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>
21 #ifdef CONFIG_HAVE_HW_BREAKPOINT
22 #include <asm/hw_breakpoint.h>
26 * User-space ABI bits:
33 PERF_TYPE_HARDWARE = 0,
34 PERF_TYPE_SOFTWARE = 1,
35 PERF_TYPE_TRACEPOINT = 2,
36 PERF_TYPE_HW_CACHE = 3,
38 PERF_TYPE_BREAKPOINT = 5,
40 PERF_TYPE_MAX, /* non-ABI */
44 * Generalized performance event event_id types, used by the
45 * attr.event_id parameter of the sys_perf_event_open()
50 * Common hardware events, generalized by the kernel:
52 PERF_COUNT_HW_CPU_CYCLES = 0,
53 PERF_COUNT_HW_INSTRUCTIONS = 1,
54 PERF_COUNT_HW_CACHE_REFERENCES = 2,
55 PERF_COUNT_HW_CACHE_MISSES = 3,
56 PERF_COUNT_HW_BRANCH_INSTRUCTIONS = 4,
57 PERF_COUNT_HW_BRANCH_MISSES = 5,
58 PERF_COUNT_HW_BUS_CYCLES = 6,
60 PERF_COUNT_HW_MAX, /* non-ABI */
64 * Generalized hardware cache events:
66 * { L1-D, L1-I, LLC, ITLB, DTLB, BPU } x
67 * { read, write, prefetch } x
68 * { accesses, misses }
70 enum perf_hw_cache_id {
71 PERF_COUNT_HW_CACHE_L1D = 0,
72 PERF_COUNT_HW_CACHE_L1I = 1,
73 PERF_COUNT_HW_CACHE_LL = 2,
74 PERF_COUNT_HW_CACHE_DTLB = 3,
75 PERF_COUNT_HW_CACHE_ITLB = 4,
76 PERF_COUNT_HW_CACHE_BPU = 5,
78 PERF_COUNT_HW_CACHE_MAX, /* non-ABI */
81 enum perf_hw_cache_op_id {
82 PERF_COUNT_HW_CACHE_OP_READ = 0,
83 PERF_COUNT_HW_CACHE_OP_WRITE = 1,
84 PERF_COUNT_HW_CACHE_OP_PREFETCH = 2,
86 PERF_COUNT_HW_CACHE_OP_MAX, /* non-ABI */
89 enum perf_hw_cache_op_result_id {
90 PERF_COUNT_HW_CACHE_RESULT_ACCESS = 0,
91 PERF_COUNT_HW_CACHE_RESULT_MISS = 1,
93 PERF_COUNT_HW_CACHE_RESULT_MAX, /* non-ABI */
97 * Special "software" events provided by the kernel, even if the hardware
98 * does not support performance events. These events measure various
99 * physical and sw events of the kernel (and allow the profiling of them as
103 PERF_COUNT_SW_CPU_CLOCK = 0,
104 PERF_COUNT_SW_TASK_CLOCK = 1,
105 PERF_COUNT_SW_PAGE_FAULTS = 2,
106 PERF_COUNT_SW_CONTEXT_SWITCHES = 3,
107 PERF_COUNT_SW_CPU_MIGRATIONS = 4,
108 PERF_COUNT_SW_PAGE_FAULTS_MIN = 5,
109 PERF_COUNT_SW_PAGE_FAULTS_MAJ = 6,
110 PERF_COUNT_SW_ALIGNMENT_FAULTS = 7,
111 PERF_COUNT_SW_EMULATION_FAULTS = 8,
113 PERF_COUNT_SW_MAX, /* non-ABI */
117 * Bits that can be set in attr.sample_type to request information
118 * in the overflow packets.
