4 * Copyright (C) 2008-2009, Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2011, Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2011, 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,
55 PERF_COUNT_HW_STALLED_CYCLES_FRONTEND
= 7,
56 PERF_COUNT_HW_STALLED_CYCLES_BACKEND
= 8,
58 PERF_COUNT_HW_MAX
, /* non-ABI */
62 * Generalized hardware cache events:
64 * { L1-D, L1-I, LLC, ITLB, DTLB, BPU } x
65 * { read, write, prefetch } x
66 * { accesses, misses }
68 enum perf_hw_cache_id
{
69 PERF_COUNT_HW_CACHE_L1D
= 0,
70 PERF_COUNT_HW_CACHE_L1I
= 1,
71 PERF_COUNT_HW_CACHE_LL
= 2,
72 PERF_COUNT_HW_CACHE_DTLB
= 3,
73 PERF_COUNT_HW_CACHE_ITLB
= 4,
74 PERF_COUNT_HW_CACHE_BPU
= 5,
76 PERF_COUNT_HW_CACHE_MAX
, /* non-ABI */
79 enum perf_hw_cache_op_id
{
80 PERF_COUNT_HW_CACHE_OP_READ
= 0,
81 PERF_COUNT_HW_CACHE_OP_WRITE
= 1,
82 PERF_COUNT_HW_CACHE_OP_PREFETCH
= 2,
84 PERF_COUNT_HW_CACHE_OP_MAX
, /* non-ABI */
87 enum perf_hw_cache_op_result_id
{
88 PERF_COUNT_HW_CACHE_RESULT_ACCESS
= 0,
89 PERF_COUNT_HW_CACHE_RESULT_MISS
= 1,
91 PERF_COUNT_HW_CACHE_RESULT_MAX
, /* non-ABI */
95 * Special "software" events provided by the kernel, even if the hardware
96 * does not support performance events. These events measure various
97 * physical and sw events of the kernel (and allow the profiling of them as
101 PERF_COUNT_SW_CPU_CLOCK
= 0,
102 PERF_COUNT_SW_TASK_CLOCK
= 1,
103 PERF_COUNT_SW_PAGE_FAULTS
= 2,
104 PERF_COUNT_SW_CONTEXT_SWITCHES
= 3,
105 PERF_COUNT_SW_CPU_MIGRATIONS
= 4,
106 PERF_COUNT_SW_PAGE_FAULTS_MIN
= 5,
107 PERF_COUNT_SW_PAGE_FAULTS_MAJ
= 6,
108 PERF_COUNT_SW_ALIGNMENT_FAULTS
= 7,
109 PERF_COUNT_SW_EMULATION_FAULTS
= 8,
111 PERF_COUNT_SW_MAX
, /* non-ABI */
115 * Bits that can be set in attr.sample_type to request information
116 * in the overflow packets.
118 enum perf_event_sample_format
{
119 PERF_SAMPLE_IP
= 1U << 0,
120 PERF_SAMPLE_TID
= 1U << 1,
121 PERF_SAMPLE_TIME
= 1U << 2,
122 PERF_SAMPLE_ADDR
= 1U << 3,
123 PERF_SAMPLE_READ
= 1U << 4,
124 PERF_SAMPLE_CALLCHAIN
= 1U << 5,
125 PERF_SAMPLE_ID
= 1U << 6,
126 PERF_SAMPLE_CPU
= 1U << 7,
127 PERF_SAMPLE_PERIOD
= 1U << 8,
128 PERF_SAMPLE_STREAM_ID
= 1U << 9,
129 PERF_SAMPLE_RAW
= 1U << 10,
131 PERF_SAMPLE_MAX
= 1U << 11, /* non-ABI */
135 * The format of the data returned by read() on a perf event fd,
136 * as specified by attr.read_format:
138 * struct read_format {
140 * { u64 time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
141 * { u64 time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
142 * { u64 id; } && PERF_FORMAT_ID
143 * } && !PERF_FORMAT_GROUP
146 * { u64 time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
147 * { u64 time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
149 * { u64 id; } && PERF_FORMAT_ID
151 * } && PERF_FORMAT_GROUP
154 enum perf_event_read_format
{
155 PERF_FORMAT_TOTAL_TIME_ENABLED
= 1U << 0,
156 PERF_FORMAT_TOTAL_TIME_RUNNING
= 1U << 1,
157 PERF_FORMAT_ID
= 1U << 2,
158 PERF_FORMAT_GROUP
= 1U << 3,
160 PERF_FORMAT_MAX
= 1U << 4, /* non-ABI */
163 #define PERF_ATTR_SIZE_VER0 64 /* sizeof first published struct */
166 * Hardware event_id to monitor via a performance monitoring event:
168 struct perf_event_attr
{
171 * Major type: hardware/software/tracepoint/etc.
