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 _IO ('$', 0)
161 #define PERF_COUNTER_IOC_DISABLE _IO ('$', 1)
162 #define PERF_COUNTER_IOC_REFRESH _IOW('$', 2, u32)
163 #define PERF_COUNTER_IOC_RESET _IO ('$', 3)
166 * Structure of the page that can be mapped via mmap
168 struct perf_counter_mmap_page {
169 __u32 version; /* version number of this structure */
170 __u32 compat_version; /* lowest version this is compat with */
173 * Bits needed to read the hw counters in user-space.
183 * count = pmc_read(pc->index - 1);
184 * count += pc->offset;
189 * } while (pc->lock != seq);
191 * NOTE: for obvious reason this only works on self-monitoring
194 __u32 lock; /* seqlock for synchronization */
195 __u32 index; /* hardware counter identifier */
196 __s64 offset; /* add to hardware counter value */
199 * Control data for the mmap() data buffer.
201 * User-space reading this value should issue an rmb(), on SMP capable
202 * platforms, after reading this value -- see perf_counter_wakeup().
204 __u32 data_head; /* head in the data section */
207 #define PERF_EVENT_MISC_KERNEL (1 << 0)
208 #define PERF_EVENT_MISC_USER (1 << 1)
209 #define PERF_EVENT_MISC_OVERFLOW (1 << 2)
211 struct perf_event_header {
217 enum perf_event_type {
220 * The MMAP events record the PROT_EXEC mappings so that we can
221 * correlate userspace IPs to code. They have the following structure:
224 * struct perf_event_header header;
234 PERF_EVENT_MUNMAP = 2,
238 * struct perf_event_header header;
247 * When header.misc & PERF_EVENT_MISC_OVERFLOW the event_type field
248 * will be PERF_RECORD_*
251 * struct perf_event_header header;
253 * { u64 ip; } && PERF_RECORD_IP
254 * { u32 pid, tid; } && PERF_RECORD_TID
255 * { u64 time; } && PERF_RECORD_TIME
256 * { u64 addr; } && PERF_RECORD_ADDR
259 * { u64 event, val; } cnt[nr]; } && PERF_RECORD_GROUP
265 * u64 ips[nr]; } && PERF_RECORD_CALLCHAIN
272 * Kernel-internal data types and definitions:
275 #ifdef CONFIG_PERF_COUNTERS
276 # include <asm/perf_counter.h>
279 #include <linux/list.h>
280 #include <linux/mutex.h>
281 #include <linux/rculist.h>
282 #include <linux/rcupdate.h>
283 #include <linux/spinlock.h>
284 #include <linux/hrtimer.h>
285 #include <linux/fs.h>
286 #include <asm/atomic.h>
290 static inline u64 perf_event_raw(struct perf_counter_hw_event *hw_event)
292 return hw_event->config & PERF_COUNTER_RAW_MASK;
295 static inline u64 perf_event_config(struct perf_counter_hw_event *hw_event)
297 return hw_event->config & PERF_COUNTER_CONFIG_MASK;
300 static inline u64 perf_event_type(struct perf_counter_hw_event *hw_event)
302 return (hw_event->config & PERF_COUNTER_TYPE_MASK) >>
303 PERF_COUNTER_TYPE_SHIFT;
306 static inline u64 perf_event_id(struct perf_counter_hw_event *hw_event)
308 return hw_event->config & PERF_COUNTER_EVENT_MASK;
312 * struct hw_perf_counter - performance counter hardware details:
314 struct hw_perf_counter {
315 #ifdef CONFIG_PERF_COUNTERS
317 struct { /* hardware */
319 unsigned long config_base;
320 unsigned long counter_base;
324 union { /* software */
326 struct hrtimer hrtimer;
329 atomic64_t prev_count;
331 atomic64_t period_left;
338 * struct pmu - generic performance monitoring unit
341 int (*enable) (struct perf_counter *counter);
342 void (*disable) (struct perf_counter *counter);
343 void (*read) (struct perf_counter *counter);
347 * enum perf_counter_active_state - the states of a counter
349 enum perf_counter_active_state {
350 PERF_COUNTER_STATE_ERROR = -2,
351 PERF_COUNTER_STATE_OFF = -1,
352 PERF_COUNTER_STATE_INACTIVE = 0,
353 PERF_COUNTER_STATE_ACTIVE = 1,
358 struct perf_mmap_data {
359 struct rcu_head rcu_head;
360 int nr_pages; /* nr of data pages */
361 int nr_locked; /* nr pages mlocked */
363 atomic_t poll; /* POLL_ for wakeups */
364 atomic_t head; /* write position */
365 atomic_t events; /* event limit */
367 atomic_t done_head; /* completed head */
368 atomic_t lock; /* concurrent writes */
370 atomic_t wakeup; /* needs a wakeup */
372 struct perf_counter_mmap_page *user_page;
376 struct perf_pending_entry {
377 struct perf_pending_entry *next;
378 void (*func)(struct perf_pending_entry *);
382 * struct perf_counter - performance counter kernel representation:
384 struct perf_counter {
385 #ifdef CONFIG_PERF_COUNTERS
386 struct list_head list_entry;
387 struct list_head event_entry;
388 struct list_head sibling_list;
390 struct perf_counter *group_leader;
391 const struct pmu *pmu;
393 enum perf_counter_active_state state;
394 enum perf_counter_active_state prev_state;
398 * These are the total time in nanoseconds that the counter
399 * has been enabled (i.e. eligible to run, and the task has
400 * been scheduled in, if this is a per-task counter)
401 * and running (scheduled onto the CPU), respectively.
