4 * Copyright (C) 2008-2009, Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009, Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009, Red Hat, Inc., Peter Zijlstra
8 * Data type definitions, declarations, prototypes.
10 * Started by: Thomas Gleixner and Ingo Molnar
12 * For licencing details see kernel-base/COPYING
14 #ifndef _LINUX_PERF_EVENT_H
15 #define _LINUX_PERF_EVENT_H
17 #include <linux/types.h>
18 #include <linux/ioctl.h>
19 #include <asm/byteorder.h>
22 * User-space ABI bits:
29 PERF_TYPE_HARDWARE = 0,
30 PERF_TYPE_SOFTWARE = 1,
31 PERF_TYPE_TRACEPOINT = 2,
32 PERF_TYPE_HW_CACHE = 3,
34 PERF_TYPE_BREAKPOINT = 5,
36 PERF_TYPE_MAX, /* non-ABI */
40 * Generalized performance event event_id types, used by the
41 * attr.event_id parameter of the sys_perf_event_open()
46 * Common hardware events, generalized by the kernel:
48 PERF_COUNT_HW_CPU_CYCLES = 0,
49 PERF_COUNT_HW_INSTRUCTIONS = 1,
50 PERF_COUNT_HW_CACHE_REFERENCES = 2,
51 PERF_COUNT_HW_CACHE_MISSES = 3,
52 PERF_COUNT_HW_BRANCH_INSTRUCTIONS = 4,
53 PERF_COUNT_HW_BRANCH_MISSES = 5,
54 PERF_COUNT_HW_BUS_CYCLES = 6,
56 PERF_COUNT_HW_MAX, /* non-ABI */
60 * Generalized hardware cache events:
62 * { L1-D, L1-I, LLC, ITLB, DTLB, BPU } x
63 * { read, write, prefetch } x
64 * { accesses, misses }
66 enum perf_hw_cache_id {
67 PERF_COUNT_HW_CACHE_L1D = 0,
68 PERF_COUNT_HW_CACHE_L1I = 1,
69 PERF_COUNT_HW_CACHE_LL = 2,
70 PERF_COUNT_HW_CACHE_DTLB = 3,
71 PERF_COUNT_HW_CACHE_ITLB = 4,
72 PERF_COUNT_HW_CACHE_BPU = 5,
74 PERF_COUNT_HW_CACHE_MAX, /* non-ABI */
77 enum perf_hw_cache_op_id {
78 PERF_COUNT_HW_CACHE_OP_READ = 0,
79 PERF_COUNT_HW_CACHE_OP_WRITE = 1,
80 PERF_COUNT_HW_CACHE_OP_PREFETCH = 2,
82 PERF_COUNT_HW_CACHE_OP_MAX, /* non-ABI */
85 enum perf_hw_cache_op_result_id {
86 PERF_COUNT_HW_CACHE_RESULT_ACCESS = 0,
87 PERF_COUNT_HW_CACHE_RESULT_MISS = 1,
89 PERF_COUNT_HW_CACHE_RESULT_MAX, /* non-ABI */
93 * Special "software" events provided by the kernel, even if the hardware
94 * does not support performance events. These events measure various
95 * physical and sw events of the kernel (and allow the profiling of them as
99 PERF_COUNT_SW_CPU_CLOCK = 0,
100 PERF_COUNT_SW_TASK_CLOCK = 1,
101 PERF_COUNT_SW_PAGE_FAULTS = 2,
102 PERF_COUNT_SW_CONTEXT_SWITCHES = 3,
103 PERF_COUNT_SW_CPU_MIGRATIONS = 4,
104 PERF_COUNT_SW_PAGE_FAULTS_MIN = 5,
105 PERF_COUNT_SW_PAGE_FAULTS_MAJ = 6,
106 PERF_COUNT_SW_ALIGNMENT_FAULTS = 7,
107 PERF_COUNT_SW_EMULATION_FAULTS = 8,
109 PERF_COUNT_SW_MAX, /* non-ABI */
113 * Bits that can be set in attr.sample_type to request information
114 * in the overflow packets.