176 * Size of the attr structure, for fwd/bwd compat.
181 * Type specific configuration information.
193 __u64 disabled
: 1, /* off by default */
194 inherit
: 1, /* children inherit it */
195 pinned
: 1, /* must always be on PMU */
196 exclusive
: 1, /* only group on PMU */
197 exclude_user
: 1, /* don't count user */
198 exclude_kernel
: 1, /* ditto kernel */
199 exclude_hv
: 1, /* ditto hypervisor */
200 exclude_idle
: 1, /* don't count when idle */
201 mmap
: 1, /* include mmap data */
202 comm
: 1, /* include comm data */
203 freq
: 1, /* use freq, not period */
204 inherit_stat
: 1, /* per task counts */
205 enable_on_exec
: 1, /* next exec enables */
206 task
: 1, /* trace fork/exit */
207 watermark
: 1, /* wakeup_watermark */
211 * 0 - SAMPLE_IP can have arbitrary skid
212 * 1 - SAMPLE_IP must have constant skid
213 * 2 - SAMPLE_IP requested to have 0 skid
214 * 3 - SAMPLE_IP must have 0 skid
216 * See also PERF_RECORD_MISC_EXACT_IP
218 precise_ip
: 2, /* skid constraint */
219 mmap_data
: 1, /* non-exec mmap data */
220 sample_id_all
: 1, /* sample_type all events */
225 __u32 wakeup_events
; /* wakeup every n events */
226 __u32 wakeup_watermark
; /* bytes before wakeup */
232 __u64 config1
; /* extension of config */
236 __u64 config2
; /* extension of config1 */
241 * Ioctls that can be done on a perf event fd:
243 #define PERF_EVENT_IOC_ENABLE _IO ('$', 0)
244 #define PERF_EVENT_IOC_DISABLE _IO ('$', 1)
245 #define PERF_EVENT_IOC_REFRESH _IO ('$', 2)
246 #define PERF_EVENT_IOC_RESET _IO ('$', 3)
247 #define PERF_EVENT_IOC_PERIOD _IOW('$', 4, __u64)
248 #define PERF_EVENT_IOC_SET_OUTPUT _IO ('$', 5)
249 #define PERF_EVENT_IOC_SET_FILTER _IOW('$', 6, char *)
251 enum perf_event_ioc_flags
{
252 PERF_IOC_FLAG_GROUP
= 1U << 0,
256 * Structure of the page that can be mapped via mmap
258 struct perf_event_mmap_page
{
259 __u32 version
; /* version number of this structure */
260 __u32 compat_version
; /* lowest version this is compat with */
263 * Bits needed to read the hw events in user-space.
273 * count = pmc_read(pc->index - 1);
274 * count += pc->offset;
279 * } while (pc->lock != seq);
281 * NOTE: for obvious reason this only works on self-monitoring
284 __u32 lock
; /* seqlock for synchronization */
285 __u32 index
; /* hardware event identifier */
286 __s64 offset
; /* add to hardware event value */
287 __u64 time_enabled
; /* time event active */
288 __u64 time_running
; /* time event on cpu */
291 * Hole for extension of the self monitor capabilities
294 __u64 __reserved
[123]; /* align to 1k */
297 * Control data for the mmap() data buffer.
299 * User-space reading the @data_head value should issue an rmb(), on
300 * SMP capable platforms, after reading this value -- see
301 * perf_event_wakeup().
