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,
107 PERF_RECORD_CONFIG = 1U << 6,
108 PERF_RECORD_CPU = 1U << 7,
112 * Bits that can be set in hw_event.read_format to request that
113 * reads on the counter should return the indicated quantities,
114 * in increasing order of bit value, after the counter value.
116 enum perf_counter_read_format {
117 PERF_FORMAT_TOTAL_TIME_ENABLED = 1,
118 PERF_FORMAT_TOTAL_TIME_RUNNING = 2,
122 * Hardware event to monitor via a performance monitoring counter:
124 struct perf_counter_hw_event {
126 * The MSB of the config word signifies if the rest contains cpu
127 * specific (raw) counter configuration data, if unset, the next
128 * 7 bits are an event type and the rest of the bits are the event
141 __u64 disabled : 1, /* off by default */
142 nmi : 1, /* NMI sampling */
143 inherit : 1, /* children inherit it */
144 pinned : 1, /* must always be on PMU */
145 exclusive : 1, /* only group on PMU */
146 exclude_user : 1, /* don't count user */
147 exclude_kernel : 1, /* ditto kernel */
148 exclude_hv : 1, /* ditto hypervisor */
149 exclude_idle : 1, /* don't count when idle */
150 mmap : 1, /* include mmap data */
151 munmap : 1, /* include munmap data */
152 comm : 1, /* include comm data */
153 freq : 1, /* use freq, not period */
157 __u32 extra_config_len;
158 __u32 wakeup_events; /* wakeup every n events */
165 * Ioctls that can be done on a perf counter fd:
167 #define PERF_COUNTER_IOC_ENABLE _IOW('$', 0, u32)
168 #define PERF_COUNTER_IOC_DISABLE _IOW('$', 1, u32)
169 #define PERF_COUNTER_IOC_REFRESH _IOW('$', 2, u32)
170 #define PERF_COUNTER_IOC_RESET _IOW('$', 3, u32)
172 enum perf_counter_ioc_flags {
173 PERF_IOC_FLAG_GROUP = 1U << 0,
177 * Structure of the page that can be mapped via mmap
179 struct perf_counter_mmap_page {
180 __u32 version; /* version number of this structure */
181 __u32 compat_version; /* lowest version this is compat with */
184 * Bits needed to read the hw counters in user-space.
194 * count = pmc_read(pc->index - 1);
195 * count += pc->offset;
200 * } while (pc->lock != seq);
202 * NOTE: for obvious reason this only works on self-monitoring
205 __u32 lock; /* seqlock for synchronization */
206 __u32 index; /* hardware counter identifier */
207 __s64 offset; /* add to hardware counter value */
210 * Control data for the mmap() data buffer.
212 * User-space reading this value should issue an rmb(), on SMP capable
213 * platforms, after reading this value -- see perf_counter_wakeup().
215 __u32 data_head; /* head in the data section */
218 #define PERF_EVENT_MISC_CPUMODE_MASK (3 << 0)
219 #define PERF_EVENT_MISC_CPUMODE_UNKNOWN (0 << 0)
220 #define PERF_EVENT_MISC_KERNEL (1 << 0)
221 #define PERF_EVENT_MISC_USER (2 << 0)
222 #define PERF_EVENT_MISC_HYPERVISOR (3 << 0)
223 #define PERF_EVENT_MISC_OVERFLOW (1 << 2)
225 struct perf_event_header {
231 enum perf_event_type {
234 * The MMAP events record the PROT_EXEC mappings so that we can
235 * correlate userspace IPs to code. They have the following structure:
238 * struct perf_event_header header;
248 PERF_EVENT_MUNMAP = 2,
252 * struct perf_event_header header;
261 * When header.misc & PERF_EVENT_MISC_OVERFLOW the event_type field
262 * will be PERF_RECORD_*
265 * struct perf_event_header header;
267 * { u64 ip; } && PERF_RECORD_IP
268 * { u32 pid, tid; } && PERF_RECORD_TID
269 * { u64 time; } && PERF_RECORD_TIME
270 * { u64 addr; } && PERF_RECORD_ADDR
271 * { u64 config; } && PERF_RECORD_CONFIG
272 * { u32 cpu, res; } && PERF_RECORD_CPU
275 * { u64 event, val; } cnt[nr]; } && PERF_RECORD_GROUP
281 * u64 ips[nr]; } && PERF_RECORD_CALLCHAIN
288 * Kernel-internal data types and definitions:
291 #ifdef CONFIG_PERF_COUNTERS
292 # include <asm/perf_counter.h>
295 #include <linux/list.h>
296 #include <linux/mutex.h>
297 #include <linux/rculist.h>
298 #include <linux/rcupdate.h>
299 #include <linux/spinlock.h>
300 #include <linux/hrtimer.h>
301 #include <linux/fs.h>
302 #include <asm/atomic.h>
306 static inline u64 perf_event_raw(struct perf_counter_hw_event *hw_event)
308 return hw_event->config & PERF_COUNTER_RAW_MASK;
311 static inline u64 perf_event_config(struct perf_counter_hw_event *hw_event)