116 enum perf_event_sample_format {
117 PERF_SAMPLE_IP = 1U << 0,
118 PERF_SAMPLE_TID = 1U << 1,
119 PERF_SAMPLE_TIME = 1U << 2,
120 PERF_SAMPLE_ADDR = 1U << 3,
121 PERF_SAMPLE_READ = 1U << 4,
122 PERF_SAMPLE_CALLCHAIN = 1U << 5,
123 PERF_SAMPLE_ID = 1U << 6,
124 PERF_SAMPLE_CPU = 1U << 7,
125 PERF_SAMPLE_PERIOD = 1U << 8,
126 PERF_SAMPLE_STREAM_ID = 1U << 9,
127 PERF_SAMPLE_RAW = 1U << 10,
129 PERF_SAMPLE_MAX = 1U << 11, /* non-ABI */
133 * The format of the data returned by read() on a perf event fd,
134 * as specified by attr.read_format:
136 * struct read_format {
138 * { u64 time_enabled; } && PERF_FORMAT_ENABLED
139 * { u64 time_running; } && PERF_FORMAT_RUNNING
140 * { u64 id; } && PERF_FORMAT_ID
141 * } && !PERF_FORMAT_GROUP
144 * { u64 time_enabled; } && PERF_FORMAT_ENABLED
145 * { u64 time_running; } && PERF_FORMAT_RUNNING
147 * { u64 id; } && PERF_FORMAT_ID
149 * } && PERF_FORMAT_GROUP
152 enum perf_event_read_format {
153 PERF_FORMAT_TOTAL_TIME_ENABLED = 1U << 0,
154 PERF_FORMAT_TOTAL_TIME_RUNNING = 1U << 1,
155 PERF_FORMAT_ID = 1U << 2,
156 PERF_FORMAT_GROUP = 1U << 3,
158 PERF_FORMAT_MAX = 1U << 4, /* non-ABI */
161 #define PERF_ATTR_SIZE_VER0 64 /* sizeof first published struct */
164 * Hardware event_id to monitor via a performance monitoring event:
166 struct perf_event_attr {
169 * Major type: hardware/software/tracepoint/etc.
174 * Size of the attr structure, for fwd/bwd compat.
179 * Type specific configuration information.
191 __u64 disabled : 1, /* off by default */
192 inherit : 1, /* children inherit it */
193 pinned : 1, /* must always be on PMU */
194 exclusive : 1, /* only group on PMU */
195 exclude_user : 1, /* don't count user */
196 exclude_kernel : 1, /* ditto kernel */
197 exclude_hv : 1, /* ditto hypervisor */
198 exclude_idle : 1, /* don't count when idle */
199 mmap : 1, /* include mmap data */
200 comm : 1, /* include comm data */
201 freq : 1, /* use freq, not period */
202 inherit_stat : 1, /* per task counts */
203 enable_on_exec : 1, /* next exec enables */
204 task : 1, /* trace fork/exit */
205 watermark : 1, /* wakeup_watermark */
210 __u32 wakeup_events; /* wakeup every n events */
211 __u32 wakeup_watermark; /* bytes before wakeup */
214 struct { /* Hardware breakpoint info */
218 __u64 __bp_reserved_1;
219 __u64 __bp_reserved_2;
228 * Ioctls that can be done on a perf event fd:
230 #define PERF_EVENT_IOC_ENABLE _IO ('$', 0)
231 #define PERF_EVENT_IOC_DISABLE _IO ('$', 1)
232 #define PERF_EVENT_IOC_REFRESH _IO ('$', 2)
233 #define PERF_EVENT_IOC_RESET _IO ('$', 3)
234 #define PERF_EVENT_IOC_PERIOD _IOW('$', 4, __u64)
235 #define PERF_EVENT_IOC_SET_OUTPUT _IO ('$', 5)
236 #define PERF_EVENT_IOC_SET_FILTER _IOW('$', 6, char *)
238 enum perf_event_ioc_flags {
239 PERF_IOC_FLAG_GROUP = 1U << 0,
243 * Structure of the page that can be mapped via mmap
245 struct perf_event_mmap_page {
246 __u32 version; /* version number of this structure */
247 __u32 compat_version; /* lowest version this is compat with */
250 * Bits needed to read the hw events in user-space.
