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,
132 PERF_SAMPLE_BRANCH_STACK = 1U << 11,
134 PERF_SAMPLE_MAX = 1U << 12, /* non-ABI */
138 * values to program into branch_sample_type when PERF_SAMPLE_BRANCH is set
140 * If the user does not pass priv level information via branch_sample_type,
141 * the kernel uses the event's priv level. Branch and event priv levels do
142 * not have to match. Branch priv level is checked for permissions.
144 * The branch types can be combined, however BRANCH_ANY covers all types
145 * of branches and therefore it supersedes all the other types.
147 enum perf_branch_sample_type {
148 PERF_SAMPLE_BRANCH_USER = 1U << 0, /* user branches */
149 PERF_SAMPLE_BRANCH_KERNEL = 1U << 1, /* kernel branches */
150 PERF_SAMPLE_BRANCH_HV = 1U << 2, /* hypervisor branches */
152 PERF_SAMPLE_BRANCH_ANY = 1U << 3, /* any branch types */
153 PERF_SAMPLE_BRANCH_ANY_CALL = 1U << 4, /* any call branch */
154 PERF_SAMPLE_BRANCH_ANY_RETURN = 1U << 5, /* any return branch */
155 PERF_SAMPLE_BRANCH_IND_CALL = 1U << 6, /* indirect calls */
157 PERF_SAMPLE_BRANCH_MAX = 1U << 7, /* non-ABI */
160 #define PERF_SAMPLE_BRANCH_PLM_ALL \
161 (PERF_SAMPLE_BRANCH_USER|\
162 PERF_SAMPLE_BRANCH_KERNEL|\
163 PERF_SAMPLE_BRANCH_HV)
166 * The format of the data returned by read() on a perf event fd,
167 * as specified by attr.read_format:
169 * struct read_format {
171 * { u64 time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
172 * { u64 time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
173 * { u64 id; } && PERF_FORMAT_ID
174 * } && !PERF_FORMAT_GROUP
177 * { u64 time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
178 * { u64 time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
180 * { u64 id; } && PERF_FORMAT_ID
182 * } && PERF_FORMAT_GROUP
185 enum perf_event_read_format {
186 PERF_FORMAT_TOTAL_TIME_ENABLED = 1U << 0,
187 PERF_FORMAT_TOTAL_TIME_RUNNING = 1U << 1,
188 PERF_FORMAT_ID = 1U << 2,
189 PERF_FORMAT_GROUP = 1U << 3,
191 PERF_FORMAT_MAX = 1U << 4, /* non-ABI */
194 #define PERF_ATTR_SIZE_VER0 64 /* sizeof first published struct */
197 * Hardware event_id to monitor via a performance monitoring event:
199 struct perf_event_attr {
202 * Major type: hardware/software/tracepoint/etc.
207 * Size of the attr structure, for fwd/bwd compat.
212 * Type specific configuration information.