303 * When the mapping is PROT_WRITE the @data_tail value should be
304 * written by userspace to reflect the last read data. In this case
305 * the kernel will not over-write unread data.
307 __u64 data_head
; /* head in the data section */
308 __u64 data_tail
; /* user-space written tail */
311 #define PERF_RECORD_MISC_CPUMODE_MASK (7 << 0)
312 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN (0 << 0)
313 #define PERF_RECORD_MISC_KERNEL (1 << 0)
314 #define PERF_RECORD_MISC_USER (2 << 0)
315 #define PERF_RECORD_MISC_HYPERVISOR (3 << 0)
316 #define PERF_RECORD_MISC_GUEST_KERNEL (4 << 0)
317 #define PERF_RECORD_MISC_GUEST_USER (5 << 0)
320 * Indicates that the content of PERF_SAMPLE_IP points to
321 * the actual instruction that triggered the event. See also
322 * perf_event_attr::precise_ip.
324 #define PERF_RECORD_MISC_EXACT_IP (1 << 14)
326 * Reserve the last bit to indicate some extended misc field
328 #define PERF_RECORD_MISC_EXT_RESERVED (1 << 15)
330 struct perf_event_header
{
336 enum perf_event_type
{
339 * If perf_event_attr.sample_id_all is set then all event types will
340 * have the sample_type selected fields related to where/when
341 * (identity) an event took place (TID, TIME, ID, CPU, STREAM_ID)
342 * described in PERF_RECORD_SAMPLE below, it will be stashed just after
343 * the perf_event_header and the fields already present for the existing
344 * fields, i.e. at the end of the payload. That way a newer perf.data
345 * file will be supported by older perf tools, with these new optional
346 * fields being ignored.
348 * The MMAP events record the PROT_EXEC mappings so that we can
349 * correlate userspace IPs to code. They have the following structure:
352 * struct perf_event_header header;
361 PERF_RECORD_MMAP
= 1,
365 * struct perf_event_header header;
370 PERF_RECORD_LOST
= 2,
374 * struct perf_event_header header;
380 PERF_RECORD_COMM
= 3,
384 * struct perf_event_header header;
390 PERF_RECORD_EXIT
= 4,
394 * struct perf_event_header header;
400 PERF_RECORD_THROTTLE
= 5,
401 PERF_RECORD_UNTHROTTLE
= 6,
405 * struct perf_event_header header;
411 PERF_RECORD_FORK
= 7,
415 * struct perf_event_header header;
418 * struct read_format values;
421 PERF_RECORD_READ
= 8,
425 * struct perf_event_header header;
427 * { u64 ip; } && PERF_SAMPLE_IP
428 * { u32 pid, tid; } && PERF_SAMPLE_TID
429 * { u64 time; } && PERF_SAMPLE_TIME
430 * { u64 addr; } && PERF_SAMPLE_ADDR
431 * { u64 id; } && PERF_SAMPLE_ID
432 * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID
433 * { u32 cpu, res; } && PERF_SAMPLE_CPU
434 * { u64 period; } && PERF_SAMPLE_PERIOD
436 * { struct read_format values; } && PERF_SAMPLE_READ
439 * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN
442 * # The RAW record below is opaque data wrt the ABI
444 * # That is, the ABI doesn't make any promises wrt to
445 * # the stability of its content, it may vary depending
446 * # on event, hardware, kernel version and phase of
449 * # In other words, PERF_SAMPLE_RAW contents are not an ABI.