120 enum perf_event_sample_format {
121 PERF_SAMPLE_IP = 1U << 0,
122 PERF_SAMPLE_TID = 1U << 1,
123 PERF_SAMPLE_TIME = 1U << 2,
124 PERF_SAMPLE_ADDR = 1U << 3,
125 PERF_SAMPLE_READ = 1U << 4,
126 PERF_SAMPLE_CALLCHAIN = 1U << 5,
127 PERF_SAMPLE_ID = 1U << 6,
128 PERF_SAMPLE_CPU = 1U << 7,
129 PERF_SAMPLE_PERIOD = 1U << 8,
130 PERF_SAMPLE_STREAM_ID = 1U << 9,
131 PERF_SAMPLE_RAW = 1U << 10,
133 PERF_SAMPLE_MAX = 1U << 11, /* non-ABI */
137 * The format of the data returned by read() on a perf event fd,
138 * as specified by attr.read_format:
140 * struct read_format {
142 * { u64 time_enabled; } && PERF_FORMAT_ENABLED
143 * { u64 time_running; } && PERF_FORMAT_RUNNING
144 * { u64 id; } && PERF_FORMAT_ID
145 * } && !PERF_FORMAT_GROUP
148 * { u64 time_enabled; } && PERF_FORMAT_ENABLED
149 * { u64 time_running; } && PERF_FORMAT_RUNNING
151 * { u64 id; } && PERF_FORMAT_ID
153 * } && PERF_FORMAT_GROUP
156 enum perf_event_read_format {
157 PERF_FORMAT_TOTAL_TIME_ENABLED = 1U << 0,
158 PERF_FORMAT_TOTAL_TIME_RUNNING = 1U << 1,
159 PERF_FORMAT_ID = 1U << 2,
160 PERF_FORMAT_GROUP = 1U << 3,
162 PERF_FORMAT_MAX = 1U << 4, /* non-ABI */
165 #define PERF_ATTR_SIZE_VER0 64 /* sizeof first published struct */
168 * Hardware event_id to monitor via a performance monitoring event:
170 struct perf_event_attr {
173 * Major type: hardware/software/tracepoint/etc.
178 * Size of the attr structure, for fwd/bwd compat.
183 * Type specific configuration information.
195 __u64 disabled : 1, /* off by default */
196 inherit : 1, /* children inherit it */
197 pinned : 1, /* must always be on PMU */
198 exclusive : 1, /* only group on PMU */
199 exclude_user : 1, /* don't count user */
200 exclude_kernel : 1, /* ditto kernel */
201 exclude_hv : 1, /* ditto hypervisor */
202 exclude_idle : 1, /* don't count when idle */
203 mmap : 1, /* include mmap data */
204 comm : 1, /* include comm data */
205 freq : 1, /* use freq, not period */
206 inherit_stat : 1, /* per task counts */
207 enable_on_exec : 1, /* next exec enables */
208 task : 1, /* trace fork/exit */
209 watermark : 1, /* wakeup_watermark */
214 __u32 wakeup_events; /* wakeup every n events */
215 __u32 wakeup_watermark; /* bytes before wakeup */
219 struct { /* Hardware breakpoint info */
223 __u64 __bp_reserved_1;
224 __u64 __bp_reserved_2;
234 * Ioctls that can be done on a perf event fd:
236 #define PERF_EVENT_IOC_ENABLE _IO ('$', 0)
237 #define PERF_EVENT_IOC_DISABLE _IO ('$', 1)
238 #define PERF_EVENT_IOC_REFRESH _IO ('$', 2)
239 #define PERF_EVENT_IOC_RESET _IO ('$', 3)
240 #define PERF_EVENT_IOC_PERIOD _IOW('$', 4, __u64)
241 #define PERF_EVENT_IOC_SET_OUTPUT _IO ('$', 5)
242 #define PERF_EVENT_IOC_SET_FILTER _IOW('$', 6, char *)
244 enum perf_event_ioc_flags {
245 PERF_IOC_FLAG_GROUP = 1U << 0,
249 * Structure of the page that can be mapped via mmap
251 struct perf_event_mmap_page {
252 __u32 version; /* version number of this structure */
253 __u32 compat_version; /* lowest version this is compat with */
256 * Bits needed to read the hw events in user-space.
266 * count = pmc_read(pc->index - 1);
267 * count += pc->offset;
272 * } while (pc->lock != seq);
274 * NOTE: for obvious reason this only works on self-monitoring
277 __u32 lock; /* seqlock for synchronization */
278 __u32 index; /* hardware event identifier */
279 __s64 offset; /* add to hardware event value */
280 __u64 time_enabled; /* time event active */
281 __u64 time_running; /* time event on cpu */
284 * Hole for extension of the self monitor capabilities
287 __u64 __reserved[123]; /* align to 1k */
290 * Control data for the mmap() data buffer.
292 * User-space reading the @data_head value should issue an rmb(), on
293 * SMP capable platforms, after reading this value -- see
294 * perf_event_wakeup().
296 * When the mapping is PROT_WRITE the @data_tail value should be
297 * written by userspace to reflect the last read data. In this case
298 * the kernel will not over-write unread data.