224 __u64 disabled : 1, /* off by default */
225 inherit : 1, /* children inherit it */
226 pinned : 1, /* must always be on PMU */
227 exclusive : 1, /* only group on PMU */
228 exclude_user : 1, /* don't count user */
229 exclude_kernel : 1, /* ditto kernel */
230 exclude_hv : 1, /* ditto hypervisor */
231 exclude_idle : 1, /* don't count when idle */
232 mmap : 1, /* include mmap data */
233 comm : 1, /* include comm data */
234 freq : 1, /* use freq, not period */
235 inherit_stat : 1, /* per task counts */
236 enable_on_exec : 1, /* next exec enables */
237 task : 1, /* trace fork/exit */
238 watermark : 1, /* wakeup_watermark */
242 * 0 - SAMPLE_IP can have arbitrary skid
243 * 1 - SAMPLE_IP must have constant skid
244 * 2 - SAMPLE_IP requested to have 0 skid
245 * 3 - SAMPLE_IP must have 0 skid
247 * See also PERF_RECORD_MISC_EXACT_IP
249 precise_ip : 2, /* skid constraint */
250 mmap_data : 1, /* non-exec mmap data */
251 sample_id_all : 1, /* sample_type all events */
253 exclude_host : 1, /* don't count in host */
254 exclude_guest : 1, /* don't count in guest */
259 __u32 wakeup_events; /* wakeup every n events */
260 __u32 wakeup_watermark; /* bytes before wakeup */
266 __u64 config1; /* extension of config */
270 __u64 config2; /* extension of config1 */
272 __u64 branch_sample_type; /* enum branch_sample_type */
276 * Ioctls that can be done on a perf event fd:
278 #define PERF_EVENT_IOC_ENABLE _IO ('$', 0)
279 #define PERF_EVENT_IOC_DISABLE _IO ('$', 1)
280 #define PERF_EVENT_IOC_REFRESH _IO ('$', 2)
281 #define PERF_EVENT_IOC_RESET _IO ('$', 3)
282 #define PERF_EVENT_IOC_PERIOD _IOW('$', 4, __u64)
283 #define PERF_EVENT_IOC_SET_OUTPUT _IO ('$', 5)
284 #define PERF_EVENT_IOC_SET_FILTER _IOW('$', 6, char *)
286 enum perf_event_ioc_flags {
287 PERF_IOC_FLAG_GROUP = 1U << 0,
291 * Structure of the page that can be mapped via mmap
293 struct perf_event_mmap_page {
294 __u32 version; /* version number of this structure */
295 __u32 compat_version; /* lowest version this is compat with */
298 * Bits needed to read the hw events in user-space.
308 * count = pmc_read(pc->index - 1);
309 * count += pc->offset;
314 * } while (pc->lock != seq);
316 * NOTE: for obvious reason this only works on self-monitoring
319 __u32 lock; /* seqlock for synchronization */
320 __u32 index; /* hardware event identifier */
321 __s64 offset; /* add to hardware event value */
322 __u64 time_enabled; /* time event active */
323 __u64 time_running; /* time event on cpu */
324 __u32 time_mult, time_shift;
328 * Hole for extension of the self monitor capabilities
331 __u64 __reserved[121]; /* align to 1k */
334 * Control data for the mmap() data buffer.
336 * User-space reading the @data_head value should issue an rmb(), on
337 * SMP capable platforms, after reading this value -- see
338 * perf_event_wakeup().
340 * When the mapping is PROT_WRITE the @data_tail value should be
341 * written by userspace to reflect the last read data. In this case
342 * the kernel will not over-write unread data.
344 __u64 data_head; /* head in the data section */
345 __u64 data_tail; /* user-space written tail */
348 #define PERF_RECORD_MISC_CPUMODE_MASK (7 << 0)
349 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN (0 << 0)
350 #define PERF_RECORD_MISC_KERNEL (1 << 0)
351 #define PERF_RECORD_MISC_USER (2 << 0)
352 #define PERF_RECORD_MISC_HYPERVISOR (3 << 0)
353 #define PERF_RECORD_MISC_GUEST_KERNEL (4 << 0)
354 #define PERF_RECORD_MISC_GUEST_USER (5 << 0)
357 * Indicates that the content of PERF_SAMPLE_IP points to
358 * the actual instruction that triggered the event. See also
359 * perf_event_attr::precise_ip.
361 #define PERF_RECORD_MISC_EXACT_IP (1 << 14)
363 * Reserve the last bit to indicate some extended misc field
365 #define PERF_RECORD_MISC_EXT_RESERVED (1 << 15)
367 struct perf_event_header {
373 enum perf_event_type {
376 * If perf_event_attr.sample_id_all is set then all event types will
377 * have the sample_type selected fields related to where/when
378 * (identity) an event took place (TID, TIME, ID, CPU, STREAM_ID)
379 * described in PERF_RECORD_SAMPLE below, it will be stashed just after
380 * the perf_event_header and the fields already present for the existing
381 * fields, i.e. at the end of the payload. That way a newer perf.data
382 * file will be supported by older perf tools, with these new optional
383 * fields being ignored.
