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 */
209 * 0 - SAMPLE_IP can have arbitrary skid
210 * 1 - SAMPLE_IP must have constant skid
211 * 2 - SAMPLE_IP requested to have 0 skid
212 * 3 - SAMPLE_IP must have 0 skid
214 * See also PERF_RECORD_MISC_EXACT_IP
216 precise_ip
: 2, /* skid constraint */
217 mmap_data
: 1, /* non-exec mmap data */
222 __u32 wakeup_events
; /* wakeup every n events */
223 __u32 wakeup_watermark
; /* bytes before wakeup */
232 * Ioctls that can be done on a perf event fd:
234 #define PERF_EVENT_IOC_ENABLE _IO ('$', 0)
235 #define PERF_EVENT_IOC_DISABLE _IO ('$', 1)
236 #define PERF_EVENT_IOC_REFRESH _IO ('$', 2)
237 #define PERF_EVENT_IOC_RESET _IO ('$', 3)
238 #define PERF_EVENT_IOC_PERIOD _IOW('$', 4, __u64)
239 #define PERF_EVENT_IOC_SET_OUTPUT _IO ('$', 5)
240 #define PERF_EVENT_IOC_SET_FILTER _IOW('$', 6, char *)
242 enum perf_event_ioc_flags
{
243 PERF_IOC_FLAG_GROUP
= 1U << 0,
247 * Structure of the page that can be mapped via mmap
249 struct perf_event_mmap_page
{
250 __u32 version
; /* version number of this structure */
251 __u32 compat_version
; /* lowest version this is compat with */
254 * Bits needed to read the hw events in user-space.
264 * count = pmc_read(pc->index - 1);
265 * count += pc->offset;
270 * } while (pc->lock != seq);
272 * NOTE: for obvious reason this only works on self-monitoring
275 __u32 lock
; /* seqlock for synchronization */
276 __u32 index
; /* hardware event identifier */
277 __s64 offset
; /* add to hardware event value */
278 __u64 time_enabled
; /* time event active */
279 __u64 time_running
; /* time event on cpu */
282 * Hole for extension of the self monitor capabilities
285 __u64 __reserved
[123]; /* align to 1k */
288 * Control data for the mmap() data buffer.
290 * User-space reading the @data_head value should issue an rmb(), on
291 * SMP capable platforms, after reading this value -- see
292 * perf_event_wakeup().
294 * When the mapping is PROT_WRITE the @data_tail value should be
295 * written by userspace to reflect the last read data. In this case
296 * the kernel will not over-write unread data.
298 __u64 data_head
; /* head in the data section */
299 __u64 data_tail
; /* user-space written tail */
302 #define PERF_RECORD_MISC_CPUMODE_MASK (7 << 0)
303 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN (0 << 0)
304 #define PERF_RECORD_MISC_KERNEL (1 << 0)
305 #define PERF_RECORD_MISC_USER (2 << 0)
306 #define PERF_RECORD_MISC_HYPERVISOR (3 << 0)
307 #define PERF_RECORD_MISC_GUEST_KERNEL (4 << 0)
308 #define PERF_RECORD_MISC_GUEST_USER (5 << 0)
311 * Indicates that the content of PERF_SAMPLE_IP points to
312 * the actual instruction that triggered the event. See also
313 * perf_event_attr::precise_ip.