313 return hw_event->config & PERF_COUNTER_CONFIG_MASK;
316 static inline u64 perf_event_type(struct perf_counter_hw_event *hw_event)
318 return (hw_event->config & PERF_COUNTER_TYPE_MASK) >>
319 PERF_COUNTER_TYPE_SHIFT;
322 static inline u64 perf_event_id(struct perf_counter_hw_event *hw_event)
324 return hw_event->config & PERF_COUNTER_EVENT_MASK;
328 * struct hw_perf_counter - performance counter hardware details:
330 struct hw_perf_counter {
331 #ifdef CONFIG_PERF_COUNTERS
333 struct { /* hardware */
335 unsigned long config_base;
336 unsigned long counter_base;
340 union { /* software */
342 struct hrtimer hrtimer;
345 atomic64_t prev_count;
347 atomic64_t period_left;
355 * struct pmu - generic performance monitoring unit
358 int (*enable) (struct perf_counter *counter);
359 void (*disable) (struct perf_counter *counter);
360 void (*read) (struct perf_counter *counter);
364 * enum perf_counter_active_state - the states of a counter
366 enum perf_counter_active_state {
367 PERF_COUNTER_STATE_ERROR = -2,
368 PERF_COUNTER_STATE_OFF = -1,
369 PERF_COUNTER_STATE_INACTIVE = 0,
370 PERF_COUNTER_STATE_ACTIVE = 1,
375 struct perf_mmap_data {
376 struct rcu_head rcu_head;
377 int nr_pages; /* nr of data pages */
378 int nr_locked; /* nr pages mlocked */
380 atomic_t poll; /* POLL_ for wakeups */
381 atomic_t head; /* write position */
382 atomic_t events; /* event limit */
384 atomic_t done_head; /* completed head */
385 atomic_t lock; /* concurrent writes */
387 atomic_t wakeup; /* needs a wakeup */
389 struct perf_counter_mmap_page *user_page;
393 struct perf_pending_entry {
394 struct perf_pending_entry *next;
395 void (*func)(struct perf_pending_entry *);
399 * struct perf_counter - performance counter kernel representation:
401 struct perf_counter {
402 #ifdef CONFIG_PERF_COUNTERS
403 struct list_head list_entry;
404 struct list_head event_entry;
405 struct list_head sibling_list;
407 struct perf_counter *group_leader;
408 const struct pmu *pmu;
410 enum perf_counter_active_state state;
411 enum perf_counter_active_state prev_state;
415 * These are the total time in nanoseconds that the counter
416 * has been enabled (i.e. eligible to run, and the task has
417 * been scheduled in, if this is a per-task counter)
418 * and running (scheduled onto the CPU), respectively.
420 * They are computed from tstamp_enabled, tstamp_running and
421 * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
423 u64 total_time_enabled;
424 u64 total_time_running;
427 * These are timestamps used for computing total_time_enabled
428 * and total_time_running when the counter is in INACTIVE or
429 * ACTIVE state, measured in nanoseconds from an arbitrary point
431 * tstamp_enabled: the notional time when the counter was enabled
432 * tstamp_running: the notional time when the counter was scheduled on
433 * tstamp_stopped: in INACTIVE state, the notional time when the
434 * counter was scheduled off.
440 struct perf_counter_hw_event hw_event;
441 struct hw_perf_counter hw;
443 struct perf_counter_context *ctx;
444 struct task_struct *task;
447 struct perf_counter *parent;
448 struct list_head child_list;
451 * These accumulate total time (in nanoseconds) that children
452 * counters have been enabled and running, respectively.
454 atomic64_t child_total_time_enabled;
455 atomic64_t child_total_time_running;
458 * Protect attach/detach and child_list:
466 struct mutex mmap_mutex;
468 struct perf_mmap_data *data;
471 wait_queue_head_t waitq;
472 struct fasync_struct *fasync;
474 /* delayed work for NMIs and such */
478 struct perf_pending_entry pending;
480 atomic_t event_limit;
482 void (*destroy)(struct perf_counter *);
483 struct rcu_head rcu_head;
488 * struct perf_counter_context - counter context structure
490 * Used as a container for task counters and CPU counters as well:
492 struct perf_counter_context {
493 #ifdef CONFIG_PERF_COUNTERS
495 * Protect the states of the counters in the list,
496 * nr_active, and the list:
500 * Protect the list of counters. Locking either mutex or lock
501 * is sufficient to ensure the list doesn't change; to change
502 * the list you need to lock both the mutex and the spinlock.