385 * The MMAP events record the PROT_EXEC mappings so that we can
386 * correlate userspace IPs to code. They have the following structure:
389 * struct perf_event_header header;
398 PERF_RECORD_MMAP = 1,
402 * struct perf_event_header header;
407 PERF_RECORD_LOST = 2,
411 * struct perf_event_header header;
417 PERF_RECORD_COMM = 3,
421 * struct perf_event_header header;
427 PERF_RECORD_EXIT = 4,
431 * struct perf_event_header header;
437 PERF_RECORD_THROTTLE = 5,
438 PERF_RECORD_UNTHROTTLE = 6,
442 * struct perf_event_header header;
448 PERF_RECORD_FORK = 7,
452 * struct perf_event_header header;
455 * struct read_format values;
458 PERF_RECORD_READ = 8,
462 * struct perf_event_header header;
464 * { u64 ip; } && PERF_SAMPLE_IP
465 * { u32 pid, tid; } && PERF_SAMPLE_TID
466 * { u64 time; } && PERF_SAMPLE_TIME
467 * { u64 addr; } && PERF_SAMPLE_ADDR
468 * { u64 id; } && PERF_SAMPLE_ID
469 * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID
470 * { u32 cpu, res; } && PERF_SAMPLE_CPU
471 * { u64 period; } && PERF_SAMPLE_PERIOD
473 * { struct read_format values; } && PERF_SAMPLE_READ
476 * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN
479 * # The RAW record below is opaque data wrt the ABI
481 * # That is, the ABI doesn't make any promises wrt to
482 * # the stability of its content, it may vary depending
483 * # on event, hardware, kernel version and phase of
486 * # In other words, PERF_SAMPLE_RAW contents are not an ABI.
490 * char data[size];}&& PERF_SAMPLE_RAW
492 * { u64 from, to, flags } lbr[nr];} && PERF_SAMPLE_BRANCH_STACK
495 PERF_RECORD_SAMPLE = 9,
497 PERF_RECORD_MAX, /* non-ABI */
500 enum perf_callchain_context {
501 PERF_CONTEXT_HV = (__u64)-32,
502 PERF_CONTEXT_KERNEL = (__u64)-128,
503 PERF_CONTEXT_USER = (__u64)-512,
505 PERF_CONTEXT_GUEST = (__u64)-2048,
506 PERF_CONTEXT_GUEST_KERNEL = (__u64)-2176,
507 PERF_CONTEXT_GUEST_USER = (__u64)-2560,
509 PERF_CONTEXT_MAX = (__u64)-4095,
512 #define PERF_FLAG_FD_NO_GROUP (1U << 0)
513 #define PERF_FLAG_FD_OUTPUT (1U << 1)
514 #define PERF_FLAG_PID_CGROUP (1U << 2) /* pid=cgroup id, per-cpu mode only */
518 * Kernel-internal data types and definitions:
521 #ifdef CONFIG_PERF_EVENTS
522 # include <linux/cgroup.h>
523 # include <asm/perf_event.h>
524 # include <asm/local64.h>
527 struct perf_guest_info_callbacks {
528 int (*is_in_guest)(void);
529 int (*is_user_mode)(void);
530 unsigned long (*get_guest_ip)(void);
533 #ifdef CONFIG_HAVE_HW_BREAKPOINT
534 #include <asm/hw_breakpoint.h>
537 #include <linux/list.h>
538 #include <linux/mutex.h>
539 #include <linux/rculist.h>
540 #include <linux/rcupdate.h>
541 #include <linux/spinlock.h>
542 #include <linux/hrtimer.h>
543 #include <linux/fs.h>
544 #include <linux/pid_namespace.h>
545 #include <linux/workqueue.h>
546 #include <linux/ftrace.h>
547 #include <linux/cpu.h>
548 #include <linux/irq_work.h>
549 #include <linux/static_key.h>
550 #include <linux/atomic.h>
551 #include <asm/local.h>
553 #define PERF_MAX_STACK_DEPTH 255
555 struct perf_callchain_entry {
557 __u64 ip[PERF_MAX_STACK_DEPTH];
560 struct perf_raw_record {
566 * single taken branch record layout:
568 * from: source instruction (may not always be a branch insn)
570 * mispred: branch target was mispredicted
571 * predicted: branch target was predicted
573 * support for mispred, predicted is optional. In case it
574 * is not supported mispred = predicted = 0.
