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,
57 PERF_COUNT_HW_REF_CPU_CYCLES
= 9,
59 PERF_COUNT_HW_MAX
, /* non-ABI */
63 * Generalized hardware cache events:
65 * { L1-D, L1-I, LLC, ITLB, DTLB, BPU, NODE } x
66 * { read, write, prefetch } x
67 * { accesses, misses }
69 enum perf_hw_cache_id
{
70 PERF_COUNT_HW_CACHE_L1D
= 0,
71 PERF_COUNT_HW_CACHE_L1I
= 1,
72 PERF_COUNT_HW_CACHE_LL
= 2,
73 PERF_COUNT_HW_CACHE_DTLB
= 3,
74 PERF_COUNT_HW_CACHE_ITLB
= 4,
75 PERF_COUNT_HW_CACHE_BPU
= 5,
76 PERF_COUNT_HW_CACHE_NODE
= 6,
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_TOTAL_TIME_ENABLED
143 * { u64 time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
144 * { u64 id; } && PERF_FORMAT_ID
145 * } && !PERF_FORMAT_GROUP
148 * { u64 time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
149 * { u64 time_running; } && PERF_FORMAT_TOTAL_TIME_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 */
213 * 0 - SAMPLE_IP can have arbitrary skid
214 * 1 - SAMPLE_IP must have constant skid
215 * 2 - SAMPLE_IP requested to have 0 skid
216 * 3 - SAMPLE_IP must have 0 skid
218 * See also PERF_RECORD_MISC_EXACT_IP
220 precise_ip
: 2, /* skid constraint */
221 mmap_data
: 1, /* non-exec mmap data */
222 sample_id_all
: 1, /* sample_type all events */
224 exclude_host
: 1, /* don't count in host */
225 exclude_guest
: 1, /* don't count in guest */
230 __u32 wakeup_events
; /* wakeup every n events */
231 __u32 wakeup_watermark
; /* bytes before wakeup */
237 __u64 config1
; /* extension of config */
241 __u64 config2
; /* extension of config1 */
246 * Ioctls that can be done on a perf event fd:
248 #define PERF_EVENT_IOC_ENABLE _IO ('$', 0)
249 #define PERF_EVENT_IOC_DISABLE _IO ('$', 1)
250 #define PERF_EVENT_IOC_REFRESH _IO ('$', 2)
251 #define PERF_EVENT_IOC_RESET _IO ('$', 3)
252 #define PERF_EVENT_IOC_PERIOD _IOW('$', 4, __u64)
253 #define PERF_EVENT_IOC_SET_OUTPUT _IO ('$', 5)
254 #define PERF_EVENT_IOC_SET_FILTER _IOW('$', 6, char *)
256 enum perf_event_ioc_flags
{
257 PERF_IOC_FLAG_GROUP
= 1U << 0,
261 * Structure of the page that can be mapped via mmap
263 struct perf_event_mmap_page
{
264 __u32 version
; /* version number of this structure */
265 __u32 compat_version
; /* lowest version this is compat with */
268 * Bits needed to read the hw events in user-space.
278 * count = pmc_read(pc->index - 1);
279 * count += pc->offset;
284 * } while (pc->lock != seq);
286 * NOTE: for obvious reason this only works on self-monitoring
289 __u32 lock
; /* seqlock for synchronization */
290 __u32 index
; /* hardware event identifier */
291 __s64 offset
; /* add to hardware event value */
292 __u64 time_enabled
; /* time event active */
293 __u64 time_running
; /* time event on cpu */
296 * Hole for extension of the self monitor capabilities
299 __u64 __reserved
[123]; /* align to 1k */
302 * Control data for the mmap() data buffer.
304 * User-space reading the @data_head value should issue an rmb(), on
305 * SMP capable platforms, after reading this value -- see
306 * perf_event_wakeup().
308 * When the mapping is PROT_WRITE the @data_tail value should be
309 * written by userspace to reflect the last read data. In this case
310 * the kernel will not over-write unread data.