315 #define PERF_RECORD_MISC_EXACT_IP (1 << 14)
317 * Reserve the last bit to indicate some extended misc field
319 #define PERF_RECORD_MISC_EXT_RESERVED (1 << 15)
321 struct perf_event_header
{
327 enum perf_event_type
{
330 * The MMAP events record the PROT_EXEC mappings so that we can
331 * correlate userspace IPs to code. They have the following structure:
334 * struct perf_event_header header;
343 PERF_RECORD_MMAP
= 1,
347 * struct perf_event_header header;
352 PERF_RECORD_LOST
= 2,
356 * struct perf_event_header header;
362 PERF_RECORD_COMM
= 3,
366 * struct perf_event_header header;
372 PERF_RECORD_EXIT
= 4,
376 * struct perf_event_header header;
382 PERF_RECORD_THROTTLE
= 5,
383 PERF_RECORD_UNTHROTTLE
= 6,
387 * struct perf_event_header header;
393 PERF_RECORD_FORK
= 7,
397 * struct perf_event_header header;
400 * struct read_format values;
403 PERF_RECORD_READ
= 8,
407 * struct perf_event_header header;
409 * { u64 ip; } && PERF_SAMPLE_IP
410 * { u32 pid, tid; } && PERF_SAMPLE_TID
411 * { u64 time; } && PERF_SAMPLE_TIME
412 * { u64 addr; } && PERF_SAMPLE_ADDR
413 * { u64 id; } && PERF_SAMPLE_ID
414 * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID
415 * { u32 cpu, res; } && PERF_SAMPLE_CPU
416 * { u64 period; } && PERF_SAMPLE_PERIOD
418 * { struct read_format values; } && PERF_SAMPLE_READ
421 * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN
424 * # The RAW record below is opaque data wrt the ABI
426 * # That is, the ABI doesn't make any promises wrt to
427 * # the stability of its content, it may vary depending
428 * # on event, hardware, kernel version and phase of
431 * # In other words, PERF_SAMPLE_RAW contents are not an ABI.
435 * char data[size];}&& PERF_SAMPLE_RAW
438 PERF_RECORD_SAMPLE
= 9,
440 PERF_RECORD_MAX
, /* non-ABI */
443 enum perf_callchain_context
{
444 PERF_CONTEXT_HV
= (__u64
)-32,
445 PERF_CONTEXT_KERNEL
= (__u64
)-128,
446 PERF_CONTEXT_USER
= (__u64
)-512,
448 PERF_CONTEXT_GUEST
= (__u64
)-2048,
449 PERF_CONTEXT_GUEST_KERNEL
= (__u64
)-2176,
450 PERF_CONTEXT_GUEST_USER
= (__u64
)-2560,
452 PERF_CONTEXT_MAX
= (__u64
)-4095,
455 #define PERF_FLAG_FD_NO_GROUP (1U << 0)
456 #define PERF_FLAG_FD_OUTPUT (1U << 1)
460 * Kernel-internal data types and definitions:
463 #ifdef CONFIG_PERF_EVENTS
464 # include <asm/perf_event.h>
465 # include <asm/local64.h>
468 struct perf_guest_info_callbacks
{
469 int (*is_in_guest
) (void);
470 int (*is_user_mode
) (void);
471 unsigned long (*get_guest_ip
) (void);
474 #ifdef CONFIG_HAVE_HW_BREAKPOINT
475 #include <asm/hw_breakpoint.h>
478 #include <linux/list.h>
479 #include <linux/mutex.h>
480 #include <linux/rculist.h>
481 #include <linux/rcupdate.h>
482 #include <linux/spinlock.h>
483 #include <linux/hrtimer.h>
484 #include <linux/fs.h>
485 #include <linux/pid_namespace.h>
486 #include <linux/workqueue.h>
487 #include <linux/ftrace.h>
488 #include <linux/cpu.h>
489 #include <asm/atomic.h>
490 #include <asm/local.h>
492 #define PERF_MAX_STACK_DEPTH 255
494 struct perf_callchain_entry
{
496 __u64 ip
[PERF_MAX_STACK_DEPTH
];
499 struct perf_raw_record
{
504 struct perf_branch_entry
{
510 struct perf_branch_stack
{
512 struct perf_branch_entry entries
[0];
518 * struct hw_perf_event - performance event hardware details:
520 struct hw_perf_event
{
521 #ifdef CONFIG_PERF_EVENTS
523 struct { /* hardware */
526 unsigned long config_base
;
527 unsigned long event_base
;
531 struct { /* software */
532 struct hrtimer hrtimer
;
534 #ifdef CONFIG_HAVE_HW_BREAKPOINT
535 struct { /* breakpoint */
536 struct arch_hw_breakpoint info
;
537 struct list_head bp_list
;
542 local64_t prev_count
;
545 local64_t period_left
;
549 u64 freq_count_stamp
;
554 * hw_perf_event::state flags
556 #define PERF_HES_STOPPED 0x01 /* the counter is stopped */
557 #define PERF_HES_UPTODATE 0x02 /* event->count up-to-date */
558 #define PERF_HES_ARCH 0x04
563 * Common implementation detail of pmu::{start,commit,cancel}_txn
565 #define PERF_EVENT_TXN 0x1
568 * struct pmu - generic performance monitoring unit
571 struct list_head entry
;
573 int * __percpu pmu_disable_count
;
574 struct perf_cpu_context
* __percpu pmu_cpu_context
;
578 * Fully disable/enable this PMU, can be used to protect from the PMI
579 * as well as for lazy/batch writing of the MSRs.
