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 attr.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 attr.sample_type to request information
98 * in the overflow packets.
100 enum perf_counter_sample_format {
101 PERF_SAMPLE_IP = 1U << 0,
102 PERF_SAMPLE_TID = 1U << 1,
103 PERF_SAMPLE_TIME = 1U << 2,
104 PERF_SAMPLE_ADDR = 1U << 3,
105 PERF_SAMPLE_GROUP = 1U << 4,
106 PERF_SAMPLE_CALLCHAIN = 1U << 5,
107 PERF_SAMPLE_CONFIG = 1U << 6,
108 PERF_SAMPLE_CPU = 1U << 7,
112 * Bits that can be set in attr.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 = 1U << 0,
118 PERF_FORMAT_TOTAL_TIME_RUNNING = 1U << 1,
119 PERF_FORMAT_ID = 1U << 2,
123 * Hardware event to monitor via a performance monitoring counter:
125 struct perf_counter_attr {
127 * The MSB of the config word signifies if the rest contains cpu
128 * specific (raw) counter configuration data, if unset, the next
129 * 7 bits are an event type and the rest of the bits are the event
142 __u64 disabled : 1, /* off by default */
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 comm : 1, /* include comm data */
152 freq : 1, /* use freq, not period */
156 __u32 wakeup_events; /* wakeup every n events */
164 * Ioctls that can be done on a perf counter fd:
166 #define PERF_COUNTER_IOC_ENABLE _IO ('$', 0)
167 #define PERF_COUNTER_IOC_DISABLE _IO ('$', 1)
168 #define PERF_COUNTER_IOC_REFRESH _IO ('$', 2)
169 #define PERF_COUNTER_IOC_RESET _IO ('$', 3)
170 #define PERF_COUNTER_IOC_PERIOD _IOW('$', 4, u64)
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 __u64 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;
251 * struct perf_event_header header;
261 * struct perf_event_header header;
266 PERF_EVENT_PERIOD = 4,
270 * struct perf_event_header header;
274 PERF_EVENT_THROTTLE = 5,
275 PERF_EVENT_UNTHROTTLE = 6,
279 * struct perf_event_header header;
286 * When header.misc & PERF_EVENT_MISC_OVERFLOW the event_type field
287 * will be PERF_RECORD_*
290 * struct perf_event_header header;
292 * { u64 ip; } && PERF_RECORD_IP
293 * { u32 pid, tid; } && PERF_RECORD_TID
294 * { u64 time; } && PERF_RECORD_TIME
295 * { u64 addr; } && PERF_RECORD_ADDR
296 * { u64 config; } && PERF_RECORD_CONFIG
297 * { u32 cpu, res; } && PERF_RECORD_CPU
300 * { u64 id, val; } cnt[nr]; } && PERF_RECORD_GROUP
306 * u64 ips[nr]; } && PERF_RECORD_CALLCHAIN
313 * Kernel-internal data types and definitions:
316 #ifdef CONFIG_PERF_COUNTERS
317 # include <asm/perf_counter.h>
320 #include <linux/list.h>
321 #include <linux/mutex.h>
322 #include <linux/rculist.h>
323 #include <linux/rcupdate.h>
324 #include <linux/spinlock.h>
325 #include <linux/hrtimer.h>
326 #include <linux/fs.h>
327 #include <linux/pid_namespace.h>
328 #include <asm/atomic.h>
332 static inline u64 perf_event_raw(struct perf_counter_attr *attr)
334 return attr->config & PERF_COUNTER_RAW_MASK;
337 static inline u64 perf_event_config(struct perf_counter_attr *attr)
339 return attr->config & PERF_COUNTER_CONFIG_MASK;
342 static inline u64 perf_event_type(struct perf_counter_attr *attr)
344 return (attr->config & PERF_COUNTER_TYPE_MASK) >>
345 PERF_COUNTER_TYPE_SHIFT;
348 static inline u64 perf_event_id(struct perf_counter_attr *attr)
350 return attr->config & PERF_COUNTER_EVENT_MASK;
354 * struct hw_perf_counter - performance counter hardware details:
356 struct hw_perf_counter {
357 #ifdef CONFIG_PERF_COUNTERS
359 struct { /* hardware */
361 unsigned long config_base;
362 unsigned long counter_base;
365 union { /* software */
367 struct hrtimer hrtimer;
370 atomic64_t prev_count;
372 atomic64_t period_left;
380 * struct pmu - generic performance monitoring unit
383 int (*enable) (struct perf_counter *counter);