260 * count = pmc_read(pc->index - 1);
261 * count += pc->offset;
266 * } while (pc->lock != seq);
268 * NOTE: for obvious reason this only works on self-monitoring
271 __u32 lock; /* seqlock for synchronization */
272 __u32 index; /* hardware event identifier */
273 __s64 offset; /* add to hardware event value */
274 __u64 time_enabled; /* time event active */
275 __u64 time_running; /* time event on cpu */
278 * Hole for extension of the self monitor capabilities
281 __u64 __reserved[123]; /* align to 1k */
284 * Control data for the mmap() data buffer.
286 * User-space reading the @data_head value should issue an rmb(), on
287 * SMP capable platforms, after reading this value -- see
288 * perf_event_wakeup().
290 * When the mapping is PROT_WRITE the @data_tail value should be
291 * written by userspace to reflect the last read data. In this case
292 * the kernel will not over-write unread data.
294 __u64 data_head; /* head in the data section */
295 __u64 data_tail; /* user-space written tail */
298 #define PERF_RECORD_MISC_CPUMODE_MASK (3 << 0)
299 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN (0 << 0)
300 #define PERF_RECORD_MISC_KERNEL (1 << 0)
301 #define PERF_RECORD_MISC_USER (2 << 0)
302 #define PERF_RECORD_MISC_HYPERVISOR (3 << 0)
304 struct perf_event_header {
310 enum perf_event_type {
313 * The MMAP events record the PROT_EXEC mappings so that we can
314 * correlate userspace IPs to code. They have the following structure:
317 * struct perf_event_header header;
326 PERF_RECORD_MMAP = 1,
330 * struct perf_event_header header;
335 PERF_RECORD_LOST = 2,
339 * struct perf_event_header header;
345 PERF_RECORD_COMM = 3,
349 * struct perf_event_header header;
355 PERF_RECORD_EXIT = 4,
359 * struct perf_event_header header;
365 PERF_RECORD_THROTTLE = 5,
366 PERF_RECORD_UNTHROTTLE = 6,
370 * struct perf_event_header header;
376 PERF_RECORD_FORK = 7,
380 * struct perf_event_header header;
383 * struct read_format values;
386 PERF_RECORD_READ = 8,
390 * struct perf_event_header header;
392 * { u64 ip; } && PERF_SAMPLE_IP
393 * { u32 pid, tid; } && PERF_SAMPLE_TID
394 * { u64 time; } && PERF_SAMPLE_TIME
395 * { u64 addr; } && PERF_SAMPLE_ADDR
396 * { u64 id; } && PERF_SAMPLE_ID
397 * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID
398 * { u32 cpu, res; } && PERF_SAMPLE_CPU
399 * { u64 period; } && PERF_SAMPLE_PERIOD
401 * { struct read_format values; } && PERF_SAMPLE_READ
404 * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN
407 * # The RAW record below is opaque data wrt the ABI
409 * # That is, the ABI doesn't make any promises wrt to
410 * # the stability of its content, it may vary depending
411 * # on event, hardware, kernel version and phase of
414 * # In other words, PERF_SAMPLE_RAW contents are not an ABI.