576 struct perf_branch_entry {
579 __u64 mispred:1, /* target mispredicted */
580 predicted:1,/* target predicted */
585 * branch stack layout:
586 * nr: number of taken branches stored in entries[]
588 * Note that nr can vary from sample to sample
589 * branches (to, from) are stored from most recent
590 * to least recent, i.e., entries[0] contains the most
593 struct perf_branch_stack {
595 struct perf_branch_entry entries[0];
601 * extra PMU register associated with an event
603 struct hw_perf_event_extra {
604 u64 config; /* register value */
605 unsigned int reg; /* register address or index */
606 int alloc; /* extra register already allocated */
607 int idx; /* index in shared_regs->regs[] */
611 * struct hw_perf_event - performance event hardware details:
613 struct hw_perf_event {
614 #ifdef CONFIG_PERF_EVENTS
616 struct { /* hardware */
619 unsigned long config_base;
620 unsigned long event_base;
624 struct hw_perf_event_extra extra_reg;
625 struct hw_perf_event_extra branch_reg;
627 struct { /* software */
628 struct hrtimer hrtimer;
630 #ifdef CONFIG_HAVE_HW_BREAKPOINT
631 struct { /* breakpoint */
632 struct arch_hw_breakpoint info;
633 struct list_head bp_list;
635 * Crufty hack to avoid the chicken and egg
636 * problem hw_breakpoint has with context
637 * creation and event initalization.
639 struct task_struct *bp_target;
644 local64_t prev_count;
647 local64_t period_left;
652 u64 freq_count_stamp;
657 * hw_perf_event::state flags
659 #define PERF_HES_STOPPED 0x01 /* the counter is stopped */
660 #define PERF_HES_UPTODATE 0x02 /* event->count up-to-date */
661 #define PERF_HES_ARCH 0x04
666 * Common implementation detail of pmu::{start,commit,cancel}_txn
668 #define PERF_EVENT_TXN 0x1
671 * struct pmu - generic performance monitoring unit
674 struct list_head entry;
677 const struct attribute_group **attr_groups;
681 int * __percpu pmu_disable_count;
682 struct perf_cpu_context * __percpu pmu_cpu_context;
686 * Fully disable/enable this PMU, can be used to protect from the PMI
687 * as well as for lazy/batch writing of the MSRs.
689 void (*pmu_enable) (struct pmu *pmu); /* optional */
690 void (*pmu_disable) (struct pmu *pmu); /* optional */
693 * Try and initialize the event for this PMU.
694 * Should return -ENOENT when the @event doesn't match this PMU.
696 int (*event_init) (struct perf_event *event);
698 #define PERF_EF_START 0x01 /* start the counter when adding */
699 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
700 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
703 * Adds/Removes a counter to/from the PMU, can be done inside
704 * a transaction, see the ->*_txn() methods.
706 int (*add) (struct perf_event *event, int flags);
707 void (*del) (struct perf_event *event, int flags);
710 * Starts/Stops a counter present on the PMU. The PMI handler
711 * should stop the counter when perf_event_overflow() returns
712 * !0. ->start() will be used to continue.
