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_TIME = 1U << 2,
104 PERF_RECORD_ADDR = 1U << 3,
105 PERF_RECORD_GROUP = 1U << 4,
106 PERF_RECORD_CALLCHAIN = 1U << 5,
110 * Bits that can be set in hw_event.read_format to request that
111 * reads on the counter should return the indicated quantities,
112 * in increasing order of bit value, after the counter value.
114 enum perf_counter_read_format {
115 PERF_FORMAT_TOTAL_TIME_ENABLED = 1,
116 PERF_FORMAT_TOTAL_TIME_RUNNING = 2,
120 * Hardware event to monitor via a performance monitoring counter:
122 struct perf_counter_hw_event {
124 * The MSB of the config word signifies if the rest contains cpu
125 * specific (raw) counter configuration data, if unset, the next
126 * 7 bits are an event type and the rest of the bits are the event
135 __u64 disabled : 1, /* off by default */
136 nmi : 1, /* NMI sampling */
137 inherit : 1, /* children inherit it */
138 pinned : 1, /* must always be on PMU */
139 exclusive : 1, /* only group on PMU */
140 exclude_user : 1, /* don't count user */
141 exclude_kernel : 1, /* ditto kernel */
142 exclude_hv : 1, /* ditto hypervisor */
143 exclude_idle : 1, /* don't count when idle */
144 mmap : 1, /* include mmap data */
145 munmap : 1, /* include munmap data */
146 comm : 1, /* include comm data */
150 __u32 extra_config_len;
151 __u32 wakeup_events; /* wakeup every n events */
158 * Ioctls that can be done on a perf counter fd:
160 #define PERF_COUNTER_IOC_ENABLE _IOW('$', 0, u32)
161 #define PERF_COUNTER_IOC_DISABLE _IOW('$', 1, u32)
162 #define PERF_COUNTER_IOC_REFRESH _IOW('$', 2, u32)
163 #define PERF_COUNTER_IOC_RESET _IOW('$', 3, u32)
165 enum perf_counter_ioc_flags {
166 PERF_IOC_FLAG_GROUP = 1U << 0,
170 * Structure of the page that can be mapped via mmap
172 struct perf_counter_mmap_page {
173 __u32 version; /* version number of this structure */
174 __u32 compat_version; /* lowest version this is compat with */
177 * Bits needed to read the hw counters in user-space.
187 * count = pmc_read(pc->index - 1);
188 * count += pc->offset;
193 * } while (pc->lock != seq);
195 * NOTE: for obvious reason this only works on self-monitoring
198 __u32 lock; /* seqlock for synchronization */
199 __u32 index; /* hardware counter identifier */
200 __s64 offset; /* add to hardware counter value */
203 * Control data for the mmap() data buffer.
205 * User-space reading this value should issue an rmb(), on SMP capable
206 * platforms, after reading this value -- see perf_counter_wakeup().
208 __u32 data_head; /* head in the data section */
211 #define PERF_EVENT_MISC_KERNEL (1 << 0)
212 #define PERF_EVENT_MISC_USER (1 << 1)
213 #define PERF_EVENT_MISC_OVERFLOW (1 << 2)
215 struct perf_event_header {
221 enum perf_event_type {
224 * The MMAP events record the PROT_EXEC mappings so that we can
225 * correlate userspace IPs to code. They have the following structure:
228 * struct perf_event_header header;
238 PERF_EVENT_MUNMAP = 2,
242 * struct perf_event_header header;
251 * When header.misc & PERF_EVENT_MISC_OVERFLOW the event_type field
252 * will be PERF_RECORD_*
255 * struct perf_event_header header;
257 * { u64 ip; } && PERF_RECORD_IP
258 * { u32 pid, tid; } && PERF_RECORD_TID
259 * { u64 time; } && PERF_RECORD_TIME
260 * { u64 addr; } && PERF_RECORD_ADDR
263 * { u64 event, val; } cnt[nr]; } && PERF_RECORD_GROUP
269 * u64 ips[nr]; } && PERF_RECORD_CALLCHAIN
276 * Kernel-internal data types and definitions:
279 #ifdef CONFIG_PERF_COUNTERS
280 # include <asm/perf_counter.h>
283 #include <linux/list.h>
284 #include <linux/mutex.h>
285 #include <linux/rculist.h>