453 * char data[size];}&& PERF_SAMPLE_RAW
456 PERF_RECORD_SAMPLE
= 9,
458 PERF_RECORD_MAX
, /* non-ABI */
461 enum perf_callchain_context
{
462 PERF_CONTEXT_HV
= (__u64
)-32,
463 PERF_CONTEXT_KERNEL
= (__u64
)-128,
464 PERF_CONTEXT_USER
= (__u64
)-512,
466 PERF_CONTEXT_GUEST
= (__u64
)-2048,
467 PERF_CONTEXT_GUEST_KERNEL
= (__u64
)-2176,
468 PERF_CONTEXT_GUEST_USER
= (__u64
)-2560,
470 PERF_CONTEXT_MAX
= (__u64
)-4095,
473 #define PERF_FLAG_FD_NO_GROUP (1U << 0)
474 #define PERF_FLAG_FD_OUTPUT (1U << 1)
475 #define PERF_FLAG_PID_CGROUP (1U << 2) /* pid=cgroup id, per-cpu mode only */
479 * Kernel-internal data types and definitions:
482 #ifdef CONFIG_PERF_EVENTS
483 # include <linux/cgroup.h>
484 # include <asm/perf_event.h>
485 # include <asm/local64.h>
488 struct perf_guest_info_callbacks
{
489 int (*is_in_guest
)(void);
490 int (*is_user_mode
)(void);
491 unsigned long (*get_guest_ip
)(void);
494 #ifdef CONFIG_HAVE_HW_BREAKPOINT
495 #include <asm/hw_breakpoint.h>
498 #include <linux/list.h>
499 #include <linux/mutex.h>
500 #include <linux/rculist.h>
501 #include <linux/rcupdate.h>
502 #include <linux/spinlock.h>
503 #include <linux/hrtimer.h>
504 #include <linux/fs.h>
505 #include <linux/pid_namespace.h>
506 #include <linux/workqueue.h>
507 #include <linux/ftrace.h>
508 #include <linux/cpu.h>
509 #include <linux/irq_work.h>
510 #include <linux/jump_label.h>
511 #include <asm/atomic.h>
512 #include <asm/local.h>
514 #define PERF_MAX_STACK_DEPTH 255
516 struct perf_callchain_entry
{
518 __u64 ip
[PERF_MAX_STACK_DEPTH
];
521 struct perf_raw_record
{
526 struct perf_branch_entry
{
532 struct perf_branch_stack
{
534 struct perf_branch_entry entries
[0];
540 * struct hw_perf_event - performance event hardware details:
542 struct hw_perf_event
{
543 #ifdef CONFIG_PERF_EVENTS
545 struct { /* hardware */
548 unsigned long config_base
;
549 unsigned long event_base
;
552 unsigned int extra_reg
;
556 struct { /* software */
557 struct hrtimer hrtimer
;
559 #ifdef CONFIG_HAVE_HW_BREAKPOINT
560 struct { /* breakpoint */
561 struct arch_hw_breakpoint info
;
562 struct list_head bp_list
;
564 * Crufty hack to avoid the chicken and egg
565 * problem hw_breakpoint has with context
566 * creation and event initalization.
568 struct task_struct
*bp_target
;
573 local64_t prev_count
;
576 local64_t period_left
;
580 u64 freq_count_stamp
;
585 * hw_perf_event::state flags
587 #define PERF_HES_STOPPED 0x01 /* the counter is stopped */
588 #define PERF_HES_UPTODATE 0x02 /* event->count up-to-date */
589 #define PERF_HES_ARCH 0x04
594 * Common implementation detail of pmu::{start,commit,cancel}_txn
596 #define PERF_EVENT_TXN 0x1
599 * struct pmu - generic performance monitoring unit
602 struct list_head entry
;
608 int * __percpu pmu_disable_count
;
609 struct perf_cpu_context
* __percpu pmu_cpu_context
;
613 * Fully disable/enable this PMU, can be used to protect from the PMI
614 * as well as for lazy/batch writing of the MSRs.
616 void (*pmu_enable
) (struct pmu
*pmu
); /* optional */
617 void (*pmu_disable
) (struct pmu
*pmu
); /* optional */
620 * Try and initialize the event for this PMU.
621 * Should return -ENOENT when the @event doesn't match this PMU.
623 int (*event_init
) (struct perf_event
*event
);
625 #define PERF_EF_START 0x01 /* start the counter when adding */
626 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
627 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
630 * Adds/Removes a counter to/from the PMU, can be done inside
631 * a transaction, see the ->*_txn() methods.
633 int (*add
) (struct perf_event
*event
, int flags
);
634 void (*del
) (struct perf_event
*event
, int flags
);
637 * Starts/Stops a counter present on the PMU. The PMI handler
638 * should stop the counter when perf_event_overflow() returns
639 * !0. ->start() will be used to continue.