418 * char data[size];}&& PERF_SAMPLE_RAW
421 PERF_RECORD_SAMPLE = 9,
423 PERF_RECORD_MAX, /* non-ABI */
426 enum perf_callchain_context {
427 PERF_CONTEXT_HV = (__u64)-32,
428 PERF_CONTEXT_KERNEL = (__u64)-128,
429 PERF_CONTEXT_USER = (__u64)-512,
431 PERF_CONTEXT_GUEST = (__u64)-2048,
432 PERF_CONTEXT_GUEST_KERNEL = (__u64)-2176,
433 PERF_CONTEXT_GUEST_USER = (__u64)-2560,
435 PERF_CONTEXT_MAX = (__u64)-4095,
438 #define PERF_FLAG_FD_NO_GROUP (1U << 0)
439 #define PERF_FLAG_FD_OUTPUT (1U << 1)
443 * Kernel-internal data types and definitions:
446 #ifdef CONFIG_PERF_EVENTS
447 # include <asm/perf_event.h>
450 #ifdef CONFIG_HAVE_HW_BREAKPOINT
451 #include <asm/hw_breakpoint.h>
454 #include <linux/list.h>
455 #include <linux/mutex.h>
456 #include <linux/rculist.h>
457 #include <linux/rcupdate.h>
458 #include <linux/spinlock.h>
459 #include <linux/hrtimer.h>
460 #include <linux/fs.h>
461 #include <linux/pid_namespace.h>
462 #include <linux/workqueue.h>
463 #include <asm/atomic.h>
465 #define PERF_MAX_STACK_DEPTH 255
467 struct perf_callchain_entry {
469 __u64 ip[PERF_MAX_STACK_DEPTH];
472 struct perf_raw_record {
480 * struct hw_perf_event - performance event hardware details:
482 struct hw_perf_event {
483 #ifdef CONFIG_PERF_EVENTS
485 struct { /* hardware */
487 unsigned long config_base;
488 unsigned long event_base;
491 struct { /* software */
493 struct hrtimer hrtimer;
495 #ifdef CONFIG_HAVE_HW_BREAKPOINT
496 union { /* breakpoint */
497 struct arch_hw_breakpoint info;
501 atomic64_t prev_count;
504 atomic64_t period_left;
516 * struct pmu - generic performance monitoring unit
519 int (*enable) (struct perf_event *event);
520 void (*disable) (struct perf_event *event);
521 void (*read) (struct perf_event *event);
522 void (*unthrottle) (struct perf_event *event);
526 * enum perf_event_active_state - the states of a event
528 enum perf_event_active_state {
529 PERF_EVENT_STATE_ERROR = -2,
530 PERF_EVENT_STATE_OFF = -1,
531 PERF_EVENT_STATE_INACTIVE = 0,
532 PERF_EVENT_STATE_ACTIVE = 1,
537 struct perf_mmap_data {
538 struct rcu_head rcu_head;
539 #ifdef CONFIG_PERF_USE_VMALLOC
540 struct work_struct work;
543 int nr_pages; /* nr of data pages */
544 int writable; /* are we writable */
545 int nr_locked; /* nr pages mlocked */
547 atomic_t poll; /* POLL_ for wakeups */
548 atomic_t events; /* event_id limit */
550 atomic_long_t head; /* write position */
551 atomic_long_t done_head; /* completed head */
553 atomic_t lock; /* concurrent writes */
554 atomic_t wakeup; /* needs a wakeup */
555 atomic_t lost; /* nr records lost */
557 long watermark; /* wakeup watermark */
559 struct perf_event_mmap_page *user_page;
563 struct perf_pending_entry {
564 struct perf_pending_entry *next;
565 void (*func)(struct perf_pending_entry *);
568 struct perf_sample_data;
570 typedef void (*perf_overflow_handler_t)(struct perf_event *, int,
571 struct perf_sample_data *,
572 struct pt_regs *regs);
575 * struct perf_event - performance event kernel representation:
578 #ifdef CONFIG_PERF_EVENTS
579 struct list_head group_entry;
580 struct list_head event_entry;
581 struct list_head sibling_list;
583 struct perf_event *group_leader;
584 struct perf_event *output;
585 const struct pmu *pmu;
587 enum perf_event_active_state state;
591 * These are the total time in nanoseconds that the event
592 * has been enabled (i.e. eligible to run, and the task has
593 * been scheduled in, if this is a per-task event)
594 * and running (scheduled onto the CPU), respectively.
596 * They are computed from tstamp_enabled, tstamp_running and
597 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
599 u64 total_time_enabled;
600 u64 total_time_running;
603 * These are timestamps used for computing total_time_enabled
604 * and total_time_running when the event is in INACTIVE or
605 * ACTIVE state, measured in nanoseconds from an arbitrary point
607 * tstamp_enabled: the notional time when the event was enabled
608 * tstamp_running: the notional time when the event was scheduled on
609 * tstamp_stopped: in INACTIVE state, the notional time when the
610 * event was scheduled off.
