu_head rcu_head
;
319 struct perf_counter_mmap_page
*user_page
;
323 struct perf_wakeup_entry
{
324 struct perf_wakeup_entry
*next
;
328 * struct perf_counter - performance counter kernel representation:
330 struct perf_counter
{
331 #ifdef CONFIG_PERF_COUNTERS
332 struct list_head list_entry
;
333 struct list_head event_entry
;
334 struct list_head sibling_list
;
336 struct perf_counter
*group_leader
;
337 const struct hw_perf_counter_ops
*hw_ops
;
339 enum perf_counter_active_state state
;
340 enum perf_counter_active_state prev_state
;
344 * These are the total time in nanoseconds that the counter
345 * has been enabled (i.e. eligible to run, and the task has
346 * been scheduled in, if this is a per-task counter)
347 * and running (scheduled onto the CPU), respectively.
349 * They are computed from tstamp_enabled, tstamp_running and
350 * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
352 u64 total_time_enabled
;
353 u64 total_time_running
;
356 * These are timestamps used for computing total_time_enabled
357 * and total_time_running when the counter is in INACTIVE or
358 * ACTIVE state, measured in nanoseconds from an arbitrary point
360 * tstamp_enabled: the notional time when the counter was enabled
361 * tstamp_running: the notional time when the counter was scheduled on
362 * tstamp_stopped: in INACTIVE state, the notional time when the
363 * counter was scheduled off.
369 struct perf_counter_hw_event hw_event
;
370 struct hw_perf_counter hw
;
372 struct perf_counter_context
*ctx
;
373 struct task_struct
*task
;
376 struct perf_counter
*parent
;
377 struct list_head child_list
;
380 * These accumulate total time (in nanoseconds) that children
381 * counters have been enabled and running, respectively.
383 atomic64_t child_total_time_enabled
;
384 atomic64_t child_total_time_running
;
387 * Protect attach/detach and child_list:
395 struct mutex mmap_mutex
;
397 struct perf_mmap_data
*data
;
400 wait_queue_head_t waitq
;
401 /* optional: for NMIs */
402 struct perf_wakeup_entry wakeup
;
404 void (*destroy
)(struct perf_counter
*);
405 struct rcu_head rcu_head
;
410 * struct perf_counter_context - counter context structure
412 * Used as a container for task counters and CPU counters as well:
414 struct perf_counter_context
{
415 #ifdef CONFIG_PERF_COUNTERS
417 * Protect the states of the counters in the list,
418 * nr_active, and the list:
422 * Protect the list of counters. Locking either mutex or lock
423 * is sufficient to ensure the list doesn't change; to change
424 * the list you need to lock both the mutex and the spinlock.
428 struct list_head counter_list
;
429 struct list_head event_list
;
433 struct task_struct
*task
;
436 * time_now is the current time in nanoseconds since an arbitrary
437 * point in the past. For per-task counters, this is based on the
438 * task clock, and for per-cpu counters it is based on the cpu clock.
439 * time_lost is an offset from the task/cpu clock, used to make it
440 * appear that time only passes while the context is scheduled in.
448 * struct perf_counter_cpu_context - per cpu counter context structure
450 struct perf_cpu_context
{
451 struct perf_counter_context ctx
;
452 struct perf_counter_context
*task_ctx
;
458 * Recursion avoidance:
460 * task, softirq, irq, nmi context
466 * Set by architecture code:
468 extern int perf_max_counters
;
470 #ifdef CONFIG_PERF_COUNTERS
471 extern const struct hw_perf_counter_ops
*
472 hw_perf_counter_init(struct perf_counter
*counter
);
474 extern void perf_counter_task_sched_in(struct task_struct
*task
, int cpu
);
475 extern void perf_counter_task_sched_out(struct task_struct
*task
, int cpu
);
476 extern void perf_counter_task_tick(struct task_struct
*task
, int cpu
);
477 extern void perf_counter_init_task(struct task_struct
*child
);
478 extern void perf_counter_exit_task(struct task_struct
*child
);
479 extern void perf_counter_do_pending(void);
480 extern void perf_counter_print_debug(void);
481 extern void perf_counter_unthrottle(void);
482 extern u64
hw_perf_save_disable(void);
483 extern void hw_perf_restore(u64 ctrl
);
484 extern int perf_counter_task_disable(void);
485 extern int perf_counter_task_enable(void);
486 extern int hw_perf_group_sched_in(struct perf_counter
*group_leader
,
487 struct perf_cpu_context
*cpuctx
,
488 struct perf_counter_context
*ctx
, int cpu
);
489 extern void perf_counter_update_userpage(struct perf_counter
*counter
);
491 extern void perf_counter_output(struct perf_counter
*counter
,
492 int nmi
, struct pt_regs
*regs
);
494 * Return 1 for a software counter, 0 for a hardware counter
496 static inline int is_software_counter(struct perf_counter
*counter
)
498 return !perf_event_raw(&counter
->hw_event
) &&
499 perf_event_type(&counter
->hw_event
) != PERF_TYPE_HARDWARE
;
502 extern void perf_swcounter_event(u32
, u64
, int, struct pt_regs
*);
504 extern void perf_counter_mmap(unsigned long addr
, unsigned long len
,
505 unsigned long pgoff
, struct file
*file
);
507 extern void perf_counter_munmap(unsigned long addr
, unsigned long len
,
508 unsigned long pgoff
, struct file
*file
);
510 #define MAX_STACK_DEPTH 254
512 struct perf_callchain_entry
{
513 u32 nr
, hv
, kernel
, user
;
514 u64 ip
[MAX_STACK_DEPTH
];
517 extern struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
);
521 perf_counter_task_sched_in(struct task_struct
*task
, int cpu
) { }
523 perf_counter_task_sched_out(struct task_struct
*task
, int cpu
) { }
525 perf_counter_task_tick(struct task_struct
*task
, int cpu
) { }
526 static inline void perf_counter_init_task(struct task_struct
*child
) { }
527 static inline void perf_counter_exit_task(struct task_struct
*child
) { }
528 static inline void perf_counter_do_pending(void) { }
529 static inline void perf_counter_print_debug(void) { }
530 static inline void perf_counter_unthrottle(void) { }
531 static inline void hw_perf_restore(u64 ctrl
) { }
532 static inline u64
hw_perf_save_disable(void) { return 0; }
533 static inline int perf_counter_task_disable(void) { return -EINVAL
; }
534 static inline int perf_counter_task_enable(void) { return -EINVAL
; }
537 perf_swcounter_event(u32 event
, u64 nr
, int nmi
, struct pt_regs
*regs
) { }
541 perf_counter_mmap(unsigned long addr
, unsigned long len
,
542 unsigned long pgoff
, struct file
*file
) { }
545 perf_counter_munmap(unsigned long addr
, unsigned long len
,
546 unsigned long pgoff
, struct file
*file
) { }
550 #endif /* __KERNEL__ */
551 #endif /* _LINUX_PERF_COUNTER_H */