641 void (*start
) (struct perf_event
*event
, int flags
);
642 void (*stop
) (struct perf_event
*event
, int flags
);
645 * Updates the counter value of the event.
647 void (*read
) (struct perf_event
*event
);
650 * Group events scheduling is treated as a transaction, add
651 * group events as a whole and perform one schedulability test.
652 * If the test fails, roll back the whole group
654 * Start the transaction, after this ->add() doesn't need to
655 * do schedulability tests.
657 void (*start_txn
) (struct pmu
*pmu
); /* optional */
659 * If ->start_txn() disabled the ->add() schedulability test
660 * then ->commit_txn() is required to perform one. On success
661 * the transaction is closed. On error the transaction is kept
662 * open until ->cancel_txn() is called.
664 int (*commit_txn
) (struct pmu
*pmu
); /* optional */
666 * Will cancel the transaction, assumes ->del() is called
667 * for each successful ->add() during the transaction.
669 void (*cancel_txn
) (struct pmu
*pmu
); /* optional */
673 * enum perf_event_active_state - the states of a event
675 enum perf_event_active_state
{
676 PERF_EVENT_STATE_ERROR
= -2,
677 PERF_EVENT_STATE_OFF
= -1,
678 PERF_EVENT_STATE_INACTIVE
= 0,
679 PERF_EVENT_STATE_ACTIVE
= 1,
683 struct perf_sample_data
;
685 typedef void (*perf_overflow_handler_t
)(struct perf_event
*,
686 struct perf_sample_data
*,
687 struct pt_regs
*regs
);
689 enum perf_group_flag
{
690 PERF_GROUP_SOFTWARE
= 0x1,
693 #define SWEVENT_HLIST_BITS 8
694 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
696 struct swevent_hlist
{
697 struct hlist_head heads
[SWEVENT_HLIST_SIZE
];
698 struct rcu_head rcu_head
;
701 #define PERF_ATTACH_CONTEXT 0x01
702 #define PERF_ATTACH_GROUP 0x02
703 #define PERF_ATTACH_TASK 0x04
705 #ifdef CONFIG_CGROUP_PERF
707 * perf_cgroup_info keeps track of time_enabled for a cgroup.
708 * This is a per-cpu dynamically allocated data structure.
710 struct perf_cgroup_info
{
716 struct cgroup_subsys_state css
;
717 struct perf_cgroup_info
*info
; /* timing info, one per cpu */
724 * struct perf_event - performance event kernel representation:
727 #ifdef CONFIG_PERF_EVENTS
728 struct list_head group_entry
;
729 struct list_head event_entry
;
730 struct list_head sibling_list
;
731 struct hlist_node hlist_entry
;
734 struct perf_event
*group_leader
;
737 enum perf_event_active_state state
;
738 unsigned int attach_state
;
740 atomic64_t child_count
;
743 * These are the total time in nanoseconds that the event
744 * has been enabled (i.e. eligible to run, and the task has
745 * been scheduled in, if this is a per-task event)
746 * and running (scheduled onto the CPU), respectively.
748 * They are computed from tstamp_enabled, tstamp_running and
749 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
751 u64 total_time_enabled
;
752 u64 total_time_running
;
755 * These are timestamps used for computing total_time_enabled
756 * and total_time_running when the event is in INACTIVE or
757 * ACTIVE state, measured in nanoseconds from an arbitrary point
759 * tstamp_enabled: the notional time when the event was enabled
760 * tstamp_running: the notional time when the event was scheduled on
761 * tstamp_stopped: in INACTIVE state, the notional time when the
762 * event was scheduled off.
769 * timestamp shadows the actual context timing but it can
770 * be safely used in NMI interrupt context. It reflects the
771 * context time as it was when the event was last scheduled in.
773 * ctx_time already accounts for ctx->timestamp. Therefore to
774 * compute ctx_time for a sample, simply add perf_clock().
778 struct perf_event_attr attr
;
782 struct hw_perf_event hw
;
784 struct perf_event_context
*ctx
;
788 * These accumulate total time (in nanoseconds) that children
789 * events have been enabled and running, respectively.