616 struct perf_event_attr attr;
617 struct hw_perf_event hw;
619 struct perf_event_context *ctx;
623 * These accumulate total time (in nanoseconds) that children
624 * events have been enabled and running, respectively.
626 atomic64_t child_total_time_enabled;
627 atomic64_t child_total_time_running;
630 * Protect attach/detach and child_list:
632 struct mutex child_mutex;
633 struct list_head child_list;
634 struct perf_event *parent;
639 struct list_head owner_entry;
640 struct task_struct *owner;
643 struct mutex mmap_mutex;
645 struct perf_mmap_data *data;
648 wait_queue_head_t waitq;
649 struct fasync_struct *fasync;
651 /* delayed work for NMIs and such */
655 struct perf_pending_entry pending;
657 atomic_t event_limit;
659 void (*destroy)(struct perf_event *);
660 struct rcu_head rcu_head;
662 struct pid_namespace *ns;
665 perf_overflow_handler_t overflow_handler;
667 #ifdef CONFIG_EVENT_PROFILE
668 struct event_filter *filter;
671 #endif /* CONFIG_PERF_EVENTS */
675 * struct perf_event_context - event context structure
677 * Used as a container for task events and CPU events as well:
679 struct perf_event_context {
681 * Protect the states of the events in the list,
682 * nr_active, and the list:
686 * Protect the list of events. Locking either mutex or lock
687 * is sufficient to ensure the list doesn't change; to change
688 * the list you need to lock both the mutex and the spinlock.
692 struct list_head group_list;
693 struct list_head event_list;
699 struct task_struct *task;
702 * Context clock, runs when context enabled.
708 * These fields let us detect when two contexts have both
709 * been cloned (inherited) from a common ancestor.
711 struct perf_event_context *parent_ctx;
715 struct rcu_head rcu_head;
719 * struct perf_event_cpu_context - per cpu event context structure
721 struct perf_cpu_context {
722 struct perf_event_context ctx;
723 struct perf_event_context *task_ctx;
729 * Recursion avoidance:
731 * task, softirq, irq, nmi context
736 struct perf_output_handle {
737 struct perf_event *event;
738 struct perf_mmap_data *data;
740 unsigned long offset;
746 #ifdef CONFIG_PERF_EVENTS
749 * Set by architecture code:
751 extern int perf_max_events;
753 extern const struct pmu *hw_perf_event_init(struct perf_event *event);
755 extern void perf_event_task_sched_in(struct task_struct *task, int cpu);
756 extern void perf_event_task_sched_out(struct task_struct *task,
757 struct task_struct *next, int cpu);
758 extern void perf_event_task_tick(struct task_struct *task, int cpu);
759 extern int perf_event_init_task(struct task_struct *child);
760 extern void perf_event_exit_task(struct task_struct *child);
761 extern void perf_event_free_task(struct task_struct *task);
762 extern void set_perf_event_pending(void);
763 extern void perf_event_do_pending(void);
764 extern void perf_event_print_debug(void);
765 extern void __perf_disable(void);
766 extern bool __perf_enable(void);
767 extern void perf_disable(void);
768 extern void perf_enable(void);
769 extern int perf_event_task_disable(void);
770 extern int perf_event_task_enable(void);
771 extern int hw_perf_group_sched_in(struct perf_event *group_leader,
772 struct perf_cpu_context *cpuctx,
773 struct perf_event_context *ctx, int cpu);
774 extern void perf_event_update_userpage(struct perf_event *event);
775 extern int perf_event_release_kernel(struct perf_event *event);
776 extern struct perf_event *
777 perf_event_create_kernel_counter(struct perf_event_attr *attr,
780 perf_overflow_handler_t callback);
781 extern u64 perf_event_read_value(struct perf_event *event,
782 u64 *enabled, u64 *running);
784 struct perf_sample_data {
801 struct perf_callchain_entry *callchain;
802 struct perf_raw_record *raw;
805 extern void perf_output_sample(struct perf_output_handle *handle,
806 struct perf_event_header *header,
807 struct perf_sample_data *data,
808 struct perf_event *event);
809 extern void perf_prepare_sample(struct perf_event_header *header,