312 __u64 data_head
; /* head in the data section */
313 __u64 data_tail
; /* user-space written tail */
316 #define PERF_RECORD_MISC_CPUMODE_MASK (7 << 0)
317 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN (0 << 0)
318 #define PERF_RECORD_MISC_KERNEL (1 << 0)
319 #define PERF_RECORD_MISC_USER (2 << 0)
320 #define PERF_RECORD_MISC_HYPERVISOR (3 << 0)
321 #define PERF_RECORD_MISC_GUEST_KERNEL (4 << 0)
322 #define PERF_RECORD_MISC_GUEST_USER (5 << 0)
325 * Indicates that the content of PERF_SAMPLE_IP points to
326 * the actual instruction that triggered the event. See also
327 * perf_event_attr::precise_ip.
329 #define PERF_RECORD_MISC_EXACT_IP (1 << 14)
331 * Reserve the last bit to indicate some extended misc field
333 #define PERF_RECORD_MISC_EXT_RESERVED (1 << 15)
335 struct perf_event_header
{
341 enum perf_event_type
{
344 * If perf_event_attr.sample_id_all is set then all event types will
345 * have the sample_type selected fields related to where/when
346 * (identity) an event took place (TID, TIME, ID, CPU, STREAM_ID)
347 * described in PERF_RECORD_SAMPLE below, it will be stashed just after
348 * the perf_event_header and the fields already present for the existing
349 * fields, i.e. at the end of the payload. That way a newer perf.data
350 * file will be supported by older perf tools, with these new optional
351 * fields being ignored.
353 * The MMAP events record the PROT_EXEC mappings so that we can
354 * correlate userspace IPs to code. They have the following structure:
357 * struct perf_event_header header;
366 PERF_RECORD_MMAP
= 1,
370 * struct perf_event_header header;
375 PERF_RECORD_LOST
= 2,
379 * struct perf_event_header header;
385 PERF_RECORD_COMM
= 3,
389 * struct perf_event_header header;
395 PERF_RECORD_EXIT
= 4,
399 * struct perf_event_header header;
405 PERF_RECORD_THROTTLE
= 5,
406 PERF_RECORD_UNTHROTTLE
= 6,
410 * struct perf_event_header header;
416 PERF_RECORD_FORK
= 7,
420 * struct perf_event_header header;
423 * struct read_format values;
426 PERF_RECORD_READ
= 8,
430 * struct perf_event_header header;
432 * { u64 ip; } && PERF_SAMPLE_IP
433 * { u32 pid, tid; } && PERF_SAMPLE_TID
434 * { u64 time; } && PERF_SAMPLE_TIME
435 * { u64 addr; } && PERF_SAMPLE_ADDR
436 * { u64 id; } && PERF_SAMPLE_ID
437 * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID
438 * { u32 cpu, res; } && PERF_SAMPLE_CPU
439 * { u64 period; } && PERF_SAMPLE_PERIOD
441 * { struct read_format values; } && PERF_SAMPLE_READ
444 * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN
447 * # The RAW record below is opaque data wrt the ABI
449 * # That is, the ABI doesn't make any promises wrt to
450 * # the stability of its content, it may vary depending
451 * # on event, hardware, kernel version and phase of
454 * # In other words, PERF_SAMPLE_RAW contents are not an ABI.