286 #include <linux/rcupdate.h>
287 #include <linux/spinlock.h>
288 #include <linux/hrtimer.h>
289 #include <linux/fs.h>
290 #include <asm/atomic.h>
294 static inline u64 perf_event_raw(struct perf_counter_hw_event *hw_event)
296 return hw_event->config & PERF_COUNTER_RAW_MASK;
299 static inline u64 perf_event_config(struct perf_counter_hw_event *hw_event)
301 return hw_event->config & PERF_COUNTER_CONFIG_MASK;
304 static inline u64 perf_event_type(struct perf_counter_hw_event *hw_event)
306 return (hw_event->config & PERF_COUNTER_TYPE_MASK) >>
307 PERF_COUNTER_TYPE_SHIFT;
310 static inline u64 perf_event_id(struct perf_counter_hw_event *hw_event)
312 return hw_event->config & PERF_COUNTER_EVENT_MASK;
316 * struct hw_perf_counter - performance counter hardware details:
318 struct hw_perf_counter {
319 #ifdef CONFIG_PERF_COUNTERS
321 struct { /* hardware */
323 unsigned long config_base;
324 unsigned long counter_base;
328 union { /* software */
330 struct hrtimer hrtimer;
333 atomic64_t prev_count;
335 atomic64_t period_left;
342 * struct pmu - generic performance monitoring unit
345 int (*enable) (struct perf_counter *counter);
346 void (*disable) (struct perf_counter *counter);
347 void (*read) (struct perf_counter *counter);
351 * enum perf_counter_active_state - the states of a counter
353 enum perf_counter_active_state {
354 PERF_COUNTER_STATE_ERROR = -2,
355 PERF_COUNTER_STATE_OFF = -1,
356 PERF_COUNTER_STATE_INACTIVE = 0,
357 PERF_COUNTER_STATE_ACTIVE = 1,
362 struct perf_mmap_data {
363 struct rcu_head rcu_head;
364 int nr_pages; /* nr of data pages */
365 int nr_locked; /* nr pages mlocked */
367 atomic_t poll; /* POLL_ for wakeups */
368 atomic_t head; /* write position */
369 atomic_t events; /* event limit */
371 atomic_t done_head; /* completed head */
372 atomic_t lock; /* concurrent writes */
374 atomic_t wakeup; /* needs a wakeup */
376 struct perf_counter_mmap_page *user_page;
380 struct perf_pending_entry {
381 struct perf_pending_entry *next;
382 void (*func)(struct perf_pending_entry *);
386 * struct perf_counter - performance counter kernel representation:
388 struct perf_counter {
389 #ifdef CONFIG_PERF_COUNTERS
390 struct list_head list_entry;
391 struct list_head event_entry;
392 struct list_head sibling_list;
394 struct perf_counter *group_leader;
395 const struct pmu *pmu;
397 enum perf_counter_active_state state;
398 enum perf_counter_active_state prev_state;
402 * These are the total time in nanoseconds that the counter
403 * has been enabled (i.e. eligible to run, and the task has
404 * been scheduled in, if this is a per-task counter)
405 * and running (scheduled onto the CPU), respectively.
407 * They are computed from tstamp_enabled, tstamp_running and
408 * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
410 u64 total_time_enabled;
411 u64 total_time_running;
414 * These are timestamps used for computing total_time_enabled
415 * and total_time_running when the counter is in INACTIVE or
416 * ACTIVE state, measured in nanoseconds from an arbitrary point
418 * tstamp_enabled: the notional time when the counter was enabled
419 * tstamp_running: the notional time when the counter was scheduled on
420 * tstamp_stopped: in INACTIVE state, the notional time when the
421 * counter was scheduled off.
427 struct perf_counter_hw_event hw_event;
428 struct hw_perf_counter hw;
430 struct perf_counter_context *ctx;
431 struct task_struct *task;
434 struct perf_counter *parent;
435 struct list_head child_list;
438 * These accumulate total time (in nanoseconds) that children
439 * counters have been enabled and running, respectively.