384 void (*disable) (struct perf_counter *counter);
385 void (*read) (struct perf_counter *counter);
386 void (*unthrottle) (struct perf_counter *counter);
390 * enum perf_counter_active_state - the states of a counter
392 enum perf_counter_active_state {
393 PERF_COUNTER_STATE_ERROR = -2,
394 PERF_COUNTER_STATE_OFF = -1,
395 PERF_COUNTER_STATE_INACTIVE = 0,
396 PERF_COUNTER_STATE_ACTIVE = 1,
401 struct perf_mmap_data {
402 struct rcu_head rcu_head;
403 int nr_pages; /* nr of data pages */
404 int nr_locked; /* nr pages mlocked */
406 atomic_t poll; /* POLL_ for wakeups */
407 atomic_t events; /* event limit */
409 atomic_long_t head; /* write position */
410 atomic_long_t done_head; /* completed head */
412 atomic_t lock; /* concurrent writes */
414 atomic_t wakeup; /* needs a wakeup */
416 struct perf_counter_mmap_page *user_page;
420 struct perf_pending_entry {
421 struct perf_pending_entry *next;
422 void (*func)(struct perf_pending_entry *);
426 * struct perf_counter - performance counter kernel representation:
428 struct perf_counter {
429 #ifdef CONFIG_PERF_COUNTERS
430 struct list_head list_entry;
431 struct list_head event_entry;
432 struct list_head sibling_list;
434 struct perf_counter *group_leader;
435 const struct pmu *pmu;
437 enum perf_counter_active_state state;
441 * These are the total time in nanoseconds that the counter
442 * has been enabled (i.e. eligible to run, and the task has
443 * been scheduled in, if this is a per-task counter)
444 * and running (scheduled onto the CPU), respectively.
446 * They are computed from tstamp_enabled, tstamp_running and
447 * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
449 u64 total_time_enabled;
450 u64 total_time_running;
453 * These are timestamps used for computing total_time_enabled
454 * and total_time_running when the counter is in INACTIVE or
455 * ACTIVE state, measured in nanoseconds from an arbitrary point
457 * tstamp_enabled: the notional time when the counter was enabled
458 * tstamp_running: the notional time when the counter was scheduled on
459 * tstamp_stopped: in INACTIVE state, the notional time when the
460 * counter was scheduled off.
466 struct perf_counter_attr attr;
467 struct hw_perf_counter hw;
469 struct perf_counter_context *ctx;
473 * These accumulate total time (in nanoseconds) that children
474 * counters have been enabled and running, respectively.
476 atomic64_t child_total_time_enabled;
477 atomic64_t child_total_time_running;
480 * Protect attach/detach and child_list:
482 struct mutex child_mutex;
483 struct list_head child_list;
484 struct perf_counter *parent;
489 struct list_head owner_entry;
490 struct task_struct *owner;
493 struct mutex mmap_mutex;
495 struct perf_mmap_data *data;
498 wait_queue_head_t waitq;
499 struct fasync_struct *fasync;
501 /* delayed work for NMIs and such */
505 struct perf_pending_entry pending;
507 atomic_t event_limit;
509 void (*destroy)(struct perf_counter *);
510 struct rcu_head rcu_head;
512 struct pid_namespace *ns;
518 * struct perf_counter_context - counter context structure
520 * Used as a container for task counters and CPU counters as well:
522 struct perf_counter_context {
524 * Protect the states of the counters in the list,
525 * nr_active, and the list:
529 * Protect the list of counters. Locking either mutex or lock
530 * is sufficient to ensure the list doesn't change; to change
531 * the list you need to lock both the mutex and the spinlock.
535 struct list_head counter_list;
536 struct list_head event_list;
541 struct task_struct *task;
544 * Context clock, runs when context enabled.
550 * These fields let us detect when two contexts have both
551 * been cloned (inherited) from a common ancestor.