458 * char data[size];}&& PERF_SAMPLE_RAW
461 PERF_RECORD_SAMPLE
= 9,
463 PERF_RECORD_MAX
, /* non-ABI */
466 enum perf_callchain_context
{
467 PERF_CONTEXT_HV
= (__u64
)-32,
468 PERF_CONTEXT_KERNEL
= (__u64
)-128,
469 PERF_CONTEXT_USER
= (__u64
)-512,
471 PERF_CONTEXT_GUEST
= (__u64
)-2048,
472 PERF_CONTEXT_GUEST_KERNEL
= (__u64
)-2176,
473 PERF_CONTEXT_GUEST_USER
= (__u64
)-2560,
475 PERF_CONTEXT_MAX
= (__u64
)-4095,
478 #define PERF_FLAG_FD_NO_GROUP (1U << 0)
479 #define PERF_FLAG_FD_OUTPUT (1U << 1)
480 #define PERF_FLAG_PID_CGROUP (1U << 2) /* pid=cgroup id, per-cpu mode only */
484 * Kernel-internal data types and definitions:
487 #ifdef CONFIG_PERF_EVENTS
488 # include <linux/cgroup.h>
489 # include <asm/perf_event.h>
490 # include <asm/local64.h>
493 struct perf_guest_info_callbacks
{
494 int (*is_in_guest
)(void);
495 int (*is_user_mode
)(void);
496 unsigned long (*get_guest_ip
)(void);
499 #ifdef CONFIG_HAVE_HW_BREAKPOINT
500 #include <asm/hw_breakpoint.h>
503 #include <linux/list.h>
504 #include <linux/mutex.h>
505 #include <linux/rculist.h>
506 #include <linux/rcupdate.h>
507 #include <linux/spinlock.h>
508 #include <linux/hrtimer.h>
509 #include <linux/fs.h>
510 #include <linux/pid_namespace.h>
511 #include <linux/workqueue.h>
512 #include <linux/ftrace.h>
513 #include <linux/cpu.h>
514 #include <linux/irq_work.h>
515 #include <linux/jump_label.h>
516 #include <linux/atomic.h>
517 #include <asm/local.h>
519 #define PERF_MAX_STACK_DEPTH 255
521 struct perf_callchain_entry
{
523 __u64 ip
[PERF_MAX_STACK_DEPTH
];
526 struct perf_raw_record
{
531 struct perf_branch_entry
{
537 struct perf_branch_stack
{
539 struct perf_branch_entry entries
[0];
545 * extra PMU register associated with an event
547 struct hw_perf_event_extra
{
548 u64 config
; /* register value */
549 unsigned int reg
; /* register address or index */
550 int alloc
; /* extra register already allocated */
551 int idx
; /* index in shared_regs->regs[] */
555 * struct hw_perf_event - performance event hardware details:
557 struct hw_perf_event
{
558 #ifdef CONFIG_PERF_EVENTS
560 struct { /* hardware */
563 unsigned long config_base
;
564 unsigned long event_base
;
567 struct hw_perf_event_extra extra_reg
;
569 struct { /* software */
570 struct hrtimer hrtimer
;
572 #ifdef CONFIG_HAVE_HW_BREAKPOINT
573 struct { /* breakpoint */
574 struct arch_hw_breakpoint info
;
575 struct list_head bp_list
;
577 * Crufty hack to avoid the chicken and egg
578 * problem hw_breakpoint has with context
579 * creation and event initalization.
581 struct task_struct
*bp_target
;
586 local64_t prev_count
;
589 local64_t period_left
;
593 u64 freq_count_stamp
;
598 * hw_perf_event::state flags
600 #define PERF_HES_STOPPED 0x01 /* the counter is stopped */
601 #define PERF_HES_UPTODATE 0x02 /* event->count up-to-date */
602 #define PERF_HES_ARCH 0x04
607 * Common implementation detail of pmu::{start,commit,cancel}_txn
609 #define PERF_EVENT_TXN 0x1
612 * struct pmu - generic performance monitoring unit
615 struct list_head entry
;
621 int * __percpu pmu_disable_count
;
622 struct perf_cpu_context
* __percpu pmu_cpu_context
;
626 * Fully disable/enable this PMU, can be used to protect from the PMI
627 * as well as for lazy/batch writing of the MSRs.
629 void (*pmu_enable
) (struct pmu
*pmu
); /* optional */
630 void (*pmu_disable
) (struct pmu
*pmu
); /* optional */
633 * Try and initialize the event for this PMU.
634 * Should return -ENOENT when the @event doesn't match this PMU.
636 int (*event_init
) (struct perf_event
*event
);
638 #define PERF_EF_START 0x01 /* start the counter when adding */
639 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
640 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
643 * Adds/Removes a counter to/from the PMU, can be done inside
644 * a transaction, see the ->*_txn() methods.
646 int (*add
) (struct perf_event
*event
, int flags
);
647 void (*del
) (struct perf_event
*event
, int flags
);
650 * Starts/Stops a counter present on the PMU. The PMI handler
651 * should stop the counter when perf_event_overflow() returns
652 * !0. ->start() will be used to continue.
654 void (*start
) (struct perf_event
*event
, int flags
);
655 void (*stop
) (struct perf_event
*event
, int flags
);
658 * Updates the counter value of the event.