791 atomic64_t child_total_time_enabled
;
792 atomic64_t child_total_time_running
;
795 * Protect attach/detach and child_list:
797 struct mutex child_mutex
;
798 struct list_head child_list
;
799 struct perf_event
*parent
;
804 struct list_head owner_entry
;
805 struct task_struct
*owner
;
808 struct mutex mmap_mutex
;
811 struct user_struct
*mmap_user
;
812 struct ring_buffer
*rb
;
815 wait_queue_head_t waitq
;
816 struct fasync_struct
*fasync
;
818 /* delayed work for NMIs and such */
822 struct irq_work pending
;
824 atomic_t event_limit
;
826 void (*destroy
)(struct perf_event
*);
827 struct rcu_head rcu_head
;
829 struct pid_namespace
*ns
;
832 perf_overflow_handler_t overflow_handler
;
834 #ifdef CONFIG_EVENT_TRACING
835 struct ftrace_event_call
*tp_event
;
836 struct event_filter
*filter
;
839 #ifdef CONFIG_CGROUP_PERF
840 struct perf_cgroup
*cgrp
; /* cgroup event is attach to */
841 int cgrp_defer_enabled
;
844 #endif /* CONFIG_PERF_EVENTS */
847 enum perf_event_context_type
{
853 * struct perf_event_context - event context structure
855 * Used as a container for task events and CPU events as well:
857 struct perf_event_context
{
859 enum perf_event_context_type type
;
861 * Protect the states of the events in the list,
862 * nr_active, and the list:
866 * Protect the list of events. Locking either mutex or lock
867 * is sufficient to ensure the list doesn't change; to change
868 * the list you need to lock both the mutex and the spinlock.
872 struct list_head pinned_groups
;
873 struct list_head flexible_groups
;
874 struct list_head event_list
;
881 struct task_struct
*task
;
884 * Context clock, runs when context enabled.
890 * These fields let us detect when two contexts have both
891 * been cloned (inherited) from a common ancestor.
893 struct perf_event_context
*parent_ctx
;
897 int nr_cgroups
; /* cgroup events present */
898 struct rcu_head rcu_head
;
902 * Number of contexts where an event can trigger:
903 * task, softirq, hardirq, nmi.
905 #define PERF_NR_CONTEXTS 4
908 * struct perf_event_cpu_context - per cpu event context structure
910 struct perf_cpu_context
{
911 struct perf_event_context ctx
;
912 struct perf_event_context
*task_ctx
;
915 struct list_head rotation_list
;
916 int jiffies_interval
;
917 struct pmu
*active_pmu
;
918 struct perf_cgroup
*cgrp
;
921 struct perf_output_handle
{
922 struct perf_event
*event
;
923 struct ring_buffer
*rb
;
924 unsigned long wakeup
;
930 #ifdef CONFIG_PERF_EVENTS
932 extern int perf_pmu_register(struct pmu
*pmu
, char *name
, int type
);
933 extern void perf_pmu_unregister(struct pmu
*pmu
);
935 extern int perf_num_counters(void);
936 extern const char *perf_pmu_name(void);
937 extern void __perf_event_task_sched_in(struct task_struct
*task
);
938 extern void __perf_event_task_sched_out(struct task_struct
*task
, struct task_struct
*next
);
939 extern int perf_event_init_task(struct task_struct
*child
);
940 extern void perf_event_exit_task(struct task_struct
*child
);
941 extern void perf_event_free_task(struct task_struct
*task
);
942 extern void perf_event_delayed_put(struct task_struct
*task
);
943 extern void perf_event_print_debug(void);
944 extern void perf_pmu_disable(struct pmu
*pmu
);
945 extern void perf_pmu_enable(struct pmu
*pmu
);
946 extern int perf_event_task_disable(void);
947 extern int perf_event_task_enable(void);
948 extern void perf_event_update_userpage(struct perf_event
*event
);
949 extern int perf_event_release_kernel(struct perf_event
*event
);
950 extern struct perf_event
*
951 perf_event_create_kernel_counter(struct perf_event_attr
*attr
,
953 struct task_struct
*task
,
954 perf_overflow_handler_t callback