714 void (*start) (struct perf_event *event, int flags);
715 void (*stop) (struct perf_event *event, int flags);
718 * Updates the counter value of the event.
720 void (*read) (struct perf_event *event);
723 * Group events scheduling is treated as a transaction, add
724 * group events as a whole and perform one schedulability test.
725 * If the test fails, roll back the whole group
727 * Start the transaction, after this ->add() doesn't need to
728 * do schedulability tests.
730 void (*start_txn) (struct pmu *pmu); /* optional */
732 * If ->start_txn() disabled the ->add() schedulability test
733 * then ->commit_txn() is required to perform one. On success
734 * the transaction is closed. On error the transaction is kept
735 * open until ->cancel_txn() is called.
737 int (*commit_txn) (struct pmu *pmu); /* optional */
739 * Will cancel the transaction, assumes ->del() is called
740 * for each successful ->add() during the transaction.
742 void (*cancel_txn) (struct pmu *pmu); /* optional */
745 * Will return the value for perf_event_mmap_page::index for this event,
746 * if no implementation is provided it will default to: event->hw.idx + 1.
748 int (*event_idx) (struct perf_event *event); /*optional */
752 * enum perf_event_active_state - the states of a event
754 enum perf_event_active_state {
755 PERF_EVENT_STATE_ERROR = -2,
756 PERF_EVENT_STATE_OFF = -1,
757 PERF_EVENT_STATE_INACTIVE = 0,
758 PERF_EVENT_STATE_ACTIVE = 1,
762 struct perf_sample_data;
764 typedef void (*perf_overflow_handler_t)(struct perf_event *,
765 struct perf_sample_data *,
766 struct pt_regs *regs);
768 enum perf_group_flag {
769 PERF_GROUP_SOFTWARE = 0x1,
772 #define SWEVENT_HLIST_BITS 8
773 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
775 struct swevent_hlist {
776 struct hlist_head heads[SWEVENT_HLIST_SIZE];
777 struct rcu_head rcu_head;
780 #define PERF_ATTACH_CONTEXT 0x01
781 #define PERF_ATTACH_GROUP 0x02
782 #define PERF_ATTACH_TASK 0x04
784 #ifdef CONFIG_CGROUP_PERF
786 * perf_cgroup_info keeps track of time_enabled for a cgroup.
787 * This is a per-cpu dynamically allocated data structure.
789 struct perf_cgroup_info {
795 struct cgroup_subsys_state css;
796 struct perf_cgroup_info *info; /* timing info, one per cpu */
803 * struct perf_event - performance event kernel representation:
806 #ifdef CONFIG_PERF_EVENTS
807 struct list_head group_entry;
808 struct list_head event_entry;
809 struct list_head sibling_list;
810 struct hlist_node hlist_entry;
813 struct perf_event *group_leader;
816 enum perf_event_active_state state;
817 unsigned int attach_state;
819 atomic64_t child_count;
822 * These are the total time in nanoseconds that the event
823 * has been enabled (i.e. eligible to run, and the task has
824 * been scheduled in, if this is a per-task event)
825 * and running (scheduled onto the CPU), respectively.
827 * They are computed from tstamp_enabled, tstamp_running and
828 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
830 u64 total_time_enabled;
831 u64 total_time_running;
834 * These are timestamps used for computing total_time_enabled
835 * and total_time_running when the event is in INACTIVE or
836 * ACTIVE state, measured in nanoseconds from an arbitrary point
838 * tstamp_enabled: the notional time when the event was enabled
839 * tstamp_running: the notional time when the event was scheduled on
840 * tstamp_stopped: in INACTIVE state, the notional time when the
841 * event was scheduled off.
848 * timestamp shadows the actual context timing but it can
849 * be safely used in NMI interrupt context. It reflects the
850 * context time as it was when the event was last scheduled in.