660 void (*read
) (struct perf_event
*event
);
663 * Group events scheduling is treated as a transaction, add
664 * group events as a whole and perform one schedulability test.
665 * If the test fails, roll back the whole group
667 * Start the transaction, after this ->add() doesn't need to
668 * do schedulability tests.
670 void (*start_txn
) (struct pmu
*pmu
); /* optional */
672 * If ->start_txn() disabled the ->add() schedulability test
673 * then ->commit_txn() is required to perform one. On success
674 * the transaction is closed. On error the transaction is kept
675 * open until ->cancel_txn() is called.
677 int (*commit_txn
) (struct pmu
*pmu
); /* optional */
679 * Will cancel the transaction, assumes ->del() is called
680 * for each successful ->add() during the transaction.
682 void (*cancel_txn
) (struct pmu
*pmu
); /* optional */
685 * Will return the value for perf_event_mmap_page::index for this event,
686 * if no implementation is provided it will default to: event->hw.idx + 1.
688 int (*event_idx
) (struct perf_event
*event
); /*optional */
692 * enum perf_event_active_state - the states of a event
694 enum perf_event_active_state
{
695 PERF_EVENT_STATE_ERROR
= -2,
696 PERF_EVENT_STATE_OFF
= -1,
697 PERF_EVENT_STATE_INACTIVE
= 0,
698 PERF_EVENT_STATE_ACTIVE
= 1,
702 struct perf_sample_data
;
704 typedef void (*perf_overflow_handler_t
)(struct perf_event
*,
705 struct perf_sample_data
*,
706 struct pt_regs
*regs
);
708 enum perf_group_flag
{
709 PERF_GROUP_SOFTWARE
= 0x1,
712 #define SWEVENT_HLIST_BITS 8
713 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
715 struct swevent_hlist
{
716 struct hlist_head heads
[SWEVENT_HLIST_SIZE
];
717 struct rcu_head rcu_head
;
720 #define PERF_ATTACH_CONTEXT 0x01
721 #define PERF_ATTACH_GROUP 0x02
722 #define PERF_ATTACH_TASK 0x04
724 #ifdef CONFIG_CGROUP_PERF
726 * perf_cgroup_info keeps track of time_enabled for a cgroup.
727 * This is a per-cpu dynamically allocated data structure.
729 struct perf_cgroup_info
{
735 struct cgroup_subsys_state css
;
736 struct perf_cgroup_info
*info
; /* timing info, one per cpu */
743 * struct perf_event - performance event kernel representation:
746 #ifdef CONFIG_PERF_EVENTS
747 struct list_head group_entry
;
748 struct list_head event_entry
;
749 struct list_head sibling_list
;
750 struct hlist_node hlist_entry
;
753 struct perf_event
*group_leader
;
756 enum perf_event_active_state state
;
757 unsigned int attach_state
;
759 atomic64_t child_count
;
762 * These are the total time in nanoseconds that the event
763 * has been enabled (i.e. eligible to run, and the task has
764 * been scheduled in, if this is a per-task event)
765 * and running (scheduled onto the CPU), respectively.
767 * They are computed from tstamp_enabled, tstamp_running and
768 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
770 u64 total_time_enabled
;
771 u64 total_time_running
;
774 * These are timestamps used for computing total_time_enabled
775 * and total_time_running when the event is in INACTIVE or
776 * ACTIVE state, measured in nanoseconds from an arbitrary point
778 * tstamp_enabled: the notional time when the event was enabled
779 * tstamp_running: the notional time when the event was scheduled on
780 * tstamp_stopped: in INACTIVE state, the notional time when the
781 * event was scheduled off.
788 * timestamp shadows the actual context timing but it can
789 * be safely used in NMI interrupt context. It reflects the
790 * context time as it was when the event was last scheduled in.
792 * ctx_time already accounts for ctx->timestamp. Therefore to
793 * compute ctx_time for a sample, simply add perf_clock().
797 struct perf_event_attr attr
;
801 struct hw_perf_event hw
;
803 struct perf_event_context
*ctx
;
807 * These accumulate total time (in nanoseconds) that children
808 * events have been enabled and running, respectively.