);
955 extern u64
perf_event_read_value(struct perf_event
*event
,
956 u64
*enabled
, u64
*running
);
958 struct perf_sample_data
{
975 struct perf_callchain_entry
*callchain
;
976 struct perf_raw_record
*raw
;
979 static inline void perf_sample_data_init(struct perf_sample_data
*data
, u64 addr
)
985 extern void perf_output_sample(struct perf_output_handle
*handle
,
986 struct perf_event_header
*header
,
987 struct perf_sample_data
*data
,
988 struct perf_event
*event
);
989 extern void perf_prepare_sample(struct perf_event_header
*header
,
990 struct perf_sample_data
*data
,
991 struct perf_event
*event
,
992 struct pt_regs
*regs
);
994 extern int perf_event_overflow(struct perf_event
*event
,
995 struct perf_sample_data
*data
,
996 struct pt_regs
*regs
);
998 static inline bool is_sampling_event(struct perf_event
*event
)
1000 return event
->attr
.sample_period
!= 0;
1004 * Return 1 for a software event, 0 for a hardware event
1006 static inline int is_software_event(struct perf_event
*event
)
1008 return event
->pmu
->task_ctx_nr
== perf_sw_context
;
1011 extern struct jump_label_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
1013 extern void __perf_sw_event(u32
, u64
, struct pt_regs
*, u64
);
1015 #ifndef perf_arch_fetch_caller_regs
1016 static inline void perf_arch_fetch_caller_regs(struct pt_regs
*regs
, unsigned long ip
) { }
1020 * Take a snapshot of the regs. Skip ip and frame pointer to
1021 * the nth caller. We only need a few of the regs:
1022 * - ip for PERF_SAMPLE_IP
1023 * - cs for user_mode() tests
1024 * - bp for callchains
1025 * - eflags, for future purposes, just in case
1027 static inline void perf_fetch_caller_regs(struct pt_regs
*regs
)
1029 memset(regs
, 0, sizeof(*regs
));
1031 perf_arch_fetch_caller_regs(regs
, CALLER_ADDR0
);
1034 static __always_inline
void
1035 perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
1037 struct pt_regs hot_regs
;
1039 if (static_branch(&perf_swevent_enabled
[event_id
])) {
1041 perf_fetch_caller_regs(&hot_regs
);
1044 __perf_sw_event(event_id
, nr
, regs
, addr
);
1048 extern struct jump_label_key perf_sched_events
;
1050 static inline void perf_event_task_sched_in(struct task_struct
*task
)
1052 if (static_branch(&perf_sched_events
))
1053 __perf_event_task_sched_in(task
);
1056 static inline void perf_event_task_sched_out(struct task_struct
*task
, struct task_struct
*next
)
1058 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES
, 1, NULL
, 0);
1060 __perf_event_task_sched_out(task
, next
);
1063 extern void perf_event_mmap(struct vm_area_struct
*vma
);
1064 extern struct perf_guest_info_callbacks
*perf_guest_cbs
;
1065 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*callbacks
);
1066 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*callbacks
);
1068 extern void perf_event_comm(struct task_struct
*tsk
);
1069 extern void perf_event_fork(struct task_struct
*tsk
);
1072 DECLARE_PER_CPU(struct perf_callchain_entry
, perf_callchain_entry
);
1074 extern void perf_callchain_user(struct perf_callchain_entry
*entry
, struct pt_regs
*regs
);
1075 extern void perf_callchain_kernel(struct perf_callchain_entry
*entry
, struct pt_regs
*regs
);
1077 static inline void perf_callchain_store(struct perf_callchain_entry
*entry
, u64 ip
)
1079 if (entry
->nr
< PERF_MAX_STACK_DEPTH
)
1080 entry
->ip
[entry
->nr
++] = ip
;
1083 extern int sysctl_perf_event_paranoid
;
1084 extern int sysctl_perf_event_mlock
;
1085 extern int sysctl_perf_event_sample_rate
;
1087 extern int perf_proc_update_handler(struct ctl_table
*table
, int write
,
1088 void __user
*buffer
, size_t *lenp
,
1091 static inline bool perf_paranoid_tracepoint_raw(void)