300 __u64 data_head; /* head in the data section */
301 __u64 data_tail; /* user-space written tail */
304 #define PERF_RECORD_MISC_CPUMODE_MASK (3 << 0)
305 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN (0 << 0)
306 #define PERF_RECORD_MISC_KERNEL (1 << 0)
307 #define PERF_RECORD_MISC_USER (2 << 0)
308 #define PERF_RECORD_MISC_HYPERVISOR (3 << 0)
310 struct perf_event_header {
316 enum perf_event_type {
319 * The MMAP events record the PROT_EXEC mappings so that we can
320 * correlate userspace IPs to code. They have the following structure:
323 * struct perf_event_header header;
332 PERF_RECORD_MMAP = 1,
336 * struct perf_event_header header;
341 PERF_RECORD_LOST = 2,
345 * struct perf_event_header header;
351 PERF_RECORD_COMM = 3,
355 * struct perf_event_header header;
361 PERF_RECORD_EXIT = 4,
365 * struct perf_event_header header;
371 PERF_RECORD_THROTTLE = 5,
372 PERF_RECORD_UNTHROTTLE = 6,
376 * struct perf_event_header header;
382 PERF_RECORD_FORK = 7,
386 * struct perf_event_header header;
389 * struct read_format values;
392 PERF_RECORD_READ = 8,
396 * struct perf_event_header header;
398 * { u64 ip; } && PERF_SAMPLE_IP
399 * { u32 pid, tid; } && PERF_SAMPLE_TID
400 * { u64 time; } && PERF_SAMPLE_TIME
401 * { u64 addr; } && PERF_SAMPLE_ADDR
402 * { u64 id; } && PERF_SAMPLE_ID
403 * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID
404 * { u32 cpu, res; } && PERF_SAMPLE_CPU
405 * { u64 period; } && PERF_SAMPLE_PERIOD
407 * { struct read_format values; } && PERF_SAMPLE_READ
410 * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN
413 * # The RAW record below is opaque data wrt the ABI
415 * # That is, the ABI doesn't make any promises wrt to
416 * # the stability of its content, it may vary depending
417 * # on event, hardware, kernel version and phase of
420 * # In other words, PERF_SAMPLE_RAW contents are not an ABI.
424 * char data[size];}&& PERF_SAMPLE_RAW
427 PERF_RECORD_SAMPLE = 9,
429 PERF_RECORD_MAX, /* non-ABI */
432 enum perf_callchain_context {
433 PERF_CONTEXT_HV = (__u64)-32,
434 PERF_CONTEXT_KERNEL = (__u64)-128,
435 PERF_CONTEXT_USER = (__u64)-512,
437 PERF_CONTEXT_GUEST = (__u64)-2048,
438 PERF_CONTEXT_GUEST_KERNEL = (__u64)-2176,
439 PERF_CONTEXT_GUEST_USER = (__u64)-2560,
441 PERF_CONTEXT_MAX = (__u64)-4095,
444 #define PERF_FLAG_FD_NO_GROUP (1U << 0)
445 #define PERF_FLAG_FD_OUTPUT (1U << 1)
449 * Kernel-internal data types and definitions:
452 #ifdef CONFIG_PERF_EVENTS
453 # include <asm/perf_event.h>
456 #include <linux/list.h>
457 #include <linux/mutex.h>
458 #include <linux/rculist.h>
459 #include <linux/rcupdate.h>
460 #include <linux/spinlock.h>
461 #include <linux/hrtimer.h>
462 #include <linux/fs.h>
463 #include <linux/pid_namespace.h>
464 #include <linux/workqueue.h>
465 #include <asm/atomic.h>
467 #define PERF_MAX_STACK_DEPTH 255
469 struct perf_callchain_entry {
471 __u64 ip[PERF_MAX_STACK_DEPTH];
474 struct perf_raw_record {
482 * struct hw_perf_event - performance event hardware details:
484 struct hw_perf_event {
485 #ifdef CONFIG_PERF_EVENTS
487 struct { /* hardware */
489 unsigned long config_base;
490 unsigned long event_base;
493 struct { /* software */
495 struct hrtimer hrtimer;
497 #ifdef CONFIG_HAVE_HW_BREAKPOINT
498 union { /* breakpoint */
499 struct arch_hw_breakpoint info;
503 atomic64_t prev_count;
506 atomic64_t period_left;
518 * struct pmu - generic performance monitoring unit
521 int (*enable) (struct perf_event *event);
522 void (*disable) (struct perf_event *event);
523 void (*read) (struct perf_event *event);
524 void (*unthrottle) (struct perf_event *event);
528 * enum perf_event_active_state - the states of a event
530 enum perf_event_active_state {
531 PERF_EVENT_STATE_ERROR = -2,
532 PERF_EVENT_STATE_OFF = -1,
533 PERF_EVENT_STATE_INACTIVE = 0,
534 PERF_EVENT_STATE_ACTIVE = 1,
539 struct perf_mmap_data {
540 struct rcu_head rcu_head;