810 atomic64_t child_total_time_enabled
;
811 atomic64_t child_total_time_running
;
814 * Protect attach/detach and child_list:
816 struct mutex child_mutex
;
817 struct list_head child_list
;
818 struct perf_event
*parent
;
823 struct list_head owner_entry
;
824 struct task_struct
*owner
;
827 struct mutex mmap_mutex
;
830 struct user_struct
*mmap_user
;
831 struct ring_buffer
*rb
;
832 struct list_head rb_entry
;
835 wait_queue_head_t waitq
;
836 struct fasync_struct
*fasync
;
838 /* delayed work for NMIs and such */
842 struct irq_work pending
;
844 atomic_t event_limit
;
846 void (*destroy
)(struct perf_event
*);
847 struct rcu_head rcu_head
;
849 struct pid_namespace
*ns
;
852 perf_overflow_handler_t overflow_handler
;
853 void *overflow_handler_context
;
855 #ifdef CONFIG_EVENT_TRACING
856 struct ftrace_event_call
*tp_event
;
857 struct event_filter
*filter
;
860 #ifdef CONFIG_CGROUP_PERF
861 struct perf_cgroup
*cgrp
; /* cgroup event is attach to */
862 int cgrp_defer_enabled
;
865 #endif /* CONFIG_PERF_EVENTS */
868 enum perf_event_context_type
{
874 * struct perf_event_context - event context structure
876 * Used as a container for task events and CPU events as well:
878 struct perf_event_context
{
880 enum perf_event_context_type type
;
882 * Protect the states of the events in the list,
883 * nr_active, and the list:
887 * Protect the list of events. Locking either mutex or lock
888 * is sufficient to ensure the list doesn't change; to change
889 * the list you need to lock both the mutex and the spinlock.
893 struct list_head pinned_groups
;
894 struct list_head flexible_groups
;
895 struct list_head event_list
;
903 struct task_struct
*task
;
906 * Context clock, runs when context enabled.
912 * These fields let us detect when two contexts have both
913 * been cloned (inherited) from a common ancestor.
915 struct perf_event_context
*parent_ctx
;
919 int nr_cgroups
; /* cgroup events present */
920 struct rcu_head rcu_head
;
924 * Number of contexts where an event can trigger:
925 * task, softirq, hardirq, nmi.
927 #define PERF_NR_CONTEXTS 4
930 * struct perf_event_cpu_context - per cpu event context structure
932 struct perf_cpu_context
{
933 struct perf_event_context ctx
;
934 struct perf_event_context
*task_ctx
;
937 struct list_head rotation_list
;
938 int jiffies_interval
;
939 struct pmu
*active_pmu
;
940 struct perf_cgroup
*cgrp
;
943 struct perf_output_handle
{
944 struct perf_event
*event
;
945 struct ring_buffer
*rb
;
946 unsigned long wakeup
;
952 #ifdef CONFIG_PERF_EVENTS
954 extern int perf_pmu_register(struct pmu
*pmu
, char *name
, int type
);
955 extern void perf_pmu_unregister(struct pmu
*pmu
);
957 extern int perf_num_counters(void);
958 extern const char *perf_pmu_name(void);
959 extern void __perf_event_task_sched_in(struct task_struct
*prev
,
960 struct task_struct
*task
);
961 extern void __perf_event_task_sched_out(struct task_struct
*prev
,
962 struct task_struct
*next
);
963 extern int perf_event_init_task(struct task_struct
*child
);
964 extern void perf_event_exit_task(struct task_struct
*child
);
965 extern void perf_event_free_task(struct task_struct
*task
);
966 extern void perf_event_delayed_put(struct task_struct
*task
);
967 extern void perf_event_print_debug(void);
968 extern void perf_pmu_disable(struct pmu
*pmu
);
969 extern void perf_pmu_enable(struct pmu
*pmu
);
970 extern int perf_event_task_disable(void);
971 extern int perf_event_task_enable(void);
972 extern int perf_event_refresh(struct perf_event
*event
, int refresh
);
973 extern void perf_event_update_userpage(struct perf_event
*event
);
974 extern int perf_event_release_kernel(struct perf_event
*event
);
975 extern struct perf_event
*
976 perf_event_create_kernel_counter(struct perf_event_attr
*attr
,
978 struct task_struct
*task