441 atomic64_t child_total_time_enabled;
442 atomic64_t child_total_time_running;
445 * Protect attach/detach and child_list:
453 struct mutex mmap_mutex;
455 struct perf_mmap_data *data;
458 wait_queue_head_t waitq;
459 struct fasync_struct *fasync;
461 /* delayed work for NMIs and such */
465 struct perf_pending_entry pending;
467 atomic_t event_limit;
469 void (*destroy)(struct perf_counter *);
470 struct rcu_head rcu_head;
475 * struct perf_counter_context - counter context structure
477 * Used as a container for task counters and CPU counters as well:
479 struct perf_counter_context {
480 #ifdef CONFIG_PERF_COUNTERS
482 * Protect the states of the counters in the list,
483 * nr_active, and the list:
487 * Protect the list of counters. Locking either mutex or lock
488 * is sufficient to ensure the list doesn't change; to change
489 * the list you need to lock both the mutex and the spinlock.
493 struct list_head counter_list;
494 struct list_head event_list;
498 struct task_struct *task;
501 * Context clock, runs when context enabled.
509 * struct perf_counter_cpu_context - per cpu counter context structure
511 struct perf_cpu_context {
512 struct perf_counter_context ctx;
513 struct perf_counter_context *task_ctx;
519 * Recursion avoidance:
521 * task, softirq, irq, nmi context
526 #ifdef CONFIG_PERF_COUNTERS
529 * Set by architecture code:
531 extern int perf_max_counters;
533 extern const struct pmu *hw_perf_counter_init(struct perf_counter *counter);
535 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
536 extern void perf_counter_task_sched_out(struct task_struct *task, int cpu);
537 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
538 extern void perf_counter_init_task(struct task_struct *child);
539 extern void perf_counter_exit_task(struct task_struct *child);
540 extern void perf_counter_do_pending(void);
541 extern void perf_counter_print_debug(void);
542 extern void perf_counter_unthrottle(void);
543 extern u64 hw_perf_save_disable(void);
544 extern void hw_perf_restore(u64 ctrl);
545 extern int perf_counter_task_disable(void);
546 extern int perf_counter_task_enable(void);
547 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
548 struct perf_cpu_context *cpuctx,
549 struct perf_counter_context *ctx, int cpu);
550 extern void perf_counter_update_userpage(struct perf_counter *counter);
552 extern int perf_counter_overflow(struct perf_counter *counter,
553 int nmi, struct pt_regs *regs, u64 addr);
555 * Return 1 for a software counter, 0 for a hardware counter
557 static inline int is_software_counter(struct perf_counter *counter)
559 return !perf_event_raw(&counter->hw_event) &&
560 perf_event_type(&counter->hw_event) != PERF_TYPE_HARDWARE;
563 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *, u64);
565 extern void perf_counter_mmap(unsigned long addr, unsigned long len,
566 unsigned long pgoff, struct file *file);
568 extern void perf_counter_munmap(unsigned long addr, unsigned long len,
569 unsigned long pgoff, struct file *file);
571 extern void perf_counter_comm(struct task_struct *tsk);
573 #define MAX_STACK_DEPTH 255
575 struct perf_callchain_entry {
576 u16 nr, hv, kernel, user;
577 u64 ip[MAX_STACK_DEPTH];
580 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
582 extern int sysctl_perf_counter_priv;
583 extern int sysctl_perf_counter_mlock;
585 extern void perf_counter_init(void);
589 perf_counter_task_sched_in(struct task_struct *task, int cpu) { }
591 perf_counter_task_sched_out(struct task_struct *task, int cpu) { }
593 perf_counter_task_tick(struct task_struct *task, int cpu) { }
594 static inline void perf_counter_init_task(struct task_struct *child) { }
595 static inline void perf_counter_exit_task(struct task_struct *child) { }
596 static inline void perf_counter_do_pending(void) { }
597 static inline void perf_counter_print_debug(void) { }
598 static inline void perf_counter_unthrottle(void) { }
599 static inline void hw_perf_restore(u64 ctrl) { }
600 static inline u64 hw_perf_save_disable(void) { return 0; }
601 static inline int perf_counter_task_disable(void) { return -EINVAL; }
602 static inline int perf_counter_task_enable(void) { return -EINVAL; }
605 perf_swcounter_event(u32 event, u64 nr, int nmi,
606 struct pt_regs *regs, u64 addr) { }
609 perf_counter_mmap(unsigned long addr, unsigned long len,
610 unsigned long pgoff, struct file *file) { }
613 perf_counter_munmap(unsigned long addr, unsigned long len,
614 unsigned long pgoff, struct file *file) { }
616 static inline void perf_counter_comm(struct task_struct *tsk) { }
617 static inline void perf_counter_init(void) { }
620 #endif /* __KERNEL__ */
621 #endif /* _LINUX_PERF_COUNTER_H */