541 #ifdef CONFIG_PERF_USE_VMALLOC
542 struct work_struct work;
545 int nr_pages; /* nr of data pages */
546 int writable; /* are we writable */
547 int nr_locked; /* nr pages mlocked */
549 atomic_t poll; /* POLL_ for wakeups */
550 atomic_t events; /* event_id limit */
552 atomic_long_t head; /* write position */
553 atomic_long_t done_head; /* completed head */
555 atomic_t lock; /* concurrent writes */
556 atomic_t wakeup; /* needs a wakeup */
557 atomic_t lost; /* nr records lost */
559 long watermark; /* wakeup watermark */
561 struct perf_event_mmap_page *user_page;
565 struct perf_pending_entry {
566 struct perf_pending_entry *next;
567 void (*func)(struct perf_pending_entry *);
570 typedef void (*perf_callback_t)(struct perf_event *, void *);
572 struct perf_sample_data;
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 void (*overflow_handler)(struct perf_event *event,
666 int nmi, struct perf_sample_data *data,
667 struct pt_regs *regs);
669 #ifdef CONFIG_EVENT_PROFILE
670 struct event_filter *filter;
673 perf_callback_t callback;
675 perf_callback_t event_callback;
677 #endif /* CONFIG_PERF_EVENTS */
681 * struct perf_event_context - event context structure
683 * Used as a container for task events and CPU events as well:
685 struct perf_event_context {
687 * Protect the states of the events in the list,
688 * nr_active, and the list:
692 * Protect the list of events. Locking either mutex or lock
693 * is sufficient to ensure the list doesn't change; to change
694 * the list you need to lock both the mutex and the spinlock.
698 struct list_head group_list;
699 struct list_head event_list;
705 struct task_struct *task;
708 * Context clock, runs when context enabled.
714 * These fields let us detect when two contexts have both
715 * been cloned (inherited) from a common ancestor.
717 struct perf_event_context *parent_ctx;
721 struct rcu_head rcu_head;
725 * struct perf_event_cpu_context - per cpu event context structure
727 struct perf_cpu_context {
728 struct perf_event_context ctx;
729 struct perf_event_context *task_ctx;
735 * Recursion avoidance:
737 * task, softirq, irq, nmi context
742 struct perf_output_handle {
743 struct perf_event *event;
744 struct perf_mmap_data *data;
746 unsigned long offset;
752 #ifdef CONFIG_PERF_EVENTS
755 * Set by architecture code:
757 extern int perf_max_events;
759 extern const struct pmu *hw_perf_event_init(struct perf_event *event);
761 extern void perf_event_task_sched_in(struct task_struct *task, int cpu);
762 extern void perf_event_task_sched_out(struct task_struct *task,
763 struct task_struct *next, int cpu);
764 extern void perf_event_task_tick(struct task_struct *task, int cpu);
765 extern int perf_event_init_task(struct task_struct *child);
766 extern void perf_event_exit_task(struct task_struct *child);
767 extern void perf_event_free_task(struct task_struct *task);
768 extern void set_perf_event_pending(void);
769 extern void perf_event_do_pending(void);
770 extern void perf_event_print_debug(void);
771 extern void __perf_disable(void);
772 extern bool __perf_enable(void);
773 extern void perf_disable(void);
774 extern void perf_enable(void);
775 extern int perf_event_task_disable(void);
776 extern int perf_event_task_enable(void);
777 extern int hw_perf_group_sched_in(struct perf_event *group_leader,
778 struct perf_cpu_context *cpuctx,
779 struct perf_event_context *ctx, int cpu);
780 extern void perf_event_update_userpage(struct perf_event *event);
781 extern int perf_event_release_kernel(struct perf_event *event);
782 extern struct perf_event *
783 perf_event_create_kernel_counter(struct perf_event_attr *attr,
786 perf_callback_t callback);
787 extern u64 perf_event_read_value(struct perf_event *event,
788 u64 *enabled, u64 *running);
790 struct perf_sample_data {
807 struct perf_callchain_entry *callchain;
808 struct perf_raw_record *raw;