,
979 perf_overflow_handler_t callback
,
981 extern u64
perf_event_read_value(struct perf_event
*event
,
982 u64
*enabled
, u64
*running
);
984 struct perf_sample_data
{
1001 struct perf_callchain_entry
*callchain
;
1002 struct perf_raw_record
*raw
;
1005 static inline void perf_sample_data_init(struct perf_sample_data
*data
, u64 addr
)
1011 extern void perf_output_sample(struct perf_output_handle
*handle
,
1012 struct perf_event_header
*header
,
1013 struct perf_sample_data
*data
,
1014 struct perf_event
*event
);
1015 extern void perf_prepare_sample(struct perf_event_header
*header
,
1016 struct perf_sample_data
*data
,
1017 struct perf_event
*event
,
1018 struct pt_regs
*regs
);
1020 extern int perf_event_overflow(struct perf_event
*event
,
1021 struct perf_sample_data
*data
,
1022 struct pt_regs
*regs
);
1024 static inline bool is_sampling_event(struct perf_event
*event
)
1026 return event
->attr
.sample_period
!= 0;
1030 * Return 1 for a software event, 0 for a hardware event
1032 static inline int is_software_event(struct perf_event
*event
)
1034 return event
->pmu
->task_ctx_nr
== perf_sw_context
;
1037 extern struct jump_label_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
1039 extern void __perf_sw_event(u32
, u64
, struct pt_regs
*, u64
);
1041 #ifndef perf_arch_fetch_caller_regs
1042 static inline void perf_arch_fetch_caller_regs(struct pt_regs
*regs
, unsigned long ip
) { }
1046 * Take a snapshot of the regs. Skip ip and frame pointer to
1047 * the nth caller. We only need a few of the regs:
1048 * - ip for PERF_SAMPLE_IP
1049 * - cs for user_mode() tests
1050 * - bp for callchains
1051 * - eflags, for future purposes, just in case
1053 static inline void perf_fetch_caller_regs(struct pt_regs
*regs
)
1055 memset(regs
, 0, sizeof(*regs
));
1057 perf_arch_fetch_caller_regs(regs
, CALLER_ADDR0
);
1060 static __always_inline
void
1061 perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
1063 struct pt_regs hot_regs
;
1065 if (static_branch(&perf_swevent_enabled
[event_id
])) {
1067 perf_fetch_caller_regs(&hot_regs
);
1070 __perf_sw_event(event_id
, nr
, regs
, addr
);
1074 extern struct jump_label_key_deferred perf_sched_events
;
1076 static inline void perf_event_task_sched_in(struct task_struct
*prev
,
1077 struct task_struct
*task
)
1079 if (static_branch(&perf_sched_events
.key
))
1080 __perf_event_task_sched_in(prev
, task
);
1083 static inline void perf_event_task_sched_out(struct task_struct
*prev
,
1084 struct task_struct
*next
)
1086 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES
, 1, NULL
, 0);
1088 if (static_branch(&perf_sched_events
.key
))
1089 __perf_event_task_sched_out(prev
, next
);
1092 extern void perf_event_mmap(struct vm_area_struct
*vma
);
1093 extern struct perf_guest_info_callbacks
*perf_guest_cbs
;
1094 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*callbacks
);
1095 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*callbacks
);
1097 extern void perf_event_comm(struct task_struct
*tsk
);
1098 extern void perf_event_fork(struct task_struct
*tsk
);
1101 DECLARE_PER_CPU(struct perf_callchain_entry
, perf_callchain_entry
);
1103 extern void perf_callchain_user(struct perf_callchain_entry
*entry
, struct pt_regs
*regs
);
1104 extern void perf_callchain_kernel(struct perf_callchain_entry
*entry
, struct pt_regs
*regs
);
1106 static inline void perf_callchain_store(struct perf_callchain_entry
*entry
, u64 ip
)
1108 if (entry
->nr
< PERF_MAX_STACK_DEPTH
)
1109 entry
->ip
[entry
->nr
++] = ip
;
1112 extern int sysctl_perf_event_paranoid
;
1113 extern int sysctl_perf_event_mlock
;
1114 extern int sysctl_perf_event_sample_rate
;
1116 extern int perf_proc_update_handler(struct ctl_table
*table
, int write
,
1117 void __user
*buffer
, size_t *lenp
,
1120 static inline bool perf_paranoid_tracepoint_raw(void)