1093 return sysctl_perf_event_paranoid
> -1;
1096 static inline bool perf_paranoid_cpu(void)
1098 return sysctl_perf_event_paranoid
> 0;
1101 static inline bool perf_paranoid_kernel(void)
1103 return sysctl_perf_event_paranoid
> 1;
1106 extern void perf_event_init(void);
1107 extern void perf_tp_event(u64 addr
, u64 count
, void *record
,
1108 int entry_size
, struct pt_regs
*regs
,
1109 struct hlist_head
*head
, int rctx
);
1110 extern void perf_bp_event(struct perf_event
*event
, void *data
);
1112 #ifndef perf_misc_flags
1113 # define perf_misc_flags(regs) \
1114 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1115 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1118 extern int perf_output_begin(struct perf_output_handle
*handle
,
1119 struct perf_event
*event
, unsigned int size
);
1120 extern void perf_output_end(struct perf_output_handle
*handle
);
1121 extern void perf_output_copy(struct perf_output_handle
*handle
,
1122 const void *buf
, unsigned int len
);
1123 extern int perf_swevent_get_recursion_context(void);
1124 extern void perf_swevent_put_recursion_context(int rctx
);
1125 extern void perf_event_enable(struct perf_event
*event
);
1126 extern void perf_event_disable(struct perf_event
*event
);
1127 extern void perf_event_task_tick(void);
1130 perf_event_task_sched_in(struct task_struct
*task
) { }
1132 perf_event_task_sched_out(struct task_struct
*task
,
1133 struct task_struct
*next
) { }
1134 static inline int perf_event_init_task(struct task_struct
*child
) { return 0; }
1135 static inline void perf_event_exit_task(struct task_struct
*child
) { }
1136 static inline void perf_event_free_task(struct task_struct
*task
) { }
1137 static inline void perf_event_delayed_put(struct task_struct
*task
) { }
1138 static inline void perf_event_print_debug(void) { }
1139 static inline int perf_event_task_disable(void) { return -EINVAL
; }
1140 static inline int perf_event_task_enable(void) { return -EINVAL
; }
1143 perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
) { }
1145 perf_bp_event(struct perf_event
*event
, void *data
) { }
1147 static inline int perf_register_guest_info_callbacks
1148 (struct perf_guest_info_callbacks
*callbacks
) { return 0; }
1149 static inline int perf_unregister_guest_info_callbacks
1150 (struct perf_guest_info_callbacks
*callbacks
) { return 0; }
1152 static inline void perf_event_mmap(struct vm_area_struct
*vma
) { }
1153 static inline void perf_event_comm(struct task_struct
*tsk
) { }
1154 static inline void perf_event_fork(struct task_struct
*tsk
) { }
1155 static inline void perf_event_init(void) { }
1156 static inline int perf_swevent_get_recursion_context(void) { return -1; }
1157 static inline void perf_swevent_put_recursion_context(int rctx
) { }
1158 static inline void perf_event_enable(struct perf_event
*event
) { }
1159 static inline void perf_event_disable(struct perf_event
*event
) { }
1160 static inline void perf_event_task_tick(void) { }
1163 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1166 * This has to have a higher priority than migration_notifier in sched.c.
1168 #define perf_cpu_notifier(fn) \
1170 static struct notifier_block fn##_nb __cpuinitdata = \
1171 { .notifier_call = fn, .priority = CPU_PRI_PERF }; \
1172 fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE, \
1173 (void *)(unsigned long)smp_processor_id()); \
1174 fn(&fn##_nb, (unsigned long)CPU_STARTING, \
1175 (void *)(unsigned long)smp_processor_id()); \
1176 fn(&fn##_nb, (unsigned long)CPU_ONLINE, \
1177 (void *)(unsigned long)smp_processor_id()); \
1178 register_cpu_notifier(&fn##_nb); \
1181 #endif /* __KERNEL__ */
1182 #endif /* _LINUX_PERF_EVENT_H */