581 void (*pmu_enable
) (struct pmu
*pmu
); /* optional */
582 void (*pmu_disable
) (struct pmu
*pmu
); /* optional */
585 * Try and initialize the event for this PMU.
586 * Should return -ENOENT when the @event doesn't match this PMU.
588 int (*event_init
) (struct perf_event
*event
);
590 #define PERF_EF_START 0x01 /* start the counter when adding */
591 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
592 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
595 * Adds/Removes a counter to/from the PMU, can be done inside
596 * a transaction, see the ->*_txn() methods.
598 int (*add
) (struct perf_event
*event
, int flags
);
599 void (*del
) (struct perf_event
*event
, int flags
);
602 * Starts/Stops a counter present on the PMU. The PMI handler
603 * should stop the counter when perf_event_overflow() returns
604 * !0. ->start() will be used to continue.
606 void (*start
) (struct perf_event
*event
, int flags
);
607 void (*stop
) (struct perf_event
*event
, int flags
);
610 * Updates the counter value of the event.
612 void (*read
) (struct perf_event
*event
);
615 * Group events scheduling is treated as a transaction, add
616 * group events as a whole and perform one schedulability test.
617 * If the test fails, roll back the whole group
619 * Start the transaction, after this ->add() doesn't need to
620 * do schedulability tests.
622 void (*start_txn
) (struct pmu
*pmu
); /* optional */
624 * If ->start_txn() disabled the ->add() schedulability test
625 * then ->commit_txn() is required to perform one. On success
626 * the transaction is closed. On error the transaction is kept
627 * open until ->cancel_txn() is called.
629 int (*commit_txn
) (struct pmu
*pmu
); /* optional */
631 * Will cancel the transaction, assumes ->del() is called
632 * for each successfull ->add() during the transaction.
634 void (*cancel_txn
) (struct pmu
*pmu
); /* optional */
638 * enum perf_event_active_state - the states of a event
640 enum perf_event_active_state
{
641 PERF_EVENT_STATE_ERROR
= -2,
642 PERF_EVENT_STATE_OFF
= -1,
643 PERF_EVENT_STATE_INACTIVE
= 0,
644 PERF_EVENT_STATE_ACTIVE
= 1,
649 #define PERF_BUFFER_WRITABLE 0x01
653 struct rcu_head rcu_head
;
654 #ifdef CONFIG_PERF_USE_VMALLOC
655 struct work_struct work
;
656 int page_order
; /* allocation order */
658 int nr_pages
; /* nr of data pages */
659 int writable
; /* are we writable */
661 atomic_t poll
; /* POLL_ for wakeups */
663 local_t head
; /* write position */
664 local_t nest
; /* nested writers */
665 local_t events
; /* event limit */
666 local_t wakeup
; /* wakeup stamp */
667 local_t lost
; /* nr records lost */
669 long watermark
; /* wakeup watermark */
671 struct perf_event_mmap_page
*user_page
;
675 struct perf_pending_entry
{
676 struct perf_pending_entry
*next
;
677 void (*func
)(struct perf_pending_entry
*);
680 struct perf_sample_data
;
682 typedef void (*perf_overflow_handler_t
)(struct perf_event
*, int,
683 struct perf_sample_data
*,
684 struct pt_regs
*regs
);
686 enum perf_group_flag
{
687 PERF_GROUP_SOFTWARE
= 0x1,
690 #define SWEVENT_HLIST_BITS 8
691 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
693 struct swevent_hlist
{
694 struct hlist_head heads
[SWEVENT_HLIST_SIZE
];
695 struct rcu_head rcu_head
;
698 #define PERF_ATTACH_CONTEXT 0x01
699 #define PERF_ATTACH_GROUP 0x02
702 * struct perf_event - performance event kernel representation:
705 #ifdef CONFIG_PERF_EVENTS
706 struct list_head group_entry
;
707 struct list_head event_entry
;
708 struct list_head sibling_list
;
709 struct hlist_node hlist_entry
;
712 struct perf_event
*group_leader
;
715 enum perf_event_active_state state
;
716 unsigned int attach_state
;
718 atomic64_t child_count
;
721 * These are the total time in nanoseconds that the event
722 * has been enabled (i.e. eligible to run, and the task has
723 * been scheduled in, if this is a per-task event)
724 * and running (scheduled onto the CPU), respectively.