811 extern void perf_output_sample(struct perf_output_handle *handle,
812 struct perf_event_header *header,
813 struct perf_sample_data *data,
814 struct perf_event *event);
815 extern void perf_prepare_sample(struct perf_event_header *header,
816 struct perf_sample_data *data,
817 struct perf_event *event,
818 struct pt_regs *regs);
820 extern int perf_event_overflow(struct perf_event *event, int nmi,
821 struct perf_sample_data *data,
822 struct pt_regs *regs);
825 * Return 1 for a software event, 0 for a hardware event
827 static inline int is_software_event(struct perf_event *event)
829 return (event->attr.type != PERF_TYPE_RAW) &&
830 (event->attr.type != PERF_TYPE_HARDWARE) &&
831 (event->attr.type != PERF_TYPE_HW_CACHE);
834 extern atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
836 extern void __perf_sw_event(u32, u64, int, struct pt_regs *, u64);
839 perf_sw_event(u32 event_id, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
841 if (atomic_read(&perf_swevent_enabled[event_id]))
842 __perf_sw_event(event_id, nr, nmi, regs, addr);
845 extern void __perf_event_mmap(struct vm_area_struct *vma);
847 static inline void perf_event_mmap(struct vm_area_struct *vma)
849 if (vma->vm_flags & VM_EXEC)
850 __perf_event_mmap(vma);
853 extern void perf_event_comm(struct task_struct *tsk);
854 extern void perf_event_fork(struct task_struct *tsk);
856 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
858 extern int sysctl_perf_event_paranoid;
859 extern int sysctl_perf_event_mlock;
860 extern int sysctl_perf_event_sample_rate;
862 extern void perf_event_init(void);
863 extern void perf_tp_event(int event_id, u64 addr, u64 count,
864 void *record, int entry_size);
865 extern void perf_bp_event(struct perf_event *event, void *data);
867 #ifndef perf_misc_flags
868 #define perf_misc_flags(regs) (user_mode(regs) ? PERF_RECORD_MISC_USER : \
869 PERF_RECORD_MISC_KERNEL)
870 #define perf_instruction_pointer(regs) instruction_pointer(regs)
873 extern int perf_output_begin(struct perf_output_handle *handle,
874 struct perf_event *event, unsigned int size,
875 int nmi, int sample);
876 extern void perf_output_end(struct perf_output_handle *handle);
877 extern void perf_output_copy(struct perf_output_handle *handle,
878 const void *buf, unsigned int len);
879 extern int perf_swevent_get_recursion_context(void);
880 extern void perf_swevent_put_recursion_context(int rctx);
883 perf_event_task_sched_in(struct task_struct *task, int cpu) { }
885 perf_event_task_sched_out(struct task_struct *task,
886 struct task_struct *next, int cpu) { }
888 perf_event_task_tick(struct task_struct *task, int cpu) { }
889 static inline int perf_event_init_task(struct task_struct *child) { return 0; }
890 static inline void perf_event_exit_task(struct task_struct *child) { }
891 static inline void perf_event_free_task(struct task_struct *task) { }
892 static inline void perf_event_do_pending(void) { }
893 static inline void perf_event_print_debug(void) { }
894 static inline void perf_disable(void) { }
895 static inline void perf_enable(void) { }
896 static inline int perf_event_task_disable(void) { return -EINVAL; }
897 static inline int perf_event_task_enable(void) { return -EINVAL; }
900 perf_sw_event(u32 event_id, u64 nr, int nmi,
901 struct pt_regs *regs, u64 addr) { }
903 perf_bp_event(struct perf_event *event, void *data) { }
905 static inline void perf_event_mmap(struct vm_area_struct *vma) { }
906 static inline void perf_event_comm(struct task_struct *tsk) { }
907 static inline void perf_event_fork(struct task_struct *tsk) { }
908 static inline void perf_event_init(void) { }
909 static inline int perf_swevent_get_recursion_context(void) { return -1; }
910 static inline void perf_swevent_put_recursion_context(int rctx) { }
914 #define perf_output_put(handle, x) \
915 perf_output_copy((handle), &(x), sizeof(x))
917 #endif /* __KERNEL__ */
918 #endif /* _LINUX_PERF_EVENT_H */