403 * They are computed from tstamp_enabled, tstamp_running and
404 * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
406 u64 total_time_enabled;
407 u64 total_time_running;
410 * These are timestamps used for computing total_time_enabled
411 * and total_time_running when the counter is in INACTIVE or
412 * ACTIVE state, measured in nanoseconds from an arbitrary point
414 * tstamp_enabled: the notional time when the counter was enabled
415 * tstamp_running: the notional time when the counter was scheduled on
416 * tstamp_stopped: in INACTIVE state, the notional time when the
417 * counter was scheduled off.
423 struct perf_counter_hw_event hw_event;
424 struct hw_perf_counter hw;
426 struct perf_counter_context *ctx;
427 struct task_struct *task;
430 struct perf_counter *parent;
431 struct list_head child_list;
434 * These accumulate total time (in nanoseconds) that children
435 * counters have been enabled and running, respectively.
437 atomic64_t child_total_time_enabled;
438 atomic64_t child_total_time_running;
441 * Protect attach/detach and child_list:
449 struct mutex mmap_mutex;
451 struct perf_mmap_data *data;
454 wait_queue_head_t waitq;
455 struct fasync_struct *fasync;
457 /* delayed work for NMIs and such */
461 struct perf_pending_entry pending;
463 atomic_t event_limit;
465 void (*destroy)(struct perf_counter *);
466 struct rcu_head rcu_head;
471 * struct perf_counter_context - counter context structure
473 * Used as a container for task counters and CPU counters as well:
475 struct perf_counter_context {
476 #ifdef CONFIG_PERF_COUNTERS
478 * Protect the states of the counters in the list,
479 * nr_active, and the list:
483 * Protect the list of counters. Locking either mutex or lock
484 * is sufficient to ensure the list doesn't change; to change
485 * the list you need to lock both the mutex and the spinlock.
489 struct list_head counter_list;
490 struct list_head event_list;
494 struct task_struct *task;
497 * Context clock, runs when context enabled.
505 * struct perf_counter_cpu_context - per cpu counter context structure
507 struct perf_cpu_context {
508 struct perf_counter_context ctx;
509 struct perf_counter_context *task_ctx;
515 * Recursion avoidance:
517 * task, softirq, irq, nmi context
522 #ifdef CONFIG_PERF_COUNTERS
525 * Set by architecture code:
527 extern int perf_max_counters;
529 extern const struct pmu *hw_perf_counter_init(struct perf_counter *counter);
531 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
532 extern void perf_counter_task_sched_out(struct task_struct *task, int cpu);
533 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
534 extern void perf_counter_init_task(struct task_struct *child);
535 extern void perf_counter_exit_task(struct task_struct *child);
536 extern void perf_counter_do_pending(void);
537 extern void perf_counter_print_debug(void);
538 extern void perf_counter_unthrottle(void);
539 extern u64 hw_perf_save_disable(void);
540 extern void hw_perf_restore(u64 ctrl);
541 extern int perf_counter_task_disable(void);
542 extern int perf_counter_task_enable(void);
543 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
544 struct perf_cpu_context *cpuctx,
545 struct perf_counter_context *ctx, int cpu);
546 extern void perf_counter_update_userpage(struct perf_counter *counter);
548 extern int perf_counter_overflow(struct perf_counter *counter,
549 int nmi, struct pt_regs *regs, u64 addr);
551 * Return 1 for a software counter, 0 for a hardware counter
553 static inline int is_software_counter(struct perf_counter *counter)
555 return !perf_event_raw(&counter->hw_event) &&
556 perf_event_type(&counter->hw_event) != PERF_TYPE_HARDWARE;
559 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *, u64);
561 extern void perf_counter_mmap(unsigned long addr, unsigned long len,
562 unsigned long pgoff, struct file *file);
564 extern void perf_counter_munmap(unsigned long addr, unsigned long len,
565 unsigned long pgoff, struct file *file);
567 extern void perf_counter_comm(struct task_struct *tsk);
569 #define MAX_STACK_DEPTH 255
571 struct perf_callchain_entry {
572 u16 nr, hv, kernel, user;
573 u64 ip[MAX_STACK_DEPTH];
576 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
578 extern int sysctl_perf_counter_priv;
579 extern int sysctl_perf_counter_mlock;
581 extern void perf_counter_init(void);
585 perf_counter_task_sched_in(struct task_struct *task, int cpu) { }
587 perf_counter_task_sched_out(struct task_struct *task, int cpu) { }
589 perf_counter_task_tick(struct task_struct *task, int cpu) { }
590 static inline void perf_counter_init_task(struct task_struct *child) { }
591 static inline void perf_counter_exit_task(struct task_struct *child) { }
592 static inline void perf_counter_do_pending(void) { }
593 static inline void perf_counter_print_debug(void) { }
594 static inline void perf_counter_unthrottle(void) { }
595 static inline void hw_perf_restore(u64 ctrl) { }
596 static inline u64 hw_perf_save_disable(void) { return 0; }
597 static inline int perf_counter_task_disable(void) { return -EINVAL; }
598 static inline int perf_counter_task_enable(void) { return -EINVAL; }
601 perf_swcounter_event(u32 event, u64 nr, int nmi,
602 struct pt_regs *regs, u64 addr) { }
605 perf_counter_mmap(unsigned long addr, unsigned long len,
606 unsigned long pgoff, struct file *file) { }
609 perf_counter_munmap(unsigned long addr, unsigned long len,
610 unsigned long pgoff, struct file *file) { }
612 static inline void perf_counter_comm(struct task_struct *tsk) { }
613 static inline void perf_counter_init(void) { }
616 #endif /* __KERNEL__ */
617 #endif /* _LINUX_PERF_COUNTER_H */