506 struct list_head counter_list;
507 struct list_head event_list;
512 struct task_struct *task;
515 * Context clock, runs when context enabled.
523 * struct perf_counter_cpu_context - per cpu counter context structure
525 struct perf_cpu_context {
526 struct perf_counter_context ctx;
527 struct perf_counter_context *task_ctx;
533 * Recursion avoidance:
535 * task, softirq, irq, nmi context
540 #ifdef CONFIG_PERF_COUNTERS
543 * Set by architecture code:
545 extern int perf_max_counters;
547 extern const struct pmu *hw_perf_counter_init(struct perf_counter *counter);
549 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
550 extern void perf_counter_task_sched_out(struct task_struct *task, int cpu);
551 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
552 extern void perf_counter_init_task(struct task_struct *child);
553 extern void perf_counter_exit_task(struct task_struct *child);
554 extern void perf_counter_do_pending(void);
555 extern void perf_counter_print_debug(void);
556 extern void perf_counter_unthrottle(void);
557 extern void __perf_disable(void);
558 extern bool __perf_enable(void);
559 extern void perf_disable(void);
560 extern void perf_enable(void);
561 extern int perf_counter_task_disable(void);
562 extern int perf_counter_task_enable(void);
563 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
564 struct perf_cpu_context *cpuctx,
565 struct perf_counter_context *ctx, int cpu);
566 extern void perf_counter_update_userpage(struct perf_counter *counter);
568 extern int perf_counter_overflow(struct perf_counter *counter,
569 int nmi, struct pt_regs *regs, u64 addr);
571 * Return 1 for a software counter, 0 for a hardware counter
573 static inline int is_software_counter(struct perf_counter *counter)
575 return !perf_event_raw(&counter->hw_event) &&
576 perf_event_type(&counter->hw_event) != PERF_TYPE_HARDWARE;
579 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *, u64);
581 extern void perf_counter_mmap(unsigned long addr, unsigned long len,
582 unsigned long pgoff, struct file *file);
584 extern void perf_counter_munmap(unsigned long addr, unsigned long len,
585 unsigned long pgoff, struct file *file);
587 extern void perf_counter_comm(struct task_struct *tsk);
589 #define MAX_STACK_DEPTH 255
591 struct perf_callchain_entry {
592 u16 nr, hv, kernel, user;
593 u64 ip[MAX_STACK_DEPTH];
596 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
598 extern int sysctl_perf_counter_priv;
599 extern int sysctl_perf_counter_mlock;
601 extern void perf_counter_init(void);
603 #ifndef perf_misc_flags
604 #define perf_misc_flags(regs) (user_mode(regs) ? PERF_EVENT_MISC_USER : \
605 PERF_EVENT_MISC_KERNEL)
606 #define perf_instruction_pointer(regs) instruction_pointer(regs)
611 perf_counter_task_sched_in(struct task_struct *task, int cpu) { }
613 perf_counter_task_sched_out(struct task_struct *task, int cpu) { }
615 perf_counter_task_tick(struct task_struct *task, int cpu) { }
616 static inline void perf_counter_init_task(struct task_struct *child) { }
617 static inline void perf_counter_exit_task(struct task_struct *child) { }
618 static inline void perf_counter_do_pending(void) { }
619 static inline void perf_counter_print_debug(void) { }
620 static inline void perf_counter_unthrottle(void) { }
621 static inline void perf_disable(void) { }
622 static inline void perf_enable(void) { }
623 static inline int perf_counter_task_disable(void) { return -EINVAL; }
624 static inline int perf_counter_task_enable(void) { return -EINVAL; }
627 perf_swcounter_event(u32 event, u64 nr, int nmi,
628 struct pt_regs *regs, u64 addr) { }
631 perf_counter_mmap(unsigned long addr, unsigned long len,
632 unsigned long pgoff, struct file *file) { }
635 perf_counter_munmap(unsigned long addr, unsigned long len,
636 unsigned long pgoff, struct file *file) { }
638 static inline void perf_counter_comm(struct task_struct *tsk) { }
639 static inline void perf_counter_init(void) { }
642 #endif /* __KERNEL__ */
643 #endif /* _LINUX_PERF_COUNTER_H */