852 * ctx_time already accounts for ctx->timestamp. Therefore to
853 * compute ctx_time for a sample, simply add perf_clock().
857 struct perf_event_attr attr;
861 struct hw_perf_event hw;
863 struct perf_event_context *ctx;
867 * These accumulate total time (in nanoseconds) that children
868 * events have been enabled and running, respectively.
870 atomic64_t child_total_time_enabled;
871 atomic64_t child_total_time_running;
874 * Protect attach/detach and child_list:
876 struct mutex child_mutex;
877 struct list_head child_list;
878 struct perf_event *parent;
883 struct list_head owner_entry;
884 struct task_struct *owner;
887 struct mutex mmap_mutex;
890 struct user_struct *mmap_user;
891 struct ring_buffer *rb;
892 struct list_head rb_entry;
895 wait_queue_head_t waitq;
896 struct fasync_struct *fasync;
898 /* delayed work for NMIs and such */
902 struct irq_work pending;
904 atomic_t event_limit;
906 void (*destroy)(struct perf_event *);
907 struct rcu_head rcu_head;
909 struct pid_namespace *ns;
912 perf_overflow_handler_t overflow_handler;
913 void *overflow_handler_context;
915 #ifdef CONFIG_EVENT_TRACING
916 struct ftrace_event_call *tp_event;
917 struct event_filter *filter;
918 #ifdef CONFIG_FUNCTION_TRACER
919 struct ftrace_ops ftrace_ops;
923 #ifdef CONFIG_CGROUP_PERF
924 struct perf_cgroup *cgrp; /* cgroup event is attach to */
925 int cgrp_defer_enabled;
928 #endif /* CONFIG_PERF_EVENTS */
931 enum perf_event_context_type {
937 * struct perf_event_context - event context structure
939 * Used as a container for task events and CPU events as well:
941 struct perf_event_context {
943 enum perf_event_context_type type;
945 * Protect the states of the events in the list,
946 * nr_active, and the list:
950 * Protect the list of events. Locking either mutex or lock
951 * is sufficient to ensure the list doesn't change; to change
952 * the list you need to lock both the mutex and the spinlock.
956 struct list_head pinned_groups;
957 struct list_head flexible_groups;
958 struct list_head event_list;
966 struct task_struct *task;
969 * Context clock, runs when context enabled.
975 * These fields let us detect when two contexts have both
976 * been cloned (inherited) from a common ancestor.
978 struct perf_event_context *parent_ctx;
982 int nr_cgroups; /* cgroup events present */
983 struct rcu_head rcu_head;
987 * Number of contexts where an event can trigger:
988 * task, softirq, hardirq, nmi.
990 #define PERF_NR_CONTEXTS 4
993 * struct perf_event_cpu_context - per cpu event context structure
995 struct perf_cpu_context {
996 struct perf_event_context ctx;
997 struct perf_event_context *task_ctx;
1000 struct list_head rotation_list;
1001 int jiffies_interval;
1002 struct pmu *active_pmu;
1003 struct perf_cgroup *cgrp;
1006 struct perf_output_handle {
1007 struct perf_event *event;
1008 struct ring_buffer *rb;
1009 unsigned long wakeup;
1015 #ifdef CONFIG_PERF_EVENTS
1017 extern int perf_pmu_register(struct pmu *pmu, char *name, int type);
1018 extern void perf_pmu_unregister(struct pmu *pmu);
1020 extern int perf_num_counters(void);
1021 extern const char *perf_pmu_name(void);
1022 extern void __perf_event_task_sched_in(struct task_struct *prev,
1023 struct task_struct *task);
1024 extern void __perf_event_task_sched_out(struct task_struct *prev,
1025 struct task_struct *next);
1026 extern int perf_event_init_task(struct task_struct *child);
1027 extern void perf_event_exit_task(struct task_struct *child);
1028 extern void perf_event_free_task(struct task_struct *task);
1029 extern void perf_event_delayed_put(struct task_struct *task);
1030 extern void perf_event_print_debug(void);
1031 extern void perf_pmu_disable(struct pmu *pmu);
1032 extern void perf_pmu_enable(struct pmu *pmu);
1033 extern int perf_event_task_disable(void);
1034 extern int perf_event_task_enable(void);