726 * They are computed from tstamp_enabled, tstamp_running and
727 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
729 u64 total_time_enabled
;
730 u64 total_time_running
;
733 * These are timestamps used for computing total_time_enabled
734 * and total_time_running when the event is in INACTIVE or
735 * ACTIVE state, measured in nanoseconds from an arbitrary point
737 * tstamp_enabled: the notional time when the event was enabled
738 * tstamp_running: the notional time when the event was scheduled on
739 * tstamp_stopped: in INACTIVE state, the notional time when the
740 * event was scheduled off.
746 struct perf_event_attr attr
;
747 struct hw_perf_event hw
;
749 struct perf_event_context
*ctx
;
753 * These accumulate total time (in nanoseconds) that children
754 * events have been enabled and running, respectively.
756 atomic64_t child_total_time_enabled
;
757 atomic64_t child_total_time_running
;
760 * Protect attach/detach and child_list:
762 struct mutex child_mutex
;
763 struct list_head child_list
;
764 struct perf_event
*parent
;
769 struct list_head owner_entry
;
770 struct task_struct
*owner
;
773 struct mutex mmap_mutex
;
776 struct user_struct
*mmap_user
;
777 struct perf_buffer
*buffer
;
780 wait_queue_head_t waitq
;
781 struct fasync_struct
*fasync
;
783 /* delayed work for NMIs and such */
787 struct perf_pending_entry pending
;
789 atomic_t event_limit
;
791 void (*destroy
)(struct perf_event
*);
792 struct rcu_head rcu_head
;
794 struct pid_namespace
*ns
;
797 perf_overflow_handler_t overflow_handler
;
799 #ifdef CONFIG_EVENT_TRACING
800 struct ftrace_event_call
*tp_event
;
801 struct event_filter
*filter
;
804 #endif /* CONFIG_PERF_EVENTS */
808 * struct perf_event_context - event context structure
810 * Used as a container for task events and CPU events as well:
812 struct perf_event_context
{
815 * Protect the states of the events in the list,
816 * nr_active, and the list:
820 * Protect the list of events. Locking either mutex or lock
821 * is sufficient to ensure the list doesn't change; to change
822 * the list you need to lock both the mutex and the spinlock.
826 struct list_head pinned_groups
;
827 struct list_head flexible_groups
;
828 struct list_head event_list
;
834 struct task_struct
*task
;
837 * Context clock, runs when context enabled.
843 * These fields let us detect when two contexts have both
844 * been cloned (inherited) from a common ancestor.
846 struct perf_event_context
*parent_ctx
;
850 struct rcu_head rcu_head
;
854 * Number of contexts where an event can trigger:
855 * task, softirq, hardirq, nmi.