1122 return sysctl_perf_event_paranoid
> -1;
1125 static inline bool perf_paranoid_cpu(void)
1127 return sysctl_perf_event_paranoid
> 0;
1130 static inline bool perf_paranoid_kernel(void)
1132 return sysctl_perf_event_paranoid
> 1;
1135 extern void perf_event_init(void);
1136 extern void perf_tp_event(u64 addr
, u64 count
, void *record
,
1137 int entry_size
, struct pt_regs
*regs
,
1138 struct hlist_head
*head
, int rctx
);
1139 extern void perf_bp_event(struct perf_event
*event
, void *data
);
1141 #ifndef perf_misc_flags
1142 # define perf_misc_flags(regs) \
1143 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1144 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1147 extern int perf_output_begin(struct perf_output_handle
*handle
,
1148 struct perf_event
*event
, unsigned int size
);
1149 extern void perf_output_end(struct perf_output_handle
*handle
);
1150 extern void perf_output_copy(struct perf_output_handle
*handle
,
1151 const void *buf
, unsigned int len
);
1152 extern int perf_swevent_get_recursion_context(void);
1153 extern void perf_swevent_put_recursion_context(int rctx
);
1154 extern void perf_event_enable(struct perf_event
*event
);
1155 extern void perf_event_disable(struct perf_event
*event
);
1156 extern void perf_event_task_tick(void);
1159 perf_event_task_sched_in(struct task_struct
*prev
,
1160 struct task_struct
*task
) { }
1162 perf_event_task_sched_out(struct task_struct
*prev
,
1163 struct task_struct
*next
) { }
1164 static inline int perf_event_init_task(struct task_struct
*child
) { return 0; }
1165 static inline void perf_event_exit_task(struct task_struct
*child
) { }
1166 static inline void perf_event_free_task(struct task_struct
*task
) { }
1167 static inline void perf_event_delayed_put(struct task_struct
*task
) { }
1168 static inline void perf_event_print_debug(void) { }
1169 static inline int perf_event_task_disable(void) { return -EINVAL
; }
1170 static inline int perf_event_task_enable(void) { return -EINVAL
; }
1171 static inline int perf_event_refresh(struct perf_event
*event
, int refresh
)
1177 perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
) { }
1179 perf_bp_event(struct perf_event
*event
, void *data
) { }
1181 static inline int perf_register_guest_info_callbacks
1182 (struct perf_guest_info_callbacks
*callbacks
) { return 0; }
1183 static inline int perf_unregister_guest_info_callbacks
1184 (struct perf_guest_info_callbacks
*callbacks
) { return 0; }
1186 static inline void perf_event_mmap(struct vm_area_struct
*vma
) { }
1187 static inline void perf_event_comm(struct task_struct
*tsk
) { }
1188 static inline void perf_event_fork(struct task_struct
*tsk
) { }
1189 static inline void perf_event_init(void) { }
1190 static inline int perf_swevent_get_recursion_context(void) { return -1; }
1191 static inline void perf_swevent_put_recursion_context(int rctx
) { }
1192 static inline void perf_event_enable(struct perf_event
*event
) { }
1193 static inline void perf_event_disable(struct perf_event
*event
) { }
1194 static inline void perf_event_task_tick(void) { }
1197 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1200 * This has to have a higher priority than migration_notifier in sched.c.
1202 #define perf_cpu_notifier(fn) \
1204 static struct notifier_block fn##_nb __cpuinitdata = \
1205 { .notifier_call = fn, .priority = CPU_PRI_PERF }; \
1206 fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE, \
1207 (void *)(unsigned long)smp_processor_id()); \
1208 fn(&fn##_nb, (unsigned long)CPU_STARTING, \
1209 (void *)(unsigned long)smp_processor_id()); \
1210 fn(&fn##_nb, (unsigned long)CPU_ONLINE, \
1211 (void *)(unsigned long)smp_processor_id()); \
1212 register_cpu_notifier(&fn##_nb); \
1215 #endif /* __KERNEL__ */
1216 #endif /* _LINUX_PERF_EVENT_H */