553 struct perf_counter_context *parent_ctx;
557 struct rcu_head rcu_head;
561 * struct perf_counter_cpu_context - per cpu counter context structure
563 struct perf_cpu_context {
564 struct perf_counter_context ctx;
565 struct perf_counter_context *task_ctx;
571 * Recursion avoidance:
573 * task, softirq, irq, nmi context
578 #ifdef CONFIG_PERF_COUNTERS
581 * Set by architecture code:
583 extern int perf_max_counters;
585 extern const struct pmu *hw_perf_counter_init(struct perf_counter *counter);
587 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
588 extern void perf_counter_task_sched_out(struct task_struct *task,
589 struct task_struct *next, int cpu);
590 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
591 extern int perf_counter_init_task(struct task_struct *child);
592 extern void perf_counter_exit_task(struct task_struct *child);
593 extern void perf_counter_free_task(struct task_struct *task);
594 extern void perf_counter_do_pending(void);
595 extern void perf_counter_print_debug(void);
596 extern void __perf_disable(void);
597 extern bool __perf_enable(void);
598 extern void perf_disable(void);
599 extern void perf_enable(void);
600 extern int perf_counter_task_disable(void);
601 extern int perf_counter_task_enable(void);
602 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
603 struct perf_cpu_context *cpuctx,
604 struct perf_counter_context *ctx, int cpu);
605 extern void perf_counter_update_userpage(struct perf_counter *counter);
607 extern int perf_counter_overflow(struct perf_counter *counter,
608 int nmi, struct pt_regs *regs, u64 addr);
610 * Return 1 for a software counter, 0 for a hardware counter
612 static inline int is_software_counter(struct perf_counter *counter)
614 return !perf_event_raw(&counter->attr) &&
615 perf_event_type(&counter->attr) != PERF_TYPE_HARDWARE;
618 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *, u64);
620 extern void perf_counter_mmap(unsigned long addr, unsigned long len,
621 unsigned long pgoff, struct file *file);
623 extern void perf_counter_comm(struct task_struct *tsk);
624 extern void perf_counter_fork(struct task_struct *tsk);
626 extern void perf_counter_task_migration(struct task_struct *task, int cpu);
628 #define MAX_STACK_DEPTH 255
630 struct perf_callchain_entry {
631 u16 nr, hv, kernel, user;
632 u64 ip[MAX_STACK_DEPTH];
635 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
637 extern int sysctl_perf_counter_priv;
638 extern int sysctl_perf_counter_mlock;
639 extern int sysctl_perf_counter_limit;
641 extern void perf_counter_init(void);
643 #ifndef perf_misc_flags
644 #define perf_misc_flags(regs) (user_mode(regs) ? PERF_EVENT_MISC_USER : \
645 PERF_EVENT_MISC_KERNEL)
646 #define perf_instruction_pointer(regs) instruction_pointer(regs)
651 perf_counter_task_sched_in(struct task_struct *task, int cpu) { }
653 perf_counter_task_sched_out(struct task_struct *task,
654 struct task_struct *next, int cpu) { }
656 perf_counter_task_tick(struct task_struct *task, int cpu) { }
657 static inline int perf_counter_init_task(struct task_struct *child) { return 0; }
658 static inline void perf_counter_exit_task(struct task_struct *child) { }
659 static inline void perf_counter_free_task(struct task_struct *task) { }
660 static inline void perf_counter_do_pending(void) { }
661 static inline void perf_counter_print_debug(void) { }
662 static inline void perf_disable(void) { }
663 static inline void perf_enable(void) { }
664 static inline int perf_counter_task_disable(void) { return -EINVAL; }
665 static inline int perf_counter_task_enable(void) { return -EINVAL; }
668 perf_swcounter_event(u32 event, u64 nr, int nmi,
669 struct pt_regs *regs, u64 addr) { }
672 perf_counter_mmap(unsigned long addr, unsigned long len,
673 unsigned long pgoff, struct file *file) { }
675 static inline void perf_counter_comm(struct task_struct *tsk) { }
676 static inline void perf_counter_fork(struct task_struct *tsk) { }
677 static inline void perf_counter_init(void) { }
678 static inline void perf_counter_task_migration(struct task_struct *task,
682 #endif /* __KERNEL__ */
683 #endif /* _LINUX_PERF_COUNTER_H */