857 #define PERF_NR_CONTEXTS 4
860 * struct perf_event_cpu_context - per cpu event context structure
862 struct perf_cpu_context
{
863 struct perf_event_context ctx
;
864 struct perf_event_context
*task_ctx
;
868 struct hrtimer timer
;
871 struct perf_output_handle
{
872 struct perf_event
*event
;
873 struct perf_buffer
*buffer
;
874 unsigned long wakeup
;
882 #ifdef CONFIG_PERF_EVENTS
884 extern int perf_pmu_register(struct pmu
*pmu
);
885 extern void perf_pmu_unregister(struct pmu
*pmu
);
887 extern void perf_event_task_sched_in(struct task_struct
*task
);
888 extern void perf_event_task_sched_out(struct task_struct
*task
, struct task_struct
*next
);
889 extern int perf_event_init_task(struct task_struct
*child
);
890 extern void perf_event_exit_task(struct task_struct
*child
);
891 extern void perf_event_free_task(struct task_struct
*task
);
892 extern void perf_event_delayed_put(struct task_struct
*task
);
893 extern void set_perf_event_pending(void);
894 extern void perf_event_do_pending(void);
895 extern void perf_event_print_debug(void);
896 extern void perf_pmu_disable(struct pmu
*pmu
);
897 extern void perf_pmu_enable(struct pmu
*pmu
);
898 extern int perf_event_task_disable(void);
899 extern int perf_event_task_enable(void);
900 extern void perf_event_update_userpage(struct perf_event
*event
);
901 extern int perf_event_release_kernel(struct perf_event
*event
);
902 extern struct perf_event
*
903 perf_event_create_kernel_counter(struct perf_event_attr
*attr
,
906 perf_overflow_handler_t callback
);
907 extern u64
perf_event_read_value(struct perf_event
*event
,
908 u64
*enabled
, u64
*running
);
910 struct perf_sample_data
{
927 struct perf_callchain_entry
*callchain
;
928 struct perf_raw_record
*raw
;
932 void perf_sample_data_init(struct perf_sample_data
*data
, u64 addr
)
938 extern void perf_output_sample(struct perf_output_handle
*handle
,
939 struct perf_event_header
*header
,
940 struct perf_sample_data
*data
,
941 struct perf_event
*event
);
942 extern void perf_prepare_sample(struct perf_event_header
*header
,
943 struct perf_sample_data
*data
,
944 struct perf_event
*event
,
945 struct pt_regs
*regs
);
947 extern int perf_event_overflow(struct perf_event
*event
, int nmi
,
948 struct perf_sample_data
*data
,
949 struct pt_regs
*regs
);
952 * Return 1 for a software event, 0 for a hardware event
954 static inline int is_software_event(struct perf_event
*event
)
956 return event
->pmu
->task_ctx_nr
== perf_sw_context
;
959 extern atomic_t perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
961 extern void __perf_sw_event(u32
, u64
, int, struct pt_regs
*, u64
);
963 #ifndef perf_arch_fetch_caller_regs
965 perf_arch_fetch_caller_regs(struct pt_regs
*regs
, unsigned long ip
) { }
969 * Take a snapshot of the regs. Skip ip and frame pointer to
970 * the nth caller. We only need a few of the regs:
971 * - ip for PERF_SAMPLE_IP
972 * - cs for user_mode() tests
973 * - bp for callchains
974 * - eflags, for future purposes, just in case
976 static inline void perf_fetch_caller_regs(struct pt_regs
*regs
)
978 memset(regs
, 0, sizeof(*regs
));
980 perf_arch_fetch_caller_regs(regs
, CALLER_ADDR0
);
984 perf_sw_event(u32 event_id
, u64 nr
, int nmi
, struct pt_regs
*regs
, u64 addr
)
986 if (atomic_read(&perf_swevent_enabled
[event_id
])) {
987 struct pt_regs hot_regs
;
990 perf_fetch_caller_regs(&hot_regs
);
993 __perf_sw_event(event_id
, nr
, nmi
, regs
, addr
);
997 extern void perf_event_mmap(struct vm_area_struct
*vma
);
998 extern struct perf_guest_info_callbacks
*perf_guest_cbs
;
999 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*callbacks
);
1000 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*callbacks
);
1002 extern void perf_event_comm(struct task_struct
*tsk
);
1003 extern void perf_event_fork(struct task_struct
*tsk
);
1006 DECLARE_PER_CPU(struct perf_callchain_entry
, perf_callchain_entry
);
1008 extern void perf_callchain_user(struct perf_callchain_entry
*entry
,
1009 struct pt_regs
*regs
);
1010 extern void perf_callchain_kernel(struct perf_callchain_entry
*entry
,
1011 struct pt_regs