1035 extern int perf_event_refresh(struct perf_event *event, int refresh);
1036 extern void perf_event_update_userpage(struct perf_event *event);
1037 extern int perf_event_release_kernel(struct perf_event *event);
1038 extern struct perf_event *
1039 perf_event_create_kernel_counter(struct perf_event_attr *attr,
1041 struct task_struct *task,
1042 perf_overflow_handler_t callback,
1044 extern u64 perf_event_read_value(struct perf_event *event,
1045 u64 *enabled, u64 *running);
1047 struct perf_sample_data {
1064 struct perf_callchain_entry *callchain;
1065 struct perf_raw_record *raw;
1066 struct perf_branch_stack *br_stack;
1069 static inline void perf_sample_data_init(struct perf_sample_data *data, u64 addr)
1073 data->br_stack = NULL;
1076 extern void perf_output_sample(struct perf_output_handle *handle,
1077 struct perf_event_header *header,
1078 struct perf_sample_data *data,
1079 struct perf_event *event);
1080 extern void perf_prepare_sample(struct perf_event_header *header,
1081 struct perf_sample_data *data,
1082 struct perf_event *event,
1083 struct pt_regs *regs);
1085 extern int perf_event_overflow(struct perf_event *event,
1086 struct perf_sample_data *data,
1087 struct pt_regs *regs);
1089 static inline bool is_sampling_event(struct perf_event *event)
1091 return event->attr.sample_period != 0;
1095 * Return 1 for a software event, 0 for a hardware event
1097 static inline int is_software_event(struct perf_event *event)
1099 return event->pmu->task_ctx_nr == perf_sw_context;
1102 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
1104 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
1106 #ifndef perf_arch_fetch_caller_regs
1107 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
1111 * Take a snapshot of the regs. Skip ip and frame pointer to
1112 * the nth caller. We only need a few of the regs:
1113 * - ip for PERF_SAMPLE_IP
1114 * - cs for user_mode() tests
1115 * - bp for callchains
1116 * - eflags, for future purposes, just in case
1118 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1120 memset(regs, 0, sizeof(*regs));
1122 perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1125 static __always_inline void
1126 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1128 struct pt_regs hot_regs;
1130 if (static_key_false(&perf_swevent_enabled[event_id])) {
1132 perf_fetch_caller_regs(&hot_regs);
1135 __perf_sw_event(event_id, nr, regs, addr);
1139 extern struct static_key_deferred perf_sched_events;
1141 static inline void perf_event_task_sched_in(struct task_struct *prev,
1142 struct task_struct *task)
1144 if (static_key_false(&perf_sched_events.key))
1145 __perf_event_task_sched_in(prev, task);
1148 static inline void perf_event_task_sched_out(struct task_struct *prev,
1149 struct task_struct *next)
1151 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, NULL, 0);
1153 if (static_key_false(&perf_sched_events.key))
1154 __perf_event_task_sched_out(prev, next);
1157 extern void perf_event_mmap(struct vm_area_struct *vma);
1158 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1159 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1160 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1162 extern void perf_event_comm(struct task_struct *tsk);
1163 extern void perf_event_fork(struct task_struct *tsk);
1166 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1168 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
1169 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
1171 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
1173 if (entry->nr < PERF_MAX_STACK_DEPTH)
1174 entry->ip[entry->nr++] = ip;
1177 extern int sysctl_perf_event_paranoid;
1178 extern int sysctl_perf_event_mlock;
1179 extern int sysctl_perf_event_sample_rate;
1181 extern int perf_proc_update_handler(struct ctl_table *table, int write,
1182 void __user *buffer, size_t *lenp,