810 struct perf_sample_data *data,
811 struct perf_event *event,
812 struct pt_regs *regs);
814 extern int perf_event_overflow(struct perf_event *event, int nmi,
815 struct perf_sample_data *data,
816 struct pt_regs *regs);
819 * Return 1 for a software event, 0 for a hardware event
821 static inline int is_software_event(struct perf_event *event)
823 return (event->attr.type != PERF_TYPE_RAW) &&
824 (event->attr.type != PERF_TYPE_HARDWARE) &&
825 (event->attr.type != PERF_TYPE_HW_CACHE);
828 extern atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
830 extern void __perf_sw_event(u32, u64, int, struct pt_regs *, u64);
833 perf_sw_event(u32 event_id, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
835 if (atomic_read(&perf_swevent_enabled[event_id]))
836 __perf_sw_event(event_id, nr, nmi, regs, addr);
839 extern void __perf_event_mmap(struct vm_area_struct *vma);
841 static inline void perf_event_mmap(struct vm_area_struct *vma)
843 if (vma->vm_flags & VM_EXEC)
844 __perf_event_mmap(vma);
847 extern void perf_event_comm(struct task_struct *tsk);
848 extern void perf_event_fork(struct task_struct *tsk);
850 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
852 extern int sysctl_perf_event_paranoid;
853 extern int sysctl_perf_event_mlock;
854 extern int sysctl_perf_event_sample_rate;
856 extern void perf_event_init(void);
857 extern void perf_tp_event(int event_id, u64 addr, u64 count,
858 void *record, int entry_size);
859 extern void perf_bp_event(struct perf_event *event, void *data);
861 #ifndef perf_misc_flags
862 #define perf_misc_flags(regs) (user_mode(regs) ? PERF_RECORD_MISC_USER : \
863 PERF_RECORD_MISC_KERNEL)
864 #define perf_instruction_pointer(regs) instruction_pointer(regs)
867 extern int perf_output_begin(struct perf_output_handle *handle,
868 struct perf_event *event, unsigned int size,
869 int nmi, int sample);
870 extern void perf_output_end(struct perf_output_handle *handle);
871 extern void perf_output_copy(struct perf_output_handle *handle,
872 const void *buf, unsigned int len);
873 extern int perf_swevent_get_recursion_context(void);
874 extern void perf_swevent_put_recursion_context(int rctx);
875 extern void perf_event_enable(struct perf_event *event);
876 extern void perf_event_disable(struct perf_event *event);
879 perf_event_task_sched_in(struct task_struct *task, int cpu) { }
881 perf_event_task_sched_out(struct task_struct *task,
882 struct task_struct *next, int cpu) { }
884 perf_event_task_tick(struct task_struct *task, int cpu) { }
885 static inline int perf_event_init_task(struct task_struct *child) { return 0; }
886 static inline void perf_event_exit_task(struct task_struct *child) { }
887 static inline void perf_event_free_task(struct task_struct *task) { }
888 static inline void perf_event_do_pending(void) { }
889 static inline void perf_event_print_debug(void) { }
890 static inline void perf_disable(void) { }
891 static inline void perf_enable(void) { }
892 static inline int perf_event_task_disable(void) { return -EINVAL; }
893 static inline int perf_event_task_enable(void) { return -EINVAL; }
896 perf_sw_event(u32 event_id, u64 nr, int nmi,
897 struct pt_regs *regs, u64 addr) { }
899 perf_bp_event(struct perf_event *event, void *data) { }
901 static inline void perf_event_mmap(struct vm_area_struct *vma) { }
902 static inline void perf_event_comm(struct task_struct *tsk) { }
903 static inline void perf_event_fork(struct task_struct *tsk) { }
904 static inline void perf_event_init(void) { }
905 static inline int perf_swevent_get_recursion_context(void) { return -1; }
906 static inline void perf_swevent_put_recursion_context(int rctx) { }
907 static inline void perf_event_enable(struct perf_event *event) { }
908 static inline void perf_event_disable(struct perf_event *event) { }
911 #define perf_output_put(handle, x) \
912 perf_output_copy((handle), &(x), sizeof(x))
914 #endif /* __KERNEL__ */
915 #endif /* _LINUX_PERF_EVENT_H */