*regs
);
1015 perf_callchain_store(struct perf_callchain_entry
*entry
, u64 ip
)
1017 if (entry
->nr
< PERF_MAX_STACK_DEPTH
)
1018 entry
->ip
[entry
->nr
++] = ip
;
1021 extern int sysctl_perf_event_paranoid
;
1022 extern int sysctl_perf_event_mlock
;
1023 extern int sysctl_perf_event_sample_rate
;
1025 static inline bool perf_paranoid_tracepoint_raw(void)
1027 return sysctl_perf_event_paranoid
> -1;
1030 static inline bool perf_paranoid_cpu(void)
1032 return sysctl_perf_event_paranoid
> 0;
1035 static inline bool perf_paranoid_kernel(void)
1037 return sysctl_perf_event_paranoid
> 1;
1040 extern void perf_event_init(void);
1041 extern void perf_tp_event(u64 addr
, u64 count
, void *record
,
1042 int entry_size
, struct pt_regs
*regs
,
1043 struct hlist_head
*head
, int rctx
);
1044 extern void perf_bp_event(struct perf_event
*event
, void *data
);
1046 #ifndef perf_misc_flags
1047 #define perf_misc_flags(regs) (user_mode(regs) ? PERF_RECORD_MISC_USER : \
1048 PERF_RECORD_MISC_KERNEL)
1049 #define perf_instruction_pointer(regs) instruction_pointer(regs)
1052 extern int perf_output_begin(struct perf_output_handle
*handle
,
1053 struct perf_event
*event
, unsigned int size
,
1054 int nmi
, int sample
);
1055 extern void perf_output_end(struct perf_output_handle
*handle
);
1056 extern void perf_output_copy(struct perf_output_handle
*handle
,
1057 const void *buf
, unsigned int len
);
1058 extern int perf_swevent_get_recursion_context(void);
1059 extern void perf_swevent_put_recursion_context(int rctx
);
1060 extern void perf_event_enable(struct perf_event
*event
);
1061 extern void perf_event_disable(struct perf_event
*event
);
1064 perf_event_task_sched_in(struct task_struct
*task
) { }
1066 perf_event_task_sched_out(struct task_struct
*task
,
1067 struct task_struct
*next
) { }
1068 static inline int perf_event_init_task(struct task_struct
*child
) { return 0; }
1069 static inline void perf_event_exit_task(struct task_struct
*child
) { }
1070 static inline void perf_event_free_task(struct task_struct
*task
) { }
1071 static inline void perf_event_delayed_put(struct task_struct
*task
) { }
1072 static inline void perf_event_do_pending(void) { }
1073 static inline void perf_event_print_debug(void) { }
1074 static inline int perf_event_task_disable(void) { return -EINVAL
; }
1075 static inline int perf_event_task_enable(void) { return -EINVAL
; }
1078 perf_sw_event(u32 event_id
, u64 nr
, int nmi
,
1079 struct pt_regs
*regs
, u64 addr
) { }
1081 perf_bp_event(struct perf_event
*event
, void *data
) { }
1083 static inline int perf_register_guest_info_callbacks
1084 (struct perf_guest_info_callbacks
*callbacks
) { return 0; }
1085 static inline int perf_unregister_guest_info_callbacks
1086 (struct perf_guest_info_callbacks
*callbacks
) { return 0; }
1088 static inline void perf_event_mmap(struct vm_area_struct
*vma
) { }
1089 static inline void perf_event_comm(struct task_struct
*tsk
) { }
1090 static inline void perf_event_fork(struct task_struct
*tsk
) { }
1091 static inline void perf_event_init(void) { }
1092 static inline int perf_swevent_get_recursion_context(void) { return -1; }
1093 static inline void perf_swevent_put_recursion_context(int rctx
) { }
1094 static inline void perf_event_enable(struct perf_event
*event
) { }
1095 static inline void perf_event_disable(struct perf_event
*event
) { }
1098 #define perf_output_put(handle, x) \
1099 perf_output_copy((handle), &(x), sizeof(x))
1102 * This has to have a higher priority than migration_notifier in sched.c.
1104 #define perf_cpu_notifier(fn) \
1106 static struct notifier_block fn##_nb __cpuinitdata = \
1107 { .notifier_call = fn, .priority = CPU_PRI_PERF }; \
1108 fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE, \
1109 (void *)(unsigned long)smp_processor_id()); \
1110 fn(&fn##_nb, (unsigned long)CPU_STARTING, \
1111 (void *)(unsigned long)smp_processor_id()); \
1112 fn(&fn##_nb, (unsigned long)CPU_ONLINE, \
1113 (void *)(unsigned long)smp_processor_id()); \
1114 register_cpu_notifier(&fn##_nb); \
1117 #endif /* __KERNEL__ */
1118 #endif /* _LINUX_PERF_EVENT_H */