1185 static inline bool perf_paranoid_tracepoint_raw(void)
1187 return sysctl_perf_event_paranoid > -1;
1190 static inline bool perf_paranoid_cpu(void)
1192 return sysctl_perf_event_paranoid > 0;
1195 static inline bool perf_paranoid_kernel(void)
1197 return sysctl_perf_event_paranoid > 1;
1200 extern void perf_event_init(void);
1201 extern void perf_tp_event(u64 addr, u64 count, void *record,
1202 int entry_size, struct pt_regs *regs,
1203 struct hlist_head *head, int rctx);
1204 extern void perf_bp_event(struct perf_event *event, void *data);
1206 #ifndef perf_misc_flags
1207 # define perf_misc_flags(regs) \
1208 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1209 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1212 static inline bool has_branch_stack(struct perf_event *event)
1214 return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1217 extern int perf_output_begin(struct perf_output_handle *handle,
1218 struct perf_event *event, unsigned int size);
1219 extern void perf_output_end(struct perf_output_handle *handle);
1220 extern void perf_output_copy(struct perf_output_handle *handle,
1221 const void *buf, unsigned int len);
1222 extern int perf_swevent_get_recursion_context(void);
1223 extern void perf_swevent_put_recursion_context(int rctx);
1224 extern void perf_event_enable(struct perf_event *event);
1225 extern void perf_event_disable(struct perf_event *event);
1226 extern void perf_event_task_tick(void);
1229 perf_event_task_sched_in(struct task_struct *prev,
1230 struct task_struct *task) { }
1232 perf_event_task_sched_out(struct task_struct *prev,
1233 struct task_struct *next) { }
1234 static inline int perf_event_init_task(struct task_struct *child) { return 0; }
1235 static inline void perf_event_exit_task(struct task_struct *child) { }
1236 static inline void perf_event_free_task(struct task_struct *task) { }
1237 static inline void perf_event_delayed_put(struct task_struct *task) { }
1238 static inline void perf_event_print_debug(void) { }
1239 static inline int perf_event_task_disable(void) { return -EINVAL; }
1240 static inline int perf_event_task_enable(void) { return -EINVAL; }
1241 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1247 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) { }
1249 perf_bp_event(struct perf_event *event, void *data) { }
1251 static inline int perf_register_guest_info_callbacks
1252 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1253 static inline int perf_unregister_guest_info_callbacks
1254 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1256 static inline void perf_event_mmap(struct vm_area_struct *vma) { }
1257 static inline void perf_event_comm(struct task_struct *tsk) { }
1258 static inline void perf_event_fork(struct task_struct *tsk) { }
1259 static inline void perf_event_init(void) { }
1260 static inline int perf_swevent_get_recursion_context(void) { return -1; }
1261 static inline void perf_swevent_put_recursion_context(int rctx) { }
1262 static inline void perf_event_enable(struct perf_event *event) { }
1263 static inline void perf_event_disable(struct perf_event *event) { }
1264 static inline void perf_event_task_tick(void) { }
1267 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1270 * This has to have a higher priority than migration_notifier in sched.c.
1272 #define perf_cpu_notifier(fn) \
1274 static struct notifier_block fn##_nb __cpuinitdata = \
1275 { .notifier_call = fn, .priority = CPU_PRI_PERF }; \
1276 fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE, \
1277 (void *)(unsigned long)smp_processor_id()); \
1278 fn(&fn##_nb, (unsigned long)CPU_STARTING, \
1279 (void *)(unsigned long)smp_processor_id()); \
1280 fn(&fn##_nb, (unsigned long)CPU_ONLINE, \
1281 (void *)(unsigned long)smp_processor_id()); \
1282 register_cpu_notifier(&fn##_nb); \
1285 #endif /* __KERNEL__ */
1286 #endif /* _LINUX_PERF_EVENT_H */